Anthropic has a strange problem for a company whose product is selling so fast. It needs more computers than its largest suppliers are prepared to provide. Those suppliers are Amazon, Google, and Microsoft. They also sell AI tools that compete with Claude. And now a draft stock-market filing suggests Anthropic’s future computing bill could pass $750 billion.
That figure appeared in reporting on October 20, 2026, about Anthropic’s confidential IPO documents. It sounds like the company has already signed a check for three-quarters of a trillion dollars. It hasn’t. The reported estimate includes deals still being negotiated, and the payments would be spread across years. But there are already hundreds of billions of dollars in commitments behind it. Reuters examined an earlier draft in September, and The Wall Street Journal’s October reporting has put a bigger possible number on the table.
What bothers me more than the headline amount is the relationship behind it. Anthropic sells intelligence through Claude, but Amazon and Google control much of the computing infrastructure that lets it operate. Microsoft sells access to Claude on Azure while investing in other AI products. Every time Anthropic gets more popular, it needs more capacity from companies that may prefer to sell their own AI. That’s an uncomfortable way to build a business, especially when some of the bills stay due even if demand slows.
Access without medium partner: Cloud Partners Are Becoming Rivals
First, the $750 Billion Number Needs Some Cleaning Up
I had to check the September and October reports twice because the $518 billion and $450 billion figures first looked like conflicting totals. They cover different types of spending, and the draft documents were reported at different stages. Mixing them together makes the situation look either better or worse than it is. So it helps to separate binding obligations from possible future purchases.
On September 29, Reuters said it had reviewed a confidential prospectus describing at least $518 billion of infrastructure spending with six partners over about a decade. The report said roughly 80% of that amount involved agreements that could not be canceled freely or required payments regardless of how much computing capacity Anthropic used. Eighty percent of $518 billion is around $414 billion. That calculation gives a sense of the money involved, but it isn’t a published schedule of exact payments.
The same report identified at least $111.1 billion in obligations to Google, $110 billion to Amazon, and $31.4 billion to Microsoft. These cloud contracts run across different periods, roughly seven to ten years. Reuters also described about $161.2 billion of Broadcom-related equipment lease obligations. Some other arrangements were more flexible, including potential Nvidia-based capacity from xAI that Reuters said could reach $84.5 billion and was largely cancelable on 90 days’ notice. A possible maximum isn’t the same thing as a fixed bill.
Then came the October 20 reports. The Journal described about $450 billion in cloud-computing agreements and a possible $750 billion in total future compute spending, including deals still under discussion. TechCrunch’s account said Anthropic was having trouble securing all the capacity it wanted. The September $518 billion figure included infrastructure and chip leasing, so it should not be added to the October $450 billion cloud figure as if they were separate new purchases. The categories overlap, and the documents were reported at different stages.
This is also a confidential filing, not a final public prospectus that everyone can inspect. Anthropic confirmed in June that it had submitted a draft S-1 to the SEC, but the detailed figures being discussed now come from reporters who saw versions of that draft. They may change before a public listing. A November IPO has been reported, but Anthropic’s own June announcement said the timing and offer terms would depend on market conditions.
Still, $518 billion is a lot of money even when spread over ten years. A simple average comes to about $51.8 billion a year. The actual payment dates will be uneven, and the full amount doesn’t represent expenses booked immediately. Yet it tells you how much of Anthropic’s future has been tied to a supply of chips, electricity, cooling, and data center space that the company cannot produce overnight.
The Revenue Is Real, but So Is the Cost of Serving It
Anthropic isn’t buying all this capacity without customers. According to Reuters’ review of its filing, its 2025 revenue reached nearly $4.6 billion, up from around $400 million in 2024. The October draft, as reported by the Journal and Reuters, put first-half 2026 revenue at $16.3 billion, compared with $1.1 billion in the same period of 2025. These are reported company figures from a confidential filing, not independently audited public results.
And the growth didn’t stop in June. At an investor event on October 14, Anthropic reportedly told attendees that its quarterly revenue had grown from $4.7 billion in Q1 to $11.6 billion in Q2, then $16.8 billion in Q3. The third-quarter increase was around 45% from the previous three months. That sounds healthy because it is healthy growth. But the rate has slowed from the huge jump between the first two quarters, and investors won’t assume that the next quarter can keep adding billions at the same speed. Reuters reported those figures after speaking with people at the event.
There is a small calculation I keep coming back to. Multiply Q3’s $16.8 billion by four and you get an annualized pace of $67.2 billion. Compare that with the simple $51.8 billion yearly average from the September infrastructure plan. It shows why Anthropic might think these commitments can work. But annualizing one quarter is not a revenue forecast, and dividing a decade-long infrastructure plan into ten equal pieces is not a cash-flow forecast either. They are rough comparisons, nothing more.
The company reportedly spent nearly $9 billion on computing capacity and infrastructure in the first six months of 2026. That compares with roughly $2.3 billion in the first half of 2025 and $7.3 billion during all of last year, according to the October Journal report summarized by Anadolu. Those figures are not enough to calculate gross margin, since infrastructure spending and accounting costs do not always line up with the period in which services are consumed. But they do show how quickly the scale of the operation has changed.
Look at what customers are paying for. Someone running Claude Code may leave several agents working across a repository. A business may send thousands of support tickets through the API every hour. Another customer may ask a long-context model to read a contract and compare it with internal records. Anthropic needs chips to process all those requests, plus spare capacity when demand suddenly jumps. If it sells access faster than it brings machines online, users get rate limits and delays. If it rents too much capacity, it may pay for hardware sitting idle. Neither outcome is cheap.
Amazon Is Both the Biggest Helper and a Direct Rival
Amazon’s relationship with Anthropic is the clearest example of the problem. The two companies have worked together since 2023. Amazon has invested billions in Anthropic, sells Claude through Bedrock, and supplies the processors and data centers used to train and run the models. Amazon is also building and selling its own Nova models. So the same company helping Claude reach business customers has a separate reason to make competing AI products successful.
In April 2026, Anthropic announced an agreement involving more than $100 billion over ten years in AWS technologies and up to five gigawatts of new capacity. It said it was already using more than one million Trainium2 chips across its work with Amazon. The company also said over 100,000 customers were running Claude on Amazon Bedrock. Amazon committed another $5 billion in investment at that time, with up to $20 billion more possible later. That potential extra investment should not be treated as money already transferred.
Five gigawatts needs some explanation. It describes the scale of electrical power associated with the planned compute capacity, not the speed of one AI model or proof that all those data centers are running today. A campus can be announced years before every building, cooling system, network link, and power connection is ready. Anthropic said its new AWS capacity would come online in stages, including close to one gigawatt of Trainium2 and Trainium3 capacity expected by the end of 2026.
Amazon benefits when Claude usage rises. It sells chip time and collects a share of distribution revenue. It can also use the work with Anthropic to improve its Trainium processors and the AWS Neuron software that runs on them. Anthropic described that engineering work in 2024, including writing low-level code for the chips. This is a close technical partnership, not a normal supplier relationship where one side delivers a box and walks away.
But a close partnership does not remove the competition. Amazon can offer a customer Claude today and its own AI model tomorrow. It can use Bedrock to present several choices in the same purchasing flow. That doesn’t prove Amazon is deliberately denying Anthropic capacity, and the reporting does not establish any such coordinated plan. The problem is that incentives are mixed. When capacity is scarce, the provider has to decide where it goes, and Anthropic cannot make that decision alone.
Google’s Deal Has the Same Tension, With Different Chips
Google sells Gemini models and AI tools through Google Cloud. It also gives Anthropic access to its TPUs, which are custom processors built for machine learning work. In April 2026, Anthropic and Google announced a new agreement for multiple gigawatts of next-generation TPU capacity, expected to start coming online in 2027. In a May funding announcement, Anthropic described the Google-Broadcom deal as five gigawatts.
That makes Google a supplier, a place where customers buy Claude, and a rival model developer. It is a useful relationship for Anthropic because businesses already using BigQuery, Cloud Run, or other Google services can add Claude without rebuilding their entire setup. But it also means Google sits close to the customer purchasing process. Anthropic needs that reach. Google has its own ideas about which models customers should try.
Microsoft complicates the story further. In November 2025, Anthropic announced a deal to buy $30 billion of Azure computing capacity, with plans for up to one gigawatt of additional infrastructure using Nvidia hardware. The later draft filing reported by Reuters put Microsoft’s infrastructure obligation at $31.4 billion. Microsoft also distributes Claude through its Foundry services and offers AI through Copilot. These arrangements let companies use models from more than one supplier, which is useful for customers even if it makes the vendor relationships messy.
The risk is wider than a single chip shortage. Reuters reported that 47% of Anthropic’s 2025 revenue, about $2.16 billion, came through Amazon and Google’s cloud sales channels. It estimated about $351 million in distribution fees, or roughly 16 cents of every dollar of revenue sold through those channels. Cloud partners also gain visibility into commercial terms and prices, according to the filing. That is a lot of dependence on companies that sell competing products.
Even worse, cloud providers are a path to the end customer. A company buying Claude inside its existing AWS or Azure bill may build its internal tools around that cloud account, billing arrangement, security approvals, and support team. Anthropic gets distribution it could not build as quickly on its own, but it doesn’t control all the steps between its model and the buyer. That makes it harder to walk away if the relationship becomes uncomfortable.
Renting More Data Centers Doesn’t Solve Every Problem
Anthropic is trying to reduce its dependence on the three cloud giants by going beyond their infrastructure. The reported filing describes direct chip leases and deals with other data center operators. In May, Anthropic announced a SpaceX agreement for more than 300 megawatts of capacity at Colossus 1, representing over 220,000 Nvidia GPUs, which it said would become available within the month. The company connected that capacity expansion with higher Claude Code and API usage limits.
It has also moved closer to Broadcom. Reuters reported on October 1 that Broadcom had agreed to lend Anthropic up to $42 billion for infrastructure financing. The earlier draft had also described about $161.2 billion of related equipment lease obligations. That Reuters account shows how the deals can link chip supply, financing, and long-term payments. A loan facility and an equipment lease are different financial arrangements, so adding their headline figures as though they were fresh cash spent today would give the wrong picture.
This approach gives Anthropic more choices, but someone still has to build the buildings and keep the machines working. Electricity connections can take years. A cooling problem can stop thousands of chips at once. Parts can arrive late. The leaked filing, as described by TechCrunch, even warns that smaller data center operators have shorter operating histories than the big clouds. A new supplier is not automatically a safer supplier.
And Anthropic cannot fix power shortages by writing better software alone. If it reserves a large campus but the local grid cannot supply the electricity on schedule, the contracts and delivery dates become a serious business issue. This is why the word “gigawatt” matters far more here than another flashy model benchmark.
Three Kinds of AI Chips Mean Three Kinds of Engineering Work
The technical side is easy to miss under the financial numbers. Anthropic uses AWS Trainium, Google TPUs, and Nvidia GPUs. That mix is a smart way to avoid relying on one chip company, and Anthropic has confirmed it publicly. But it does not mean its engineering team can move every training job or inference service from one platform to another with a single setting change.
Nvidia software commonly uses CUDA. Google’s TPUs depend on a different compiler and execution setup. AWS Trainium uses its Neuron tools. Teams have to test kernels, memory use, communication between accelerators, and model behavior for each platform. They also need to watch whether a feature available on one chip works well on another. The same model can have different cost and speed profiles depending on the hardware.
For a normal developer, this is the kind of headache that happens when a project works on one machine but refuses to run cleanly on another. Scale that problem to a training job spread across thousands of accelerators, and even a small mismatch can become costly. Every hour spent finding a compiler problem is an hour the cluster may not be doing the work it was rented for.
Still, I think Anthropic is right to spread the work across different chips. Relying on Nvidia alone would leave it exposed to price changes and hardware availability. Relying on one cloud would give a single supplier too much control. The trade is more engineering complexity. The company is choosing to pay that cost because the alternative could be worse.
The $80.7 Billion Loss Is Real Accounting, but It Isn’t $80.7 Billion of Cash Burn
This is the number most likely to be repeated badly. The October 20 draft showed a $80.7 billion net loss for the first half of 2026, according to Reuters. Read that without the next sentence and you’d think Anthropic had somehow spent more than $80 billion in cash in six months.
About $72 billion of the reported loss came from an accounting charge linked to changes in the value of convertible notes and other fundraising instruments. Those instruments may eventually turn into company shares, and changes in their estimated value can create large charges in the accounts. That charge is not the same as a $72 billion payment for chips or electricity. Subtracting it from the net loss also does not give you the company’s exact cash burn. For that you need a proper cash-flow statement and details on how infrastructure payments and leases are recorded.
This distinction matters because the risks are already serious without exaggerating them. Reuters reported that Anthropic’s operating loss was more than $8 billion in 2025, despite nearly $4.6 billion in revenue. In October, attendees at its investor event said the company reported an adjusted profit in Q3 2026. We do not yet have a public reconciliation showing exactly which costs were removed from that adjusted figure. Both statements can be true under different accounting measures and periods.
The bigger question is how the spending behaves if revenue stops growing so quickly. Consider a hypothetical year in which Anthropic’s sales flatten while its committed compute payments keep rising. It may still owe money on reserved capacity, even if customers aren’t using enough tokens to cover it. Now consider the opposite: demand keeps growing, inference gets cheaper per request, and Anthropic fills most of the machines it has reserved. Then the same contracts could help protect its ability to serve customers and keep costs under control. Neither future has happened yet.
A customer concentration problem sits on top of that. The draft says two customers accounted for about 20% of first-half 2026 revenue, according to the Journal’s report. Their identities were not disclosed in the reporting. It would be wrong to guess. But losing either one, or having either reduce usage, would matter more when expenses cannot be reduced at the same speed.
What Could Happen if the Cloud Companies Stop Being So Friendly?
The optimistic case is fairly easy to understand. Claude demand keeps increasing, Anthropic gets its ordered hardware on schedule, and the model becomes cheaper to run as software and chip efficiency improve. Amazon and Google earn money from hosting Claude even when they also sell competing models. Anthropic’s multi-cloud setup gives customers a choice, and the partnerships hold because both sides make enough money. The commitments look huge today but become manageable over their long lives.
The middle case may be more realistic, though it is still only a scenario. Demand grows at a slower pace, the company has to renegotiate how much capacity it adds, and some new sites arrive later than planned. Anthropic keeps using all three major clouds and signs more deals with other operators. Prices for AI models fall as competition gets stronger. Revenue goes up, yet earnings and cash flow remain under pressure because hardware costs don’t fall on the same schedule.
The difficult case is a combination of weaker demand and less friendly supply terms. Suppose a large cloud provider favors its own product in sales or capacity planning. Anthropic can’t assume it can move a large training run elsewhere overnight. If it also has to pay under non-cancelable contracts, the pressure is doubled. The confidential filing warns about conflicts of interest and reduced access to compute, but it does not prove that Amazon, Google, or Microsoft is currently withholding chips to hurt Claude. That’s an important line between a disclosed risk and an event that has occurred.
There is another possible outcome that doesn’t require any partner to behave badly. AI inference could become much more efficient, with customers needing fewer chips for the same work. This would be good for software users. It could also leave companies holding expensive capacity deals signed under older assumptions. Whether cheaper inference leads to enough additional usage to fill that capacity is still an open question. Pricing and demand will matter just as much as model quality.
And then there’s the IPO. Anthropic raised $65 billion in May at a reported $965 billion post-money valuation. News reports have discussed much higher possible public valuations, but no final share price has been set. An IPO could bring in more capital, yet selling shares does not make long-term operating commitments disappear. Public investors may ask for clearer details about cash expenses, chip leases, customer concentration, and the exact schedule of compute payments than private investors have demanded so far.
The Number I’d Watch After the $750 Billion Headline
If Anthropic publishes its S-1, I would look first for the cash payment schedule attached to the non-cancelable agreements. Not the biggest possible total, and not just the annual revenue figure. I want to see what is due in 2027, what is due in 2028, and how much flexibility remains if AI usage changes. That is where a risky promise turns into a bill someone actually has to pay.
I would also look at the proportion of customer sales that still passes through Amazon and Google. It was 47% in 2025, based on the leaked figures Reuters reviewed. If Anthropic can build more direct customer relationships while keeping its cloud partners happy, it gets a little more freedom. If that percentage rises while the providers expand their own models, the relationship gets harder to manage.
The uncomfortable part is that Anthropic probably needs these competitors. Amazon has chips and customers. Google has TPUs and data centers. Microsoft has Azure and enterprise buyers. Anthropic has Claude, which is attracting people willing to pay for AI work. Each side offers something the other wants, and each side would prefer to keep more of the value for itself.
For now, the confirmed story is a company growing very fast and committing enormous sums to stay ahead of demand. The possible $750 billion total is a warning about how far that plan could go, not proof that the money has already been spent. When the public filing arrives, the less glamorous pages explaining leases, minimum payments, and contract dates may tell us more about Anthropic’s future than its next model launch.