A Trillion Dollars on a Six-Month Lease

00:00:04 Today's headline is the number. Anthropic raised 65 billion dollars at a 965 billion dollar post-money valuation. That puts them ahead of OpenAI, whose last private mark was 852 billion. The company says its annualized revenue crossed 47 billion dollars this month, up from about 7 billion across all of calendar 2025.

00:00:24 Their own press release calls this Series H. Altimeter Capital, Dragoneer, Greenoaks, and Sequoia are named as leads. Capital Group, Coatue, D1 Capital Partners, GIC, and ICONIQ are co-leading. And 15 billion of the headline number is previously committed hyperscaler money, including 5 billion from Amazon announced in April.

00:00:44 A few things to set up front. A 965 billion dollar private company is essentially a pre-IPO trillion-dollar firm. Anthropic and OpenAI are both reportedly preparing public-market filings in the coming months — OpenAI is said to be weeks away from filing its confidential prospectus.

00:01:02 So today's number lands as a price tag being walked across the threshold from private to public, where the markets get to decide whether the next leg of valuation gets added or chipped off. The investor list reads as the same names you'd expect — there's nothing novel in the cap table.

00:01:20 What's novel is the speed. From a 7 billion dollar annual figure for all of last year to a 47 billion dollar run rate by this month is roughly a seven-times jump in seventeen months. If you believe the run-rate number, Anthropic is one of the fastest revenue ramps in private-company history.

00:01:39 I have one note of skepticism, and I'll name it. Annualized run rate isn't the same as 47 billion dollars of money received in 2026. It's typically the most recent monthly number times twelve. So we're being told that May ran at about 4 billion dollars. That's an enormous figure.

00:01:56 Before I treat it as settled fact, I'd want to see the breakdown by enterprise contract, by application programming interface consumption, by hyperscaler reseller revenue, and by government work. Anthropic isn't a public company, so they don't have to show you that breakdown, and the New York Times piece that broke the number didn't include it either.

00:02:19 Treat the 47 billion as the company's framing, not as audited reality. Here's the world-facing piece. The price the markets are now putting on frontier AI capacity assumes those firms will keep extracting fees from the world's economy at a pace previously associated with the very largest software platforms — but with a cost structure of chips, power, and data centers that nobody has solved.

00:02:44 That's the bet. Today the bet got bigger.

00:02:46 The companion story to the valuation came earlier today, and it sits oddly next to the price tag. Earlier this month, xAI announced a compute deal with Anthropic for capacity on the Colossus 2 system. SpaceX's recently released S-1 filing — yes, SpaceX is going public too — described the arrangement as Anthropic agreeing to pay a monthly fee through May 2029, with capacity ramping in May this year.

00:03:11 The filing repeats the figure 1.25 billion dollars per month, through May 2029, on multiple pages. That's a contract on the order of 45 billion dollars over its stated term. Quite a lot of money. Today, Elon Musk got on X and said the public read of his own company's S-1 was wrong.

00:03:28 He wrote that SpaceX hasn't committed to leasing Colossus for years, and described the Anthropic arrangement as a 180-day lease with 90-day mutual cancellation rights thereafter. He added, and this is a direct quote, "The short term was our request, not Anthropic's.

00:03:44 We won't leave them hanging and will provide a reasonable off-ramp, but if compute gets super tight I said we might need it back at some point." Russell Brandom at TechCrunch describes the gap as "opinions vary." Both readings can be technically true — there's a six-month firm period followed by an at-will continuation through 2029 — but the way the contract gets described to public-market investors does different work than the way it gets described to compute counterparties.

00:04:28 Here's why this matters for the rest of the show. If Musk's version is right, Anthropic's compute floor at a major partner is essentially six months. If the S-1 version is right, SpaceX has a three-year guaranteed revenue stream from a counterparty that just printed a 965-billion-dollar valuation.

00:04:46 The difference between those two stories is who has leverage over whom in the compute market for the next three years. My read, held loosely. Musk gets to walk it back in public because the option in the contract gives him room. He also wants the market to read xAI as a buyer-of-last-resort for compute, not a vendor locked into a long lease with a competitor.

00:05:08 A three-year guaranteed-revenue framing flatters SpaceX's S-1. A six-month posture flatters xAI's competitive position against Anthropic. Same dollars, different audience. Place this next to the valuation news. The company that just got marked at a near-trillion-dollar valuation today is, by its largest compute counterparty's public framing, on a six-month lease for some of the capacity it's selling against.

00:05:34 The capacity story and the price story aren't as settled as the headline implies.

00:05:39 JD Vance delivered the commencement address at the U.S. Air Force Academy in Colorado Springs this morning, and the segment that ended up everywhere is about AI and war. I'll quote him at some length because the framing matters. "As AI transforms the battlefield — in some ways positively, in some ways not — I ask that you be jealous and selfish about your role as the decision-maker in warfare.

00:06:03 Use technology to make you better, but never submit to it. You are the masters of warfare. And both your minds, but also your hearts, are the opposite of artificial." "Your fellow Americans are, understandably, they're worried about AI, about how it will affect the labor market, how it will distribute resources, and how it has fundamentally changed how we interact with one another — our social lives.

00:06:32 But the thing I worry about most with AI is how it will change warfare." Then the line that will get clipped most: "If the warfare of the future is to live up to the moral values of our ancestors, decisions over life and death must be made by humans and not machines."

00:06:50 Vance has been one of the Trump administration's most enthusiastic pro-AI voices. Last year in Paris he told a roomful of European regulators to look at AI "with optimism rather than trepidation." He's the venture-capitalist vice president. So a speech in which he tells future Air Force officers to never submit to the technology marks a meaningful shift in posture, even if it's only at the level of public language.

00:07:16 The second piece is institutional. Vance praised the recent encyclical from Pope Leo the Fourteenth on AI, work, and human dignity in an NBC News interview earlier this week, and referenced it on stage. The Pope and the U.S. vice president are now reading from compatible scripts on lethal autonomous weapons.

00:07:35 That doesn't decide policy, but it changes the political room available to whoever's writing the next executive order. The third piece is the specific fight Vance didn't name. Defense Secretary Pete Hegseth has been in a running dispute with Anthropic over what military uses are permitted for the Mythos model — including, per NBC, in lethal autonomous weapons.

00:07:57 Anthropic has resisted Pentagon pressure to lift those red lines. When the vice president stands at a service academy and says decisions over life and death must be made by humans, he's saying something the company that just printed a 965-billion-dollar valuation today has been saying for months — against pressure from his own administration's defense secretary.

00:08:19 The political weather on autonomy moved a little today. One last point. President Trump declined at the last minute to sign an executive order this week that would have set up a voluntary mechanism for government testing of frontier models. Susie Wiles, Sean Cairncross, and Scott Bessent were reportedly pushing for it.

00:08:38 David Sacks and the lighter-touch faction won the round. Vance's speech today hints that the fight inside the building isn't over, and that the voluntary-review version may come back wearing different clothes.

00:08:51 While Washington argues over voluntary model testing, Brussels is reaching for something much heavier. The Financial Times reported today, citing a draft text, that the European Commission is preparing emergency powers to intervene in semiconductor supply chains during a shortage.

00:09:08 The draft is expected to be formally presented on June 3. The draft does three specific things. First, it lets the Commission force semiconductor manufacturers operating in Europe to prioritize orders for crisis-critical products, overriding existing private contracts.

00:09:24 Second, it adds a fine of up to 300,000 euros for companies that fail to disclose supply-chain capacity when asked. And third, the Commission gets the centralized-purchasing role it used during the pandemic for vaccines — acting as the single buyer for member states.

00:09:41 The trigger is a declared semiconductor crisis tied to shortages of inputs to weapons, medical devices, or digital infrastructure. This is the European Chips Act growing teeth. The original 2023 law was mostly about subsidy — billions to attract fabs to Europe.

00:09:56 The new draft is about intervention, actually moving silicon between counterparties in a crisis. Bruegel published an analysis recently arguing Europe's posture was shifting from self-sufficiency to indispensability — making sure other people need you, rather than trying to make everything yourself.

00:10:15 The contract-override power is what indispensability looks like when the lawyer writes it down. The chip industry will hate this. Forcing TSMC or Intel or Infineon to break a customer commitment in Europe in order to serve a Commission-declared priority is the kind of legal power that scares procurement teams everywhere.

00:10:34 If you're a German automaker or a French satellite firm, you might love it. If you're a downstream customer outside Europe trying to plan twelve months out, the existence of the power changes how you price counterparty risk on European fabs. The same day, the same Commission fined Temu 200 million euros under the Digital Services Act for failing to assess the systemic risks of illegal products on its platform — chargers that failed safety tests, baby toys with chemicals above EU limits, and detachable parts that pose suffocation hazards.

00:11:07 The Commission ran a mystery-shopping exercise to gather evidence and is giving Temu until August 28 to file an action plan. Temu calls the fine disproportionate. Two enforcement actions in one day, against two different kinds of foreign concentration. The first says: we will reach into your semiconductor contracts if our weapons and hospitals need them.

00:11:29 The second says: we will reach into your e-commerce risk assessments if your safety story isn't sourced. The pattern, if you want one, is the Commission stress-testing how far its regulatory authority extends into transnational supply chains. The U.S. used to be where that line got drawn.

00:11:46 Now Brussels is drawing one of its own.

00:11:48 The same day Brussels was drafting contract-override powers, France's Mistral announced two industrial partnerships that say something about what European AI sovereignty actually looks like in deployment. Mistral signed a five-year deal with Airbus and a separate partnership with BMW.

00:12:05 The contracts target very different problems, but the underlying logic is the same. The Airbus piece extends across commercial aircraft, helicopters, defense, and space. Priority areas, per the company, include onboard AI in aircraft and spacecraft. The list also covers automating technical documentation, accelerating engineering design through model-driven simulation, and edge models for automatic object recognition that support flight safety.

00:12:30 The defense piece is the operative line. An onboard model performing automatic object recognition for an Airbus military platform helps define how the next generation of European-flagged combat aircraft sees the world. The BMW deal is narrower and more interpretable.

00:12:46 BMW has an enormous archive of crash-simulation data, and the partnership uses Mistral models to improve the accuracy and speed of safety testing. Crash simulation is one of those backend industrial workloads that consumes vast compute, gets refined over decades, and benefits from a model trained on the right archive of data.

00:13:05 BMW gets faster product cycles. Mistral gets a sovereign reference customer in German auto, plus the proprietary data to keep training on. Behind both deals sits the political claim Mistral's chief executive Arthur Mensch has been making for two years: Europe's biggest obstacle to tech independence is the scale of investment necessary, and the company's response is to keep raising money and to keep arguing publicly that the only acceptable outcome is European frontier capacity.

00:13:32 The Wall Street Journal piece today quotes Mistral as accelerating "superintelligence" development to ensure Europe's independence from U.S. tech giants. The S-word does a lot of marketing here. The contracts handle the substance. The harder question — and I'll name it as opinion — is whether two industrial deals and an open-weights brand are enough to keep a European company at the frontier when its U.S.

00:13:56 and Chinese competitors are spending 65 billion dollars a round on capital. Probably not, on their own. But they create the conditions under which the next contract — a sovereign cloud deal, a Commission procurement, a defense ministry framework — can be signed without an American counterparty in the loop.

00:14:14 That's the indispensability play in action. Brussels writes the law that says European fabs must serve European priority orders. Paris writes the contracts that say European defense platforms run on European models. The two pieces do the same job from opposite ends.

00:14:29 Take the Anthropic and Mistral stories for a moment and place them next to what's happening on the buy side of enterprise AI. Today, Xavier Rivera on X compiled a thread that's been circulating all week: corporate America is in AI bill shock. The specific examples are documented elsewhere — Fortune, The Information, TheStreet, and BigGo Finance — and the pattern repeats across them.

00:14:53 Uber's chief technology officer told The Information in April that the company burned through its entire 2026 AI coding budget in four months. Engineers reported individual monthly application programming interface costs between 500 and 2,000 dollars per person.

00:15:10 Uber rolled out Claude Code to its engineering team in December last year. By February usage had doubled as developers discovered multi-step capabilities. By April the annual budget was gone. Uber's chief operating officer is now publicly questioning whether the AI spending is worth it, despite a reported 3.4 billion dollar cumulative spend with Anthropic.

00:15:32 Microsoft is more dramatic. On May 14, the company began revoking internal licenses for Claude Code across its Experiences and Devices division — the team responsible for Windows, Microsoft 365, Outlook, Teams, and Surface. The cutoff deadline is June 30. Engineers are being redirected to GitHub Copilot's command-line tool, Microsoft's own.

00:15:54 Six months earlier, Microsoft had rolled Claude Code out to thousands of those same engineers with enthusiasm. Then the bill came in. There's a specific story circulating, and I'll flag it as unverified, that one company burned 500 million dollars in a single month after leaving employee AI licenses uncapped.

00:16:14 I can't trace that to a primary source today, so take it as rumor rather than fact. But the structural point holds across the cases that are documented. Enterprises are layering usage-based AI fees on top of fixed payroll without cutting headcount. Better tools get used more.

00:16:31 More usage means a bigger bill. No natural ceiling exists unless someone caps it manually. What does this mean for the Anthropic valuation we opened with? The 47-billion-dollar run rate is being assembled, in part, out of contracts that finance teams at Anthropic's largest customers are now actively trying to cut.

00:16:51 The same week the company prints a near-trillion-dollar valuation on the strength of that run rate, the largest enterprise customer in tech is moving its engineers off the product. That's the normal lag between a usage explosion and the procurement renegotiation that follows.

00:17:09 But the second derivative of enterprise AI revenue might bend before the IPO prospectus does, and that's what public-market investors will price. The Microsoft retreat also carries a second motive. Tren Griffin has argued for the past two weeks that Microsoft is moving engineers from Claude Code to GitHub Copilot in part to dogfood its own harness, not just to cut Anthropic spend.

00:17:33 Both can be true. If the Copilot harness improves because Microsoft engineers are now stuck on it, Microsoft gains a product. If the bill drops at the same time, Microsoft gains a finance win. Anthropic loses internal advocacy at the world's largest software company either way.

00:17:51 Reuters broke a story this morning that fits the financial picture I've just described. The Shanghai Futures Exchange is in early design work on futures contracts tied to AI tokens — specifically the per-token pricing of AI services. At the same time, U.S. exchanges including CME Group and the Intercontinental Exchange are preparing to launch graphics-processor compute futures, tied to the cost of renting computing capacity.

00:18:16 Two markets, two designs, and the same underlying problem. Companies whose costs depend on inference can't hedge. Today, if you're Uber and you've committed to a 3.4 billion dollar Anthropic spend, you have no instrument to lock in the per-token price. If you're a Chinese cloud provider exposed to fluctuating compute costs, same problem.

00:18:36 Futures markets exist to let firms transfer that risk to speculators in exchange for a known price. We have oil futures because nobody wants to run a refinery without one. We're about to have token and compute futures for similar reasons. The Shanghai design is interesting because it picks tokens as the unit rather than chip-hours.

00:18:55 A token is the smallest unit of information a model processes — and crucially, the unit a company actually pays for when it buys application programming interface access. Chip-hours are the supplier-side abstraction. Tokens are the buyer-side abstraction. The U.S.

00:19:11 picked the supplier side. China picked the buyer side. Those are two different theories about which actor in the supply chain most needs the hedge, and which has the political weight to anchor a domestic financial market. One piece of context. Reuters notes that China's daily token usage has risen roughly a thousand-fold since the start of 2024, to more than 140 trillion by the end of March this year.

00:19:35 That's the demand curve a futures market is being built on top of. It's also a number to hold onto when people argue that Chinese AI is lagging — the consumption is enormous, even if the model leadership debate continues. What changes once these markets exist? A few things.

00:19:51 The per-token price stops being purely a vendor decision; it becomes a market quote that vendors negotiate against. Large customers can show their boards a hedged cost for next year's AI budget, which makes signing the contract easier and may extend enterprise AI growth past the bill-shock cliff.

00:20:09 A derivative market on AI capacity also invites speculation, which means somebody, somewhere, will eventually blow up trying to corner inference cost during a model release. One more piece I'll name. Once tokens have a futures curve, AI capacity becomes a financial asset the way oil and electricity are financial assets.

00:20:28 That changes how labs make capacity decisions, how governments think about national reserves, and how lenders price loans to data-center developers. The financialization of compute is a serious institutional event, and the fact that Shanghai and Chicago are designing it the same week makes today the day it began.

00:20:46 A reminder that everything I've discussed so far runs on physical infrastructure that sits somewhere. The Wall Street Journal reported today on Google's 15-billion-dollar AI data center hub being built in Visakhapatnam, in the Indian state of Andhra Pradesh. The build forms part of Google's five-year India commitment, and the Visakhapatnam campus is set to be the company's largest data center outside the United States.

00:21:11 The foundation stone was laid last month in Tarluvada village. The local concern, and the reason the Journal piece exists, is water. A typical one-megawatt data center system consumes roughly 25 to 26 million liters of water annually for cooling. The Visakhapatnam district has the lowest level of available groundwater for domestic, agricultural, and industrial use in the state, at 2.12 thousand million cubic feet as of April 1 this year.

00:21:37 Rights groups and local councils have raised what the World Resources Institute classifies as "extremely high" water stress. Mongabay's reporting today adds context. More than half of India's data centers sit in water-stressed regions. Data-center water consumption in India is projected to rise from about 150 billion liters in 2025 to 358 billion liters by 2030.

00:21:58 Google says it's partnering with an organization called Sponge Collaborative on a watershed management plan that integrates clean drinking-water systems and reverse-osmosis installations in nearby communities. Whether that's enough to offset the draw is the question rights groups are pressing.

00:22:15 This is the part of the AI build-out that doesn't show up on the Anthropic balance sheet or in the Shanghai futures contract. Someone has to host the compute. The countries offering the most aggressive incentives — billions in subsidy and fast-tracked approvals — are disproportionately the countries with the least water margin.

00:22:34 India is one. Saudi Arabia is another. So are parts of northern Mexico. The decision to put a hyperscale campus in Visakhapatnam, instead of Hyderabad or Pune, is a decision that local farmers and municipal water authorities will live with for thirty years. The institutional answer Google offers is community partnership and reverse-osmosis plants.

00:22:54 The institutional question Visakhapatnam locals are asking is whether the watershed plan will hold up in the third or fourth year of drought, when the campus and the city are competing for the same aquifer. We don't know the answer yet. The world's largest AI infrastructure builds are landing on the world's most water-stressed land, because that's where the subsidies are.

00:23:16 The cost is being absorbed by whoever happens to live downstream.

00:23:20 Five threads to track over the next week. Anthropic's IPO prospectus is the obvious one — sometime in the next sixty days, with the question of whether the 47-billion-dollar run-rate gets broken out by customer concentration when it lands. If a quarter of that revenue is one or two hyperscaler reseller arrangements, the picture differs sharply from what the headline suggests.

00:23:41 SpaceX's S-1 sits alongside it. Does it get amended to reconcile with Musk's public version of the Anthropic Colossus terms, or do SEC comments force a clearer disclosure? A clash between a filing and a tweet rarely survives an S-1 review. Then there's the executive order.

00:23:57 Trump declined to sign the rewritten version this week, and Vance's pro-human-decision speech at the Air Force Academy hints at where the next draft goes. The Hegseth-versus-Anthropic fight over Mythos in lethal autonomy is the test case for whether the administration's words on human judgment translate into procurement terms.

00:24:16 Brussels has a deadline. The chip-emergency draft has to survive June 3 with the contract-override power intact. If it does, and the Commission ever uses it, every fab operator in Europe will have to reprice counterparty risk — and that ripples into every chip contract signed in 2027.

00:24:32 One more: the water audit on Visakhapatnam. Andhra Pradesh's state government has a chance to publish a real groundwater study tied to the Google build. If it does, and the numbers are bad, the precedent for the next round of Indian data-center approvals shifts.

00:24:47 If it doesn't, the playbook stays "approve fast, manage later." Three different institutions, all reaching for control of the same arc. The next test of who actually has it is procurement — which contracts get signed, which orders get prioritized, which budgets get cut.

00:25:17 Jonas.