Analysis•

Brad Gerstner Argues 500 Megawatts Is Enough for Anthropic to Hit Its 2027 Revenue Forecast

Investor Brad Gerstner says roughly 500 megawatts serving paying customers could carry Anthropic's 2027 revenue forecast, a fraction of the 3.5 gigawatts the company is expected to run.

Written By :MINRK
Brad Gerstner Argues 500 Megawatts Is Enough for Anthropic to Hit Its 2027 Revenue Forecast

Photo by Raymond Sime on Unsplash

1. A Small Number in a Gigawatt Industry

Investor Brad Gerstner, the founder of Altimeter Capital, has argued that Anthropic could meet its 2027 revenue forecast with about 500 megawatts (MW) of computing power devoted to paying customers. The claim, which drew attention on November 17, 2026, puts the Anthropic compute requirement for customer-facing work at roughly one seventh of the more than 3.5 gigawatts (GW) the company is expected to operate this year.

That contrast is what makes the remark notable. Artificial intelligence developers now describe their infrastructure in gigawatts, a scale more commonly associated with power plants than with software companies, and Anthropic has taken on spending obligations so large that many national governments would struggle to fit them into a budget. Against that backdrop, half a gigawatt is a modest figure.

The argument does not come from Anthropic. It is one investor's reading of how much of the company's hardware is needed to produce the sales it books, and it rests on a revenue target that has not been made public by the company. It is nonetheless a useful window into the debate over whether the economics of frontier AI are better or worse than the headline spending suggests.

2. What Gerstner Is Saying, and What He Is Not

At its core, the point is about efficiency. If the workloads of paying customers can be handled on about 500 MW, and those workloads generate enough sales to reach next year's target, then every megawatt assigned to customers is producing more revenue than critics of the sector tend to assume.

Several limits apply to that reasoning. The 500 MW figure has not been independently verified. The 2027 revenue forecast it leans on has not been published, so the target itself cannot be checked from the outside. Without that number, nobody beyond the company and its investors can divide revenue by megawatts and test whether the implied productivity is realistic.

The wording matters as well. Gerstner said Anthropic could support the forecast at that level of compute. He did not say that it will, and he did not present the figure as a commitment or as guidance from management. The statement is a conditional estimate about capacity, not a projection of results.

It is also narrow in scope. The claim concerns only the computing power required to serve customers who pay for access to Anthropic's Claude models. It says nothing about the hardware needed to train new models, run research or support other internal work, which is where much of a frontier lab's capacity goes.

3. Why Power Has Become the Yardstick

Measuring an AI company in megawatts can seem odd, but it has become the standard shorthand for computing capacity. Data centers are limited less by floor space than by how much electricity they can draw and cool, so the power rating of a site is a practical proxy for how many chips it can run. A megawatt figure therefore describes the quantity of processing a company can put to work at any moment.

Using that yardstick, a claim about megawatts per dollar of revenue is really a claim about unit economics. It asks how much infrastructure must be switched on to earn a given amount of money. A company that needs less power per dollar of sales has more room to cover the cost of chips, buildings and electricity, while one that needs more is closer to selling its service at or below cost.

The measure has limits. Two sites with the same power rating can deliver different amounts of useful work depending on the chips installed and the software running on them, and a megawatt used for one kind of task is not interchangeable with a megawatt used for another. Capacity figures are best read as a rough indication of scale, not as a precise count of output, which is one more reason the 500 MW estimate should be treated as an approximation.

This is why a single number from an investor can attract attention. Much of the skepticism around AI spending centers on whether customer revenue can ever justify the capital being poured into data centers. A figure suggesting that the revenue side of the business runs on a relatively small slice of that capacity speaks directly to that concern, even if it cannot yet be confirmed.

4. Anthropic Compute Costs and Commitments in Context

The financial backdrop explains why the question is pressing. Anthropic's revenue for the third quarter of 2026 is estimated at about $17 billion, and its revenue for the first nine months of the year at $33.2 billion. Those are large sums for a company that sells access to AI models, and they show how quickly demand has grown.

Costs are substantial too. Spending on compute came to nearly $10 billion in the third quarter of 2026 alone. Set beside quarterly revenue of roughly $17 billion, that means well over half of each dollar of sales in the period was matched by a dollar of computing expense, before salaries, sales costs or anything else are counted.

The longer-term obligations are larger still. Anthropic's future commitments for compute and infrastructure are estimated at about $895 billion. These are agreements to buy or rent capacity over the coming years, and they dwarf the revenue the company currently brings in.

Taken together, the figures describe a business with strong sales, a heavy and rising bill for computing, and a very long list of obligations still to be paid. Whether that combination is sustainable depends on how productive the hardware turns out to be, which is precisely the issue Gerstner's estimate addresses.

5. How Can 500 MW and 3.5 GW Both Be True?

The two numbers measure different things, so they do not contradict each other. Anthropic is expected to have more than 3.5 GW of compute capacity in 2026, about seven times the 500 MW that Gerstner ties to paying customers. The difference is capacity that is used for purposes other than answering customer requests.

A frontier AI lab runs its hardware for several reasons at once. Some of it handles the requests that customers pay for, the activity usually called inference. A great deal more is consumed by training new models, by research and experimentation, and by other uses that do not produce revenue directly in the period in which they occur.

Read that way, the 500 MW figure is an estimate of the cost of serving customers rather than the cost of running the whole company. It isolates the part of the infrastructure that is tied to current sales and sets aside the part that is an investment in future products.

That distinction cuts both ways. It supports the view that the revenue-generating side of Anthropic is leaner than its total footprint implies. It also means the remaining capacity, the larger share, has to be paid for out of the same revenue, or out of capital raised from investors, even though it is not directly earning money today.

6. What the Claim Means for Investors

For investors, the estimate is an argument that Anthropic's revenue engine is more efficient than its spending suggests. If serving customers really does require only about 500 MW, then the company's sales are being produced by a small fraction of its hardware, and the rest of its outlay can be viewed as discretionary investment in research and future models rather than as the unavoidable cost of doing business.

The third-quarter figures show why that framing is contested. A compute bill of nearly $10 billion in a single quarter, alongside commitments of about $895 billion, leaves little margin for error. If the cost of infrastructure keeps climbing faster than the company can turn it into sales, margins come under pressure regardless of how efficient the customer-serving portion is.

The central question is therefore one of pace. Anthropic needs to improve the efficiency of its models and the money it earns from them quickly enough to stay ahead of what it has promised to spend. Gerstner's view is an optimistic answer to that question, but it remains one investor's estimate rather than a figure disclosed by the company, and it should be weighed accordingly.

The debate also reaches beyond a single firm. Much of the capital flowing into chips, data centers and power is justified by the expectation that AI services will produce revenue on a matching scale. Evidence about how much capacity a leading developer needs to serve its customers, even indirect evidence, feeds into how investors judge that wider bet.

7. What to Watch Next

Three developments would help settle whether the 500 MW estimate holds up. The first is whether a public figure for Anthropic's 2027 revenue emerges. Until a target is on the record, the arithmetic behind the claim, revenue divided by megawatts, cannot be checked.

The second is the relationship between compute spending and revenue in coming quarters. If sales grow faster than the computing bill, the case for improving unit economics strengthens. If the bill keeps growing faster than sales, the efficiency of the customer-serving portion matters less, because the overall cost base is still outrunning income.

The third is any detail on how Anthropic divides its capacity between paying customers and everything else. A breakdown of that kind would show directly how much of the more than 3.5 GW is tied to revenue and how much is devoted to training and research, replacing an outside estimate with the company's own numbers.

Until then, the claim stands as an informed but unverified argument: that the part of Anthropic which earns money is small relative to the part that spends it, and that the gap reflects a choice to invest in future models rather than a weakness in the business of selling the current ones.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY