Tunguz Sees 80 GW AI Ceiling; Anthropic Compute 58% of Sales

Tomasz Tunguz of Theory Ventures puts the ceiling for AI in 2027 at “about 80gw : doubling all the compute built in the history of the country in twelve months,” in a 4 November essay and post on X. His figure adds 43 gigawatts coming online in 2027 to about 38 gigawatts installed in the United States by the end of 2026, both as he quotes the investor Brad Gerstner: 81 gigawatts, about 2.1 times the base, by our arithmetic.

Anthropic’s own filing shows what scarce compute costs a lab. Its Form S-1/A, filed on 2 November, puts compute and infrastructure expense for the July-September quarter at $9,863 million against revenue of $16,956 million. That is 58% of revenue, up from 48% in the April-June quarter and the first rise in the seven quarters the filing lists, by our arithmetic.

Business Pill · Margin Squeeze

An evergreen Business Engineer explainer on margin squeeze: when input costs rise and selling prices fall. It teaches the general idea only and does not discuss Anthropic, Tunguz or gigawatts.

What Tunguz Says

Tunguz argues that power, not demand, sets the limit. He writes that even at a record build pace, “short-term rental rates have jumped from $3 to $12-13 per hour, pushing Anthropic’s blended GPU-hour cost up 82% in just two quarters.” His footnote cites Anthropic’s S-1/A for a blended cost rising from $3.10 to $5.65 an hour. In a 3 November essay he says those per-hour rates “combine Anthropic S-1/A contract commitments” with public capacity disclosures, such as SpaceX’s prospectus.

The gigawatt figures come from Gerstner’s remarks on the All-In podcast’s election-night livestream, as Tunguz quotes them: “The consensus forecast is a total of 43 gigawatts coming online in 2027,” and the cumulative total “in the United States, is about 38.” Gerstner added: “I think we could see at least 10 to 15 gigawatts of the expected build not come online.”

If 10 to 15 gigawatts slip, 2027 additions would be 28 to 33 gigawatts, and the year-end total 66 to 71 gigawatts instead of 81, by our arithmetic on the quoted figures. Our earlier piece on Morgan Stanley’s 43 GW power gap covered a different measure, a net shortfall, which is not the same 43 gigawatts.

What Was Posted on X

Tomasz Tunguz posted this on 4 November. It says the ceiling for AI in 2027 is about 80gw, which he describes as doubling all the compute built in the history of the country in twelve months, and links an essay on five downstream impacts.

His Revenue Math, Checked

Tunguz says today’s 38-gigawatt fleet earns about $140 billion, and that the 43 gigawatts added in 2027 “should deliver an annualized earnings capability of about $350b, for a total of nearly $500b.” He assumes 40% of GPU hours go to inference, citing SemiAnalysis, and a 50/50 mix of $31 billion per gigawatt at frontier labs and $9.4 billion at the broader market.

His sums hold on his own inputs. Forty percent of 43 gigawatts is 17.2 gigawatts of new inference; the 50/50 mix gives $20.2 billion per gigawatt; 17.2 times $20.2 billion is about $347 billion, by our arithmetic. On the same per-gigawatt figure, Gerstner’s 10 to 15 gigawatts of slippage would take about $81 billion to $121 billion off that annualized revenue figure, also our arithmetic.

What Anthropic’s Filing Shows

The S-1/A’s quarterly table runs from January-March 2025 to July-September 2026. Compute and infrastructure expense was 251% of revenue in the first quarter of 2025, 144% in the fourth, 72% in the first quarter of 2026 and 48% in the second, before rising to 58% in the third, by our arithmetic.

In the third quarter, compute and infrastructure expense rose about 77% from the second quarter, while revenue rose about 47%, by our arithmetic. The filing says the swing from an operating income of $974 million to an operating loss of $567 million was “primarily driven by $5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments, and $2.0 billion of non-cash charges related to our donation matching program.”

It says Adjusted Operating Margin fell from 19% to 15% “primarily driven by the $5.0 billion of short-term compute noted above.” That $5.0 billion is about 51% of the quarter’s compute and infrastructure expense, by our arithmetic.

The company also states: “We expect our use of short-term compute to continue to increase in the near term and to in turn compress our operating margin and Adjusted Operating Margin as a result.” Our earlier pieces covered the same filing’s obligations table by year and the duration of its compute bills.

The Five Impacts Tunguz Lists

First, he writes that “foundation model companies may need to shift the fraction of GPU hours to inference from training.” Second, “latency becomes the release valve for too much demand.” Third, “limitations in compute availability will encourage algorithmic innovation,” and he cites Chinese models such as DeepSeek, Qwen and Kimi.

Fourth, “the harness becomes critical to deliver state-of-the-art performance on limited GPU availability.” Fifth, “the economics of vibe-coding face a reality check.” He writes that when inference becomes more expensive, off-the-shelf software can be less expensive than bespoke internal software.

Connecting the Coverage

Our related pieces on AI power and compute costs:

• Patel’s Data Puts Nvidia H100 Rent Up 65% in a Year at $2.80: GPU rental prices.
• Data Center Limits Lead 40 of 41 Votes on Unofficial Counts: local votes on data-center siting.
• Morgan Stanley’s Net US Data Center Power Gap Hits 43 GW: a different 43 GW measure, the net power shortfall.

The Structural Read

As we read Tunguz’s essay, his revenue figures follow from his inputs: 17.2 gigawatts of new inference at $20.2 billion per gigawatt gives about $347 billion, by our arithmetic.

As we read Anthropic’s quarterly table, compute and infrastructure expense grew about 77% in the third quarter while revenue grew about 47%, by our arithmetic.

As we read the two sources together, Tunguz’s ceiling is an estimate about the whole market, while the S-1/A is one company’s reported costs; neither proves the other.

Three Implications

FOR AI LABS Anthropic’s filing says it expects its use of short-term compute to increase in the near term and to compress its operating margin.

FOR GIGAWATT FORECASTS Gerstner, as Tunguz quotes him, said at least 10 to 15 gigawatts of the expected 2027 build may not come online.

FOR INFERENCE REVENUE Tunguz puts the 2027 total at nearly $500 billion on his own assumptions, including 40% of GPU hours going to inference.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of AI Data Centers.

The framework says: “Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”

As we read it, Tunguz’s 80-gigawatt ceiling is an estimate of how much of that power equation the US can solve in 2027. The framework does not discuss Tunguz, Gerstner or Anthropic’s filing.

What Is Not Established

The 80, 43 and 38 gigawatt figures are a forecast and an estimate that Tunguz quotes from Gerstner; we have not checked them against grid operator or utility data. The $3-to-$12 rental rates and the 82% rise in Anthropic’s blended GPU-hour cost are Tunguz’s figures, built, he says, from S-1/A commitments and other public disclosures; we have not rebuilt them. The S-1/A ties its third-quarter loss to short-term compute; it does not tie that compute to a shortage of power, and we draw no such link.

The Bottom Line

Tunguz’s case is that powered GPUs, not demand, cap AI in 2027, and his own revenue sums add up on his inputs. Anthropic’s filing shows the cost side for one lab: in the third quarter, its compute bill grew faster than its revenue for the first time in the seven quarters it reports, by our arithmetic, and the filing ties the drop in its Adjusted Operating Margin to $5.0 billion of short-term compute.