Someone Made Fire
Anthropic filed the first big pure-play AI IPO this week. But the model is the lotion, not the cure. What they're really taking public is the operation.

THE NUMBER: $1.25 billion a month — what Anthropic pays xAI, a direct competitor, to rent the Colossus supercomputer through 2029. The most valuable AI startup on earth can’t make enough of its own fire, so it leases the forge from the man it’s racing. That tells you where the real power sits, and it isn’t printed on the S-1.
In Duplicity, Tom Wilkinson plays a CEO named Howard Tully who’s sitting on a formula to regrow hair. He knows it’ll get stolen. He’s known from the first frame. So he doesn’t guard it. He runs an elaborate operation that lets his rival’s spies steal the wrong thing, a common skin lotion, while he keeps the lead time the real formula bought him. Then he gives the best speech in the movie. Someone made fire, he says. Someone was first. And until that knowledge got stolen, that person was the most powerful one alive. That’s not biology. That’s corporate evolution: the company that’s first and most gets to manufacture with the highest frequency, in the most desirable markets.
Anthropic filed a confidential S-1 on Monday. A $965 billion valuation, a $47 billion annualized run-rate, the first of the pure-play model labs to walk the paperwork into the SEC. SpaceX got there first and will probably price first, but SpaceX is a rocket company that rents out compute, not a lab that lives or dies on a frontier model. Among the two that do, Anthropic and OpenAI, Anthropic went first. Every newsletter on earth led with the valuation. Wrong fire.
Because the model is the lotion. Opus 4.8, the leaderboards, the “4x fewer coding errors,” that’s the formula everyone’s racing to copy, and they will. MiniMax shipped M3 this week and closed the gap again. The fire gets stolen every quarter now, and Anthropic knows it better than anyone, because half the time they’re the ones doing the stealing.
Here’s the part that should stop you cold. If the most powerful person is the one who makes the fire everyone wants, then right now that person isn’t Dario Amodei. It’s Jensen Huang. Trace Anthropic’s $1.25 billion a month upstream and it lands on 220,000 of Jensen’s GPUs. Two fire-makers run through this whole story. One sells the literal fire and owns the room today. The other is betting a trillion dollars that “today” is the one thing nobody gets to keep.
💲 Anthropic Filed the Paperwork. Follow the Money and It Ends in Santa Clara.
Start with the number that ran every front page. Anthropic’s Series H put the company at $965 billion, more valuable than OpenAI, on a run-rate that crossed $47 billion. The confidential S-1 makes it the first pure-play lab moving toward the public markets. Wilson Sonsini, the firm that ran Google’s 2004 IPO, is reportedly steering it. By the letter of Tully’s speech, this is what being first and most looks like.
Now read the cash-flow statement instead of the headline. Anthropic pays xAI $1.25 billion a month to rent Colossus, a contract that runs through 2029 and totals north of $40 billion. They lease from SpaceX too. The most valuable AI startup on the planet can’t make enough of its own fire. It buys the fire from the two companies it’s racing, and those companies buy their edge from Jensen.
That’s the tell finance people learn to read first. Whoever sits at the bottom of everyone’s cost stack, taking a toll no matter who wins the race upstairs, is the one actually getting paid. I used to short the stock the day a company announced its big SAP installation. Same instinct here, run in reverse: when every frontier lab is wiring $15 billion a year out the door to rent compute, you don’t bet on the labs. You look at who owns the meter.
None of that makes the IPO a bad bet. It makes it an honest one. Anthropic isn’t pretending it owns the forge. It’s raising public money precisely because the forge is rented and the fire is perishable, and the only way to win that game is to out-deploy everyone while the lead still exists.
The strategic read: When you evaluate an AI company, skip the valuation and find the toll. Ask who it pays, how much, and whether that counterparty also competes with it. Anthropic’s answer is “xAI, $1.25B a month, yes.” That’s not a reason to fade the stock. It’s the reason the IPO exists. Read every AI balance sheet this way and the power map looks nothing like the headlines.
⚡ Jensen’s Running the Exact Same Play
While Anthropic filed, NVIDIA (NASDAQ: NVDA) ran its GTC Taipei keynote and reminded everyone who actually makes the fire. Jensen Huang unveiled Vera, the first CPU built for AI agents, 1.8 times faster than x86, already spoken for by Anthropic, OpenAI, ByteDance, and Oracle (NYSE: ORCL). He shipped the RTX Spark, a chip he’s calling the most efficient PC silicon ever built, and the N1X, a Windows machine he’s pitching as a trillion-parameter desktop. The man whose data centers made him the most valuable company on earth spent his biggest night of the year selling you a computer that doesn’t need a data center.
Read that the way Tully would. Jensen knows the data-center fire gets commoditized eventually. Custom silicon is coming. Qualcomm (NASDAQ: QCOM) just cut its largest non-NVIDIA inference deal ever with ByteDance. China is openly building to route around him. Google has its own TPUs. So the fire-maker isn’t standing still and guarding the flame. He’s already moving inland, into the PC, into robotics with the open Isaac GR00T humanoid blueprint, into anywhere the next fire will be wanted. That’s not a hardware story. That’s a man who read the same script Anthropic did and reached the same conclusion: being first only pays if you’re already building the next thing before the current one gets copied.
We’ve seen this movie before. Cisco owned the plumbing of the internet in 1999 and printed money while everyone else fought over websites. The toll-taker won the decade. The difference this time is that Jensen isn’t waiting to be disrupted. He’s disrupting his own monopoly on purpose, on stage, in a leather jacket.
Why it matters: The most powerful company in AI right now sells picks and shovels, and it’s behaving like it knows the gold rush has a clock on it. If your strategy assumes today’s compute bottleneck is permanent, you’re planning for a world Jensen is personally dismantling. Watch where he expands next. He’s telling you which markets are about to matter, for free, in the keynote.
🦞 The Lotion Is the Model. The Cure Is the Operation.
So if everyone gets fire, and the fire-maker himself is hedging, what’s left to own? The operation. The harness, the forward-deployed engineers sitting inside Fortune 500s, the tie-ups with the big PE shops, and a deployment order that reads like a campaign map. Coding first. Then consulting. Then finance. Then whatever’s inland from there.
Coding isn’t the destination. It’s the beachhead, the market with the cleanest ROI and the fastest reproduction rate, the place you land before you move. And the beach is filling up. Anthropic’s Claude, OpenAI’s Codex, and a SpaceX holding a $60 billion option on Cursor are all fighting over the same developers. Three forges, one prize. Competition bends that margin curve down. Not this morning, but it bends, the way it always does once a market gets this crowded and this good.
Which is exactly why you don’t plant a flag in coding and sit there admiring it. You use the lead time to reach the next market before this one turns into a commodity. AI is Clayton Christensen’s nightmare with the safety off. Whatever you thought your moat was, somebody’s about to build a hundred cheap bridges across it. The only defense is to already be somewhere else, building the next position, when the bridges go up.
That reframes the whole inference-layer story too. The week Anthropic raised its round, the delivery layer quietly raised $763 million. Groq is rebuilding as an inference cloud after NVIDIA bought its IP for $20 billion. OpenRouter pulled $113 million after running its token volume from 5 trillion to 25 trillion in six months. When the model commoditizes, the money moves to whoever routes and delivers the tokens. Operations, again. Not the fire. The pipes that carry it.
What this means for your business: Stop buying the model and start buying the operation. When you pick an AI vendor, the benchmark score is the lotion, it’ll be stale in a quarter. Ask which vendor is actually building distribution into your industry, which one shows up with engineers instead of a sales deck. Then ask the same question about yourself: what’s your beachhead, and what are the three markets you reach from it before the first one commoditizes?
🔒 The Pentagon Fight Anthropic Lost on Purpose
Here’s the move that separates a good operation from a great one. In February, Anthropic refused the Department of Defense unrestricted access to Claude. Two red lines: no fully autonomous weapons, no domestic mass surveillance. Defense Secretary Pete Hegseth responded by branding Anthropic a supply-chain risk, a label normally reserved for foreign adversaries, and the administration ordered federal agencies off the company’s tech. Then a federal judge threw the designation out, calling it “classic illegal First Amendment retaliation.”
Walk the chessboard. Anthropic gave up a federal contract, took a public branding as a national-security risk, and won. It got martyred by the Pentagon and vindicated by a judge, and it did the whole thing on the values it’s about to sell to public-market investors. For a company that needs to attract the best safety researchers in the world and convince the enterprise that its AI is the trustworthy one, getting persecuted by Hegseth is worth more than the contract ever was. That’s the lotion-and-the-cure trick run on reputation. Let them take the thing that looks valuable. Keep the thing that actually is.
It rhymes with Project Glasswing, the cyber-defense effort Anthropic launched in April on Mythos, a model it still hasn’t fully released. The public ships Opus. The real capability sits in the basement. And it rhymes with the photo from two weeks ago, co-founder Chris Olah standing beside Pope Leo XIV at the Vatican as the Church released its AI encyclical. None of that is a model release. All of it is plumage, the display that pulls in the most desirable partners.
Watch what Dario himself is doing, or rather not doing. His X account has gone silent since April 7. Nothing on the $65 billion raise, nothing on the S-1. The filing went out under the corporate handle only. That’s a quiet period run with real discipline, and it lands as the exact opposite of his rival. Because on the very same Monday Anthropic filed, Florida Attorney General James Uthmeier sued OpenAI and Sam Altman personally, an 83-page complaint seeking to hold Altman individually liable over ChatGPT-linked deaths. One founder is being sued for what he said. The other has stopped saying anything at all.
The action item: Audit your own plumage. Every company now has a values story, and most of them are laminated posters in the break room. Anthropic’s cost it a Pentagon contract and it still didn’t blink. Ask what your stated values would survive if they were actually tested in public this quarter. If the answer is “we’d fold,” your competitors already know it, and so will the talent you’re trying to hire.
What This Means For You
Two fire-makers, one playbook. Jensen makes the literal fire and owns the room today. Anthropic is wagering a trillion dollars that the room doesn’t stay owned, because the fire always gets stolen, and the only durable thing is the machine that deploys it faster than the theft happens. Everything this week pointed at the same lesson: the idea is table stakes, the operation is the game.
Pick the market, not the model. The model you standardize on is a depreciating asset the day you sign. Build so you can swap it underneath in an afternoon, and spend your real energy choosing which markets you land in and in what order.
Find the toll, then decide if you’re paying it or collecting it. Anthropic pays $1.25 billion a month to rent its forge. Somewhere in your stack, you’re renting fire from someone who competes with you. Price that risk now, while it’s a line item and not a crisis.
Treat your values like an operating asset, because they are one. The companies winning the talent and the trust aren’t the ones with the best mission statement. They’re the ones whose mission survived a public test. Most won’t. That’s the opening.
The fire will be stolen. What you build with it before then is the only thing that was ever yours.
Three Questions We Think You Should Be Asking Yourself
Who makes the fire you depend on, and do they compete with you? Map every AI capability your business runs on down to the layer beneath it: the model, the inference, the silicon. Anywhere you’re renting from a company that also wants your market, you have a counterparty problem with a clock on it.
What’s your beachhead, and what are the three markets inland from it? Coding was never Anthropic’s destination, it was the landing zone. If you can name the market you’re winning today but not the next three you reach from it, you’re planting a flag instead of running a campaign.
If your current model were free and universal tomorrow, what would you still own? Assume the fire gets stolen, because it will. Strip out the model and look at what’s left: the distribution, the harness, the relationships, the values that survive a test. If nothing’s left, you don’t have a business yet. You have a lotion.
“Someone made fire. Someone was first.”
— Howard Tully, Duplicity (2009)
— Harry and Anthony
Sources
- Anthropic — confidential draft S-1 to the SEC
- CNBC — Anthropic IPO S-1 prospectus, $965B valuation, $47B run-rate
- TechCrunch — Anthropic to pay xAI $1.25B/month for compute through 2029
- CNBC — Anthropic, SpaceX compute deal
- TechCrunch — SpaceX’s $60B option on Cursor
- NVIDIA — Vera, the CPU for agents (GTC Taipei)
- StartupHub — the inference layer raised $763M the week Anthropic raised $65B
- GovConWire — the Pentagon’s view of Anthropic’s Mythos and Project Glasswing
- Anthropic — Project Glasswing
- CNBC — Florida sues OpenAI and Sam Altman personally
- Duplicity (2009), dir. Tony Gilroy — Howard Tully’s “corporate evolution” speech
Past Briefings
Knowing Where To Hit It
THE NUMBER: $500,000,000, spent twice in eight days, in opposite directions. One company torched it on AI tools in a single month because nobody set a spending cap, the corporate version of a lab rat pressing a lever until it forgets to eat. Kirkland & Ellis — $10.56 billion in revenue last year, $11.1 million in profit per partner across 595 of them — committed the same half-billion on purpose, to build its own AI instead of renting Harvey's. Harvey is worth $11 billion now. Legora, $5.55 billion. Same figure, spent by a rat and by a master. The whole...
May 28, 2026Your Company Needs A Harness, Not An Upgraded chatbot
THE NUMBER: 152% — the quarter-over-quarter growth of agent tokens running through Salesforce's Agentforce in Q1 (28.6 trillion tokens, 3.8 billion "agentic work units"). That is what real production AI looks like, on an actual income statement, when the buyer's harness is built and tuned. The same week, under best-case assumptions, the Financial Times put hyperscaler AI ROI at minus nine percent for Microsoft, minus thirty-five at Oracle. Same models. Same vendors. The variable is the harness — and exactly who built it. Every Formula One car on the grid is built to the same rules. A thousand pages of...
May 27, 2026I Know Kung Fu and AI
THE NUMBER: 11 — the seconds your doctor listens before interrupting you, on average. In 1984 the number was eighteen, so we've spent forty years getting worse at the one thing we keep insisting machines can't do. The thing American medicine ran out of is the thing the machine has in unlimited supply. The catch is that it only hands it over if you know to ask. Tank jacks the disk in. Trinity has to fly a military helicopter she has never touched. "Can you fly that thing?" "Not yet." A phone call, a few seconds of upload, and she...