Anthropic in 2036: The Cost of Staying Ahead

An examination of Anthropic's financial disclosures, the economics of frontier AI, and what ten years of infrastructure investment might mean for shareholders.

An examination of Anthropic's financial disclosures, the economics of frontier AI, and what ten years of infrastructure investment might mean for shareholders.

Anthropic generated $16.3 billion in revenue during the first six months of 2026, a 1,300% increase over the same period last year. For perspective, the company generated $4.6 billion in all of 2025. It is a remarkable acceleration in the commercialization of artificial intelligence, and one that would have been difficult to imagine even a few years ago.

Behind every interaction with Claude is a growing industrial system of semiconductors, data centers, power infrastructure, networking, and long-term financial obligations. Anthropic sells access to software, but its ability to deliver that software depends on securing enormous amounts of physical infrastructure, often years before the demand it is intended to serve can be forecast with confidence.

The company's October 2026 S-1 offers an unusually revealing look at this undertaking. Revenue is expanding at an extraordinary pace, but the infrastructure commitments needed to support that expansion extend well into an uncertain future.

I once had a job forecasting the financial requirements of hyperscale cloud infrastructure investments. Much of the work involved connecting expectations of future customer demand to the physical capacity that would have to be constructed or reserved in advance. The forecasts were rarely difficult because we lacked a theory about where demand was going. They were difficult because infrastructure had to be financed and built before we could know whether those expectations were correct.

Anthropic is confronting a considerably larger version of the same problem, complicated by a technology that is changing faster than the infrastructure supporting it can be built.

I wanted to understand what its current trajectory might imply ten years from now. How much revenue could Anthropic generate if enterprise adoption continues? What would it cost to develop and operate successive generations of frontier models? How much of the necessary infrastructure could be financed without issuing additional equity? And after all of that investment, what might remain for shareholders?

To explore those questions, I separated Anthropic's computing expenses into two categories with very different economic characteristics: inference and training.

Two Compute Businesses Under One Roof

When a customer uses Claude, the model processes a request, produces an answer, and generates revenue for Anthropic. The computation required to serve the customer is called inference.

Training involves developing the models themselves. Researchers assemble computational resources, train and evaluate new models, and attempt to produce improvements substantial enough to justify the expenditure. These investments typically occur well before the resulting models generate meaningful revenue.

Both activities require access to increasingly sophisticated computing infrastructure, but their costs behave differently.

Inference is relatively closely tied to customer activity. More usage requires more computation, although improvements in hardware efficiency, software, and model architecture can reduce the cost of delivering a given amount of useful output. Anthropic's pricing and customer mix determine how much of those savings the company retains.

Training involves much greater uncertainty. Large investments must be made in anticipation of capabilities that have not yet been demonstrated, for markets whose ultimate size and competitive structure remain unsettled.

The first-half financial results illustrate the difference.

Anthropic generated approximately $16.3 billion in revenue against $2.4 billion in reported cost of revenue. If we treat most of that cost as associated with customer-serving inference, the implied gross margin is approximately 85%.

That is an attractive starting point for the economics of delivering AI services. It is also incomplete because the calculation excludes much of the spending required to develop the underlying models.

A supplemental analysis of Anthropic's disclosures and contemporaneous financial reporting places first-half model-development compute expenditure near $9.7 billion. These estimates draw from different financial classifications and should not be understood as a company-reported training-versus-inference income statement.

The implication is that Anthropic may have a very profitable business distributing intelligence that depends on a very expensive business creating it.

Once a model has been developed, Anthropic can deliver its capabilities to paying customers at what appears to be an attractive incremental margin. Remaining competitive, however, requires repeated investments in new models, with no guarantee that each successive generation will produce a proportionate economic return.

The company's future profitability depends partly on whether the economics of serving customers improve faster than the cost of developing the models those customers want to use.

The Infrastructure Behind the Interface

The physical requirements of frontier model development increasingly resemble those of large industrial projects.

Consider a hypothetical training cluster containing one million advanced accelerators. Each requires electrical power, cooling, networking, physical space, and supporting infrastructure. The cluster must also function as a coordinated computational system, rather than simply a collection of individual computers.

At this scale, purchasing the chips is only one part of the undertaking. They must be installed, powered, interconnected, cooled, maintained, and brought into productive operation. The supporting investments extend across semiconductor manufacturing, electrical generation and transmission, data-center construction, equipment financing, and cloud-service contracts.

These are long-lead-time investments in a field where the technology can change substantially within a single development cycle.

Anthropic's S-1 disclosed approximately $266 billion in minimum purchase commitments and another $321 billion in leases not yet commenced as of June 30, 2026. The two categories represent different contractual arrangements with different payment schedules, commencement dates, and conditions. They cannot simply be combined and treated as an immediate financial liability.

Nevertheless, the scale is substantial. Anthropic reported approximately $132 billion in cash, cash equivalents, and short-term investments at the same date. Its liquidity is enormous by conventional corporate standards, but so are the commitments it has entered into to secure future computing capacity.

There is a useful historical analogy in the development of railroads, telecommunications networks, and electrical utilities. All three required substantial capital expenditure before the full extent of customer demand could be established. Investors had to finance infrastructure whose economic value depended on eventual utilization.

Anthropic faces a similar utilization problem, although its equipment operates on a much shorter technological cycle.

A railroad built in the nineteenth century could remain economically useful for generations. A data center may have a long physical life, but the accelerators installed within it can become technologically or economically obsolete much sooner. Hardware that is indispensable for one generation of AI models may be less competitive for the next.

The problem becomes more complicated when infrastructure is financed through long-term leases and minimum purchase commitments. Anthropic may be contractually obligated to pay for computing resources even if a subsequent technological development changes their economic value.

The company has committed to a substantial portion of its future computing requirements under assumptions about technology, demand, and competitive conditions that will almost certainly change before those commitments expire.

Whether those changes favor Anthropic will depend on its ability to adapt the infrastructure and commercial business together.

Reading the Financial Results

Anthropic's operating performance helps explain why investors and infrastructure partners have been willing to finance an expansion of this magnitude.

The company reported an operating loss of approximately $693 million for the first half of 2026, a relatively modest figure compared with its revenue and infrastructure spending.

Its reported net loss was much larger, approximately $80.7 billion.

About $72.4 billion of that loss reflected changes in the fair value of convertible notes, derivative liabilities, and similar financing instruments. As Anthropic's estimated enterprise value increased, certain financing obligations became more valuable to the counterparties entitled to conversion or other contractual benefits. Accounting rules required the company to recognize much of that change as an expense.

The charge did not represent an equivalent expenditure of cash during the period. It did, however, reveal the consequences of Anthropic's financing arrangements.

Early capital providers negotiated claims whose economic value increased substantially as the company's valuation rose. Some of those claims will eventually be settled or converted according to their contractual terms, affecting the distribution of economic value among the company's investors.

We cannot infer from the comparatively modest operating loss that Anthropic is already generating large amounts of unrestricted cash. The development of its computing infrastructure involves financing, leases, commitments, and capital requirements that are not fully described by a single income-statement figure.

For an investor contemplating the IPO, the financing arrangements deserve as much attention as the growth rate. Anthropic could generate enormous future enterprise value while distributing a meaningful portion of that value to infrastructure providers, creditors, and holders of convertible securities.

The business might perform exceptionally well without every class of investor experiencing an equally exceptional return.

Modeling Anthropic in 2036

I constructed two financial scenarios to examine the consequences of different rates of customer adoption and infrastructure spending over the coming decade.

These are conditional exercises rather than predictions. Their purpose is to explore how revenue growth, inference economics, training expenditure, and the cost of capital interact.

Both begin with a quarterly revenue assumption of $14 billion in the third quarter of 2026. Inference expenses are held at 15% of revenue, approximately consistent with the implied first-half relationship between cost of revenue and sales.

I would not expect that relationship to remain unchanged for ten years. Hardware costs, model efficiency, customer mix, competitive pricing, and the composition of AI workloads will all change. Holding it constant allows us to examine other variables, but it is a consequential simplifying assumption.

Scenario One: Sustained Adoption

In the first scenario, Anthropic continues to capture expanding enterprise demand for AI services.

Quarterly revenue increases by $1.75 billion during each subsequent quarter. Training expenditure rises from approximately $20 billion annually toward $60 billion as the infrastructure supporting successive generations of models expands.

The model produces approximately $319 billion in annual revenue by the tenth year and $911 billion in cumulative after-tax free cash flow over the modeled period.

Discounting those cash flows at 8% produces a reported present value of approximately $561 billion.

The scenario illustrates the operating leverage available to a company whose inference costs remain comparatively low while revenue expands. Training expenditure can increase substantially without consuming the full economic benefit of a growing customer base.

There are several substantial assumptions embedded in this result.

Enterprise demand must continue to expand at a rate sufficient to support hundreds of billions of dollars in annual revenue. Anthropic must preserve attractive economics on customer-serving inference despite competition and potential price reductions. Its research organization must continue developing models that customers consider valuable enough to pay for, and sufficient physical capacity must become available to support those models.

The infrastructure must also be financed on terms that leave an appropriate share of the resulting cash flows with the company's investors.

The scenario requires a sustained combination of technological progress, commercial adoption, and financial execution. The company has demonstrated considerable ability in the first two areas. Its longer-term record on the third is still being established.

Scenario Two: Slower Adoption

The second scenario assumes that enterprise AI adoption continues, but at a substantially slower pace.

Quarterly revenue increases by $500 million rather than $1.75 billion. Inference expenses remain at 15% of revenue, while training expenditure rises from approximately $20 billion toward $40 billion annually.

The modeled company reaches approximately $131 billion in annual revenue by the tenth year and generates $164 billion in cumulative after-tax free cash flow.

Using a 25% discount rate, the reported present value declines to approximately $61 billion.

The two scenarios differ in more than their assumptions about revenue growth. They also apply substantially different discount rates, so the difference in present value reflects both operating performance and the price assigned to uncertainty. The higher discount rate penalizes future cash flows much more heavily.

Consequently, the comparison does not isolate the effect of slower customer adoption.

It does illustrate how a company whose research and infrastructure expenditure cannot readily contract with revenue growth may experience much weaker cash generation, even while continuing to become an enormous business.

A reduction in training investment could reduce financial pressure, but it might also affect Anthropic's ability to remain competitive. The company could be forced to choose between preserving near-term cash flow and funding the next generation of models.

Comparing the Scenarios

Measure Sustained adoption Slower adoption
Quarterly revenue increase $1.75B $0.5B
Annual revenue, year 10 ~$319B ~$131B
Annual training expense, year 10 ~$60B ~$40B
Assumed inference cost 15% of revenue 15% of revenue
Cumulative after-tax FCF ~$911B ~$164B
Discount rate 8% 25%
Reported NPV ~$561B ~$61B

The figures are rounded outputs of the supplied scenario model. The original quarterly cash-flow workbook was not available for independent recalculation. These present values are illustrative and should not be interpreted as independently verified common-equity valuations.

I find the sensitivity of the financial outcomes more interesting than either specific valuation.

A relatively small change in the rate at which an increasingly large business adds revenue each quarter compounds into a substantial difference over ten years. Meanwhile, the cost of remaining competitive continues to rise in both scenarios.

Anthropic could produce many times its present revenue and still generate disappointing returns for shareholders if the cost of developing frontier models rises faster than the commercial value those models create.

There is also a potentially much better outcome than either scenario captures. Improvements in training efficiency could reduce the resources needed to develop more capable models. If those improvements occur while enterprise demand continues expanding, the resulting cash generation could exceed these modeled outcomes.

We have too little evidence about the long-term economics of model development to confidently extrapolate today's expenditure trends through 2036.

Who Pays for the Infrastructure?

Anthropic does not have to purchase every accelerator or finance every data center directly. Its infrastructure arrangements include equipment leases, agreements with cloud providers, partner-financed facilities, and financing instruments that can defer certain cash expenditures.

The distinction between ownership and economic exposure becomes important here.

A partner may finance construction and own the facility, allowing Anthropic to avoid the initial capital expenditure. Anthropic may nevertheless be obligated to make payments for the capacity under a long-term contract.

Similarly, convertible promissory notes may be used to finance certain lease obligations under the arrangements described in the S-1. This can preserve liquidity, but the resulting securities create contractual claims that may eventually be settled through conversion into equity.

The cash expenditure has been deferred or financed. It has not ceased to be an economic cost.

These arrangements can be advantageous. Capital-intensive businesses have long used leasing, project financing, and partnerships to fund infrastructure that supports future revenue. Properly structured, such financing allows a business to increase productive capacity without requiring shareholders to provide all the necessary capital immediately.

But the economics depend on the cost of financing relative to the return generated by the assets.

If Anthropic can deploy additional computing capacity profitably and finance it at an attractive cost, the strategy may increase the value attributable to existing shareholders.

If the assets earn inadequate returns, the financing obligations will still have to be satisfied. Future cash flow may be absorbed by creditors and infrastructure partners, or additional equity may have to be issued.

The potential for dilution is particularly relevant to Anthropic's planned IPO.

What Will Shareholders Own?

The capitalization analysis accompanying the financial model suggests approximately 15.4 billion fully diluted shares following the conversion of existing preferred securities, notes, and equity awards.

More fundamentally, the share count at the IPO provides only a starting point.

Anthropic's infrastructure expansion could require substantial additional financing over the next ten years. If some of that financing takes the form of securities convertible into common shares, the ownership interests of existing investors may decline.

The financial consequences depend on the conversion prices, contractual provisions, and the value of the company when those securities are issued or converted.

Dilution is not necessarily evidence of financial failure. Issuing equity to finance an investment that earns more than its cost can increase the economic value of an investor's remaining ownership.

An investor buying shares in Anthropic will be purchasing an ownership claim on the results of a prolonged infrastructure investment program, including whatever financing is required to complete it.

The eventual return will depend on the price paid for that claim and the proportion of future economic value it represents.

The Difficulty of Forecasting a Moving Technology

There is a problem with projecting training expenditures to 2036 that deserves more attention than the financial model can provide.

We do not know whether developing frontier models will require the same kind of computational expansion that has characterized the past several years.

Current training strategies rely heavily on assembling large amounts of computing power. Extrapolating that approach produces progressively larger clusters, greater electricity demand, and enormous infrastructure requirements.

It is a plausible development path, but it is not inevitable.

New model architectures, algorithmic improvements, specialized systems, and more efficient training methods could change the relationship between computing expenditure and model capability.

The market for AI services could change just as dramatically. If model performance becomes increasingly interchangeable, competition may push prices toward the cost of inference. Efficiency improvements would then benefit customers more directly than providers.

Alternatively, if frontier models continue to deliver differentiated capabilities that enterprises value, companies such as Anthropic may retain considerable pricing power.

An especially difficult possibility for infrastructure investors is that a technological breakthrough could reduce the value of capacity already contracted under an earlier development strategy.

The same commitments that protect Anthropic against shortages of computing resources could become financially restrictive if the technology changes in an unexpected direction.

This is familiar territory for industrial investors. The economics of a large capital project depend on assumptions about demand, production technology, operating costs, and competition. Those assumptions can change before the project reaches its intended level of utilization.

AI compresses the technological cycle while leaving much of the physical investment cycle intact.

Anthropic must finance infrastructure on the timescale of industrial construction while developing products on the timescale of software research.

There may be no greater challenge in its business model.

Looking Toward 2036

Anthropic's financial disclosures describe a company growing at a rate that few businesses have ever sustained, supported by a computing infrastructure program of extraordinary scale.

Its apparent inference economics offer considerable potential for operating leverage. As enterprise usage expands, the company may be able to generate enough cash to finance continuing investments in frontier model development and still produce substantial returns for shareholders.

Yet there are several ways for that outcome to disappoint.

Inference pricing could decline faster than serving costs. Training expenditures might rise without producing proportionate improvements in capability or revenue. Infrastructure could become more expensive, arrive later than expected, or become less useful before the underlying financial obligations expire. Financing arrangements could also distribute more of the resulting economic value to creditors, equipment providers, and new investors than today's shareholders anticipate.

My interest in Anthropic's next decade therefore centers on three relationships: the growth of revenue relative to inference costs; the commercial value of successive model generations relative to the cost of developing them; and the return on infrastructure investment relative to its cost of financing.

The company's extraordinary growth gives it considerable room to succeed. It also creates the temptation to assume that enough revenue growth will eventually solve every other financial problem.

Industrial history offers plenty of examples in which demand forecasts proved correct but investors still lost money. Railroads, telecommunications networks, semiconductor manufacturing, and energy infrastructure have all experienced periods when the economic value delivered to customers exceeded the financial returns captured by the capital providers.

Anthropic is entering a period when its capital commitments will have consequences for many years, while the technological and competitive assumptions behind them may change several times.

By 2036, we will have a much better understanding of how valuable artificial intelligence has become to the economy. Whether Anthropic's shareholders capture a proportionate share of that value is a separate question.

Sources and Method

The historical financial information and infrastructure commitments referenced here are drawn from Anthropic's Form S-1 filed October 26, 2026, supplemented by contemporaneous financial reporting and estimates of the allocation of computing expenses between inference and model development.

The valuation scenarios use illustrative assumptions concerning revenue growth, inference and training expenses, capital expenditure, working capital, financing, taxation, and discount rates. Their reported present values have not been independently reproduced from a complete quarterly cash-flow workbook.

The analysis represents my personal views and is not investment advice or a recommendation to purchase or sell securities. It does not represent the information, industrial strategy, or position of any current or former employer.

Definitions