Brad Gerstner says Anthropic could support its 2027 revenue forecast with about 500 MW of compute
Anthropic / Wikimedia Commons (Public domain)
The investor's efficiency claim lands as Anthropic's revenue surges and its infrastructure commitments climb toward the trillion-dollar range
Investor Brad Gerstner has a bold claim about Anthropic. He says the AI company could support its 2027 revenue forecast using about 500 MW of compute for paying customers.
In an industry where data-center appetites are measured in gigawatts, that is a surprisingly modest number. It also lands as Anthropic’s spending commitments run into figures most countries would struggle to budget for.
What Gerstner is actually claiming
The core of the argument is efficiency. If Anthropic can serve its paying customers and hit its revenue targets with roughly 500 MW of capacity, each megawatt is earning a lot more money than skeptics might assume.
There are caveats worth keeping on the table. The claim has not been verified by independent sources. The exact 2027 revenue forecast it rests on has also not been made public.
That second point matters more than it might seem. Without the target number, nobody outside the company can calculate the revenue per megawatt Gerstner is implying, let alone check whether it holds up.
The phrasing is also hedged. Gerstner says Anthropic could support the forecast at that level, not that it will. That leaves a meaningful gap between a plausible scenario and a guaranteed outcome.
The numbers behind the efficiency debate
Anthropic’s top line has been moving fast. The company’s revenue reached an estimated $17 billion in Q3 2026, up from $1.4 billion a year earlier.
Its nine-month revenue total came in at $33.2 billion.
Costs, however, have been climbing faster than sales. In Q3 2026, Anthropic’s computing expenses rose nearly 77% to nearly $10 billion, while revenue increased approximately 47%.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Then there is the long-term tab. Anthropic’s future compute and infrastructure commitments are estimated at around $895 billion to $901 billion, with $451.6 billion expected before 2030.
Other estimates put the range between $900 billion and $1 trillion.
Why 500 MW and 3.5 GW can both be true
At first glance, Gerstner’s number seems to clash with Anthropic’s broader footprint. The company is expected to have over 3.5 GW of compute capacity in 2026.
That is roughly seven times the 500 MW figure tied to paying customers. The gap suggests the company’s capacity is spread across different uses, not all of them directly serving customer workloads.
So the 500 MW claim is narrower than it sounds. It speaks to the cost of serving customers, not the full cost of running a frontier AI lab.
What this means for investors and the industry
Gerstner’s claim is, at heart, an argument about unit economics. If serving paying customers really takes only around 500 MW, then Anthropic’s revenue engine may be leaner than its headline spending suggests.
But the Q3 2026 figures point to real pressure. Compute expenses growing at nearly 77% against revenue growth of approximately 47% is not a trend that can continue indefinitely without squeezing margins.
For investors, the key question is whether Anthropic can improve model efficiency and customer monetization fast enough to outpace its rising infrastructure costs. Gerstner’s view is optimistic on that front, but it is one investor’s estimate rather than a disclosed company figure.
There are a few things worth watching. First, whether Anthropic or anyone else puts a public number on the 2027 revenue forecast, which would let outsiders test the 500 MW math.
Second, the ratio between compute spending growth and revenue growth in coming quarters.
Third, how the company’s capacity mix evolves. With over 3.5 GW expected in 2026, the share devoted to paying customers versus everything else will say a lot about where the money is really going.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

