Anthropic's IPO Reveals the Hidden Risk in AI's Compute Boom

Editorial illustration generated by the author.

Anthropic's first operating profit, an $80 billion accounting loss, and hundreds of billions in computing commitments reveal a new challenge for artificial intelligence: how much of the future must AI companies buy before they know what it will be worth?

Anthropic's October 2026 IPO prospectus presents three numbers that appear difficult to reconcile. In the second quarter, the developer of Claude made nearly $1 billion in operating profit. Across the first half of the year, it reported an $80.7 billion net loss. And according to the Financial Times, its broader computing arrangements involve more than $800 billion of prospective spending extending into the 2040s. Each number is meaningful, but each measures a different kind of financial exposure.

The loss largely reflects changes in the value of financing instruments rather than cash consumed by operations. The quarterly profit reflects a genuine improvement in the economics of selling AI services. The commitments, meanwhile, are a mixture of minimum purchases, leases and agreements with varying conditions. Their differences matter more than the headline total.

Anthropic has established substantial demand for Claude. Yet serving that demand obliges it to reserve computing capacity long before the corresponding revenue can be forecast with confidence. Success increases the pressure to secure more supply; each expansion makes the consequences of a forecasting error larger.

The capacity that protects Anthropic from shortages today may become a burden if the pace or economics of AI adoption change. The prospectus is a case study in how a successful technology business can acquire the risks of an infrastructure operator.

The $80 Billion Loss That Wasn't an Operating Loss

The largest headline number in Anthropic's October 26 prospectus is its $80.7 billion net loss during the first six months of 2026. Without the accompanying accounts, a reader might mistake it for cash spent developing and operating AI systems.

But most of that loss did not come from running models, training new systems, paying employees or operating data centres. Approximately $72.4 billion resulted from changes in the fair value of convertible notes, derivative liabilities and similar financing instruments, primarily driven by increases in Anthropic's valuation.

The mechanism is counterintuitive. When investors provide financing through instruments that can later convert into shares, the value of those conversion rights may increase as the company becomes more valuable. Depending on their accounting classification, the resulting increase in the liability can appear as an expense in the income statement.

A company can therefore report an enormous accounting loss partly because investors believe it has become much more valuable. That does not mean the economic consequences are imaginary: conversion rights may dilute existing shareholders, and contractual liabilities still matter. But the expense is not equivalent to spending $72.4 billion in cash on AI development.

Anthropic also recorded approximately $8.8 billion in other net non-operating expenses during the period. The combined effect was a reported net loss dramatically larger than its underlying operating loss.

Financial measure	H1 2026	Interpretation
Revenue	$16.3bn	Commercial income from AI services
Operating loss	$693mn	Loss after operating expenses
Fair-value expense on convertible instruments	$72.4bn	Largely non-cash financing valuation changes
Reported net loss	$80.7bn	Dominated by non-operating expenses
Operating cash flow	+$3.9bn	Cash generated by operating activities

Source: Anthropic's SEC Form S-1, filed October 26, 2026. Figures rounded. Operating cash flow is affected by payment timing and working-capital movements and is not equivalent to free cash flow.

Anthropic's first-half operating loss was $693 million, not $80.7 billion. It also generated $3.9 billion in operating cash flow, a figure influenced by working-capital movements and the timing of customer receipts and supplier payments. Neither result is a measure of returns on the infrastructure it has reserved for future years.

Neither figure should be mistaken for evidence that Anthropic can finance all of its future expansion internally. The company also received approximately $110.8 billion in financing cash inflows during the same period. Its growth continues to depend heavily on access to capital.

Separating accounting losses, operating profitability and external financing changes the picture. Anthropic is not burning $80 billion every six months to keep Claude running. It is a rapidly expanding business whose operating economics have improved dramatically, while its future infrastructure requirements are growing on a scale that even those improvements may struggle to accommodate.

Profitable for a Quarter, Committed for Years

Anthropic's second-quarter operating profit of $974 million is more informative about operations than its reported net loss. Revenue reached approximately $11.6 billion during the quarter, producing an operating margin of about 8%. It was the first quarterly operating profit shown in the prospectus.

The improvement becomes clearer when compared with early 2025. During the first quarter of that year, compute and infrastructure expenses amounted to approximately 251% of revenue. By the second quarter of 2026, they represented about 48%.

Anthropic did not achieve this by spending less on computing. Quarterly compute expenses increased from approximately $944 million to $5.6 billion over the period. Revenue, however, rose from $376 million to $11.6 billion. Revenue grew much faster than the computing costs required to support it.

The improvement may reflect more efficient inference, better models, stronger demand or greater use of reserved infrastructure. The filing cannot isolate the precise contribution of each factor. It does show revenue outrunning a compute cost base that was previously much larger than sales.

There is nevertheless a danger in extrapolating the result. Anthropic's first-half operating result was still negative, and the company warns that margins may fluctuate as additional infrastructure becomes available and the corresponding expenses begin to appear.

Revenue growth is also beginning to encounter the arithmetic of a larger business. According to Reuters reporting on Anthropic's October investor presentation, third-quarter 2026 revenue reached approximately $16.8 billion, representing 45% growth over the second quarter. That remains extraordinary expansion, but it follows sequential growth of roughly 144% between the first and second quarters.

These third-quarter figures were reported from investor presentations rather than a complete set of quarterly GAAP financial statements in the preliminary prospectus. They therefore cannot establish whether the second quarter's operating profit was sustained on the same accounting basis.

More importantly, the deceleration illustrates why the next stage of growth is difficult to forecast. Increasing revenue from $4.7 billion to $11.6 billion in one quarter does not establish that the company can repeatedly multiply its business at comparable rates. Nor does slowing percentage growth mean that demand is collapsing. It means the company must make increasingly consequential infrastructure decisions as its revenue base grows.

The profitable quarter shows that Anthropic can earn money selling AI services under current conditions. The harder test begins as more of the computing capacity ordered in previous years is delivered, and the related expenses appear.

Contracts That Turn Forecasts into Obligations

Frontier AI companies cannot simply purchase additional computing capacity whenever customer demand increases. Advanced accelerators must be manufactured, data centres constructed, electricity connections secured and networking infrastructure installed. The relevant lead times can extend far beyond the period over which an AI developer can reliably forecast its revenue.

Anthropic says it typically procures computing capacity 12 to 18 months before it becomes available. This means substantial future expenditure must be decided before the company knows which models it will be operating, how efficiently those models will run or how much customers will be willing to pay for their services.

The Financial Times reported that Anthropic's broader infrastructure spending commitments exceed $800 billion and extend through 2045. The figure captures the extraordinary scale of the expansion, but it should not be interpreted as a single amount of debt or an unconditional obligation to pay that entire sum.

The preliminary prospectus itself leaves the headline aggregate commitment amount blank, while disclosing several categories of obligations with different contractual characteristics. As of June 30, 2026, Anthropic reported approximately $265.6 billion in future fixed or minimum guaranteed purchase commitments under agreements spanning one to twenty years. It separately disclosed $321.4 billion in estimated undiscounted lease payments under arrangements that had not yet commenced, some subject to contractual conditions.

Additional agreements entered into after June 30 further expand the company's prospective spending. Some involve fixed minimum payments, others depend on capacity delivery or deployment, and some can be cancelled under specified conditions. The categories should not be added indiscriminately, particularly where measurement dates and contractual definitions differ.

Provider	Disclosed amount	Nature of commitment
Google	At least $114.5bn	Minimum spending through July 2033, largely regardless of actual usage, with true-up provisions
Amazon	At least $110bn	Minimum spending through April 2036, regardless of usage, subject to specified exceptions
Microsoft	At least $44.9bn	Minimum spending through May 2033, generally non-cancelable subject to contractual exceptions
SpaceXAI	Up to $120.5bn	Agreements extending into 2029, cancellable on 90 days' notice after initial commitment periods ending in early 2027

Source: Anthropic S-1, Infrastructure Service Arrangements. This is a selection of contracts, not a comprehensive commitment total. Some amounts are included in broader disclosed obligation categories; they must not be added to those categories as separate liabilities.

The contrast between Google and SpaceXAI illustrates why the structure of the agreements matters. With Google, Anthropic has committed to substantial minimum spending even if its actual usage falls short. The company may have to pay the difference between what it consumes and what it promised to purchase. The SpaceXAI agreements provide substantially more flexibility after their initial non-cancelable periods, although their maximum disclosed value is also enormous.

That difference is economically significant. The value of an infrastructure commitment depends not just on its size and unit price, but on when capacity becomes available, what technology is delivered, whether payments depend on utilisation and how easily the buyer can adjust its obligations.

Anthropic is effectively assembling a portfolio of future computing resources, some secured by relatively rigid contracts and others with greater flexibility. The company's success will depend partly on how well that portfolio matches a market whose demand, pricing and technology remain unusually difficult to predict.

When Growth Magnifies Exposure

Approximately 80% of revenue during the first half of 2026 came from consumption-based services. Customers pay primarily according to how much AI they use. Some enterprise agreements include non-cancelable annual or multiyear minimum spending commitments, so consumption-based revenue does not mean that every dollar can disappear immediately. Nevertheless, a substantial portion of future income remains dependent on customer behavior rather than fixed long-term purchasing obligations.

That creates an asymmetry. Anthropic must secure large amounts of computing capacity in advance, often under arrangements that impose minimum payments. Customers, meanwhile, may have considerably more flexibility to alter their usage, negotiate prices, switch providers or abandon applications that fail to justify their cost.

At June 30, 2026, Anthropic reported $10.8 billion in remaining performance obligations, representing revenue expected from existing non-cancelable customer commitments. On the supplier side, it disclosed $265.6 billion in future minimum purchase commitments.

These are not directly comparable measures. The supplier obligations span much longer periods, while remaining performance obligations exclude substantial potential revenue from future usage, contract renewals and new customers. Comparing the amounts as though they were matching assets and liabilities would produce a misleading impression of financial coverage.

They nevertheless illustrate the underlying problem: Anthropic has committed to purchasing far more future computing capacity than its existing customer contracts guarantee will be used.

Consider a hypothetical AI company that reserves enough computing capacity to support 100 units of expected customer activity. If demand reaches that level, the company can utilise the infrastructure, distribute its fixed costs across a large revenue base and potentially earn attractive margins. If demand reaches only 70 units, but the company remains liable for most of the original spending commitment, its revenue declines relative to the expense it has already accepted.

The problem is not necessarily that the unused servers have become technologically worthless. They may be redeployed to other workloads or sold as capacity to another customer. The financial question is whether that alternative use generates enough income to justify the contracted cost.

Suppose Anthropic's demand doubles, putting pressure on its existing computing capacity. To avoid turning customers away, it must reserve another large block of infrastructure. But by the time that capacity arrives, customer demand may be growing at a different rate. If the company again expands successfully, the next procurement decision becomes even larger.

Each successful stage of expansion creates the need to make a larger forecast about the next stage. The business can become more commercially established while simultaneously exposing more capital to mistakes about the timing of future demand.

This is not necessarily evidence of irrational management. A frontier AI provider that runs out of capacity may lose customers to competitors, weaken its developer ecosystem and surrender opportunities that cannot easily be recovered. Waiting until demand is certain could be more damaging than reserving too much infrastructure.

Overbuying exposes Anthropic to expensive idle capacity; underbuying risks losing demand to competitors. In a market where a capacity shortfall could weaken customer relationships and developer adoption, some deliberate overprovisioning may be rational.

The compute commitment dilemma: advance capacity purchases can create scale advantages when demand is strong, but leave an expensive surplus when utilisation disappoints. Illustration generated by the author; conceptual, not to scale.

The Efficiency Paradox: What If AI Becomes Much Cheaper?

Demand is only one source of uncertainty. Anthropic must also estimate what computing power will cost years from now and how much of it future AI systems will require.

That is a difficult forecast because both the hardware and the algorithms are improving rapidly. New accelerators can deliver more computation per watt, model architectures can become more efficient, and inference techniques can reduce the cost of completing a given task. Smaller models may also acquire capabilities that previously required considerably larger systems.

Three variables matter: the cost of producing a unit of useful AI work, the price customers pay for it, and the total amount of work customers choose to purchase.

If Anthropic reduces its cost per task while maintaining prices, margins can improve. If competition forces prices to decline just as quickly, customers may capture much of the benefit. If lower prices create enormous additional demand, total revenue and computing consumption may continue growing despite declining prices per task.

Economists sometimes describe this possibility through Jevons paradox: cheaper use of a resource may stimulate enough new applications to increase total consumption. But the analogy is a possibility, not a forecast; lower costs do not automatically produce higher spending.

AI agents make the argument particularly interesting. A simple chatbot response may require a limited amount of computation, while an agent that researches a problem, writes code, executes tests, identifies errors and repeatedly revises its work can consume substantially more. As these systems become cheaper, companies may find it economical to automate activities that previously required too much machine computation to justify the expense.

But whether greater efficiency increases total demand depends on how strongly customers respond to falling prices and improved capabilities. It is not a law that guarantees growing compute consumption or rising profits.

For Anthropic, the danger is not that cheaper inference automatically makes its infrastructure worthless. It is that market prices for AI services might fall faster than the contracted cost of delivering them, leaving profitability dependent on a future increase in demand that has not yet materialized.

Equally, technological improvements could make Anthropic's commitments more valuable. If models become substantially more useful while demand grows faster than efficiency reduces the compute required per task, having secured large quantities of capacity could give the company an advantage that competitors cannot easily replicate.

Infrastructure contracts thus incorporate assumptions about much more than customer numbers. They also embed expectations about technological progress, utilisation, competitive pricing and the relationship between computing expenditure and useful output.

The more rapidly those factors change, the harder it becomes to know whether a contract signed today will look attractive several years later.

Who Captures the Returns?

The scale of Anthropic's commitments also illustrates a broader feature of the AI economy. Amazon and Google are not merely suppliers of its computing infrastructure. They are also investors and distribution partners, while operating competing AI businesses of their own.

These relationships help finance expansion and give Anthropic access to infrastructure that would be difficult to construct independently. They also create an unusually interconnected system in which AI developers, cloud providers and hardware suppliers depend on overlapping expectations about future demand.

A previous Journal essay on the circular economics of the AI boom examined how investment, infrastructure spending and supplier revenue can reinforce one another without establishing that the resulting capacity will earn an adequate return.

Its long-term contracts are not simply declarations of confidence in artificial intelligence. They are financial decisions that will produce costs at particular times and under particular conditions, regardless of whether the broader technology eventually transforms the economy.

AI services could become indispensable while particular investments in supplying them disappoint. A data centre may remain useful but earn less than its financing costs require. The technology can be valuable while the price customers pay for its output falls faster than the cost of providing it.

It would also be mistaken to assume that every cloud provider and AI developer faces the same exposure. Large technology companies may finance infrastructure from diversified existing businesses, while more specialized companies depend on a narrower set of products, customers and financing arrangements. Anthropic is unusual in combining frontier-model development with a rapidly growing direct commercial business and extremely large long-term commitments.

Its balance sheet provides substantial protection. At the end of June 2026, it held approximately $131.7 billion in cash, cash equivalents and short-term investments. That is a formidable financial reserve, particularly for a company founded only five years earlier.

But even that balance must be understood relative to the scale and timing of future commitments. The company is not deciding whether to spend a fixed sum once. It is arranging a sequence of large infrastructure purchases, leases and capacity deliveries while competing in a market where both capabilities and prices may change substantially before those arrangements have run their course.

How to Judge the Compute Bet

The prospectus offers no basis for declaring Anthropic's strategy either brilliant or reckless. But it does suggest observable tests of whether the commitments are paying off.

The first measure is revenue relative to computing expenditure. Anthropic has improved this relationship enormously, but the crucial test will come as newly committed capacity enters service. If revenue continues growing faster than the cost of providing the computing resources it uses, the contracts may support an increasingly profitable business. If expenses jump ahead of revenue, the pressure on margins will become visible.

The second is the growth of customer commitments. Consumption-based revenue can expand rapidly, but enterprise agreements with meaningful minimum spending requirements provide additional visibility. They do not remove demand risk, particularly when contracts can expire or be renegotiated, but they help bring the timing of contracted revenue closer to that of committed expenditure.

The third is the flexibility of Anthropic's infrastructure portfolio. The ability to shift workloads between providers, redirect capacity between training and inference, or terminate particular arrangements can reduce the cost of forecasting errors. Two companies with identical headline infrastructure commitments may have very different risk profiles if one can adjust its purchasing decisions and the other cannot.

Finally, the company must demonstrate that it can finance continued expansion without permanently depending on ever-larger injections of external capital. Operating cash flow and cash reserves matter, but they should be examined alongside investment spending, lease obligations, financing terms and the timing of committed payments.

These indicators are more useful than the headline net loss or the sheer size of the announced compute agreements. The $80.7 billion accounting loss tells us relatively little about the economics of serving Claude users. The $800 billion-plus spending figure tells us a great deal about the scale of Anthropic's ambitions, but not enough about their ultimate profitability.

The decisive test is whether each wave of reserved capacity produces enough additional profitable AI work to cover its cost, without requiring the next financing round to conceal a shortfall.

The Risk Created by Success

Anthropic's prospectus makes it harder to dismiss generative AI as a technology searching for a business model. The revenue is substantial, the improvement in operating economics is measurable, and customers are increasingly willing to pay for what its models can do.

Commercial success has changed the uncertainty rather than resolved it. During the industry's experimental phase, the danger was spending billions on technology that might never produce sufficient revenue. Anthropic now faces a different problem. Having demonstrated substantial demand, it must decide how much infrastructure to secure before that demand becomes predictable. Underinvesting could surrender customers and competitive advantage. Overinvesting could burden the company with years of commitments made on assumptions that no longer hold.

The difficulty is intensified by the speed of technological change. The future capacity being purchased must serve models that may not yet exist, workloads whose economics are still developing, and customers whose willingness to pay will depend partly on how much better and cheaper AI becomes.

Neither technological success nor growing revenue will necessarily settle which infrastructure decisions were correct. AI may become cheaper, more capable and more widely used while particular investments in its development and delivery generate disappointing returns.

Long-term contracts could make Anthropic unusually difficult to compete with if demand absorbs the capacity. They could constrain its choices if the expected workloads, prices or customer spending fail to appear. The same strategy supports both outcomes because it requires decisions about supply before the value of that supply is known.

Anthropic does not need to be wrong about the future of artificial intelligence to be wrong about how much of that future it should buy in advance.

Sources and Further Reading

• Anthropic, PBC — Form S-1 Registration Statement, filed October 26, 2026. Primary source for financial results, contractual obligations, risk disclosures and cash flows.
• Financial Times — Anthropic Begins Countdown to IPO with Prospectus Filing, October 26, 2026. Reporting on the proposed IPO and aggregate infrastructure commitments.
• Reuters — Anthropic's Q3 Revenue Surges 45% to Nearly $17 Billion, October 14, 2026. Reporting on the investor presentation and third-quarter revenue.

Financial information reflects disclosures and reporting available as of October 28, 2026. The IPO remains proposed, with final offer terms not yet announced. This article examines business economics rather than providing investment advice.