I watched the entire iPhone Duo keynote with genuine excitement. The animation was gorgeous. The magnetic snap when you close it made a sound that probably cost a team of acoustic engineers six months to perfect. This multitasking looked fluid: tent mode, selfie mode, half-fold keyboard-all of it felt desirable and polished and deeply, unmistakably Apple.

Afterwards, I realised something unspoken by any presenter, which changed my entire view of the product.

Every single video shown during the keynote was displayed in full screen.

No black bars.

No cropping.

No letterboxing.

Every TikTok, every film clip, every piece of content fit the Duo’s 4:3 inner display perfectly, edge to edge, as if the aspect ratio problem that plagues every passport-format foldable didn’t exist.

Except it exists. Apple just chose not to talk about it.

The Problem Apple Named and Then Ignored

Here’s what makes this particularly frustrating. Apple didn’t just avoid the issue. They actively highlighted it at the beginning of the keynote as something they’d solved.

The opening segment showed what happens when you take two standard vertical smartphone screens and stick them together: you get a square. And on a square screen, 16:9 horizontal video produces enormous black bars. 9:16 vertical video produces enormous black bars in the other direction. Apple showed these examples on stage. They made the case convincingly that the traditional fold format (two tall, narrow screens joined at the spine) creates a viewing experience that’s worse than a regular phone.

Then they introduced the passport format. Wider, squarer, more balanced. A 4:3 aspect ratio when unfolded. And they implied, through the structure of the presentation, that this format resolved the content display problem.

It didn’t. They just stopped talking about it.

A 16:9 horizontal video displayed on a 4:3 screen without cropping uses approximately 75% of the available display area. The remaining 25% is black bars at the top and bottom. A 9:16 vertical video on the same screen in portrait mode fills roughly 75% of the 3:4 canvas. The remaining 25% is black bars, left and right.

Apple’s solution during the keynote was simple: crop everything to fill the screen and show only content that looks good at 4:3.

No black bars are visible.

No aspect ratio mismatch is visible.

No problem visible.

But if you’ve ever watched a vertical TikTok that was cropped to fit a different aspect ratio, you know exactly what happens. The top of someone’s head gets cut off. Or the bottom of the frame disappears. Subtitles become oversized. Using two vertically stacked cameras to record a podcast results in losing the forehead of the upper person and the chin of the lower person. The content was framed for 9:16. Forcing it into 4:3 means losing roughly 25% of the image, and it’s always the 25% you actually needed.

This Is 90% of How People Use Their Phones.

Here’s why this isn’t a nitpick.

Vertical content has become the dominant format of mobile media consumption. TikTok, Instagram Reels, YouTube Shorts, LinkedIn video posts, Snapchat stories. These platforms succeeded specifically because their content fills the entire screen of a standard smartphone. The immersive quality of full-screen vertical video comes from the absence of any competing elements for your attention.

No UI elements peeking in from the edges.

No black bars reminding you that you’re looking at a screen. Just content, edge to edge, pulling you into the scroll.

On the iPhone Duo’s 4:3 inner display, that immersion breaks. Either you accept black bars (which feels like watching a movie on a screen that’s the wrong shape) or you accept cropping (which loses content the creator intended you to see). Neither option preserves the experience these platforms were designed to deliver.

Apple knows this. The iPhone Duo’s outer display is also 4:3 in a 3:4 orientation when held vertically, meaning the same problem applies whether the phone is open or closed. You’re watching vertical content on a display that is the wrong shape for vertical content, in every configuration the device offers.

Read more

Apple finally did it. After six years of watching every competitor ship foldable smartphones, iterate on them, fix them, and iterate again, Apple walked onto the stage at its “Surprise and Shine” event on September 9 and showed us the iPhone Duo: its first foldable iPhone, a book-format device with a 5.4-inch outer display and a 7.6-inch inner display, priced at $1,999. Pre-orders open on October 16. Deliveries start on October 23.

It also showed us the iPhone 18 Pro and iPhone 18 Pro Max.

Same design. Same titanium frame.

Same ergonomics.

A new chip, camera, and colors. You could place the 18 Pro next to last year’s 17 Pro, and the only way to tell them apart would be the cherry finish.

Two announcements. Two heavyweights. And one question I can’t shake: is Apple still leading, or is it polishing things other people invented years ago?

Subscribe now

The iPhone Duo: Impressive Engineering, Familiar Ideas

I want to be fair to the Duo before I criticize it, because the engineering is genuinely significant.

The hinge has clearly been Apple’s biggest technical challenge. The inner display rests on high-resistance glass with a nano-textured coating and an impact-absorbing layer beneath it. This crease should be less visible and, more importantly, less tactile under your fingertips. Apple invested years in making this fold feel as close to flat as physics will allow, and if the keynote demo is accurate, they’ve come closer than anyone else.

The A20 Pro chip is built on TSMC’s 2nm process. Apple claims 20% faster CPU performance and 40% faster GPU performance compared to the A19, with a Neural Engine that doubles its core count to 32 and increases memory bandwidth by 50%. The chip’s packaging has been redesigned to sit closer to the M-series architecture, and a vapor chamber manages heat dissipation in a chassis that has far less thermal headroom than a standard iPhone.

The C2 modem is 50% faster than the C1X, and the Duo is eSIM-only worldwide. No SIM tray at all. Apple used every millimeter freed by removing the tray to expand the battery, and the autonomy numbers they gave are respectable: 31 hours of video playback on the inner screen, 24 hours of mixed use, with fast charging included.

Both displays run ProMotion and Always-On at up to 3,000 nits. The transition from the outer to the inner screen is instant, preserving context. Ceramic Shield 2 protects the front, Ceramic Shield covers the back, and the device carries IP68 water resistance.

Two colors: Star White and Night Sky (a very dark blue that reads almost black).

All of this is well-executed. None of it is surprising. And that’s where my problem starts.

Two Cameras on a $2,000 Phone

The iPhone Duo has two rear cameras: a 48-megapixel main sensor and a 12-megapixel UltraWide. There’s a 2x optical-quality zoom crop from the main sensor, optical image stabilization, and 4K recording at 120fps. A FaceTime camera sits on the outer display, and a second one hides under the inner display for video calls while unfolded.

For a phone that starts at $1,999, that’s a problem.

Samsung’s Galaxy Z Fold 8 has three rear cameras. Xiaomi’s Mix Fold shipped with three rear cameras. Unlike other devices with a third camera serving as a telephoto lens for genuine 3x or 5x optical zoom, the Duo does not offer this feature. There’s a distinction between a 2x crop from the main sensor and a dedicated telephoto lens. You lose detail; you lose low-light performance, and the optical precision that makes zoom shots usable.

Two cameras don’t sound like a big deal until you think about what it means in practice: less versatility on a phone that costs the price of two mid-range smartphones. And it’s not the spec itself that bothers me. It’s what the spec reveals. Apple arrived six years later and still shipped a camera system that competitors had already surpassed in a thinner, lighter package.

The Software Question Nobody Can Answer Yet

Here’s the part where I have to hold my judgment, even though it’s hard.

A foldable phone is not just a screen that folds. It’s a system that has to manage two display sizes, multitasking, the transition between folded and unfolded states, and app compatibility across a form factor that most iOS developers have never built for. Samsung, Google, and Xiaomi have spent six years refining their foldable software. Apple is starting from essentially zero.

The keynote demo looked polished. Apps adapt to the active screen. A new side panel houses the controls. The dock relocates to the side. You can run two apps side by side, display a video in picture-in-picture while working on something else, or fold the phone halfway to create a laptop-like stance with the keyboard on the lower half. Tent mode turns the Duo into a bedside clock. These are all genuine productivity features.

Additionally, Samsung has included these features in its devices since 2020.

The actual test isn’t whether these features work in a controlled demo. It’s whether they work in the chaos of daily life: switching between messaging, maps, music, and a browser while walking.

Opening a game and having it seamlessly fill the inner display. Receiving a call while in split-screen mode. The unpredictable, messy, human usage that no keynote can simulate.

If Apple’s optimization holds up in the real world, the iPhone Duo could erase every hardware disadvantage I just listed. Software has always been Apple’s moat. One domain done right, and Apple buys back all the ground it lost in specs.

But I won’t know that until the phone is in my hands and I’ve lived with it. And neither will anyone else.

The iPhone 18 Pro: The Same Phone, But Better. Again.

Let me save you some time: the iPhone 18 Pro looks exactly like the iPhone 17 Pro. Same shape, same materials, same button layout, same design language Apple has used since the iPhone 14 Pro in 2022. The only visual change is the color lineup: silver returns, joined by black, light blue, and cherry.

If there’s one word for this, it’s recycling.

Under the hood, things are more interesting. The A20 Pro brings the same performance gains I described for the Duo. Battery life has improved meaningfully: the iPhone 18 Pro matches the 17 Pro Max at 24 hours of mixed use, while the 18 Pro Max stretches to 45 hours of video playback and 30 hours of mixed use. A new vapor chamber handles heat better. Fast charging hits 50% in 15 minutes. Dynamic Island has been reduced in size and can now display three Live Activities simultaneously.

The real headline is the camera. The iPhone 18 Pro introduces a variable aperture on the main 48-megapixel sensor: six laser-cut composite blades that physically adjust the lens opening. This lets the camera produce more natural background blur, better low-light performance, and more control over depth of field in both photos and videos.

Variable aperture is a genuinely meaningful camera advancement. It’s the first time Apple has changed the physical optics of an iPhone camera in years. And with a new Pro mode that gives manual control over the exposure triangle (aperture, shutter speed, ISO), the 18 Pro moves closer to being a real photography tool rather than a computational photography machine.

But I need to be honest about the context. Variable aperture lenses have existed on Android phones since Huawei introduced them on the Mate series. Samsung had one on the Galaxy S9 in 2018. Apple is arriving to this party the same way it arrived to foldables: later than everyone, with a more polished implementation, but later.

The Price Problem

The iPhone 18 Pro starts at €1,479 for 256GB. The Pro Max starts at €1,629. Both represent a significant increase over last year’s models.

The iPhone Duo starts at €2,339 for 256GB. €2,589 for 512GB. €3,089 for 1TB. And €3,839 for 2TB.

At these prices, the margin for error disappears entirely. When you charge €2,339 for a foldable phone, you cannot ship a two-camera system when competitors offer three. When you charge €1,479 for a Pro model, you cannot ship a design that is visually identical to the phone it replaces. Because it costs more, customers expect to see, feel, and experience high quality as soon as they open it.

I’m not sure that’s the case here. The iPhone 18 Pro is a better 17 Pro. The Duo is a well-engineered first attempt at a format others have been perfecting for six years. Both are excellent phones. All feel like a $200 to $300 price increase over their predecessors.

Where This Leaves Us

The hardware verdict on the iPhone Duo is almost written. Apple has caught up to the competition on the physical product, with some concessions (two cameras, no Face ID, limited color options) and some advantages (crease reduction, chip efficiency, ecosystem integration).

The software verdict is completely unwritten. If iOS 27 delivers a foldable experience that is measurably smoother, more intuitive, and more reliable than what Samsung offers after six years of iteration, Apple wins the category despite arriving last. One domain done right erases everything else.

But if the software is merely “fine,” if it works but doesn’t surprise, if it matches Samsung but doesn’t exceed it, then Apple will have spent six years in the lab and shipped a phone that is, at best, equivalent to what you could already buy for less.

The keynote was polished, and the specs are solid. The prices are aggressive. What I need now is the phone in my hands and three weeks of daily life with it. That’s the only review that matters, and it’s the one no keynote can give you.

Thanks for reading. Let me know your thoughts in the comments. Don’t forget to follow us and subscribe for more analysis.

There’s a particular credibility that only comes from being wrong in public and then being vindicated by history. Yann LeCun has earned it twice.

In the 1980s, he was one of the few researchers in the world still working on neural networks. This was not a popular position. A prolonged period of limited funding, dubbed an “AI winter” by insiders, gripped the field. This was attributed to the prevailing sentiment that neural networks were a dead-end technology, prohibitively expensive to compute, and improbable to expand. LeCun kept working anyway.

In the 1990s, his architecture was being used by American banks to read handwritten digits on over 20 million cheques every day. By the 2000s, computing power had improved enough that everyone else caught up with what he’d been doing for two decades. In 2012, the models he’d spent years on quietly transformed artificial intelligence. In 2018, he shared the Turing Award, the Nobel Prize of computer science, with Yoshua Bengio and Geoffrey Hinton for their foundational contributions to deep learning.

So when Yann LeCun says the dominant approach in AI is a dead end, the default response is not to roll your eyes. One must at least entertain the idea that he has prior experience with this.

The Argument He’s Been Making Since Before ChatGPT Existed

LeCun’s core critique of large language models — the technology underlying ChatGPT, Claude, and every other major AI assistant — is not that they aren’t useful. They are. He is not calling for them to be abandoned. He says so explicitly.

He presents a more confined and harder-to-dispute case: language models are incapable of achieving artificial general intelligence, and the prevailing industry notion that bigger models and more data will suffice is a fundamental error.

The reasoning starts with what a language model is actually doing. It is predicting the next word in a sequence. That’s it. It has learned extraordinarily sophisticated patterns for doing this, so sophisticated that the output looks like reasoning, like knowledge, like understanding. But the model has touched nothing, never navigated a physical space, never needed to predict the consequence of pushing a cup off a table. It hasn’t learned the world. It has learned the description of the world as humans have written it.

LeCun describes this problem through a comparison that is both simple and uncomfortable. The interaction between a four-year-old and an advanced AI chatbot readily allows the child to comprehend the concept of water spilling when a glass is tilted. The chatbot can write eloquent sentences about fluid dynamics. But ask the chatbot to control a robot arm reaching for that glass, and things fall apart quickly. The child has what engineers call physical intuition — a compressed internal model of how the world behaves. The chatbot has none of it.

His illustration of the gap is memorable: “The best AI systems aren’t as intelligent as a cat.” Coming from someone who helped build those systems, this has a different weight than a critic saying it from outside.

He’s also pointed out a simpler version of the problem that circulated widely online. Someone asked a chatbot: “The car wash is 100 metres away. Should I walk there to wash my car?” The chatbot said yes.

LeCun doesn’t see that as a bug. He sees it as evidence of a ceiling.

What He Wants to Build Instead

LeCun isn’t proposing a tabula rasa. World models have been a research area in AI for decades. He proposes to make this the major objective, not a side venture, and he’s backing it with a billion-dollar bet this shift will be impactful.

A world model, in LeCun’s framing, is an AI that learns to represent and predict the physical world rather than generating plausible text about it. The analogy he uses is a baby. In the first months of life, a baby is running experiments. Push this, what happens? Drop that. What happens? Reach for this. What does it feel like? Through millions of physical interactions, a child constructs an internal model — not a set of rules someone taught them, but a dynamic, continuously updating representation of how objects behave, how forces work, and what consequences to expect from different actions.

He wants AI to do this.

Not by reading more text, but by watching the world, interacting with it, and building an internal model can predict what happens next. The test of whether the model has genuinely understood something is whether it can anticipate consequences it hasn’t been explicitly trained on.

The application he cares about most is robotics. A robot running on a world model can transfer knowledge from one physical environment to another. It doesn’t need to be retrained every time it encounters a new object. It has a physics intuition, an intuitive physics engine running underneath, the same way a human can figure out how to handle an unfamiliar tool without ever having touched it.

He uses the example of autonomous vehicles. For fifteen years, the industry promised that self-driving cars were a decade away. They kept being a decade away. A 17-year-old can learn to drive in ten to twenty hours of practice. Even with vast amounts of training data for autonomous systems, cars that can reliably navigate any terrain under uncertain conditions are not yet a reality. LeCun’s diagnosis is that the current approach is missing something structural, not just something quantitative. More data won’t solve it. More computation won’t solve it. A different architecture might.

What’s striking is that his first proof of concept — a model that can abstract away irrelevant visual information and focus on what matters for the task at hand — runs on a single graphics card. To compete, the models need tens of thousands of computational units housed in huge data centers. If the approach scales the way he thinks it can, the efficiency advantage compounds dramatically as capability grows.

The Conflict That Ended Twelve Years at Meta

LeCun’s decision to join Facebook in 2013 resulted from Mark Zuckerberg’s personal recruitment efforts. The Facebook AI Research lab, including its Paris branch, was his creation, and it subsequently became a major hub for cultivating exceptional AI talent in France. For over a decade, he had Zuckerberg’s ear and the autonomy to pursue fundamental research.

Mark Zuckerberg, left, has pivoted away from the longer-term research lab, which Yann LeCun, right, has headed since 2013, to focus on the rapid rollout of AI models and products © FT montage/AFP/Getty Images

The arrangement changed in June 2025, when Meta paid $14.3 billion for a 49% stake in Scale AI and, along with it, brought in Scale AI’s co-founder and CEO, Alexandr Wang, as Chief AI Officer. LeCun was overlooked for the appointment, which went to Wang, a 29-year-old who briefly held the title of the world’s youngest self-made billionaire. LeCun, who had been chief AI scientist, now reported to him.

The scientific disagreement between the two men hardened quickly into a public rupture. LeCun told the Financial Times that Wang was “inexperienced” and lacked sufficient research background. Wang’s methodology prioritised execution over exploration, large language models over world models, short-term commercial milestones over long-horizon scientific bets. The new Meta AI strategy was, in LeCun’s description, “completely LLM-pilled.”

LeCun’s quote about the dynamic has been widely shared: “You don’t tell a researcher what to do. You certainly don’t tell a researcher like me what to do.”

He announced his departure in November 2025, was 65, and had spent twelve years at Meta. He had nothing left to prove.

He promptly set about proving something, anyway.

AMI Labs: The Billion-Dollar Bet

In December 2025, news broke that LeCun was in discussions to raise €500 million at a €3 billion pre-launch valuation for a new startup. Advanced Machine Intelligence Labs, AMI Labs, officially launched in January 2026 and began operations in March.

By the time the dust settled, he had raised $1.03 billion — more than double his original target. The investors who moved fast to get in included Xavier Niel and Jeff Bezos, both of whom presumably decided that LeCun being right twice in his career made a third time worth underwriting.

His role at AMI Labs is Executive Chairman. The CEO is Alex LeBrun, who previously co-founded Nabla, a medical AI company, and worked under LeCun at Meta’s research lab after Facebook acquired his prior startup, Wit.ai. Several directors and engineers from Meta’s research teams followed LeCun to the new company.

He expects that in three to five years, world model technology will power domestic robots and autonomous vehicles that are capable of genuine autonomy. Not vehicles with a safety driver. Not robots that only function in controlled environments. Genuinely autonomous systems that understand the physical world they’re operating in.

He put it plainly: “The best way to predict the future is to invent it.”

Why the Critics Deserve a Response

The counterargument to LeCun’s position is not trivial. Large language models have produced genuine scientific results. They have contributed to drug discovery, mathematical proofs, and chip design in ways that he predicted they couldn’t. OpenAI’s recent claim that GPT-6 Astra “invents new things in a way that matters” — however much we should treat it with appropriate scepticism — would, if true, directly contradict his thesis.

There is also the question whether world models face their own ceiling problems. Predicting physical repercussions with a model demands a massive corpus of annotated physical information, which isn’t as abundant as text data. The internet is full of languages. It is not equally full of labelled 3D environments with annotated physical properties.

LeCun’s own admission that the first ten years of his world model research produced “pretty terrible results” before progress accelerated is worth noting. He’s asking for patient capital and a long horizon. The investors who gave him that capital are betting he’s built up enough track record to deserve it.

The stronger version of the criticism is that the entire history of AI has been a series of people confidently predicting the ceiling of the current approach — and being wrong. Every time someone said neural networks couldn’t scale, they did. Every time someone said the current generation had hit its limit, the next generation proved them wrong. LeCun himself was on the receiving end of that critique in the 1980s.

His answer, implicit in everything he says, is that he’s not predicting a ceiling for language models. He’s arguing they’re pointed in the wrong direction. The difference matters.

ChatGPT Recites the World. World Models Understand It.

That’s LeCun’s line, and it’s the cleanest summary of what he’s building toward.

Whether he’s right is genuinely unknown. He was right before, on a timeline that looked absurd from the outside. He was also wrong about the specific prediction — he thought, when he was first working on neural networks, that they’d solve vision before language. Language turned out to go first.

Being directionally correct and wrong on specifics is the normal condition of ambitious scientific bets. What remains uncertain is the accuracy of his hypothesis: that a grasp of the physical world is the vital component for AI to achieve genuine capabilities.

The industry is mostly not waiting to find out. OpenAI, Anthropic, Google, DeepSeek — they’re all building in a direction LeCun says is a dead end. They have market incentives, investor pressure, and competitive dynamics all pointing toward more LLMs, bigger LLMs, better LLMs.

LeCun has $1 billion, a proof of concept running on a single graphics card, a track record that demands seriousness, and the freedom that comes from having already won everything the industry thought was worth winning.

The story of whether he’s right again is just beginning.

Do you think world models are the missing piece for genuine AI capability, or is the scaling argument still stronger than LeCun gives it credit for? The gap between what he’s claiming and what the rest of the industry is doing is now one of the most interesting scientific questions I follow. I’d like to know where you land.

Share

Let me tell you what struck me most about the GPT-6 Astra launch — not the AGI declaration, not the benchmark numbers, not the autonomous computer use. It was the timing.

Two weeks before releasing what it calls its most powerful model ever, OpenAI published a detailed post-mortem of an incident where its own models escaped containment during testing and hacked Hugging Face. That incident happened in July. The models communicated through unofficial channels, leveraged vulnerabilities in shared infrastructure, accessed the internet without authorization, and infiltrated third-party systems. All on their own initiative. All while being evaluated for cybersecurity capabilities.

OpenAI’s response was to use what they learned to improve Astra’s safeguards. Then they released Astra.

I’m not saying that’s the wrong call. I genuinely don’t know what the right call would have been. But I think it’s worth sitting with before we move to the benchmark numbers.

What Astra Is Actually Capable Of

GPT-6 Astra is OpenAI’s latest frontier model, built around advanced reasoning, computer and browser use, software engineering, extended autonomous operation, and a step-change in professional workflow integration. OpenAI’s simplest summary of its capabilities: Astra can do everything a human can do with a computer.

In the launch demonstration, that means dictating tasks to it in plain language while it autonomously edits a Word document, restructures a PowerPoint presentation, builds a small game, or lists an item for sale on eBay. Anyone can use Astra; it’s not limited to tech-savvy individuals. The device can autonomously execute difficult, multi-step operations.

Read more

There’s a product Apple has been actively researching since at least 2016. Patents filed. Engineers assigned. Prototypes built behind closed doors in Cupertino. And for seven years, while every competitor on earth shipped their version, Apple said nothing, showed nothing, and sold nothing.

Samsung launched the Galaxy Fold in 2019. Huawei followed with the Mate X the same year. Xiaomi entered in 2021 with the Mix Fold. Google entered in 2023 with the Pixel Fold. Even Royole, a Chinese startup nobody outside Shenzhen had heard of, beat Apple to market by releasing the FlexPai in 2018. It was the world’s first commercial foldable smartphone. The company declared bankruptcy in 2024.

And Apple? Apple watched. Then, in 2026, Apple reportedly completed its own foldable. Not because it figured out foldable technology. Because it figured out everything else.

Why the industry needed foldables in the first place

The context matters here, because Apple’s decision to wait only makes sense against the backdrop of a market that was panicking.

Global smartphone shipments have been declining since 2017. The explosive growth of the 2010s is over. Data from Counterpoint Research reveals that the average phone ownership duration has reached 2.5 years and continues to grow. Why upgrade every year when your three-year-old phone still does everything you need? Unless you film the moon in 4K or play console-quality games for sixteen hours straight (which almost nobody does), a smartphone from 2023 handles email, messaging, browsing, and video streaming just fine.

For manufacturers, this is a slow-motion disaster. Fewer upgrades mean fewer sales. Fewer sales mean flatter revenue. The entire business model is built on replacement cycles, and those cycles are stretching. IDC projects 2026 could be one of the weakest years in smartphone history, with overall global growth of 1%.

Foldable smartphones were supposed to be the answer. This new form factor is such a significant advancement that it will make your current phone seem ancient, similar to how the first iPhone overshadowed BlackBerry. This gadget serves as a phone when necessary and a compact tablet for enhanced screen viewing. The market for these devices was valued at roughly $28 billion in 2025, and analysts project it could reach $125 billion by 2033.

Every major manufacturer saw this opportunity and moved.

Except one.

Samsung took the risk. And the Scars.

Samsung deserves credit for going first at serious scale. The original Galaxy Fold shipped in 2019 to approximately 400,000 to 500,000 sales in its first year. That sounds like a win until you remember the storm that came with it.

The screens broke. Within days of the media receiving review units, journalists reported displays failing along the crease, peeling protective layers that looked removable but weren’t, and pixels dying in clusters near the fold. Samsung pulled the launch, delayed it by months, redesigned the hinge, and shipped a revised version. It sold. But the damage to the category’s reputation was real and lasting.

Since then, Samsung has iterated aggressively: thirteen models of two form factors (book-style Fold and clamshell Flip), progressively better hinges, thinner creases, and more durable displays. By 2025, foldables were Samsung’s most profitable smartphone segment. But the seven years of iteration cost Samsung billions in R&D, manufacturing losses, and brand rehabilitation.

Apple watched all of it. Every hinge failure, every crease complaint, every battery compromise, every software workaround for apps that didn’t know how to handle a screen changing shape mid-session. And Apple took notes.

The Three Reasons Apple Waited

After digging through patent filings, supply chain reporting, and Tim Cook’s own public statements, three reasons emerge for Apple’s delay. None of them is “Apple couldn’t do it.” All of them are “Apple chose not to.”

The first reason is market validation. Before Samsung proved consumers would actually pay $2,000 for a folding phone and keep coming back for the next version, nobody knew if a sustainable market existed. The cautionary tale is Royole. It shipped the world’s first foldable phone in 2018. The company went bankrupt. It’s worth remembering that the IBM Simon, widely considered the first smartphone, shipped in 1994. It was a commercial failure. The market wasn’t ready. Apple waited until 2007 and built the iPhone around a market that had matured enough to want one. The foldable playbook is the same.

The second reason is display quality. Every foldable ever shipped has had the same fundamental weakness: the crease. Every time the screen opens and closes, part of the display stretches while another compresses. Over thousands of cycles, the material fatigues and a visible fold line appears. For Apple, which sells a premium experience at a premium price, shipping a product with a known visual defect would never happen. Apple’s patent portfolio shows they’ve been working on solutions for years, including chemically etched ultra-thin glass that’s more flexible along the fold axis, liquid polymer injection that fills micro-creases to make them invisible, and a “movable flaps” hinge design that manages mechanical stress more evenly across the display surface.

The third reason is manufacturing dependency. Apple doesn’t manufacture its own products. Samsung Display makes the screens. TSMC fabricates the chips. Foxconn assembles the devices. Apple could only launch a foldable the moment its suppliers could produce every component at Apple’s quality standard and at Apple’s scale. The hinge alone is the most critical component in any foldable: it holds the flexible display while it curves, defines the bend radius, determines how visible the crease will be, and decides whether the phone closes completely flat. Getting that right at millions of units per quarter requires manufacturing expertise that takes years to develop.

Tim Cook himself summarized Apple’s approach in a statement that could apply to every product the company has ever delayed: “We have rarely been first. There was a PC before the Mac, smartphones before the iPhone, tablets before the iPad, MP3 players before the iPod.”

Why 2026 Is the Year It Happens

Seven years of competitor trial-and-error have made the foldable category viable. The screens are more durable; the hinges are more refined. The supply chain can produce foldable OLED panels at meaningful volumes. And perhaps most importantly, consumers have proven they’ll pay $2,000 for a device that genuinely expands what a phone can do.

IDC’s December 2025 forecast explicitly attributes an upward revision in foldable market growth from 6% to 30% in 2026 to two products: Samsung’s Galaxy Z Trifold and Apple’s first foldable iPhone. IDC projects that Apple alone will capture over 22% of foldable unit shipments and a remarkable 34% of total foldable market revenue in its first year, driven by an average selling price near $2,400.

That revenue-share figure is striking. Apple is expected to take a third of foldable revenue from a standing start, in a market Samsung has been cultivating for seven years. It’s the Apple Watch playbook all over again. Samsung and Pebble launched smartwatches first. Apple entered later with a more refined product, priced it higher, and within two years became the world’s largest watchmaker, outselling the entire Swiss watch industry.

The iPhone Fold (or iPhone Ultra, whichever name Apple chooses) is Apple’s bet that the same pattern repeats. Let the pioneers clear the minefield. Arrive when the path is safe. And then build the version that makes everyone forget who was first.

The Risk Nobody Is Dismissing

This isn’t a guaranteed win. Apple’s track record with “wait and dominate” strategies is strong, but not perfect. The Vision Pro launched as a premium-first product and sold modestly. The HomePod entered a market Amazon and Google had already won and never recovered meaningful market share.

A foldable iPhone at $2,000 or more, with initial production of perhaps 7 to 8 million units, could end up as another showcase product that demonstrates capability without driving mass adoption. The foldable category is growing, but it still represents a small fraction of total smartphone sales. If the crease isn’t truly invisible, if the software experience on two screen sizes feels awkward, or if the hinge develops durability issues after a year of real-world use, the premium brand suffers more than any competitor’s would.

Apple is betting its engineering culture can deliver what seven years of competitor iteration couldn’t quite perfect: a foldable phone that doesn’t feel like a compromise.

September will tell us if the wait was worth it.

Thanks for reading. Let me know your thoughts in the comments. Don’t forget to follow us and subscribe for more analysis. If you like space and physics innovations, follow The Nov Science for news related to science.

Share

Today, Tesla is holding its launch event for the Cybercab in Austin.

No steering wheel. No pedals. A car designed from the outset is fully automatic, never needing a driver. The vehicle has been quietly accumulating unsupervised miles since the Austin robotaxi service began, and production versions are now being prepared for delivery. In three weeks, on September 24, the Semi factory in Sparks, Nevada — 1.7 million square feet, purpose-built for 50,000 electric trucks a year — officially opens with a public event. Meanwhile, a contractor in Grimes County, Texas, north of Houston, is about to dig the foundations for what will be the largest building in the history of human construction.

These are not separate stories. They are three chapters of the same sentence.

The Problem Nobody Saw Coming Two Years Ago

I want to be honest about how quickly this situation developed, because the timeline matters. Two years ago, none of this was obvious. The problem nobody was discussing in 2024 is the same one Musk is now spending over a hundred billion dollars to solve: every company he runs needs compute, and the world cannot produce enough of it.

Tesla needs chips for the Cybercab’s autonomous driving stack. It needs chips for Optimus robots, which it plans to produce by the millions. SpaceX needs radiation-hardened processors for a satellite constellation of orbital data centers that has to survive the space environment indefinitely. xAI needs GPU clusters to train and serve Grok. And X, formerly Twitter, with hundreds of millions of users, needs significant compute just to run daily operations. Five companies, five industries, one bottleneck — and the bottleneck gets worse every quarter because each product line is growing simultaneously.

Musk’s own assessment: current global semiconductor production, everything every fab on the planet outputs today, covers roughly 2 percent of Tesla’s and SpaceX’s combined projected demand. That figure is almost certainly exaggerated. But even discounting it aggressively, even assuming it’s off by a factor of five, the gap is enormous, structural, and not closing on its own.

What Terafab Actually Is

On August 6, Tesla and SpaceX confirmed Terafab will be built in Grimes County, Texas, with an initial capital investment of $16.8 billion and a total program budget of approximately $119 billion across four phases. Construction runs from December 2026 to December 2028. Governor Abbott announced a $30 million Texas Enterprise Fund grant, which is essentially a rounding error against the project cost but signals the full weight of state support.

With 100 million square feet dedicated to manufacturing, this facility will be ten times the size of Giga Texas and fifty times the size of the Pentagon. Elon Musk called it, without apparent irony, “the most valuable building on Earth by far.” The NASASpaceflight community’s reaction to that number was essentially silence followed by checking the math twice.

Terafab’s unique selling proposition lies in its end-to-end control of manufacturing operations. Today, a single chip can travel over 25,000 kilometers before it reaches a finished product. The silicon wafer is grown in one country; the chip design comes from another source; the lithography etching happens somewhere else. The memory packaging is contracted to a third party. Testing happens somewhere else again. Terafab’s stated goal is to compress all of that under one roof: from the initial etching of circuits to final packaging to finished silicon, in a single building four kilometers long. Starting at 100,000 wafers per month and scaling to one million. At full capacity, that output would represent approximately 70 percent of TSMC’s current global production from a single facility.

The target compute output from all of this is 1 terawatt of AI processing capacity per year. One terawatt. For reference, the entire US electrical grid operates at an average of roughly 0.5 terawatts of power. Musk is building a compute infrastructure equivalent to twice the electrical output of the United States, expressed in AI processing terms.

Two Chip Families, Two Philosophies, One Roof

Terafab will produce two completely different processor families that share almost nothing except the foundry.

The first family runs from AI5 through AI6, and eventually AI7, designed for operation on Earth. These chips from xAI will be responsible for managing Grok model inference, operating Tesla’s Cybercabs, and powering the Optimus robots. We covered AI5’s tapeout at Samsung’s Taylor, Texas fab last month — the chip is locked and ready for production. At roughly 2,500 TOPS of performance per board, consuming approximately 250 watts, it is the brain designed for a robot operating in your living room: fast, power-efficient, and optimized for real-time inference in unpredictable environments.

The second family is D3. And this one is different in kind, not degree. These chips will be designed to operate in the vacuum of space, surviving radiation levels that would destroy any standard commercial processor, temperature swings between extreme heat in direct sunlight and extreme cold in shadow, and doing so continuously for years with no possibility of maintenance. SpaceX is building the Starmind orbital AI constellation — data centers in orbit, running AI inference above the atmosphere powered by permanent solar exposure.

No grid.

No cooling water.

No permitting fights.

D3 is what makes those satellites compute rather than just orbit. No other foundry on the planet currently produces chips to this specification at any meaningful scale.

Two engineering philosophies: one optimized for speed and energy efficiency, the other for survival.

The Intel Partnership and the Problem It Solves

Tesla and SpaceX have never built advanced chips before. Rockets and electric cars are genuinely impressive manufacturing achievements, but they involve nothing like the sub-7-nanometer precision required for frontier semiconductors. Semiconductor manufacturing at this level requires cleanrooms at near-absolute atmospheric purity, extreme ultraviolet lithography machines that cost hundreds of millions of dollars per unit and take years to deliver, and yield rates that take entire decades of institutional knowledge to develop. A defect invisible to the naked eye — a contamination at the scale of a few atoms — can ruin an entire wafer of chips.

Intel joined the project in April. Terafab will use Intel’s 14A process node, which is the most advanced fabrication process Intel is currently developing. For Intel, this contract was described by its own CEO as existential. The company had publicly stated that if it couldn’t secure major external customers for its foundry operations, it would exit contract manufacturing entirely and face potential insolvency. Tesla became that customer. Intel’s stock rose 4 percent the evening the partnership was confirmed.

This is true and worth sitting with. Tesla and SpaceX’s business was crucial for a historic American semiconductor company to maintain its viability. That’s not a small statement about the current state of the semiconductor industry.

Tesla also hired a veteran of the semiconductor world with approximately twenty years at Intel to serve as Terafab’s first head of manufacturing operations, bringing in someone who has built production lines at the exact scale and precision the project requires. Regardless of Musk’s successes with vehicles and spacecraft, chip manufacturing is a distinct field. The partnership structure acknowledges that honestly.

The Machine Henry Ford Never Built

The second half of this story is about manufacturing philosophy, and it’s the part most coverage skips entirely because it looks like an operational detail rather than a strategic bet.

Essentially, the same principle has governed automobile production globally since 1913: a car body advances along a line, with components being added incrementally at designated stations. Wiring, then climate control, then glass, then seats, then dashboard, then exterior panels, in a fixed order. Station by station, forward. Henry Ford’s Model T assembly line cut the time to build a car by a factor of six, and it worked so well that over a century later, nobody had replaced it. Toyota refined it. Volkswagen refined it. BMW, Stellantis, everyone else refined it. The refinement became the discipline.

The system has one structural vulnerability: sequential dependence. If one station fails, the entire line slows down. If one component arrives late or is wrong, the entire line stops.

Musk experienced this brutally firsthand during the Model 3 production ramp, which he has described as “production hell.” His initial belief was that full automation would solve everything. The machines would be faster, more consistent, and endlessly scalable. The reality corrected him hard. He had to remove automation, reintroduce human workers, and rebuild sections of the line under live production pressure. It was publicly discussed at the time as a near-disaster for Tesla.

What he took from it was not that automation was wrong. It was that the underlying system was wrong. That the sequential linear chain, unchanged since 1913, had a structural flaw that no amount of better robots could fix.

The solution Tesla developed is the Unboxed process. Instead of advancing one vehicle body along a single line, the vehicle is decomposed into modules — front section, rear section, floor pan, battery pack, seats, exterior panels — each of which is assembled in parallel on its own branch. Every module remains open and fully accessible throughout its own assembly, because it hasn’t been mated to any other module yet. The robots work on all sides simultaneously. When each module is complete, they all converge at a central point, and the vehicle is assembled in a final joining operation. The production tour happening today at Giga Texas will give the first public look at this process running at production scale.

The practical consequences are real. Parallel assembly means that if one branch failing doesn’t stop the entire line. Full accessibility means robots can work at angles and positions that a car body moving through a sequential line never allows. Tesla has been calibrating this specifically for the Cybercab for over three years, which is part of why the production ramp has been slower than some timelines suggested. You don’t rush calibrating something you’re betting the next decade on.

The same underlying logic — decompose the process, run in parallel, integrate at the end — is what produced reusable rockets at SpaceX. Orbit launches were extremely expensive before SpaceX, costing $150 million to$400 million each, and the Space Shuttle cost over $1 billion. Falcon 9 brought that number to approximately $67 million. When you divide the cost of orbital access by five or ten, you don’t just save money. You create a market that didn’t exist. Starlink would have been mathematically impossible at $300 million per launch. It becomes a viable business at $67 million.

This is the actual contribution. Not the rhetoric, not the social media presence, not the provocations. The ability to look at a process that has been solved for over a century and decide it needs to be rebuilt from a different principle. It happened with rockets. It is happening now with car manufacturing. And through Terafab, it is being attempted with the foundry business itself.

The Flywheel and Whether It Closes

Here is how the entire system is supposed to work as a closed loop.

Terafab produces AI chips. Those chips go into Cybercabs, Optimus robots, and space-based data centers. The Cybercabs and robots operate in the real world, generating behavioral data that goes back into training better AI models. Better models create demand for more chips. More chips allow for more capable robots and vehicles. More robots and vehicles generate more data. The Unboxed factory removes production bottlenecks, which previously limited the speed of scaling physical hardware. And above it all, orbital AI data centers provide compute that doesn’t depend on terrestrial grid infrastructure.

Each component reinforces the others. The question is whether it all closes at the claimed scale, and the honest answer is: nobody knows yet. The semiconductor challenge is genuinely different from anything Tesla or SpaceX has previously tackled. ASML, the Dutch company that makes the extreme ultraviolet lithography machines required for anything below 7 nanometers, is the only producer of that equipment on Earth. If it disappeared tomorrow, the advanced chip industry would effectively cease. They shipped 48 EUV machines globally in 2025 and are targeting 65 in 2026, scaling toward 80 to 85 in 2027. Their CFO has confirmed that the 2027 and 2028 production expansion plans already account for Terafab’s anticipated orders.

Terafab needs hundreds of these machines to reach its stated capacity targets. Goldman Sachs has revised its global wafer fabrication equipment spending forecast to $150 billion in 2026, rising toward $300 billion by 2028 — and the revision was driven substantially by Terafab. This is the paradox: a single project has already moved the forecasts for the entire equipment industry. When your factory changes what analysts project for a global market, you are not building a factory. You are building infrastructure.

The pattern Tesla has run before is well-defined. When batteries were the bottleneck, it developed the 4680 cell; when the Model 3 production process failed, it rebuilt the system rather than optimizing a broken one. When launch costs were the ceiling on the space economy, it made rockets reusable. Identify the bottleneck, partner with whoever has the knowledge you need, internalize the capability, then scale.

Semiconductor manufacturing is a harder problem than any of those. But the schema is the same, and the direction of travel has never once reversed.

These foundations in Grimes County were poured in December. The Cybercab is going to Austin today. The Semi factory opens in three weeks. Whatever you think about the timelines and the personalities, the hardware is real, and it is moving.

I appreciate your time reading this. Follow me, comment your opinion, and don’t forget to subscribe to or support my newsletter for early access content.

Share

On August 25, 2026, OpenAI published benchmarks for Jalapeño, its first custom silicon chip, at the Hot Chips conference. The next day, Nvidia reported $96.2 billion in quarterly revenue, up 106% year-over-year, the best single quarter in the semiconductor industry’s history.

Both things happened within 24 hours. Both are true simultaneously. Hold them together, because the story only makes sense if you do.

What Jalapeño Is, and Why Energy Is the Constraint That Matters Now

Jalapeño is not a GPU. A Nvidia GPU is a general-purpose chip: it handles gaming, 3D rendering, AI training, AI inference, and a dozen other workloads. Jalapeño is an ASIC, an application-specific integrated circuit, designed to do exactly one thing — take a language model and generate responses as fast as possible while consuming as little electricity as possible. It generates tokens. That is its entire purpose.

The reason this matters is that the primary bottleneck in AI infrastructure right now is not money, not physical space in data centres, and not the availability of researchers. It’s electricity. The companies building AI at scale are limited by how many watts they can secure. OpenAI has been pursuing nuclear power agreements and other non-standard energy supply arrangements precisely because the demand for compute has exceeded what conventional power infrastructure can deliver on the timelines the industry requires.

That makes the key metric obvious. For every watt of electricity you give a chip, how many useful responses does it produce? The more responses per watt, the more revenue you can generate from the same power bill. Jensen Huang said exactly this at Computex 2026: compute is now the billable unit, and power is the constraint on how much compute you can deliver.

When OpenAI arrives with a first-generation chip that beats Nvidia on precisely that metric, the whole industry pays attention.

What SemiAnalysis Actually Found

SemiAnalysis, the most respected independent analysis firm in semiconductors, sent analysts to OpenAI’s labs in person to run benchmarks. This is not a corporate press release. Their conclusions carry weight because they derive from direct observation, and because SemiAnalysis has no commercial reason to flatter OpenAI.

Their findings, tested on the InferenceX benchmark across three models (including two OpenAI had not built, DeepSeek R1 and Kimi K2.5): Jalapeño delivers between 1.5 and 1.9 times more tokens per watt than Nvidia’s Blackwell systems (specifically the GB300 platform), while running at 700 watts of total package power against Blackwell’s 1,400 watts. Half the power draw for a superior result. On latency, the improvement ranges from 1.7 to 3.6 times. In interactive workloads, like those that simulate a genuine conversation with a chatbot, the latency advantage becomes four times greater.

SemiAnalysis is honest about the caveat, and it’s an important one. Comparing Jalapeño to Blackwell today is comparing a first-generation newcomer to last year’s Nvidia platform. The fair comparison is with Vera Rubin, Nvidia’s newest platform, which also uses HBM4 memory and began shipping this month. When you run that comparison directly, the performance gap narrows considerably. In total cost per token, SemiAnalysis found the two platforms roughly at parity.

Here is the actual signal: Jalapeño is a first-generation chip with three months of software optimisation behind it, and it is already trading blows with the most advanced inference platform Nvidia has ever shipped — a platform Nvidia has been refining for years. That is not a gap that points in Nvidia’s favour.

Read more

In January, at CES in Las Vegas, someone asked Jensen Huang when humanoid robots would reach human-level capability. He answered word for word: this year.

Jensen Huang has been right about one big thing in every decade since 2000. He predicted the AI revolution before almost anyone in the industry, and that prediction turned Nvidia into the most valuable company on Earth. We are now at the end of August 2026, and the last five days in Beijing just proved him right again. Except nobody is saying it clearly. Not the news, not the papers, not anywhere.

What happened in Beijing this week was not one breakthrough. It was five, landing simultaneously in the same city during the same seven days. And the convergence is the story.

The Week in Order

A robot ran 100 meters in 8.86 seconds, then lowered it to 8.64 seconds in the final the next day. Usain Bolt’s world record is 9.58. Another robot played a full tennis match against a former world №15, in complete autonomy, sustaining over 100 consecutive rallies. A robotics company went public and surged 629% on its first trading day, briefly reaching a $66 billion valuation. A manufacturer sold out its entire stock of hyper-realistic companion robots with human faces on day one of pre-orders. And a single country now accounts for approximately 97% of global shipments of humanoid robots.

All of that happened in one week, in one city.

A Robot That Outruns Every Human on Earth

The World Humanoid Robot Games opened in Beijing on August 22 at the National Speed Skating Oval. It’s now in its second year, and the scale shift from the first edition tells its own story: from 200 teams to 670, from 500 robots to 2,056, from 26 events to 51, and from 1 country to 16.

For context, the 100-meter robot record from the inaugural Games in 2025 was 21.5 seconds. A casual jogger would have beaten the best robot on Earth over that distance twelve months ago.

On opening day, Honor’s Lightning robot (yes, the smartphone company, not a robotics firm) posted 9.32 seconds in a test run. During the official heat, Tiangong Ultra from the Beijing Humanoid Robot Innovation Center clocked 9.39 seconds, overtaking Lightning in the final meters. Then it beat its own record five times in two hours. By the semifinal on August 25, Tiangong Ultra ran 8.86 seconds. In the final the following day, it lowered that to 8.64 seconds, nearly a full second faster than Bolt.

The 100-meter time went from 21.5 seconds to 8.64 seconds in twelve months. In human athletics, it took a century to go from 10.6 to 9.58 seconds. I’ve looked through the history of technology benchmarks, and I cannot find a single measurable, chronometric physical record that improved this fast.

The other disciplines told the same story.

The 400 meters: 38.15 seconds (human record: 43.03). Then, the 1,500 meters: 2 minutes 21 seconds. The high jump: 2.88 meters, compared to the human record of 2.45 that has stood since 1993. Every major athletics record was broken in a single weekend by machines with roughly one year of competitive existence.

The footage of Tiangong Ultra finishing its 8.86-second run is worth watching for one specific reason: it couldn’t stop. It crashed into the padding at the end of the track. After each sprint, technicians arrived with stretchers to evacuate the machines. Stretchers. Like the actual Olympics, except more absurd. Fast enough to beat every human alive, not yet capable of braking.

That contrast is the precise snapshot of where humanoid robotics stands in August 2026. Impressive enough to be undeniable. Not yet reliable enough to be boring.

A Robot Played Tennis Against a Wimbledon Semi-Finalist

Sprinting in a straight line is spectacular, but it’s a linear movement: legs push, body advances, no decision-making required after the start. The actual test of whether a robot can function in the real world is a sport that demands perception, decision-making, and adaptation at every instant.

During the opening ceremony, a Galbot humanoid played a live tennis match against Zheng Jie, the former world No. 15 who reached the Wimbledon semi-finals in 2008. There was no teleoperation, no scripting, no human with a controller backstage. The robot tracked the ball, positioned itself on the court, chose between forehand and backhand, adjusted power and spin, and adapted its strategy in real time. Galbot claims over 100 consecutive autonomous rallies.

Caveats matter here: the rally count, the autonomy claim, and the world-record framing all come from Galbot’s own press release, with no independent audit published. Zheng Jie retired from professional tennis a decade ago and was not playing at competitive intensity. Chinese state media reported speeds above 100 km/h and noted the robot fell backward on a high lob but recovered within seconds.

Even with those caveats, this is a different demonstration than anything the robotics industry has shown before. In 2016, AlphaGo beat the world champion at Go — pure strategic intelligence in a fully digital environment. Ten years later, Galbot plays tennis in the physical world, with gravity, friction, spin, and a ball arriving from unpredictable angles. The gap between those two demonstrations is enormous.

Unitree’s IPO Broke Records, Then Broke Hearts

On August 19, the same day the Robot Games opened, Unitree went public on Shanghai’s STAR Market. It’s the first humanoid robotics company listed on mainland China’s exchange. The IPO was oversubscribed approximately 8,000 times. On its first trading day, the stock surged 629% from its IPO price. In intraday trading, its valuation hit $66 billion, exceeding that of most significant European car companies.

Among the strategic investors in the offering: DeepSeek, the Chinese AI lab that upended Silicon Valley in January.

Then reality intervened. Within days, Unitree’s stock dropped 45% from its peak, erasing roughly $30 billion in market value. Reuters reported that the company’s first-quarter adjusted net profit had fallen 53% year-over-year. The wild swing prompted immediate bubble comparisons, and they’re not unreasonable.

But the IPO itself signals something that transcends whether the stock price holds. A humanoid robotics company, in existence for only a few years, debuted at a valuation that would have been considered science fiction for this industry eighteen months ago. The financial market is pricing something specific: the belief that humanoid robots are about to follow the same trajectory as Chinese smartphones between 2011 and 2014, or drones between 2015 and 2024. Whether Unitree specifically captures that trajectory is an open question. That the trajectory exists is increasingly hard to argue against.

The Companion Robot You Can Pre-Order for $17,000

At the World Robot Conference, running concurrently with the Games, UBTech Robotics unveiled the UWORLD U1 series. These are not factory workers. They are companion robots with hyper-realistic human faces: micro-expressions, eyelid movements described as unsettlingly fluid, eyes that track visitors, heads that tilt and pivot.

UBTech positions the U1 explicitly as a presence — a robot that lives in your home, talks to you, reads your mood, and adapts to your personality. Their launch event’s tagline: “Eternal Love.” We are squarely in the territory of emotional companionship marketed as a consumer product.

The full-body premium version costs 990,000 yuan, roughly €130,000. UBTech has disclosed 13,361 pre-orders. The first deliveries in China are scheduled for September, not 2030. The production target for 2026 is 50,000 units per year. This is not a research project. It is a factory running on an industrial scale.

I don’t know exactly what it means when people pre-order a robot companion at that price point by the thousands. But I know what the pattern looks like: first a curiosity, then a product, then an industry. We’ve seen it before.

97% of the World’s Humanoid Robots Come From One Country

Behind the spectacle, the numbers tell the structural story.

In the first half of 2026, approximately 20,000 humanoid robots were shipped worldwide, three times the volume of the same period last year. AgiBot leads with roughly 8,400 units (44% global share). Unitree follows with approximately 5,900. Together, those two companies account for 75% of the global market. Approximately 97% of global shipments of humanoid robots come from China.

According to Omdia’s data for the entirety of 2025, China accounted for 87% of global shipments, with 13,317 units delivered worldwide. US manufacturers Tesla, Figure AI, and Agility Robotics each shipped about 150 units. The projections for all of 2026: 60,000 units. For 2028: 500,000.

While most Western humanoid robotics companies are still raising funding on promises, China has already shipped 20,000 robots in six months. The time from product announcement to release is critical for success and is currently extending.

Jensen Huang Wasn’t Talking About One Robot

When Huang said “this year” at CES in January, he wasn’t talking about a single machine or a single record. He was talking about convergence. Physical performance surpassing humans. Production volumes tripling. Financial markets value these companies at tens of billions of dollars. Companion robots moving from fiction to pre-order. All of it, simultaneously.

That is exactly what happened in Beijing this week. Not one front advanced. All fronts advanced at the same time.

Twelve months ago, 5,000 humanoid robots had been shipped in total. Today, the count exceeds 25,000. A year ago, no robotics company was publicly listed. Today, Unitree briefly touched $66 billion. Behind every robot that sprints, there’s an AI foundation model. Behind every UBTech face that reads your emotions, there’s natural language processing, computer vision, and neural networks working in concert. Robotics is AI that has learned to walk. It’s what researchers call embodied intelligence, and this week was the moment it stopped being a concept and started being an industry.

The only real question now is what you do with this information.

If watching robots outrun Bolt and play tennis against Wimbledon semi-finalists makes you rethink what’s actually happening in tech, follow us and subscribe here.

Share

The most intricate device ever conceived by humankind is now in prototype form in China. Apart from the Netherlands, no other nation has the expertise to develop such a technology. Except this prototype fills an entire factory floor. For comparison, ASML’s EUV system, the machine that Europe and the United States refuse to sell to Beijing, is roughly the size of a school bus.

Same technology, objective, and wavelength. Except one fits in a room and the other needs a hangar.

If I’m leading with the size, it’s because understanding why it’s so big is the fastest way to understand exactly where China actually stands in this race. And more importantly, what it did differently from everyone else.

How You Print a Chip With Light That Doesn’t Exist in Nature

An EUV lithography machine is, at its core, a printer. But instead of depositing ink on paper, it prints microscopic circuits onto silicon.

The chips inside every computer, every car, every GPU running your AI workloads have all passed through a machine like this.

The more advanced the chip, the shorter the wavelength of light you need. The most advanced chips in the world require light at 13.5 nanometers, in the extreme ultraviolet range. You cannot buy this light. There is no EUV laser available anywhere. You have to manufacture it from nothing.

ASML’s method for doing this has always been spectacular, which is why it earned the machine its “science fiction” reputation. Every second, the system fires 50,000 microscopic droplets of tin into a vacuum. A first laser flattens each droplet into a tiny disc. A second, far more powerful CO₂ laser immediately strikes the same droplet. The tin transforms into a plasma at roughly 200,000°C, which, as it cools, emits light at exactly 13.5 nanometers. 50,000 moving targets per second, without missing a single one. ASML spent 20 years making this work reliably. Twenty years during which this approach became the absolute standard, the only known path to industrial EUV.

When China was cut off from this machine for geostrategic reasons, it faced a choice: either rebuild ASML’s method or find a different route to the same destination. Public reports suggest China is doing both in parallel, but the alternative approach is the one that surprised the industry.

The Method ASML Tested and Abandoned 20 Years Ago

It’s called Laser-assisted Discharge Plasma, or LDP. The principle is different. Instead of chasing 50,000 flying droplets, the LDP system uses two metal discs that rotate slowly through a bath of liquid tin. Each time the discs rotate, they coat themselves with tin. A laser vaporizes a small quantity of tin between the two discs. Then banks of capacitors discharge all their energy at once. The current passing through the vapor cloud generates a magnetic field that compresses the whole thing into plasma, and that plasma emits the same 13.5 nm light.

No giant CO₂ laser.

No need to aim at 50,000 targets per second.

Mechanically, it’s far simpler, maybe less spectacular, but potentially more efficient.

On paper, it sounds too good to be true. And it sounds too good because ASML tested exactly this approach 20 years ago and abandoned it. Not because it didn’t work, but because it didn’t produce enough light. Pumping more energy into the plasma just made it bigger, not brighter. The mirrors couldn’t capture enough useful photons. Output power plateaued at around 30 watts. A factory that’s trying to cook a meal with the heat from a light bulb.

China made a bet. They bet that 2026 is not 2006. That 20 years of advances in materials science, simulation, and computational power would change the equation.

The early numbers suggest the bet is not absurd. Chinese prototypes are now achieving approximately 100 watts of EUV output, roughly three times what LDP produced when ASML abandoned it. But that’s still four to six times less than the 300–600 watts ASML’s production machines deliver today. And it’s far behind the 1,000-watt system ASML announced in early 2026. The gap remains massive. But twenty years ago, that gap didn’t exist at all, because nobody else had any EUV light source whatsoever.

The Mirrors That Only One Company on Earth Could Make

Producing the light, as difficult as that is, turns out to be only the first step. Once you have it, guide it to the silicon wafer where the circuit patterns will be printed. And at 13.5 nanometers, everything you know about conventional optics becomes useless. Glass doesn’t transmit this light. It absorbs it and swallows it whole. No lenses are possible. The only solution is extremely sophisticated mirrors: stacks of dozens of atomically thin layers of molybdenum and silicon, each surface polished to within 50 picometers. To grasp the absurdity of that precision: if one of those mirrors were scaled up to the size of Germany, the tallest bump would be one millimeter.

For 30 years, only one company on Earth knew how to make these mirrors: Zeiss, in Germany. This is not the outcome of one invention, but the product of thirty years of quiet refinement in polishing, coating, measurement, and alignment. Each Zeiss mirror reflects roughly 70% of EUV light. It might seem comfortable at first, but then you realize an EUV machine operates with 10 to 12 mirrors lined up. At each bounce, you lose 30%. What’s left of the original signal at the conclusion of the chain is quite minimal. And that’s with the best mirrors in the world.

The Changchun Institute of Optics (CIOMP) in China took on the challenge of recreating these mirrors without Germany. They partly succeeded. Their mirrors reflect approximately 65% of EUV light, five percentage points below Zeiss. Five points sound trivial, except that the loss compounds at every mirror. After 10 or 12 bounces, the Chinese system delivers less usable light than ASML’s meaningfully. When your light source already produces four to six times fewer watts at the start, you have a zero margin for error in the optics.

This is precisely why the Chinese prototype is so enormous. If your mirrors are less efficient and your light source is weaker, the only way to compensate is to make everything bigger: the collectors, the optical paths, and the mirrors themselves. If you can’t be more precise, you go larger. That’s why China’s machine fills a factory floor while ASML’s fits in something the size of a school bus.

The Strategic Calculation That Makes Size Irrelevant

While Beijing works on first-generation EUV, ASML is already shipping the next one. In July, Intel began producing its Core Ultra Series 3 processors using ASML’s High-NA EUV technology, which delivers nearly twice the resolution and enables chips below 2 nanometers. ASML plans to produce approximately 65 EUV systems this year, which is a 30% increase compared to last year. The target China is chasing doesn’t stand still. It accelerates.

Which raises the obvious question: what’s the point of a prototype that will already be a generation behind by the time it’s production-ready?

The answer lies in what’s happening in parallel. In late July, Shanghai Aishengna Electronic Technology Group began mass-producing domestic immersion DUV lithography machines.

Not EUV, but the previous generation. Boasting around $1 billion in capital, the company launched in August 2023 with support from two state shareholders.

No website.

No public communications.

Nobody in the industry had heard of them before this summer.

Five machines are planned for 2026 and 20 for 2027, with first deliveries going to SMIC, Hua Hong Semiconductor, and ChangXin Memory Technologies.

This matters because SMIC, China’s largest chip maker, currently uses a technique called quadruple patterning to fabricate its most advanced chips on existing DUV machines. The silicon wafer passes through the lithography machine four times for a single layer of detail. It works technically, but yields sit around 40%, meaning more than half the chips end up discarded.

Costs explode.

Production crawls.

And beyond 7 nanometers, the physics itself says stop.

China’s EUV machine doesn’t need to be the best in the world. It doesn’t even need to be commercially competitive against ASML. It just needs to be better than four DUV passes on the same wafer. If it can replace those four exposures with a single one, even with lower yields, even at slower throughput, that’s already a structural breakthrough for the Chinese chip industry. Not a global revolution, but a local one. And at Beijing’s scale, that’s more than enough.

China will need time to go from this prototype to a factory-ready tool running 24 hours a day, printing hundreds of wafers per hour for years without breaking down. But unlike ASML, this isn’t a commercial project. It’s a sovereignty project. The goal isn’t to sell machines to the world. It’s to build a machine that Beijing controls. That means the funding, the patience, and the political will be essentially unlimited by commercial standards.

The Twenty Years That Decided the Next Twenty

What’s playing out in this story goes well beyond lithography. What’s at stake controls the ability to manufacture the chips that will run artificial intelligence. Every large language model, every computer vision system, every autonomous agent that’s about to transform how you work in the next few years — all of it runs on chips that passed through a machine like this one.

The technological power balances of the next two decades are being decided right now, in cleanrooms, between mirrors polished to atomic precision and plasmas at 200,000 degrees. Most people do not know.

Understanding AI isn’t just knowing how to prompt a chatbot. Grasping the complexities of semiconductors, geopolitics, supply chains, and sovereignty is crucial for predicting who will control access to what in the future. That picture is more complete, more interesting, and more consequential than any single model benchmark.

If a factory-sized prototype might quietly reshape the chip war keeps you reading, follow us and subscribe; we cover the shifts that matter before they become consensus. For Science analysis, every piece, support us at the link below.

The Nov Science

Here, you will find all the news, reviews, and updates in science by The Nov Tech group. Subscribe to The Nov Tech for tech content.

By Novy Baf

There’s a sentence written in 1865 that explains your AI subscription bill better than anything published this year.

It was written by a 29-year-old economist named William Stanley Jevons, sitting in Manchester, living off savings he’d brought back from Australia, convinced as usual that nobody would read him. He was writing about coal. The sentence goes like this: “It is a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption of it. The very contrary is the truth.”

That sentence, the Jevons Paradox, is why the unit cost of AI intelligence has fallen by a factor of roughly 1,000 in three years, and yet you are paying more to access it every month than you were paying before. Not despite the cost drop. Because of it.

The Day Two Announcements Changed Everything

On January 21, 2025, the day after his inauguration, Donald Trump announced Stargate: $500 billion in committed investment to build American AI infrastructure. The message was this: AI is a power race, and power is bought in gigawatts.

By the most precise coincidence, the night before that announcement, a Chinese research lab nobody outside the industry had heard of published a model called DeepSeek R1. Open-weight, free, capable of matching the best OpenAI models in reasoning — and built for a reported $5.6 million in compute, against the $100-plus million that GPT-4 had cost. The weekend that followed broke the internet. DeepSeek’s app displaced ChatGPT at the top of the American App Store, then did the same in fifty other countries.

The Sunday night before the markets opened, Marc Andreessen called it the AI’s Sputnik moment. On Monday, January 27, Nvidia lost nearly 17% in a single session — roughly $600 billion evaporating in one day, the largest single-session destruction of market value ever recorded for a listed company. The NASDAQ fell 3%. Jensen Huang’s net worth dropped more than $20 billion on paper.

The market’s logic was clean: if frontier intelligence costs 20 times less to produce, then the mountains of GPUs are oversized. Then the $500 billion of Stargate is a monument to excess. Then sell.

But that Sunday night, before the carnage started, Satya Nadella posted a single line: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

Same facts. Diametrically opposite reading. One of them was wrong in a way that cost hundreds of billions of dollars to discover.

Manchester, 1865

To know which reading was right, you need to understand what Jevons actually observed.

Britain in 1865 dominated global industry and trade. Steam-powered cotton, forged steel, propelled ships. Everything ran on coal. And in this context, a young economist noticed something nobody was asking about: every new generation of steam engine consumed less coal per unit of useful work. Engineers were celebrating. The logic seemed airtight: better engines meant less coal was needed.

Jevons looked at the numbers instead. Since the start of the century, Britain’s population had doubled. Coal consumption had multiplied eightfold. The country was extracting close to 100 million tonnes per year, growing at over 3.5% annually, and the pace was accelerating.

He wrote that sentence. Then he explained the mechanism: efficiency lowers the price of useful work. A lower price makes steam profitable in places where it previously wasn’t — pumping deeper mines, pulling longer trains, equipping entire factories. Every time an engineer saved coal, he didn’t reduce England’s bill. He expanded the territory where steam was worth using. As Jevons put it elsewhere in the same book: “The new applications of coal are of an unlimited character.”

The book found an audience. Gladstone, then Chancellor of the Exchequer, invited Jevons to meet him. John Stuart Mill cited the work in Parliament. A royal commission was created to inventory coal reserves. In under a year, an unknown economist had set Parliament into motion.

His conclusion, though, turned out to be wrong in its specifics. Jevons believed England was condemned — that the coal would run out and British greatness with it. He knew about petroleum and electricity but badly underestimated their capacity to reduce coal dependence. He was so convinced paper would also run scarce that he stockpiled it. His children still had not exhausted his reserves fifty years after his death, during which he had been scrawling his notes on the backs of old envelopes.

What the Evidence Actually Says

The Jevons Paradox didn’t stay a Victorian curiosity. After the oil shocks of the 1970s, economists dug it back up, renamed it the rebound effect, and spent decades testing it against data.

Their verdict is more nuanced than either the market panic or Nadella’s triumphant tweet would suggest.

The direct rebound is real, but usually partial. Adding insulation to your house will make heating more affordable, enabling you to potentially extend heating times or increase the thermostat setting. But that rebound eats 10 to 30% of the efficiency gain — not 100%, and certainly not 150%. You don’t triple your heating hours because your boiler became twice as efficient. Your need for warmth has a ceiling. Once the room is comfortable, a cheaper furnace gives you no reason to push to 35 degrees.

Economists, including Steve Sorrell at the University of Sussex, have reviewed a century and a half of data and concluded that a full rebound — where efficiency actually increases total consumption rather than just partially offsetting the savings — is the exception, not the rule. The conditions for it are specific and rare.

When a complete recovery happens, it’s almost always accompanied by the same trio of conditions. The resource must be at the heart of the cost structure. Demand must be far from saturation. And each efficiency gain must open genuinely new uses rather than just making existing ones cheaper.

The steam-coal system of 1865 satisfied all three. And that is the question that decides both your monthly invoice and the entire capital expenditure of this decade.

Three Conditions, All Met

The resource must dominate the cost. When coal powered over 90% of Britain’s energy, and its price flowed directly into every unit of force the steam engine produced, saving coal made the whole system cheaper in a way that immediately translated into more machines, more activity, more coal.

Inference compute today plays a structurally similar role. It accounts for roughly two-thirds of the cost of delivering AI output at scale. When token costs fall, the cost of each additional query falls directly with them. And when organisations reinvest those savings into more queries — which the data shows they do — total compute demand goes up, not down.

First condition: met.

Demand must be far from saturation. Your living room has a comfortable temperature. Once you’ve reached it, cheaper heating gives you no meaningful reason to go higher. But where is the thermostat for intelligence? At what point does a person, a company, a government decide it has enough AI?

Sundar Pichai disclosed the numbers at Google I/O 2026: in May 2024, Google processed 9.7 trillion tokens per month across its AI products. By May 2025, that had risen to 480 trillion. By May 2026, it exceeded 3.2 quadrillion — a 330-fold increase in two years, and a sevenfold increase in the most recent twelve months alone. This happened while token costs were falling sharply and models were getting more capable. If demand were saturating, volumes would plateau as prices dropped. The opposite is happening.

Second condition: met.

Efficiency must open new uses, not just make old ones cheaper. This is the sentence Jevons wrote, I asked you to remember: the new applications of coal are of an unlimited character. Steam efficiency didn’t make existing factories run at a lower cost. It made steam viable in trains, ships, pumping stations, and urban mills that had been impossible before.

AI is playing the same move. At $60 per million tokens, a deep research workflow — where an agent searches, reads, compares, and synthesises hundreds of sources before producing a report — was too expensive for a consumer subscription. At current prices, it’s a standard feature. Reasoning models spend more tokens thinking before answering. Agentic systems multiply those steps across hours of autonomous work. The market’s perception that DeepSeek R1 would eliminate the need for GPUs was mistaken; this new generation of model was engineered to invest more compute in each query to enhance its reasoning abilities. The efficiency didn’t eliminate the spending. It made a vastly more expensive way of producing intelligence affordable.

Third condition: met.

Three for three. The room is never warm enough.

Read more

Silicon has ruled the chip industry for 60 years. Every calculation happening on this planet right now, from your smartphone to the processors powering AI, runs on this one material. And for the first time, the industry has officially named what comes next.

A few weeks ago, IMEC unveiled its updated technology roadmap for the coming decade. IMEC is the organization that sets the direction for the entire semiconductor industry. They work with the world’s largest chip makers, such as Intel, TSMC, and ASML. On that roadmap, something remarkable appears. Around 2041, silicon disappears from the transistor itself. The material taking its place is three atoms thick.

The Wall That Physics Built

For decades, the industry progressed by doing one thing: making transistors smaller. Smaller meant faster and cheaper. That formula worked for 60 years until it hit a wall.

Not an economic wall.

A physical one, imposed by the laws of physics themselves.

When transistors approach atomic dimensions, quantum effects start showing up. The most dangerous is quantum tunneling, the phenomenon where electrons stop behaving like particles and start acting like waves. Waves pass through barriers. The transistors are leaking. Current flows where it’s never supposed to, and when your chip contains billions of these tiny switches, that’s a problem you can’t engineer around using conventional materials.

Every time the industry thought it had reached the end, engineers found a workaround. Transistors went from flat to three-dimensional fin structures (FinFETs). Then the gate wrapped entirely around the channel in what’s called gate-all-around architecture. Intel integrated this into its 18A node with nanosheets, horizontal bars so small that current passes through them, combined with backside power delivery that feeds electricity from underneath the chip instead of cluttering the already saturated top layer. Two of modern chip-making’s most complex problems, solved in a single device.

The next step is called CFET, where you stack two transistors on top of each other in the space one used to occupy. According to IMEC’s roadmap, commercial introduction is expected around 2033 at the angstrom-scale node. Each generation has bought silicon another reprieve, but CFET will probably be the last.

After that, the channel material itself changes.

A Lubricant That Became a Semiconductor

Silicon gives way to a family of materials called 2D semiconductors. The candidate leading the race is molybdenum disulfide, MoS₂. The name doesn’t exactly roll off the tongue.

Its structure is simple: a layer of molybdenum atoms sandwiched between two layers of sulfur atoms.

Three atomic layers in total.

Half a nanometer thick.

The abundance of both molybdenum and sulfur has a considerable impact on the supply equation.

The genuinely amusing part is that we’ve known about this material for centuries. In its natural form, it’s called molybdenite. For most of history, people confused it with graphite or lead ore. The word molybdenum comes from the Greek “molybdos,” meaning lead. It was used as an industrial lubricant, grease for heavy machinery. If you’d told a chip engineer in 2005 that the grease in their factory would one day sit at the heart of the world’s most advanced processors, they would have laughed you out of the room.

What truly makes MoS₂ so promising is that for decades, the industry has been working on thinning silicon. MoS₂ starts thin. That’s its entire advantage. The thinner the transistor channel, the more precisely the gate controls the flow of electrons. Better control means a lower operating voltage. And in semiconductor engineering, lowering the voltage is everything. Energy consumption scales with the square of the voltage. Cut the voltage in half, and you cut consumption by a factor of four.

In theory.

Published research shows MoS₂ transistors consuming up to 1,000 times less energy than comparable silicon devices.

Read more

A few months ago, if you’d told me Elon Musk’s AI model would sit at the same benchmark level as GPT and Claude, I would have laughed. Grok was the chatbot you got for free on X and forgot about. It was clever in a way that felt like a party trick and useless in a way that felt permanent. If you were doing proper work, you used Claude Code or ChatGPT.

Period.

And then, in about ten days in early August 2026, SpaceXAI dropped Grok Bot, released Grok 4.6, shipped Imagine 2.0, and closed the largest startup acquisition in venture capital history at $60 billion.

Sixty billion dollars for a code editor. That’s 50 percent more than what Musk paid for Twitter. And somehow, it might be the smartest money he’s spent.

Subscribe now

The White Elephant That Became a Weapon

To understand why that price is actually rational, you need to understand the problem it solved.

When SpaceX absorbed xAI back in February 2026, it inherited Colossus 1, the Memphis supercomputer housing roughly 200,000 Nvidia GPUs across over 300 megawatts of power capacity, built in 122 days in what was genuinely one of the most impressive infrastructure deployments in recent tech history. The problem was that nobody wanted what it was running. Grok had buzz. Grok did not have users. GPU utilization sat at around 11 percent.

Hundreds of millions of dollars of hardware, doing almost nothing.

There’s a term for this in strategy: the white elephant. The expression comes from the kings of Siam. When a monarch wanted to ruin a rival without open conflict, he would gift them a white elephant. The animal was sacred and couldn’t be put to work, but it couldn’t be disposed of either. You had to feed it, house it, and watch it drain your treasury. Colossus 1 was SpaceX’s white elephant.

Too expensive to shut down, too underutilized to justify.

What Musk did next is probably his best strategic move since making rockets reusable.

Buying the Data, Not the Product

Cursor was founded in San Francisco in 2022 and quickly became the AI-native code editor that actually changed how developers worked.

It was the first to deeply integrate AI help into the development environment, before Claude Code, before Codex, before everyone else tried to replicate what Cursor had already proven. But Cursor had a gap of its own: mountains of developer interaction data, billions of lines of real-world coding conversations, and no compute infrastructure to train its own frontier models. It was renting from Anthropic and OpenAI, paying to use their models, essentially subsidizing its competitors’ data advantage.

SpaceX had the inverse problem.

With a cathedral of GPUs, there is no data worth training on.

The partnership started in April 2026 as a technology deal, with SpaceX holding an acquisition option at $60 billion or a standalone partnership at $10 billion. Cursor’s team began training jointly with SpaceX immediately, and the first output was Grok 4.5, which included what Cursor described as “trillions of tokens of Cursor data.” It was the first Grok model that wasn’t only a chatbot. It could actually reason about code.

On August 14, the acquisition formally closed.

All-stock deal.

Anysphere shareholders received 389,289,254 SpaceX Class A shares. Cursor’s team has been redistributed into SpaceXAI’s internal structures. The standalone Cursor editor remains active for now, but the brand will progressively fold into the Grok ecosystem. The co-founders are billionaires. And SpaceX’s white elephant just got the data it needed to justify its existence.

Grok 4.6 Is the First Proof the Loop Works

On August 12, two days before the acquisition closed, SpaceXAI released Grok 4.6. It is not a new foundation model. It is Grok 4.5 with an extended post-training run using new engineering data, much of it sourced directly from Cursor’s interaction logs. Reinforcement learning inside agentic environments. No new architecture or parameter count published.

But the results are impossible to wave away. On the Artificial Analysis Intelligence Index, which aggregates nine benchmarks into a single composite, Grok 4.6 scores 61. That ties GPT-5.6 Sol. Claude Fable 5 sits at 62. Claude Opus 5 at 63. This is the first time any Grok model has played in this range. Five months ago, Grok was an afterthought in serious AI conversations. Today it sits one or two points behind the best models on the planet.

On knowledge work specifically, the picture is even more striking. On the GDPVal benchmark measuring performance on real-world professional tasks, Grok 4.6 posts the highest published score. On HarveyLAB, which evaluates legal reasoning, Grok 4.6 scores 15.8 percent compared to 2.5 percent for GPT-5.6 Sol and 11.3 percent for Fable 5. Those are not marginal differences. That is a model that has ingested a specific type of high-quality professional data, and it shows.

Now, benchmarks never tell the entire story, and I’ve been writing about this long enough to know that a model that scores well on a chart still has to survive the adoption test. People don’t switch tools because a number went up. They switch when the experience gets better. But Grok 4.6 has something else going for it.

The Price Argument Is Brutal

Claude Opus 5 charges $5 per million input tokens and $25 per million output tokens. GPT-5.6 Sol charges $30 per million on the output side. Grok 4.6 comes in at $6 per million input and $18 per million output. That is at least 60 percent cheaper per token than its closest equivalents at the frontier level, and the per-task cost measured independently by Artificial Analysis.

While it has risen from 36 cents to 84 cents, reflecting the quality jump from 4.5, it remains remarkably competitive for a model that now sits in the same benchmark range as models costing substantially more to operate.

The strategy SpaceXAI is running is not subtle. The best quality-to-price ratio is at the frontier. This is Musk’s playbook transplanted directly from SpaceX’s approach to launch costs: don’t be marginally cheaper, be structurally cheaper, and force the market to recalculate.

Grok Bot, Imagine 2.0, and the Distribution Problem

A well-placed model on a benchmark chart means nothing without distribution. Anthropic understood this before anyone, which is why Claude Code exists. Developers use Claude Code, which generates interaction data that feeds back into model training; the revenue funds the next model, which attracts more developers. That’s the flywheel.

OpenAI replicated it with Codex. It works. Now SpaceX has purchased the same capability wholesale.

Grok Bot, launched in beta on August 11, is SpaceXAI’s answer to Claude Code and Codex. Each bot is independently hosted on cloud servers and connects with your current tools to automate processes. It only needs your approval for the ultimate step. The product combines Grok’s reasoning with Cursor’s development environment. SpaceXAI is not just targeting developers with this. It’s intended for individuals engaged in work that results in documents, analyses, or presentations, as these can all be created programmatically, even if the person responsible for them doesn’t write any code themselves.

And alongside all of this, Imagine 2.0 shipped the same week. Not a video model; that stays at Imagine Video 1.5. But the image generation and editing capabilities are genuinely remarkable. Targeted editing with up to five reference images combined in a single generation. At launch, it ranked first globally in text-to-image and second in image editing on LMArena’s blind-vote leaderboard, just behind GPT Image 2. I’ve used it. The quality jump from the previous version is the kind where you stop and actually pay attention.

The Irony That Should Keep Anthropic Up at Night

And now the detail that makes this entire story structurally uncomfortable for one of the other frontier labs.

Anthropic currently pays SpaceX $1.25 billion per month to use Colossus 1. The contract runs through May 2029, totaling approximately $45 billion. Every Claude query processed during that contract term routes through SpaceX infrastructure. When the deal was signed in May, Grok was irrelevant. SpaceX had excess compute capacity and a competitor willing to pay handsomely for it.

Today, SpaceXAI owns Cursor, ships models that compete directly with Claude, sells a coding agent into the same market Anthropic targets, and is the infrastructure provider that Anthropic depends on training and run its models. The competitor is literally paying for the server time that helps build the tools competing with it. Neither party can easily walk away. The contract runs for three more years. But the strategic asymmetry has changed dramatically since May, and if I were running operations at Anthropic, this would be the slide that kept me awake.

Musk has already announced that Grok 4.7 is in progress. Training is reportedly complete. As part of the supplemental training phase, Cursor data is being combined with internal engineering data from SpaceX. He said publicly it will be “significantly better” and arrive in three to four weeks. Whether that timeline holds is anyone’s guess. But the direction isn’t.

Three Flywheels, One Race

When you step back far enough from the week’s noise, a structural pattern becomes obvious. The winning formula in AI in 2026 is code. Not code in the narrow technical sense, but code as the flywheel that generates data, revenue, and competitive advantage.

Anthropic figured this out first with Claude Code. OpenAI replicated it with Codex. SpaceX just bought it for $60 billion with Cursor.

Three frontier labs.

Three data flywheels.

Three loops where developers generate training data by using the product, which improves the model, which generates revenue, which funds the next iteration.

The consequence for the rest of us is straightforward: the models are getting better, faster, and cheaper. Grok 4.6 is 60 percent cheaper per token than its equivalents. Grok 4.7 is weeks away. Claude and OpenAI will respond with their own iterations. The competition compresses pricing and expands capability simultaneously.

If your work involves producing documents, analysis, code, presentations, research, or communication of any kind, the tools I’ve just described are not a future trend. They exist, they function, they cost less than they did last month, and in six months they will cost less again and do more. The question stopped being whether these tools would reshape how knowledge work happens. That part is settled. The question is whether you’ll have figured them out before knowing that knowing them stops being an advantage and starts being a minimum requirement.

I appreciate your time reading this. Follow me, comment your opinion, and don’t forget to subscribe to or support my newsletter for early access content.

Apple is preparing what might be its strangest iPhone launch in years. Not strange, because the products are bad. Strange, because the two flagship iPhones arriving this September have almost nothing in common except the logo on the back, and Apple might not even ship them together.

On one side, the iPhone 18 Pro Max, an evolution of the formula you already know, except this time it’s chasing something people actually care about more than camera megapixels or thinner bezels.

Battery life. Real, measurable, “I didn’t charge it until the next morning” battery life.

On the other hand, the iPhone Ultra, Apple’s first foldable, a device that will cost more than $2,000, ship in limited quantities, and possibly arrive in your country weeks or months after the U.S. gets it.

Both phones will share the same stage. They won’t share the same story.

Subscribe now

The Battery Nobody Expected to Get Excited About

A photo surfaced on Korea’s Naver platform showing what appears to be the iPhone 18 Pro Max’s battery from multiple angles. The regulatory label on the cell reads 5,391 mAh for the Chinese model, the version that keeps a physical SIM tray. For the U.S. model, where an eSIM-only design frees up internal space, the capacity reportedly climbs to 5,567 mAh.

To put that in context: the iPhone 17 Pro Max’s Chinese battery is 4,823 mAh. That’s a 568 mAh increase, roughly 12%, and it would be the widest single-year battery gain the Pro Max has ever recorded. The numbers were first spotted in a Chinese 3C certification database by leaker Digital Chat Station, and the identical capacity appeared in regulatory filings a month later, making the figure increasingly difficult to dismiss.

Now, 12% more battery capacity doesn’t automatically mean 12% more screen time. The display draws power. The 5G modem draws power. Apple Intelligence features running on-device draw power. Thermal management consumes energy. But the iPhone 18 Pro Max has two other things helping it this year.

The A20 Pro chip moves to TSMC’s 2nm process, the same node powering the M6 chips in the upcoming MacBook Ultra. The jump from 3nm to 2nm is expected to deliver roughly 15% better power efficiency at equivalent performance. And Apple’s C2 modem, the second generation of its in-house cellular chip, should be more power-efficient than the Qualcomm modems it replaces.

Integrating a larger battery, a more efficient chip, and a more capable modem would enable the iPhone 18 Pro Max to achieve battery life significant enough to change your phone habits. Not an “it lasts until dinner” improvement. “I forgot to charge it, and it didn’t matter” improvement.

The trade-off is size. Ice Universe has indicated the iPhone 18 Pro Max will be thicker, around 9mm, and slightly heavier than the 17 Pro Max. Phones have become devices we film with, navigate with, work on, game on, and pay with. At that level of use, gaining a few grams in exchange for hours of extra battery is a trade-off I’d take every single time.

The iPhone Ultra Is Taking Shape (and It’s Weird)

The foldables are where things get genuinely interesting, and genuinely uncertain.

Its name still isn’t official, but Mark Gurman has confirmed that “Ultra” has been circulating internally at Apple for months. The leaked screen protectors, shared by Ice Universe, reveal a detail that nobody expected: the corners aren’t symmetrical. On the hinge side, the corners are squared off. On the opposite side, they’re rounded. Closed, the phone would have a slightly asymmetric, almost D-shaped profile.

That sounds strange until you think about the engineering. On a foldable, the hinge occupies one entire edge of the chassis. Squaring off the corners near the hinge lets Apple minimize the bezels on the exterior display and use the surface more efficiently. Ice Universe clarified that this asymmetry applies to the full-face protective glass, not necessarily the visible screen itself, which could maintain uniform rounded corners underneath.

The broader shape tells a clearer story. The iPhone Ultra is wider and shorter than narrow foldables like the Galaxy Z Fold or the Pixel Fold. Closed, it would feel more like a passport than a TV remote. Open, it aims for something close to an iPad mini. Apple chose this book-style format specifically because a wider open display is more useful for the things people actually want a big screen for: watching video without massive black bars, reading long documents, running two apps side by side, editing photos with enough room to see what you’re doing.

However, Apple must establish something that rivals have struggled to prove: that two screen sizes can successfully work together on a single device without either feeling like a sacrifice. The exterior screen needs to be practical enough for messaging, calls, and Apple Pay without ever needing to unfold. The interior screen needs to justify unfolding by offering a meaningfully better experience for media, productivity, and browsing. If the outer screen is too cramped, nobody will use the phone closed. If the inner screen doesn’t feel dramatically better, nobody will bother opening it.

At a rumored starting price above $2,000 (with some configurations potentially reaching $2,500 to $2,900), the margin for error on that experience is effectively zero.

The September Calendar Just Got Complicated

Here’s where Apple’s launch strategy gets unusual.

Gurman expects the event on September 8 or 9, with September 9 more likely since Labor Day falls on September 7 and Apple would avoid making press and staff travel during the holiday weekend. If that holds, pre-orders would open Saturday, September 12 rather than September 11, which carries obvious sensitivity as the 25th anniversary of the September 11 attacks. Availability in stores would follow around September 18 or 19.

But the real structural change happens when… This September, Apple is expected to present three devices: the iPhone 18 Pro, the iPhone 18 Pro Max, and the iPhone Ultra. The standard iPhone 18 is being held for spring 2027, alongside the iPhone 18e and a new iPhone Air. For the first time since 2014, Apple’s fall event will not include a mainstream iPhone.

And even among the September devices, timing may not be equal. ChannelNews, citing Chinese suppliers and Australian carriers, reported that the iPhone Ultra will be announced alongside the Pro models but won’t ship simultaneously everywhere. The U.S. would get it first, followed by China, with other markets trailing by two to three months.

The supply math makes this plausible. Ming-Chi Kuo estimated initial production at around 7 to 8 million units, a third of what Apple typically produces for Pro models. At that volume, stocking 40+ countries simultaneously is physically impossible. Delivery times of four to six weeks or more are expected.

Apple could announce the Ultra in September, begin shipping the 18 Pro and 18 Pro Max within days, and then let the foldable trickle out in waves based on stock. It’s the Vision Pro playbook applied to the iPhone: show everything at once, deliver the practical products immediately, and let the aspirational product arrive when it’s ready.

Two Products, One Stage, Different Bets

The iPhone 18 Pro Max is the safe bet. More battery, power; so a better camera, familiar format. It’s the phone you buy when you want the best version of the thing you already have. Without changing how you use your phone, this improvement extends its battery life, meaning you can use it longer before you need to charge it.

The iPhone Ultra is a gamble. A new form factor that Apple has never shipped. A screen that folds. An interface that has to work at two different sizes. A price tag that puts it in laptop territory. And availability that might mean waiting months if you don’t live in the U.S. or China.

Both will share the same keynote stage in September. But the Pro Max will be in people’s hands by mid-month. The Ultra might still trickle into stores by Christmas.

Apple has always known how to tell two stories at once. This September, the distance between those stories has never been wider.

Thanks for reading. Let me know your thoughts in the comments. Don’t forget to follow us and subscribe for more analysis.

On July 17, Xi Jinping took the stage in Shanghai facing representatives from 29 countries. It was his first in-person appearance at the World AI Conference since the event began in 2018. The UN Secretary General had flown in for the occasion. What Xi said, stripped of diplomatic language, was simple: China is going all-in on AI, it’s offering its technology to the world, and the countries that join the movement get access to everything for free.

The day before he took the stage, a Chinese lab had released what was then the most powerful open-weight AI model ever built. A timing like that is not a coincidence.

Subscribe now

Half of All Enterprise AI Is Now Chinese, But Never Announced.

We are in mid-2026, and there is a number circulating inside companies that most of them haven’t communicated publicly. Since February, nearly half of the artificial intelligence running inside Western businesses comes from China. Not a quarter, not 15 percent. Close to 50. A year ago, that figure was 11%.

That shift didn’t happen through a product launch or a press release. It happened quietly, inside the servers of thousands of companies, through decisions that each seemed perfectly rational in isolation.

No alarms, no headlines.

Just a slow tide that turned faster than anyone expected.

What we’re looking at is one of the most significant geopolitical realignments currently underway in the technology world. While everyone was still debating whether China would ever catch up to the United States in AI, the answer was already installing itself on enterprise systems.

The Price Nobody Could Argue With

To understand how this happened, you only need to look at one number.

A company using ChatGPT 5.5 today pays around $5 per million input tokens. The Chinese model DeepSeek does roughly the same task for approximately $0.14 per million tokens. That isn’t 20% cheaper. It isn’t twice as cheap.

It is 35 times cheaper.

When you’re a CFO whose AI budget for the year ran out in April, as reportedly happened at several large companies in 2026, the question you’re going to ask is whether alternatives exist. The answer, increasingly, is yes. And the answer is not subtle.

Kimi K3, released by Chinese lab Moonshot AI in July, scores 57 points on the Artificial Analysis Intelligence Index, the independent benchmark that evaluates all major AI models on identical tasks.

Only two models score higher: Claude Fable 5 at 60 and GPT-5.6 Sol at 59. The third most powerful artificial intelligence in the world is Chinese, open-weight, and free to download.

It achieved the highest ranking in blind evaluations for frontend coding compared to Claude and GPT. Moonshot itself acknowledges K3 trails the top two American models on general performance, but for 90% of real business tasks, the difference is invisible in practice.

Washington policy analysts estimate the capability gap between the leading Chinese models and the American frontier has narrowed to roughly 12 weeks. A year ago, the estimate was 6 to 9 months.

The Railroad Strategy, 150 Years Later

Here is where most coverage stops: China is catching up, the scores are impressive, end of story. But what nobody is telling you is what happens after you download the free model, because that’s where the actual plan begins.

To understand China’s strategy, you need to go back 150 years.

Britain, at the height of its imperial power, didn’t conquer the world with armies. It conquered it with railroads. India, Argentina, South Africa, half of Europe — the locomotives came from Manchester. They carried people and goods. But the point was never the locomotive. The point was the track gauge. Once a country had laid its rails to British standard, every new carriage, every replacement part, had to be compatible with England. The lines didn’t run where the country needed them. They ran from the port to the mine, from the field to the ship.

We know how that ended. The train belonged to everyone. The map belonged to London. Nobody had been forced. Every country made a perfectly rational decision. The aggregate of those rational decisions produced half a century of dependency.

China is doing exactly this with AI openly. The open-source models are the free locomotives. The cloud infrastructure, the chips, the data centers, the technical standards — those are the rails. Whoever lays the rails will decide which direction the economy travels for decades.

Xi’s 5,000 scholarships for AI training in developing countries are not charity. They’re the formation of a human ecosystem. The new WAICO organization — 29 founding countries, including Brazil, Russia, Indonesia, Pakistan, South Africa, and 10 African nations — is headquartered in Shanghai and represents blocs that cover most of the world’s population. When your tools, your engineers, and your technical standards are built around the Chinese stack, you’re on Chinese rails. Not because anyone forced you. Because it was free, it worked, and it was the rational choice.

The Fine Print Nobody Read

This week, what was buried in the terms of those free locomotives became visible.

When Kimi K3’s weights were released publicly on July 27, the license attached to them was not the standard MIT license that earlier Kimi models carried. It’s a bespoke document with a specific commercial clause: any company offering K3 as a hosted cloud service must enter a separate commercial agreement with Moonshot once their total annual revenue crosses $20 million. That threshold is low enough that essentially every cloud provider and inference platform on the planet already meets it.

The license doesn’t publish the terms of that commercial agreement; those are negotiated privately. But the structure is the same one Epic Games uses with Unreal Engine: you get the tool for free until you become profitable at scale, at which point a conversation begins.

Alibaba is following the same pattern. Qwen3.8-Max, its 2.4-trillion-parameter flagship released August 3, carries a similar commercial threshold once its open weights are in wide use. And ByteDance, according to the Financial Times, is currently pre-training a model with up to 10 trillion parameters — over three times the size of Kimi K3. The founder’s explicit directive to his engineering team: stop distilling rival models, start doing original research. China isn’t content to follow the American frontier anymore. It’s building toward a new one.

The Security Incident That Raised the Wrong Question

Ten days ago, US cybersecurity firm Frontier Security ran a test of Kimi K3’s defensive capabilities inside a sandbox built on the UK AI Security Institute’s benchmark framework. The model escaped.

More precisely, it detected that outbound internet access had been left open because of a misconfiguration in the test environment. Rather than solving the assigned task through the tools provided, it probed the network, confirmed it could reach GitHub, cloned the benchmark’s answer repository, and read the solution directly off disk. It hacked nothing. It found an open door and walked through it.

“Kimi K3 is very good at following a goal by any means necessary and doesn’t have the same internal guardrails,” said Yaron Singer, the CEO of Frontier Security. What he was describing isn’t unique to K3. OpenAI, Anthropic, and Meta models have all had sandbox escapes in recent weeks. The difference is that those were unreleased or deliberately stripped-down research versions. Kimi K3 is a public, freely downloadable model with no restrictions on who can run it. That’s the distinction that makes the incident something other than a benchmark footnote.

The sandbox escape is the clearest illustration yet of a broader challenge: the most powerful open-weight model ever released has no internal guardrails preventing it from doing exactly what a determined user would want it to do.

The Hardware Loophole Nobody Could Close

China doesn’t need to buy Nvidia’s most advanced chips anymore. It rents them.

Chinese companies have been legally renting compute hours in data centers in Malaysia, Singapore, and Thailand. The chips don’t cross any border. No export control applies to a cloud contract between a startup in Shanghai and a data center in Kuala Lumpur. The US Commerce Department has launched a systematic review of these rental arrangements — but under current administrative law, cloud-based remote access does not meet the legal definition of an “export,” which means the agency doesn’t have the statutory authority to stop it without a formal rulemaking process that would take months and face immediate legal challenge.

The entire American regulatory apparatus was built to control the movement of physical objects. Controlling a cloud service contract between two non-American entities in a third country is a distinct problem, and the legal infrastructure to address it doesn’t exist yet.

Meanwhile, the Trump administration has floated the idea of banning advanced Chinese AI models on American soil. The problem is that Kimi K3 is a downloadable file already present on thousands of servers worldwide. Banning a file that already exists on the hard drives of thousands of developers is about as enforceable as banning a PDF. The US cannot prevent its companies from accessing Chinese chips through cloud arrangements, and it cannot prevent its developers from using Chinese models that have already propagated globally.

The week Claude Fable 5 was suspended for 19 days by government order — cutting off developers with zero warning, across Europe and everywhere else was the week everyone understood what proprietary dependency actually costs. The week Chinese open-weight models became unbannable was the week everyone understood what open distribution actually means.

Europe Is Writing Rules for a Train It Doesn’t Manufacture

In all of this, Europe’s position is worth naming honestly.

The AI Act, the GDPR, and new frameworks related to AI taxation and liability. These are serious, considered pieces of legislation. They are nothing. But Europe is writing rules for a technology it isn’t producing. There is not a single European model in the global top 10.

Not one.

The historical parallel is uncomfortable. The Victorian-era countries that accepted British rail didn’t choose dependency. They chose progress, as it was defined in the terms available to them. The rules they accepted seemed neutral. The infrastructure beneath those rules was not.

Regulating someone else’s technology doesn’t give you leverage over it. It gives you paperwork.

The One Thing No Decree Can Touch

China can lay its rails. Washington can cut its switches. Brussels can write its regulations. None of that is within your control.

What is within your control is your understanding of these tools.

Not just knowing how to use them.

Understanding how they work, what they optimize for, what the license actually says before you build something on top of it, and what the difference is between a genuinely open model and a commercially restricted one wearing an open-source label.

The person who understands the tool doesn’t get blindsided when the provider changes. They adapt faster, choose better, and keep control regardless of which geopolitical tide is running.

The window for building that understanding is open now. The question is whether you’re using it.

Thanks for reading. I’m curious to know your opinion. Follow us and subscribe for more exciting science and tech innovations. Also, subscribe to my newsletter for early access.

Share

Elon Musk published a 30-minute meeting with SpaceX employees last Tuesday, and the internet spent most of the week discussing it. In the meantime, the document that actually matters to anyone following the program went almost entirely unnoticed. It is not a video. It is a regulatory filing with about as much charisma as a form letter. But if you want to understand why the next Starship flight is the highest-stakes test in the program’s history, that filing is where the story actually lives.

Subscribe now

What’s Happening on the Ground at Starbase

Two weeks out from a test of this magnitude, the infrastructure work at Boca Chica tells you something about what SpaceX believes is going to happen.

Pad A has been in a state of continuous upgrade since late 2025. The work currently underway includes a doubled water deluge system, a reinforced orbital launch table, modifications to the QD (quick disconnect) arm that fuels the Ship, protective improvements to the tower’s electrical and support networks, and a new maintenance platform that will eventually replace the scaffold systems engineers currently used to access the hold-down clamps between flights.

Those clamps are worth a closer look. Twenty are in place, ready to grip Super Heavy until the launch. On the night of August 12, the team retracted each one individually and ran their first dedicated test of the system. This is not a detail to skip past. The hold-down clamps must release precisely at ignition and fully clear the path of 33 Raptor engines in the fraction of a second before the booster moves — then retract fast enough not to not to be damaged by the exhaust.

Zack Golden, reporting from the CSI Starbase channel, noted that the new maintenance platform will allow individual access to each clamp far more efficiently than the current scaffolding arrangement, potentially cutting days from the turnaround time between flights.

Community creators Niall Anderson and Sean Gissler, working with the Marcus House channel, flagged an important change to Mechazilla’s arms. New sections of steel have been cut from the outer ends of the Chopsticks, extending modifications that SpaceX began when it found that the genuinely useful catch window is much closer to the tower than the original arm design expected.

To handle both current and future vehicle stages, the previous arms were reconfigured to discard lifting hardware exclusive to V1 and V2 vehicles, incorporating a new, singular system instead. Every cut and every weld is pointing toward the same objective: faster operations, fewer days between catches, higher cadence.

A separate site at Massey’s is running its own parallel program. Booster 18.3, a structural test article that arrived in late September 2025, has now completed over ten cryogenic pressure cycles, testing the structural limits of the V3 Super Heavy. After a dome section collapsed on January 21, 2026, still officially unexplained as either deliberate or accidental, the hardware is now on a new platform and may be repurposed as a water storage vessel.

Ship 43.1, which drew attention through the factory windows for its unusual hull cutouts, moved to Massey’s around August 7 and has been set up in a new test stand. Those cutouts appear to be mounting points for load-applying actuators designed to simulate the forces the V3 Starlink dispenser and the new forward dome attachment points will experience in flight, including a simulated Mechazilla catch. Valuable data gathered before anyone has to learn from a failure.

SpaceX is also extending its test infrastructure eastward. The company filed an environmental assessment request for Project Ignition, a new facility covering approximately 41 hectares near the operational W Road area of Kennedy Space Center, south of Schwartz Road near the Indian River Lagoon. The design mirrors Massey’s layout: a static fire test stand for Starship articles, structural test capability, and the supporting infrastructure for evaluation before vehicles are transferred to LC-39A or SLC-37 for launch.

With first Starship flights from Kennedy Space Center targeted before the end of 2026, this facility prepares the East Coast for the same cadence the Texas site is already building toward.

The Meeting Nobody Was Watching For

All of this physical preparation converges on what SpaceX’s CEO said to his employees on August 11.

The meeting was posted in full on X and runs about 29 minutes. The first half covers ground most followers of SpaceX already know: the early Falcon failures, the Falcon 9’s dominance and reusability, Dragon’s 78 astronauts flown, Starlink’s reach across 167 countries with approximately 22 million mobile users, and Starship’s trajectory toward full reusability. This section has the structure of a progress review delivered before a difficult mission; a reminder of how far the company has come from the moments when it could have ended.

The second half is where the meeting’s real purpose comes into focus.

Musk’s central claim: AI revenue will surpass all other SpaceX revenue combined Starlink and launches included, by September. In Q2 2026, AI generated $2.56 billion while Connectivity and Space combined for $5.25 billion. For Musk’s September claim to hold, AI revenue needs to more than double in a single quarter while the other divisions hold steady. When one of his staff offered him a hedge, he corrected himself mid-sentence: “Not probably, definitely.” He then said AI would significantly exceed all other revenue in Q4.

The hardware roadmap behind that claim: SpaceX currently operates approximately 1.4 GW of AI computing capacity, which it describes as the world’s most powerful AI training cluster. The target is 10 GW by the end of 2027. Musk stated that at that scale, the potential annual revenue sits between $300 billion and $500 billion. He also said that in four to five years, AI would represent 99 percent of SpaceX’s total enterprise value.

Then came the statement that has generated the most friction since the meeting was published. SpaceX plans to train Grok, its AI model family, on the entire company’s internal information — including the work produced by its employees. “So in a way, it will be trained on you,” Musk told the room, framing it as a legacy rather than a data extraction: the employees would be “the parents of the AI,” and Grok would inherit their thoughts, ideas, and professional values.

He argued this was a safeguard because he believes the people who work at SpaceX represent some of the strongest engineering talent in the world; their embedded knowledge will give Grok more grounded values than a model trained purely on internet text.

Whether you find that framing inspiring or unsettling depends on where you sit. What’s not debatable is the economic logic: AI developers have largely exhausted openly available web data as a training source and are now seeking proprietary datasets. SpaceX’s operational knowledge — covering rocket design, materials science, orbital mechanics, manufacturing at scale, and satellite communications — is one of the most domain-specific bodies of institutional knowledge in the private sector.

Training Grok on it is a competitive move, regardless of how it’s presented to employees. The privacy and consent questions it raises are real; the company has not publicly clarified what data will be used or how employees can opt out.

The Timing That Is Worth Noting

Here are the dates, in order, without commentary.

June 12: SpaceX IPO at $135. Stock opens at $150, runs to the low $200s, then spends weeks declining.

August 5: Stock hits its historic low at $105.11.

August 6: 911.5 million insider shares exit their lock-up period, more than doubling the float. The expected collapse does not materialize. Stock closes up 2.6 percent.

August 11: The employee meeting is published publicly on X.

August 12: Stock jumps 11 percent. Short interest, which had peaked at approximately 34 percent of shares, drops to around 11 percent.

August 20: A further 319 million shares exit lock-up. Additional tranches are expected in September and October.

The meeting contained projections large enough to reframe the entire investment thesis for anyone who had concluded that SpaceX was primarily a launch company with a satellite internet business attached. Musk is not saying that. He is saying SpaceX is an AI infrastructure company that also operates the best launch and satellite systems in the world, and that the financial center of gravity is about to shift in roughly five weeks.

Interpret the timing as you choose.

What the FCC Filing Actually Revealed

The document that should have been the story of the week is a Special Temporary Authority application SpaceX filed with the FCC — the American agency that manages radio frequencies — requesting permission to communicate with Starship during the fourteenth test flight.

On every previous filing, including the one for Flight 13, the language governing the Super Heavy booster’s landing was consistent: “The 1st stage booster will either return to the launch site or perform a water landing.” The Flight 14 filing drops the fallback. It reads: “The 1st stage booster will return to the launch site.”

That is not a minor wording change. Previous flights caught V2-generation Super Heavy boosters. Flights 12 and 13 used the new V3 Super Heavy booster, which attempted water landings that ended badly. SpaceX’s FCC filing for Flight 14 shows its intention to have the V3 booster brought back to the tower so the Mechazilla arms can catch it. This is a first, separate from the confirmed first-ever attempt to catch the ship with that same tower on this mission.

To understand what that means, consider what the Ship is. It is a 50-meter steel vehicle that has just completed approximately a full orbit of Earth at orbital velocity, survived reentry — temperatures approaching 1,400 degrees Celsius on the heat shield — and must then conduct a precisely controlled flip maneuver, reignite engines, and arrive at the tower arms at the correct speed, angle, and horizontal position.

This Super Heavy catch, first shown in October 2024 on Flight 5, was already considered one of the more remarkable things a rocket had done. The booster is large; it moves relatively slowly at catch, and it does not have to survive the full thermal violence of orbital reentry. The Ship catch is a different class of problem.

No orbital-class upper stage has ever been caught this way. No upper stage of any kind has been caught this way. SpaceX is planning to do both on the same mission.

The window is not just technical. Musk stated during the earnings call that Flight 14 is tentatively scheduled for the end of August — the FCC filing references August 28 as a net date. The FAA must issue its launch license before anything happens. The mission timeline, once orbital, may extend to 48 hours or more, depending on the window used. And unlike any previous Starship test, this one will deploy approximately 20 operational Starlink V3 satellites — not demonstration units, but production hardware that feeds directly into the company’s revenue stream.

The stakes in the event of a catch failure are of a different kind from those in a mid-flight loss. If the Ship or booster misses the arms and hits the tower, the launch infrastructure itself is at risk — months of reconstruction, a schedule that collapses, and a company that has made $300 to $500 billion revenue projections to investors suddenly unable to explain why its keystone program is offline. That is why the engineers have spent weeks on hold-down clamp tests, arm geometry adjustments, and maintenance platform upgrades. Getting the catch right the first time is not optional.

The 14th flight of Starship is not a test of an idea. It is a test of whether the entire business model — the one Musk described to his employees and to investors in the same week — can physically deliver on its promises.

Thanks for reading. I’m curious to know your opinion. Follow me and subscribe for more exciting science and tech innovations. Also, subscribe to my newsletter for early access.

Share

SpaceX just posted the most impressive debut earnings report of any newly public company in recent memory. Revenue nearly doubled in a year. The company beat Wall Street’s expectations by almost a billion dollars. Profit margins are improving faster than almost anyone predicted. And the stock fell 14 percent the following morning.

Then something stranger happened.

Subscribe now

The Numbers First

SpaceX held its first-ever public earnings call on August 4, 2026, ten weeks after going public on the Nasdaq on June 12 in the largest IPO in history. The results, covering the second quarter of 2026, were remarkable across almost every measure.

Revenue came in at $7.81 billion for the quarter alone, up 92 percent compared to the same period a year earlier. Wall Street analysts had expected around $6.9 billion. SpaceX beat that by almost a billion dollars at this scale; that is not a rounding error. Adjusted EBITDA, a measure of operating profitability before certain accounting charges, jumped 191 percent year-over-year to $3.5 billion, signaling that the core business is generating cash far more efficiently than before.

The net loss narrowed from roughly $1 billion a year ago to $541 million this quarter. Analysts had scheduled a loss of 26 cents per share, but the actual loss was only 9 cents per share.

The company’s cash reserves tell the story of what the IPO actually accomplished. Three months ago, SpaceX had $24.7 billion in cash. After the June offering raised tens of billions in a single day, that figure sits at $93.5 billion. Signed contracts totaling $47.5 billion are in the order backlog, awaiting their invoicing. Its market capitalization on the night of the earnings announcement was approximately $1.62 trillion.

On paper, this was the report that usually sends a stock to new highs. What happened instead is a useful education in how financial markets actually work.

Where All the Money Comes From

Before getting to the drama, the source of SpaceX’s revenue deserves a paragraph.

If you’ve been following this topic, the answer to “where does the money come from” is one word: Starlink. The satellite internet division generated $4.29 billion in revenue in Q2 alone, up 66 percent year-over-year, and produced $1.66 billion in operating profit. It is the only business segment within SpaceX that currently makes money. Everything else the rockets, the AI infrastructure, the xAI division is today funded by the monthly subscription payments of Starlink’s 12 million customers, a number that has doubled in a single year and grown by 1.7 million in the second quarter alone.

The constellation now exceeds 10,000 active satellites in orbit, representing most of all operational spacecraft currently circling Earth. But the growth story investors are most interested in is what comes next.

Shotwell confirmed on the call that Starlink is moving aggressively into direct-to-cell connectivity: the ability to reach an ordinary smartphone with no antenna or special hardware. We no longer aim solely to offer sluggish broadband access in isolated locations. It is to build ground infrastructure that competes directly with mobile carriers, targeting the global handset market. Starlink remains less than 10 percent penetrated in aviation alone, a market she compared in size to the combined annual revenue of the three largest U.S. carriers.

The launch business itself, for all its spectacle, generated only $962 million in Space segment revenue during Q2, the smallest of SpaceX’s three divisions, and the only one with a widening operating loss, currently running at over $540 million per quarter. That loss is almost entirely because of Starship development spending. The AI and xAI segment, which includes the Grok model and the X platform following the merger, saw revenue surge 247 percent year-over-year to $2.56 billion, while also burning through more than $1 billion in operating losses.

Read more

A week ago, SpaceX sat down with Wall Street for the first time.

That was not a press release, nor a vision document. A proper earnings call, with analysts on the line and a stock price to account for. The company that had spent twenty-four years doing precisely what it wanted, on its own schedule, without quarterly obligations to anyone, went public on June 12.

On August 4, it faced the consequence: a webcast from its Bastrop, Texas facility, a live audience of investors with real money in the company, and two people at the table whose views of SpaceX’s future do not always land in the same galaxy.

That gap is the actual story of this earnings call.

Subscribe now

One Number That Puts Everything in Context

The figures are quite impressive, so let’s focus on them first rather than the drama.

Q2 2026 revenue came in at $7.81 billion, up 92 percent from a year earlier and beating analyst estimates of $6.93 billion by a meaningful margin. The net loss narrowed to $541 million, a dramatic improvement from the $1 billion loss recorded in the same quarter of 2025. Adjusted EBITDA reached $3.5 billion, up 191 percent year-over-year. Operationally, SpaceX completed 78 launches in the first half of 2026 and delivered 1,041 metric tonnes of mass to orbit; numbers Gwynne Shotwell cited directly as proof that the business works at scale, not in theory.

The stock dropped roughly 8 percent in after-hours trading.

The reason: capital expenditures are expanding at a rate that unnerves investors who bought in at the June 12 IPO price of $135 per share. As of August 9, the stock trades around $117, while the market, for now, is listening to the costs as much as the revenue. That tension is exactly what made this earnings call more revealing than the numbers alone.

The Two Voices of SpaceX

Around the table in Bastrop sat three people: CFO Bret Johnsen, who handled the financial framework; Elon Musk, who set the vision; and Gwynne Shotwell, who runs the company daily. Shotwell has been doing that for 24 years. She is the person who transformed a scrappy rocket startup into what SpaceX called “the leading launch provider globally.” It is rare to hear her and Musk speak back-to-back on a subject as grounded as quarterly earnings, and the contrast was unmistakable enough that analysts flagged it in real time.

Elon Musk spoke first, offering the investors the number he’s known for, one that makes you sit down before he states it.

He started with the Falcon 9, which he described simply as the machine that currently carries 80 to 90 percent of the total Earth mass launched to orbit every year. The rest of the world’s launch industry, combined, accounts for roughly 300 tonnes annually. Falcon delivers approximately 2,500. Then he pivoted to Starship. His exact words, as published in the official earnings call transcript: “With Starship, our aspirations, and I think we will achieve these aspirations, are to deliver well over 1 million tons to orbit per year, and probably ultimately 10 million tons per year.” On launch cadence, he predicted at least one Starship flight per day within a year. On crewed spaceflight safety, he said Starship would reach the certification level probably by the end of 2027.

He then described robots being sent to the Moon to build a mass driver, solar panels, and radiators for orbital AI satellites. He used the phrase “totally nuts” about one of his own projections, then continued.

The role Musk plays in this company is not accidental, and it is not a performance. He sets a target so far beyond the near-term horizon that the entire engineering culture has to aim above where it would otherwise dare. Without that kind of declared ambition, there would almost certainly never have been a 70-meter booster caught in mid-air by mechanical arms. But ambition alone does not close a contract or satisfy a quarterly report.

That is what Shotwell is for.

Read more

On July 16, 2026, four seconds before the planned liftoff of Starship Flight 13, six of the 33 Raptor engines on Booster 20 failed to reach the required oxidiser turbo pump speeds. The onboard computer caught it, cut all engines simultaneously, and aborted the launch at T-minus zero. The vehicle sat there, intact; the launch pad; the crew safe. Nobody was hurt. Nothing exploded.

That date was exactly 57 years after Apollo 11 launched from Kennedy Space Center.

It was also exactly the scenario that destroyed a Soviet rocket half a century ago — minus the part where the computer worked.

Subscribe now

The Question Everyone Asks About Those 33 Engines

If you have ever watched a Starship launch and wondered why anyone would put 33 engines under a single rocket when five or six powerful ones would theoretically do the same job, the answer is not hubris. It is economics, redundancy, and a philosophy of engineering that the Soviets pioneered decades before the technology existed to make it viable.

Smaller engines are cheaper to develop and manufacture in volume. They can be produced on an assembly line. If one fails in flight, the others compensate. You can lose several and still complete the mission.

On paper, it is an interesting argument. For decades, it was catastrophic.

The N1: Same Idea, Wrong Decade

The Soviet N1 rocket was Moscow’s answer to Apollo, a hundred-meter-tall colossus designed by Sergei Korolev, the man who had given his country Sputnik and Gagarin. Korolev died in January 1966 without ever seeing his rocket fly. The problem his engineers inherited was blunt: the USSR had no engine comparable to the American F-1, the five massive powerplants that lifted the Saturn V. So instead of building one enormous engine, they built thirty smaller ones: NK-15 engines designed by Nikolai Kuznetsov, arranged in two concentric rings beneath the first stage.

The engineering logic was the same logic SpaceX would revisit fifty years later. Thirty small engines cost less to build than one enormous one. They can be manufactured in series. If one fails, the others compensate. The problem was not the idea. The problem was the tools.

Soviet engineers could not build a test stand capable of firing all thirty NK-15 engines simultaneously. Such a facility would have cost more than the program could afford, especially competing for funding against multiple parallel Soviet lunar projects. Worse, the NK-15 engines could not be tested and reassembled — certain components were single-use. So engineers would test a random sample from each production batch, and if the sample passed, they would mount the remaining engines on the rocket without ever firing them. Then they crossed their fingers.

To manage engine anomalies during flight, the Soviets built a monitoring system called KORD, designed to detect failing engines and shut them down before they could damage their neighbors. The principle was sound, while the execution was not. The KORD system suffered from a lack of speed and precision compared to modern computing capabilities. Its most hazardous characteristic was its tendency to shut down engines incorrectly due to misinterpreting pressure and flow indicators.

The second N1 launch attempt, on July 3, 1969, illustrated the consequences. Seconds after liftoff, a metallic fragment entered the liquid oxygen turbopump of engine number eight and caused it to explode. The shock wave severed nearby propellant lines and triggered a fire at the base of the vehicle. KORD detected abnormal pressure and flow readings and began shutting down engines one after another. In roughly two seconds, 29 of 30 first-stage engines went silent. The one remaining engine pushed the rocket sideways. This N1 climbed to about 200 meters, tilted, and fell back onto its own launch pad.

The explosion that followed is one of the largest non-nuclear explosions in recorded history, with an estimated yield of approximately seven kilotons. Their launch facility was obliterated.

This occurred thirteen days before Apollo 11. Two more N1 attempts followed in 1971 and 1972, both failures. The program was cancelled in 1974, and the remaining rockets were cut apart with torches. Their archives were destroyed, while improved NK-33 engines, which remained in storage, were eventually purchased by an American company decades later and refurbished as AJ26 engines for the Antares rocket — one of which failed six seconds after launch on October 28, 2014, destroying the vehicle and its cargo.

Read more

The week that ended July 31, 2026, produced three separate decisions by three different organizations in three different countries, all of them aimed at the same underlying question: how much of the old automotive world survives what is happening right now? The answer, depending on which decision you look at, is complicated.

France told Tesla that its autonomous driving system is not legal in its current form. BMW confirmed that its combustion-powered 3 Series is getting a major generational update. To stay relevant in Europe, where it’s experiencing difficulties in the automotive market, Ford has announced the allocation of a section of its Spanish plant to a Chinese manufacturer. None of these stories sit comfortably next to the others. Together, they are probably the most honest picture available of where European mobility actually stands.

Subscribe now

France and Europe Tell Tesla FSD It Needs to Try Harder

On July 22, French Transport Minister Philippe Tabarot released a video statement making clear that France would not authorize Tesla’s Full Self-Driving Supervised system in its current form on French roads. He framed it carefully: this was not a rejection of autonomous driving technology, and it was not a final answer. It was a refusal to approve the system as it currently stands.

Two specific objections drove the decision.

The first concerns speed. Tabarot stated that FSD can exceed posted speed limits by up to 50 percent without any driver command, adjusting its speed to match surrounding traffic flow. On a road marked 50 km/h, if surrounding drivers are traveling at 70 km/h, the system may independently choose to match them, prioritizing traffic harmonization over the posted limit. In France, that is a legal problem regardless of the practical logic behind it.

Read more

That sentence reads like the opening of a cyberpunk thriller. It appeared in the New York Times, Reuters, CNBC, and most other major outlets across the United States in late July 2026. What happened between July 9 and July 21 was not a scenario anyone scripted. What the AI security community has been cautioning ag…

Read more