When Anthropic filed its IPO prospectus on Monday, the document did far more than mark another step toward the public markets. It pulled back the curtain on one of the most extraordinary financial build-outs in the short history of artificial intelligence, and the numbers are almost numbing. Anthropic is a company that only came into existence in 2021, yet it has already spent more than $19 billion on AI infrastructure—chips, data centers, networking gear, and the software needed to keep it all humming. That sounds enormous on its own, but it is only the opening act. As of June, the company was contractually on the hook for roughly $587 billion in chip and data center commitments stretching out to 2045. In the four months since, it signed another $266 billion in new and amended agreements. On top of that, there is about $20 billion in expected computing capacity from AMD, and an additional $125 to $150 billion for more Google-designed chips. Put it all together, and Anthropic’s obligations could approach $1 trillion by 2045. To give that some perspective, the consulting firm PwC estimates that the entire global AI buildout through 2050 will total about $31.6 trillion, meaning Anthropic alone could account for roughly 3 percent of all of it. Its biggest rival, OpenAI, is reportedly budgeting $750 billion for computing power through 2030. So Anthropic is not just participating in the AI infrastructure boom; it is making a wager big enough to reshape the industry, the economy, and the company itself. The prospectus reveals that existential scale, but it also lays bare the uncomfortable truth that such ambition rests on a mountain of promises that must be paid off long before any of us know whether the AI revolution will justify the cost.
For all the staggering expenditures, the revenue side of the ledger does give Anthropic a credible story to tell. Between January 2023 and June 2026, the company paid out $19.2 billion for compute and infrastructure, and over that same period it generated $21.3 billion in revenue. Most of that money came in very recently: $16.3 billion was recorded in just the first half of 2026, a jaw-dropping 1,300 percent increase compared to the same period a year earlier. Meanwhile, its compute and infrastructure expenses rose by a comparatively modest 283 percent. That means Anthropic is not spending faster than it can grow—at least not yet. It is beginning to show the kind of operating leverage that investors dream about, where a company builds enormous fixed infrastructure and then scales revenue on top of it without needing to duplicate the physical plant at every step. But there’s a catch. Roughly 80 percent of Anthropic’s revenue in the first half of the year came from consumption-based users, meaning people and businesses pay according to how much they actually use Claude, its AI assistant and API. Only about 20 percent came from subscriptions. That makes the company extremely sensitive to customer behavior, pricing pressure, and the pace of adoption. If customer usage slows, or if they defect to cheaper models, the revenue machine can sputter much faster than the cost base can be adjusted. Anthropic also faces a competitive landscape that is more crowded and more aggressive by the day, not just from OpenAI but from a growing array of open-weight models that are inexpensive, high-quality, and available to anyone. Columbia Business School economist Stijn van Nieuwerburgh put it plainly: their fortunes hinge on charging premium prices for the latest model. But in AI, there is always a next model from a competitor that does a little better. That makes the business inherently more volatile than a traditional software company with sticky enterprise contracts. Anthropic may be growing at an extraordinary rate, but it is growing in a very unforgiving environment, where each new release resets expectations and yesterday’s moat can become today’s also-ran.
A large part of the risk comes from the fact that Anthropic’s biggest contracts are not with neutral, arm’s-length suppliers. They are with its own investors, allies, and sometimes even its would-be competitors. The company has a $114.5 billion cloud computing and chip contract with Google, a $110 billion cloud computing contract with Amazon, and a $45 billion contract with Microsoft for computing and access to Nvidia systems. All three companies own stakes in Anthropic, though the S-1 filing does not list them on the cap table. Its largest single agreement is a $161 billion deal with Broadcom for chips and infrastructure, and Broadcom also holds $42 billion in convertible notes that could be converted into Anthropic shares. Anthropic expects to spend more than $20 billion on chips from AMD, and AMD agreed in July to buy up to $5 billion of Anthropic common stock. It also buys memory and storage from Micron, Samsung, and SK hynix, all of which participated in Anthropic’s Series H funding round in May, a round that valued the company at $965 billion. The exact stakes of those firms were not disclosed, but the pattern is unmistakable: the people lending, building, and selling to Anthropic are also the people betting on its stock. Private markets platform Tessera PE founder Chan Ahn calls this “circular financing,” and he notes it is becoming more prominent. Many of these relationships are intertwined in complex ways, so that if Anthropic stumbles, the pain would not stop at Anthropic’s shareholders. It would ripple through some of the largest technology companies in the world. In that sense, Anthropic may be too big to simply collapse—too many powerful players have an interest in keeping it afloat. But that interdependence cuts both ways. It also means that the company’s decisions, pricing, and growth targets are constrained by a network of obligations that are not purely commercial. They are financial, strategic, and sometimes personal. One of the most striking examples is Anthropic’s $120.5 billion agreement with Elon Musk’s SpaceX to access Nvidia GPU chips that SpaceX stockpiled. Once SpaceX delivers all that capacity, expected in December, Anthropic will owe $3.1 billion per month for it. That single deal would account for roughly 25 percent of SpaceX’s projected first-quarter revenue in 2027. It ties the fortunes of Anthropic to a rocket company, and it ties SpaceX’s financial projections to Anthropic’s ability to keep growing. They are, in effect, partners in a giant bet: that there will be enough demand for AI compute to justify one of the largest procurement deals ever signed.
The detailed risk factors in Anthropic’s prospectus read less like a standard legal disclosure and more like a warning about the fragility of large-scale infrastructure projects. Anthropic lists power shortages, construction delays, and growing local opposition to data centers as genuine threats to its expansion. Those are not abstract concerns. Many of its agreements contain take-or-pay clauses, meaning Anthropic has to make payments for twenty years whether or not it ever uses the capacity. The contracts include annual price increases and pass-through fuel costs with no cap, leaving the company exposed to rising energy prices. In some cases, it could face equipment bills before it can even use the hardware. Under its leases with Broadcom, rent payments can begin even if the data centers are not ready yet. That means Anthropic can be paying for machines and facilities that are still behind chain-link fences while construction crews wait for transformers and permits. Its commitments with Amazon, Google, AMD, and Microsoft total up to 13 gigawatts of data center capacity—enough electricity to power all of New York City—but much of that capacity is not yet deployed. The company’s existing data center footprint is mostly in the United States, with additional sites under development in Australia, Canada, and Norway. All of this is happening off the balance sheet. None of the nearly $1 trillion in leases and contracts appear in Anthropic’s traditional accounting, because the company only recognizes them as liabilities when the leases actually begin. In the S-1, Anthropic reported $117 billion in liabilities, mostly convertible notes, but the true scale of its future obligations is much larger and largely invisible using standard metrics. Investors are likely to struggle to value a company whose true leverage is hidden under layers of financing structures and long-term contracts. That opacity is not necessarily deceptive; it is the modern way of funding enormous infrastructure projects. But it adds another layer of risk. If power prices spike, if construction timelines slip, if the demand for AI compute comes in lower than expected, the company’s cash flows could be squeezed by obligations that were signed in a more optimistic moment.
It is not just Anthropic that is exposed to these risks; a whole ecosystem of smaller companies has built its business around Anthropic’s appetite for compute. The prospectus reveals just how dependent the so-called neoclouds are on Anthropic. Nscale, a neocloud that rents out chips and is currently pursuing its own IPO, relies on Anthropic for 43 percent of its total $103 billion in contracts. TeraWulf, a publicly traded bitcoin miner that is expanding into data centers, has leased nearly half of its contracted computing capacity to Anthropic. Hut 8, another data center operator, gets about 25 percent of its AI leases from Anthropic. These companies are not diversified like Google or Microsoft. They are leveraged specialists, buying huge amounts of AI chips and renting them out to whoever needs compute, with Anthropic often their biggest customer. That creates a delicate chain of dependencies. If Anthropic grows as expected, these neoclouds prosper. If Anthropic hits a rough patch, they could face empty data centers and unserviceable debt. On Sunday, Elon Musk made the point in characteristically blunt fashion. He wrote that SpaceX had closed $40 billion of debt for GPUs, but that SpaceX was investment grade and had Starlink and launch revenue to support it. He then added that the more than 200 neoclouds would see “guppy-level death and consolidation. Some will make it through, but not many.” Investors have already started to price in that possibility. Many neocloud stocks have slid in recent weeks as concerns mount about the huge amounts of debt needed to finance their buildouts. None has been hit harder than CoreWeave, the most prominent neocloud, which relied on three customers for 72 percent of its revenues as of June. It is not entirely clear how much of that is Anthropic specifically, but the two firms signed a multi-year agreement in April. If the AI trade begins to wobble, these tightly wound companies may find themselves unable to refinance, forced to sell assets at distressed prices, or simply absorbed by larger players. The result could be a painful consolidation that reshapes the entire AI infrastructure sector.
For now, Anthropic remains a high-growth story with enormous momentum, but its future is no longer just about model quality or engineering talent. It is about whether the company can keep minting tens of billions of dollars in revenue, quarter after quarter, without slowing down. If it does, it may well become the anchor of the next great technology wave. If it does not, the consequences could be severe not just for its own shareholders but for the broader tech industry, given how many companies and contracts are tied to its growth. Michael Cuggino, president of the San Francisco-based mutual fund Permanent Portfolio Family of Funds, captured the market’s cautious mood when he noted that Anthropic and OpenAI are not really profitable yet, and that if revenues start to be questioned and the go-go growth fades, it becomes a difficult time to go public. The IPO may still happen and may still raise enormous amounts of capital, but the price will reflect that uncertainty. Anthropic’s prospectus is, in many ways, a document about faith. It asks investors to believe that artificial intelligence will be as transformative as the internet, that Anthropic can stay ahead of OpenAI and open-source rivals, that data centers will be built on time and at cost, that power will be available, and that customers will keep paying premium prices for smarter models. It is a very human story of ambition, confidence, and enormous risk. The numbers are so large that they almost cease to feel real, but they will become real in monthly payments, electricity bills, and quarterly revenue reports for decades to come. Anthropic is betting that it can build the future faster than anyone else and that the world will be willing to pay for it. The next few years will show whether that bet was visionary or overreach, and the answer will matter not only for Anthropic but for every company and worker caught up in the AI revolution.
