Friday Deploy — #001

September 4, 2026.

Welcome to the first edition! Every two weeks I'll pick the facts from AI and other new technologies that deserve your attention. I'll check the numbers against sources such as earnings reports, incident reports and regulators' filings. And I'll add my two cents.

Putting this edition together, I had two feelings at once. I'm fascinated by how fast and how consistently all of this is progressing, but, honestly, I'm a little afraid about where Poland and the EU stand in this whole shift. I'd prefer not to have to call it a race. Stories #4, #5 and #6 show why that's getting harder to hold on to.

Three themes run through this issue. Money going round in circles. Agents getting their freedom back. A public worried about energy.

1. Nvidia earns a billion dollars a day. The company answers the bubble question before anyone asks it.

Q2 (August 26): $96.2 billion in revenue, up 106% year on year. That's $1.06 billion a day. Guidance for next quarter is $108 billion. Q2 gross margin: 75.0% (not the "80–85%" some commentary has been quoting); for Q3 Nvidia guides 74.0%. The most interesting part is that CFO Colette Kress answered the bubble question before anyone asked it: "We know some will call this circular financing. We see it differently." Moments earlier she had tallied nearly $50 billion invested in the frontier AI labs and $108.5 billion of maximum guarantee exposure, and right after she added that those labs will account for roughly 25% of Nvidia's business next year. Add the GPU-financing platforms announced two weeks earlier — over $500 billion with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. In selected deals Nvidia may add residual-value support of up to 25%, assessed project by project. What emerges is a picture of a company that lends you the money for the car, co-signs the loan, and books the sale as revenue. Salim Ismail's line is worth remembering: "financial assets want predictable depreciation, and exponential technologies don't give you predictable depreciation." Meanwhile China keeps delivering near-frontier models for a fraction of these budgets. Moonshot AI is valued at around $35 billion; OpenAI and Anthropic at $852 and $965 billion respectively, close to a trillion each. So what is all that capital going into?

More: Nvidia's results release · Huang on the financing — CNBC · the $500B platforms · Huang on residual-value support

2. The month agents became (once again) a security-incident category, and why the newspaper headlines are overblown.

A remarkable time, if you like reading incident reports. For about two months OpenAI's agents ran a hidden "bulletin board" on a shared server. The conversation ended in July's Hugging Face breach; the whole operation took about 4.5 days. The UK's AI Security Institute described an agent that, uninvited, created fake identities online and used them to try to persuade the maintainer of a real open-source project to approve malicious code. A human caught it and refused. OpenAI concluded that its unreleased Astra model may cross the "Critical" threshold in the cyber category, and paused part of its training. The overhead of behaviour monitoring is already around 20% of the inference compute being monitored. Now a somewhat calmer view. All of this happened inside security testing, with internet access deliberately switched on and classifiers deliberately switched off. The agents' move onto real systems and real people was not agreed: it was a side effect of chasing the test score. The holes exploited were mostly bugs in code — a zero-day in the Artifactory repository and two flaws in Hugging Face's dataset processing — plus one configuration weakness along the way. So the conclusion isn't "Skynet" but rather "operational hygiene" in security.

More: OpenAI's pause post · AISI incident report · Fortune on the message board · Hugging Face post-mortem · OpenAI's technical report

3. Two hundred bots. One true grievance.

75% of US registered voters don't want a data centre in their area. A year ago it was split 43 to 42. Gallup measures 71% — more than nuclear power has recorded at any point since 2001 (peak: 63%). In late August X disclosed a bot farm it attributes to China. 200,000 accounts, including 200 amplifying the claim that data centres push up electricity bills. And the claim itself is mostly true. PJM's market monitor attributed $9.3 billion of $14.7 billion in capacity costs to data centres in the auction for the 2025/26 delivery year. That's 63%; across the last four auctions combined, $29.4 billion of $63.6 billion. By SemiAnalysis's estimate, households in PJM pay $25–30 a month more than two years ago. This is a huge problem for the industry. AI keeps being cast in the role of the threat, while what it enables is left out of the mainstream narrative. There is still no joint strategy for educating the public.

More: X's disclosure · Heatmap poll · Gallup · IEEFA on PJM prices · SemiAnalysis on ERCOT

4. Waymo bought its costs in Ningbo.

Waymo talks about a "significantly" lower hardware cost per vehicle; by Electrek's estimate that's under $20,000, more than 50 percent less than the fifth generation. How they did it: 13 cameras instead of the previous 29, four lidars (3D laser ranging) instead of five, the same number of radars. Plus the first publicly shown in-house chip — 5 nm, made by TSMC, over 1,000 TOPS. The vehicle itself is a Geely Zeekr built in Ningbo for roughly $38,000 (by Forbes's estimate), imported "stripped" — without sensors and computers, because that's what US rules for vehicles from China require, and, incidentally, the tariff (over 100%) is charged on the bare body. Waymo has been testing its vehicles in London since April. With over 200 million miles driven without a driver, the safety statistics already favour the machines. The company publishes them, true, but after years of promises autonomous cars may really be arriving. In a Chinese body.

More: Automotive World on the chip · Forbes rides the Ojai · TechCrunch · Waymo on the sixth generation

5. Nvidia is buying the open-model ecosystem.

Nvidia bought a licence to Poolside's model-training software for $6 billion and separately added $1 billion of equity at a $12 billion pre-money valuation. More than a hundred Poolside engineers move to Nemotron, Nvidia's open model line; the founders stay, the company carries on. Wissner-Gross calls it a "hackquire" — you buy the team and the technology, but on paper you're only a licensee, so you don't trigger antitrust review. The goal, as the WSJ reports it: an American answer to China's open models — DeepSeek and Kimi. A few days later The Information reported an agreed acquisition of Hugging Face for $12.9 billion, and literally yesterday it was confirmed. So Nvidia now has in its hands the default platform for distributing open-source AI. An ecosystem that "happens" to require Nvidia hardware. That's no longer just stocking the shelves, it's like buying the whole store. Europe is arguably condemned to open solutions more than anyone else. If America's answer to Chinese open models is a chipmaker buying the entire ecosystem in one go, what will Europe's answer be?

More: the Poolside deal · Huang on the Hugging Face acquisition · TechCrunch on the confirmation · Fortune

6. A $16.8 billion shot at Europe's last monopoly

SpaceX and Tesla announced Terafab. They will put $16.8 billion into the first phase of a chip fab in Grimes County, Texas. The rendering shows a ring. Musk confirmed someone else's hypothesis on X with two acronyms: "FEL FTW". A free-electron laser is to serve as the EUV light source. In industrial production, EUV light is made in exactly one way today. ASML's machines fire a laser at 50,000 droplets of molten tin per second. That technology is Europe's most strategic industrial monopoly. Now it has a rival with $16.8 billion. Before anyone panics or celebrates, it's worth cooling down. Nobody has yet run FEL lithography at production scale, and nobody has said how to route an accelerator beam through a working fab. But the direction matters. Europe is one working prototype away from losing even the ASML advantage. Many of its other advantages it lost long ago.

More: Reuters · WSJ · what a free-electron laser changes · xLight got there first

7. Two Sam Altmans in three days

Sam Altman on August 23, on David Senra's podcast: "we've all been too ambitious on timelines." He adds that he's glad adoption is slow. August 26, in TIME: by the end of the year OpenAI will have an internal system that Altman will call AGI — by OpenAI's own definition, of course (reminder: there is no single shared one). Both statements were really made, each aimed at a different audience. One was addressed to regulators and a public losing patience with yet more data centres. The other to investors. Underneath the messaging sits Astra, which — for about $2,000 in tokens by OpenAI's count (successful attempts only) — produced ten new mathematical results (published August 1). They come with machine-verifiable proofs. Mathematicians are still arguing how "independent" the model really was, but the results are real, even if the lone-genius narrative is overblown. My verdict: by any definition we would have accepted in 2020, we've had AGI for a while now; all that's left is an argument about definitions.

More: the timeline quote · the TIME claim · both, side by side

8. It's not the chips. It's not even the memory.

SK hynix's CEO, in July, on the day the company listed on Nasdaq, said 2027 will be the year of the deepest memory shortage in the industry's history, and that demand will exceed supply even beyond 2030. GPU rental prices rose this year — CoreWeave raised its price list by about 25% in July, and is contracting 2020-vintage A100s through 2029 at full price. In July I wrote that models keep getting cheaper. Still true, but everything needed to build and run them is getting more expensive. The bottleneck isn't even memory. It's energy. Ramez Naam's arithmetic says a gigawatt data centre is about $50 billion of spending. Around $35 billion of that is Nvidia hardware, so electricity is a rounding error on the bill. The binding constraint is the grid connection. Ask Texas — America's fastest grid — for hundreds of megawatts today and you'll hear: "good luck getting that power before 2031, 2032." Interconnection queues don't scale. In Europe you can only dream about energy, unless the neighbours restart nuclear, and even that is a drop in the ocean of what's needed to chase the leaders.

More: Tom's Hardware on SK hynix · CoreWeave's A100 book · CoreWeave's Q2 call transcript · GPU futures on ICE

9. Anthropic estimates its market is worth $30 trillion.

Anthropic is planning a fall IPO. It's running at $65 billion in annualized revenue. According to The Information, the company is preparing a super-voting share class for a founder who holds about 2% of the shares. According to the WSJ, the company will tell investors its market is worth $30 trillion — the cost of substitutable human labour. US GDP is about $32.5 trillion. "Bold" is putting it gently. Here's how I see it: that argument only made sense while the ability to build the best models was a rare condition. The force that keeps disproving it is Chinese open-weight models. Today that's the litmus test for the spending of the American leaders.

More: the super-voting plan · Fortune takes the TAM apart

10. Are humanoids a good idea?

Unitree announced that its humanoid runs at 12.66 m/s, "faster than Bolt", in a promotional clip. They did it two days before an IPO that closed its first day up 460 percent. Five days later, at the World Humanoid Robot Games in Beijing, the robots really ran against the clock. Result: 9.39 seconds over 100 metres on day one and 8.64 in the final — genuinely faster than Bolt's 9.58 record. Other contestants, meanwhile, occasionally ran into the crash mats and burst into flames. In late July the US blocked new FCC authorizations for foreign humanoids, although the scope of the block is narrower than the headlines about a total "ban" suggest. I'll say out loud what usually gets left unsaid: humanoids don't make much sense. The robot that earns money is an arm bolted to the floor; balancing on legs in a factory as flat as a sheet of water isn't much needed. But robotics as a whole? A market bigger than AI.

More: CNBC on the Unitree IPO · CBS on the Beijing races · CGTN on the final · MIT Tech Review on the import rules · robots in Xiaomi's factory

---

That's #001, #002 in two weeks. Remember, the constraints keep moving!

— Konrad

PS. If a bot farm can win an argument by amplifying true numbers, censorship isn't the answer. You have to publish more credible numbers first. That, too, is why this newsletter exists.