DEV Community

Cover image for Anthropic's $3.4B gas generator deal: what $2,000 per kW buys, and what the reports leave out
ai maya
ai maya

Posted on

Anthropic's $3.4B gas generator deal: what $2,000 per kW buys, and what the reports leave out

TL;DR

On November 8 the Financial Times reported that Anthropic agreed in late September to pay $3.4 billion for 1.7 gigawatts of heavy-duty gas-fired generators, taking over orders that a Texas equipment supplier, OnePWR Solutions, had placed with Caterpillar about two years earlier. Anthropic and OnePWR did not comment, and Reuters said it could not immediately verify the report, so everything below is "as reported." The headline division is simple: $3.4B / 1.7 GW = $2,000 per kW. That number sits at the top of the range of published gas-plant cost figures, but it is a bundled price that includes installation work, so it is not an equipment price and should not be compared to one directly. My arithmetic, with stated assumptions, puts the capital share of delivered electricity at roughly half the cost per kWh, and the fuel burn at somewhere between 0.17 and 0.32 billion cubic feet of gas per day depending on how hard the machines run. The deal is about 3 to 4 percent of what Nvidia's CEO says a gigawatt of AI infrastructure costs, which is the more useful way to read its size.

Everything marked "my arithmetic" or "illustrative" is a calculation from published figures plus assumptions I state. The deal terms belong to the FT and the outlets that summarized it. I have not seen the contract.

What exactly was reported?

I checked the FT-derived summaries from several outlets (The Next Web, MLQ, Cellcog and others). They agree on the core facts and differ on the details, which is normal for a story that sits on one primary report.

Item As reported Notes
Buyer Anthropic Declined to comment
Seller OnePWR Solutions, San Antonio, Texas Equipment supplier, not acquired
Equipment Caterpillar heavy-duty gas-fired generators Models and unit count not reported
Capacity 1.7 GW Nameplate, per the reports
Price $3.4B Includes the work needed to install the machines
Timing Commitment in late September, public November 8 No delivery schedule published
Sites Not known Reports say air permits are required
Structure Anthropic took over OnePWR's Caterpillar orders Anthropic first planned to buy power from OnePWR, then took over the orders instead

Two details matter for everything that follows. First, the price covers installation capability, and no source I found splits equipment from services. Second, nobody has said where the machines will run or whether they are for primary power, bridge power, or backup. Those are inferences in the secondary coverage, not reported facts.

Is $2,000 per kW expensive for gas power?

Short answer: it is at the top of the published range, but the comparison is loose. Here are the reference points I could verify.

  • The U.S. Energy Information Administration's AEO2025 cost study, prepared by Sargent & Lundy, puts a 1x1 F-class combined-cycle plant at $921 per kW in 2023 dollars.
  • A September 2025 GridLab report, using project data from Halcyon, says gas turbine projects due in 2026 and 2027 were reported at $1,116 to $1,427 per kW, and that recent combined-cycle projects "are routinely reporting costs of $2,000/kW or more." It also argues that public datasets lag the market because turbine makers have order backlogs and pricing power.

USD per kW across published gas-plant cost figures and the reported Anthropic deal

Read the chart with three cautions. The EIA figure is in 2023 dollars and is a reference estimate, not a quote. The GridLab "$2,000 or more" is a floor from recent projects, so the bar understates it. And the Anthropic figure is a quotient of two reported round numbers. If the real figures were $3.35B and 1.72 GW, the answer would be $1,948 per kW, and nobody outside the contract would know.

What the chart does support: a buyer paying roughly $2,000 per kW all in, for equipment ordered about two years ago and already in a manufacturing queue, is paying about what the market reports for new projects today. It is not an obvious overpayment, and it is not a bargain. What the buyer is really purchasing is time. Turbine and engine slots are scarce, and an order that is already placed is worth more than a price list suggests.

How much gas does 1.7 GW burn?

This is my arithmetic and depends on a heat rate I am assuming, not one that was reported. I use 8.5 MMBtu per MWh, which corresponds to roughly 40 percent electrical efficiency, a plausible figure for gas-fired generators of this class but not a specification of the machines in the deal.

mw = 1700            # reported nameplate, MW
heat_rate = 8.5      # MMBtu per MWh, assumed (about 40% efficient)
bcf_per_mmbtu = 1 / 1.037e6   # 1 Bcf is about 1.037 million MMBtu
us_gas_2025 = 92.0   # Bcf per day, EIA record annual average

for cf in (0.5, 0.7, 0.85, 0.95):
    mmbtu_h = mw * heat_rate * cf
    bcf_d = mmbtu_h * 24 * bcf_per_mmbtu
    print(f"CF {cf:.0%}: {bcf_d:.3f} Bcf/d = {bcf_d / us_gas_2025:.2%} of US demand")
Enter fullscreen mode Exit fullscreen mode

Output:

CF 50%: 0.167 Bcf/d = 0.18% of US demand
CF 70%: 0.234 Bcf/d = 0.25% of US demand
CF 85%: 0.284 Bcf/d = 0.31% of US demand
CF 95%: 0.318 Bcf/d = 0.35% of US demand
Enter fullscreen mode Exit fullscreen mode

Illustrative gas burn of 1.7 GW at four capacity factors

So the fuel appetite is real but small against a national market: at full output about 0.33 Bcf per day, roughly 0.36 percent of the 92.0 Bcf per day the EIA reported as the 2025 U.S. average. The EIA reports 2025 as a record consumption year.

The carbon arithmetic follows from the same assumption. EPA's standard emission factor for natural gas is 53.06 kg of CO2 per MMBtu, so 8.5 MMBtu per MWh gives about 451 kg of CO2 per MWh. At 100 percent output for a year, 1.7 GW produces 14.9 TWh and about 6.7 million tonnes of CO2. At 50 percent it is about 3.4 million tonnes, at 85 percent about 5.7 million. These are upper-bound style figures in the sense that they assume simple-cycle-like efficiency and no emissions controls beyond what the factor already reflects. Real numbers depend on the machines, the duty cycle, and the permits that, according to one summary of the reports, have not yet been obtained.

What does this do to the cost of a kilowatt-hour?

Capital cost per kW is only half of a power cost. Here is an illustrative levelized view using the reported $2,000 per kW, an 8 percent cost of capital over 20 years (my assumption), and natural gas at $3.50 per MMBtu (also my assumption, and it moves the answer a lot).

capex_kw = 2000.0
r, n = 0.08, 20
crf = r / (1 - (1 + r) ** -n)       # capital recovery factor = 0.1019
cap_per_kw_year = capex_kw * crf    # about $203.70 per kW-year

gas_price = 3.50     # USD per MMBtu, assumed
heat_rate = 8.5      # MMBtu per MWh, assumed
fuel_per_kwh = gas_price * heat_rate / 1000   # 2.975 cents

for cf in (0.5, 0.7, 0.85, 0.95):
    kwh_per_kw_year = 8760 * cf
    cap = cap_per_kw_year / kwh_per_kw_year
    print(f"CF {cf:.0%}: capital {cap*100:.2f}c + fuel {fuel_per_kwh*100:.2f}c = {(cap+fuel_per_kwh)*100:.2f} cents/kWh")
Enter fullscreen mode Exit fullscreen mode
Capacity factor Capital (cents/kWh) Fuel (cents/kWh) Total before O&M
50% 4.65 2.98 7.63
70% 3.32 2.98 6.30
85% 2.74 2.98 5.71
95% 2.45 2.98 5.42

Illustrative cents per kWh by capacity factor

The shape matters more than the decimals. Capital cost per kWh falls by almost half as you move from a 50 percent to a 95 percent capacity factor, while fuel stays flat. A buyer who runs these machines as backup or peaking capacity pays much more per kWh than one who runs them as baseload, which is one reason the unanswered question of how Anthropic will use them is not trivia. These figures exclude operations and maintenance, interconnection, redundancy, and any emissions equipment, so treat them as a floor for an on-site plant, not a quote.

How big is this next to the compute it would power?

Nvidia CEO Jensen Huang has said repeatedly that one gigawatt of AI data center capacity costs $50 billion to $60 billion in total, with about $35 billion of that going to Nvidia chips and systems. That figure bundles chips, buildings, cooling, and power infrastructure.

If all 1.7 GW could be turned into facility capacity (it cannot be, since nameplate generator capacity usually needs spare units for maintenance, and not every watt reaches the racks), the implied AI build would be 1.7 x $50B to 1.7 x $60B, which is $85B to $102B. Against that, $3.4B is 4.0 percent to 3.3 percent. This is a ratio of two round numbers and a caveat, but the direction is robust: the generators are a small line item next to the hardware they would feed. What they buy is not compute but the ability to turn that compute on without waiting for a utility interconnection.

That is the real story under the headline. For a lab, a few billion dollars is cheap insurance against billions of dollars of idle accelerators.

What do the reports not tell us?

  • Where the machines go. No site has been reported, and without a site there is no way to check permits, local air rules, or grid interconnection.
  • The equipment-to-services split. If installation and engineering are a large share of the $3.4B, the equipment price per kW is lower than $2,000 and the comparison with turbine-only figures changes.
  • The machine type. "Heavy-duty gas-fired generators" from Caterpillar does not say whether these are reciprocating engines or turbines. Efficiency, ramp rates, and cost per kW differ.
  • Whether the number reconciles with OnePWR's earlier disclosures. MLQ's write-up notes a gap between the reported 1.7 GW and the 1.2 GW of available inventory that OnePWR announced in February 2026. I have not verified that announcement, so treat the gap as an open question rather than a finding.
  • Delivery dates. One summary says the equipment would arrive "near term and over the next two years." No primary source for the schedule has been published.

FAQ

How much did Anthropic pay per kilowatt?
As reported, $3.4 billion for 1.7 gigawatts, which is $2,000 per kW. The price includes the work needed to install the generators, so it is not a pure equipment price.

Is $2,000 per kW high for gas generation?
It is above EIA's AEO2025 reference of $921 per kW for a combined-cycle plant in 2023 dollars and above the $1,116 to $1,427 per kW GridLab reports for 2026 and 2027 projects. It is about equal to what GridLab says recent combined-cycle projects routinely report. The deals are not like-for-like, so use it as a range check, not a verdict.

How much natural gas would 1.7 GW use?
Under an assumed heat rate of 8.5 MMBtu per MWh, about 0.33 billion cubic feet per day at full output and about 0.17 at a 50 percent capacity factor. U.S. consumption averaged a record 92.0 billion cubic feet per day in 2025, per the EIA, so full output is about 0.36 percent of national demand.

Why would an AI lab buy generators instead of buying power?
The reports say Anthropic first planned to buy power from OnePWR and then took over the orders instead. A plausible reading, which is mine and not reported, is that owning scarce generator slots and installation capacity shortens the time between delivered accelerators and powered racks.

Are the generators confirmed to power Anthropic's data centers?
The FT report describes them as for AI data centers, but no sites have been disclosed. Whether they would serve as primary, bridge, or backup power has not been reported.

Has the deal been independently confirmed?
Not as of this writing. Reuters said it could not immediately verify the FT report, and Anthropic and OnePWR did not respond to requests for comment.

Sources

Further reading

Top comments (0)