Anthropic takes on uncapped gas bills as turbine makers build their backlog

The S-1 says Anthropic pays fuel on a pass through basis with no cap, for up to 20 years, whatever the utilization. A day later Baker Hughes disclosed 1.1 GW of gas turbines for a US technology company.

Anthropic has told investors it will pay whatever gas costs at some planned data centers, with no cap. The line sits in a risk factor of its S-1 filing of 26 October. AP reported that Anthropic will list on Nasdaq as ANT next month. Reports say the offering could raise up to $100 billion.

Turbine makers book the equipment orders, while the lab carries the fuel bill. A day later Baker Hughes disclosed the other side of that kind of contract. It is 1.1 GW of gas turbines for an unnamed US technology company. The power order was booked in the third quarter.

What the fuel terms require

The S-1 says certain planned data centers rely on dedicated on site gas fired generation. That exposes the company to fuel price swings and supply disruption "for which we bear the cost on a pass-through basis". Pass through means the lab pays the actual cost.

The filing then removes any ceiling. The arrangements "generally provide for fuel and related costs to be borne by us on a pass-through basis without cost caps". So Anthropic "remain[s] exposed to increases in the cost of natural gas and its transportation".

Rivals for the same gas push the price up. The filing expects "increased competition among users for the same natural gas supply and pipeline capacity" in the areas where on site generation is allowed. That competition "could drive up the price of natural gas". A site can also be delayed or abandoned if firm gas supply and transport cannot be secured during development.

The terms also run long. Certain power and data center arrangements have initial terms of 20 years with annual escalators. They include take or pay compute payments and uncapped fuel pass throughs that apply "regardless of actual utilization". Take or pay means the lab owes the money whether or not it uses the capacity.

Turbine orders and the equipment queue

Baker Hughes shows where the equipment money lands. Its order release covers 34 Frame 5 gas turbines, nine NovaLT16 turbines, generators and gears. Together they provide about 1.1 GW.

The buyer is "a major U.S. technology company", and Anthropic is not named. The power order was booked in the third quarter and the Chart dry coolers in the fourth.

The equipment figures support the 1.1 GW total. Nine NovaLT16 units at 16.9 MW each give 152 MW. The third quarter results list a separate order of 18 Frame 5 packages at about 500 MW, so 27.8 MW each.

At that rate 34 Frame 5 packages give about 944 MW. The two sets sum to about 1,096 MW on our calculation.

That separate order is the second data center gas deal of the quarter, for "a U.S. energy company". Buyers are not limited to one name or one kind of company, so the turbine queue is not hostage to a single lab.

Baker Hughes has a substantial backlog, though the slowdown in new orders is an early signal to watch. Baker Hughes booked $5.9 billion of industrial and energy technology orders, up 42% on the year. That was down 17% from the second quarter, though Baker Hughes still called it the unit's second largest order quarter.

Gas Technology Equipment orders fell 36% from the second quarter to $3.1 billion. Orders still ran at 1.7 times sales across the unit, which lifted its record backlog to $39.4 billion. Management raised its 2026 to 2028 order outlook to "$50 billion plus".

Management credits "data centers, space, LNG and gas infrastructure" for that demand. AI sites are one source among several.

Anthropic describes the same equipment queue. Turbines, generators and transformers have faced "extended delivery timelines, price increases, tariffs" and allocation limits. It may have to buy long lead equipment above market rates.

Where a developer buys and finances the equipment, Anthropic may repay the cost and financing "over time". In the filing's words, "we may bear these costs even if the underlying capacity is delayed or never realized".

Commitments, cash flow and the next test

Fixed bills are growing fast. Purchase commitments rose from $54.6 billion at the end of 2025 to about $265.6 billion at 30 June 2026. The note ties these mainly to cloud hosting and other third party infrastructure services.

Leases sit outside that figure. Since 31 December the company has signed new and amended lease and commercial agreements with minimum payments of about $417 billion. They are mostly compute equipment, data center leases and cloud purchases that start in 2026 and 2027.

The chart shows the fixed bills around these sites, not the fuel bill itself. The fuel clause lives in the power and data center arrangements, which include leases of up to 20 years.

Credit support points the same way. In March 2026 Anthropic guaranteed up to $500 million of a data center provider's obligations under a power purchase contract. In April it put $2.5 billion into escrow for one compute and infrastructure lease.

Revenue is growing faster than the bills so far, making the fuel clause easier to overlook. First half revenue was $16.3 billion against $1.2 billion a year earlier. Compute and infrastructure expense fell to 55% of revenue, and operating cash flow was positive at $3.9 billion.

The operating loss for the half was $693 million. The filing says the shift is toward leases and capital spending "that require significant upfront payments well in advance of the related compute becoming operational". Some of these commitments run to 2045.

The danger is a flat revenue year against fixed obligations and an uncapped fuel bill. A fuel collar or index cap in the lease would limit that bill. So would firm fuel transport contracted before construction starts.

For turbine makers, the question is whether data center slots for 2028 and beyond keep selling. Baker Hughes takes questions on its 28 October call, at 9:30 a.m. Eastern. Its figures on turbine slots and data center orders will show how much of the $39.4 billion backlog comes from AI sites.