$518 Billion of AI Capacity. Most of Anthropic’s Bill Cannot Be Canceled.

Anthropic is not merely betting that demand for artificial intelligence will grow. It is committing money as if access to computing power will become the scarce asset that decides who wins.

According to a confidential IPO prospectus seen by Reuters, the company expects to spend at least $518 billion over roughly a decade on cloud services, chips, leases and dedicated infrastructure involving six partners. About 80% of those commitments is described as non-cancelable or payable regardless of how much capacity Anthropic actually uses.

The company has not publicly released the prospectus. Anthropic declined to comment to Reuters. That matters: the figures below come from Reuters’ review of confidential documents, not from a public SEC filing that every investor can inspect.

IN 30 SECONDS

• Anthropic expects at least $518 billion in AI infrastructure commitments over the coming years, according to a confidential prospectus seen by Reuters.
• About 80% is non-cancelable or must be paid regardless of usage.
• Minimum commitments include $111.1 billion to Google, $110 billion to Amazon and $31.4 billion to Microsoft.
• The company also disclosed roughly $161.2 billion in largely non-cancelable Broadcom-related equipment leases.
• Revenue reportedly reached nearly $4.6 billion in 2025, while operating loss widened to $8.06 billion.
• A future IPO could target a valuation above $2 trillion, but that valuation is not confirmed and the timing may change.

THE NUMBER THAT MATTERS

80%.

The headline number is $518 billion. The more consequential number is the share that cannot easily disappear when demand disappoints.

Cloud spending is often presented as flexible: use more capacity when customers arrive, use less when they do not. Anthropic’s disclosed obligations look different. Reuters reported that the company must spend at least $111.1 billion with Google between April 2026 and July 2033 and $110 billion with Amazon between May 2026 and April 2036. The agreements require payment even if actual usage falls short. A $31.4 billion Microsoft commitment runs from November 2026 to May 2033 and is described as non-cancelable except after an uncured material breach.

That transforms compute from a variable expense into something closer to long-duration financial infrastructure. The strategic upside is guaranteed access. The financial downside is a bill that survives weaker demand.

THE REAL PRODUCT MAY BE CAPACITY

AI companies sell intelligence through software, but the production system is physical. Models need chips. Chips need data centers. Data centers need power, cooling, land, network connections and years of procurement planning.

Anthropic’s prospectus reportedly argues that future demand will be limited mainly by available compute. If that view is correct, reserving capacity early can protect the company from shortages and give it room to train larger models, serve more customers and compete with OpenAI, Google, Meta and xAI.

If the view is wrong, the same contracts become a burden. Capacity can be scarce and still be overpriced. Demand can grow and still grow more slowly than commitments. Technology can become more efficient. Customers can choose rivals. Regulation can slow deployment. A new architecture can change how much computing power produces a useful result.

The bet is not simply “AI will be important.” It is “AI demand will grow fast enough, for long enough, to justify hundreds of billions of dollars in largely fixed capacity.”

MNS LENS: THE CLOUD PROVIDERS ARE ON BOTH SIDES

Amazon, Google and Microsoft are not ordinary suppliers. They are investors, infrastructure providers, distribution channels and competitors developing their own AI systems.

That is commercially powerful and strategically awkward.

Anthropic gains access to capital, customers and cloud capacity. The cloud companies gain a huge buyer for infrastructure and exposure to one of the largest independent AI labs. But the prospectus reportedly warns that these relationships create incentives that may not be fully aligned with Anthropic’s interests.

A supplier can reprice infrastructure. A distributor can privilege its own product. An investor can want growth while a competitor wants market share. The same partnership that accelerates Anthropic can deepen its dependence.

THE $2 TRILLION QUESTION

Reuters reported that a future public offering could value Anthropic above $2 trillion, more than double the company’s own estimated $965 billion valuation in May. “Could” is doing essential work in that sentence. An IPO price has not been set. Market conditions, disclosure, timing and investor demand can change.

The potential valuation is startling because the company is only five years old. It is also a test of how public markets will value pure-play AI businesses after years in which venture capital, sovereign wealth funds and Big Tech funded the race privately.

Revenue reportedly increased twelvefold to nearly $4.6 billion in 2025. That is extraordinary growth. It arrived beside an operating loss of $8.06 billion, up from $2.98 billion in 2024. Net loss was almost $42 billion, although roughly $34 billion came from a noncash accounting charge linked to the rising value of financing instruments that may convert into shares.

That distinction is important. A $42 billion net loss does not mean $42 billion left the bank account. It also does not make the operating loss irrelevant.

Anthropic reportedly held $20.28 billion in cash, cash equivalents and short-term investments at the end of 2025. The company is proposing commitments many times larger than that liquidity position because the obligations stretch over years and because future revenue, financing and public-market capital are expected to carry the plan.

DO THE MATH

Divide $518 billion by ten years and the simple average is $51.8 billion a year. That is not a forecast of annual cash spending: contract schedules differ, some commitments are leases, and the exact timing matters. But the comparison is useful.

Against nearly $4.6 billion of reported 2025 revenue, the average annual commitment would be more than eleven times that revenue base. Anthropic does not need 2025 revenue to fund the entire decade. It needs growth, capital and cash generation to converge before obligations overwhelm flexibility.

Another comparison: the disclosed minimum commitments to Google, Amazon and Microsoft total $252.5 billion. Broadcom-related leases add about $161.2 billion. Together, those four relationships represent roughly $413.7 billion before considering other partners.

The scale explains why the IPO matters. Public investors would not merely fund a software company. They would help underwrite one of the largest infrastructure commitments in corporate history.

WHAT DOES THIS CHANGE?

For AI investors

The discussion has to move beyond model benchmarks and revenue growth. Investors need to understand take-or-pay obligations, lease accounting, contract duration, pricing resets, customer concentration and the point at which compute produces durable margins.

Growth can be real while the economics remain unsettled.

For chip and cloud companies

Long commitments improve visibility. Google, Amazon, Microsoft, Broadcom, AMD and Nvidia-linked capacity can benefit even before the final AI applications prove profitable. In a gold rush, infrastructure vendors are paid earlier than the prospectors.

But concentration cuts both ways. A customer promising enormous spending becomes economically important. If the customer later struggles, the exposure does not stay isolated.

For businesses buying AI

Infrastructure commitments influence product pricing. If AI labs carry huge fixed costs, they need high usage, premium enterprise contracts, better monetization or all three. Falling token prices may continue as models become more efficient, but the industry still has to recover the cost of capacity somewhere.

“AI gets cheaper” and “AI infrastructure gets more expensive” can be true at the same time. Efficiency, competition and contract structure decide which force wins.

For the wider credit market

AI infrastructure is increasingly financed through bonds, leases and long-term contractual obligations. That places the AI boom inside the same capital market used by governments, utilities and traditional corporations. At high interest rates, the competition for money becomes part of the story.

WHAT TO WATCH NEXT

• A public filing: investors need the complete prospectus, audited detail and risk language rather than excerpts from a confidential document.
• IPO timing: Reuters reported that the debut may move until after the November U.S. midterm elections.
• Contract accounting: the difference between cash spending, leases, minimum purchases and cancelable options will shape the real annual burden.
• Revenue concentration: nearly one-quarter of 2025 revenue reportedly came from two customers, while many major clients lacked long-term commitments.
• Compute efficiency: better models per dollar could help margins — or reduce the value of capacity reserved under older assumptions.
• Cash generation: the key milestone is not another funding round. It is evidence that revenue can eventually carry the infrastructure.

THE MNS TAKE

The AI race is becoming less like a software launch and more like building a power grid before knowing exactly how much electricity customers will buy.

Anthropic may be right that compute is the bottleneck. It may also be paying today for a future that arrives differently from the plan. Both possibilities can coexist.

The $2 trillion valuation will attract attention. The 80% that cannot easily be canceled deserves the harder questions.

ASK NOVA

Ask Nova: “When a fast-growing company has huge non-cancelable commitments, which numbers should I compare before considering the investment?”

BRAZIL VIEW: Read the independent MNS Brazil analysis of what the infrastructure bill means for energy, pricing and local businesses.

SOURCES

Financial and contractual figures are attributed to confidential IPO materials reviewed by Reuters. See Reuters’ September 29 report on the infrastructure commitments and Reuters’ report on the prospectus, valuation and financial results.

The prospectus was confidential at publication time. This article is informational and does not constitute investment advice.