Anthropic commits $3.4B to take over 1.7GW of Caterpillar gas generator orders
The largest known power-equipment purchase by an AI lab turns energy into a first-class compute dependency, and a scheduling problem for GCC AI builds.
What happened
Summary of reporting by Financial Times (via Techmeme)Anthropic has taken over a book of gas-fired generator orders from OnePWR, the equipment supplier, in a deal the Financial Times values at about $3.4B. The package covers roughly 1.7GW of Caterpillar generation capacity plus the engineering and construction capability to install it. Anthropic did not buy OnePWR itself, and no OnePWR staff moved across.
The structure matters more than the price. OnePWR had placed the Caterpillar order about two years earlier, before equipment prices climbed, and then needed financing to complete the purchase. Anthropic had looked at buying electricity from the supplier, and instead stepped into its position in Caterpillar's order book. The FT calls it the largest known single purchase of power-generation equipment by an AI lab. Piper Sandler, named in the report as OnePWR's exclusive financial adviser, describes a $3.39B cash sale of about 1.74GW of contracted equipment, part of it available for near-term deployment and the rest over two years.
The backdrop is a scramble for generation. Anthropic has said it will invest $50B in data centres with Fluidstack in Texas and New York, and its S-1/A filed with the SEC on 2 November 2026 discloses $895.4B of purchase commitments and lease agreements, substantially all of it compute and infrastructure, with $447.5B expected before the end of 2030. Peers are making the same move: xAI bought mobile gas turbines for about $2B, and Alphabet acquired the energy developer Intersect. Equipment makers are now signing data-centre generation deals directly, including Caterpillar supplying 2GW of Cat G3516 fast-response gas generator sets to AIP Corp's planned Monarch Compute Campus in West Virginia, with deliveries scheduled from September 2026 through August 2027, and a 500MW gas-plus-carbon-capture project with OnePWR and Vero3.
Reporting: Financial Times (Martha Muir), 8 Nov 2026, https://www.ft.com/content/06be395b-8d82-4d7a-b5d5-1f753674d80f. Also indexed by Techmeme: https://www.techmeme.com/261108/p8#a261108p8.
The Azrty take
Power procurement is now compute procurement, and GCC AI plans that treat it as a facilities problem will lose 2027 slots
For anyone in the UAE or the wider GCC sizing AI capacity, the read-through is blunt: your critical path is megawatts and interconnection, not GPU SKUs. Anthropic just committed $3.4B to generators rather than chips, against $895.4B of purchase commitments and leases that are substantially all compute and infrastructure, per the S-1/A it filed with the SEC on 2 November 2026: $447.5B of that falls before the end of 2030, and $69.8B in the twelve months after 30 September 2026 (https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm). Divide the deal's two reported figures and you get roughly $2M per megawatt of generation capacity, before land, cooling, GPUs or staff. If your 2027 and 2028 AI plan has a GPU line but no written energisation date and no generation story, it is a wish list, not a plan. On the demand side of the same ledger sits Anthropic's $50B, two-site data-centre programme with Fluidstack in Texas and New York (https://www.anthropic.com/news/anthropic-invests-50-billion-in-american-ai-infrastructure).
The mechanics are the interesting part. Anthropic did not buy a power company. It bought a position in Caterpillar's order book that OnePWR had secured about two years earlier at pre-surge prices, plus the engineering, procurement and construction capability to install the kit, with no staff transfer. That is a direct answer to manufacturing lead times measured in years. The comparables are stacking up: xAI's roughly $2B of mobile gas turbines, Alphabet's acquisition of Intersect, and Anthropic's own Fluidstack programme. Caterpillar, which makes the units, is signing data-centre generation deals of its own: 2GW of Cat G3516 fast-response gas generator sets for AIP Corp's Monarch Compute Campus in West Virginia, delivering from September 2026 through August 2027 against a site target of 8GW, with battery storage for load swings, a ramp from zero to full load in about seven seconds, and selective catalytic reduction for emissions (https://www.caterpillar.com/en/news/corporate-press-releases/h/aip-boyd-cat.html); and a 500MW gas-plus-carbon-capture project with OnePWR and Vero3. Gas is the bridge fuel of this build-out, and frontier labs are now stockpiling it as equipment. (Financial Times: https://www.ft.com/content/06be395b-8d82-4d7a-b5d5-1f753674d80f; Techmeme index: https://www.techmeme.com/261108/p8#a261108p8)
In the GCC the constraint differs in degree, not in kind. Abundant gas and strong utilities do not equal delivered megawatts at your campus fence on your timeline, and the same equipment queues apply. Two failure modes show up again and again. Power stays with facilities and procurement while the AI team signs GPU contracts, so the two never share a schedule. And teams treat grid availability as usable campus capacity, ignoring that liquid-cooled racks in the 100 kW class change the electrical and cooling picture, and that gas generator output derates in Gulf summer ambient. From Dubai, we see the gap: compute roadmaps with no energy roadmap behind them.
How Azrty would approach it technically: treat power as a schedulable resource in the AI platform, not a facilities document. Keep a capacity register per site (contracted kVA, generator nameplate derated to design ambient, N+1 status, fuel hours). Scrape PDU, BMC and CDU telemetry into Prometheus, publish headroom per site, and put admission control in the platform layer so new workloads cannot land where headroom is gone. In our own stack, Kuvryn AI gives each team its own space for agents and models on your GPU servers, and FastLLM Proxy keeps routing and budgets in one place, so you can defer or shed non-urgent batch traffic when a site is tight. Start with the telemetry rule:
- name: site-power
rules:
- record: site:power_headroom_pct
expr: |
100 * (1 -
(sum by (site) (node_power_watts{env="ai"}) / 1000)
/ on(site) group_left
site_contracted_kw
)
- alert: SitePowerHeadroomLow
expr: site:power_headroom_pct < 15
labels: {severity: page}
The deeper risk for GCC organisations is buying nameplate without duty cycle, or filing this news under American curiosity. Unabated gas generation will attract emissions scrutiny too, which is why the Caterpillar, OnePWR and Vero3 pairing adds carbon capture and sequestration to a 500MW first project. Teams that negotiate power, cooling and GPU allocation on one schedule, and price the bridge generation now, will get 2027 capacity. Teams that do not will queue.
What to do now
- Create a site power register in your AI capacity repo this month: contracted kVA per site, written energisation dates from the utility or operator, generator nameplate derated to your design ambient (use 45C for Gulf sites), and N+1 status. Review it in the same meeting as GPU procurement.
- Stand up the Prometheus recording rule and alert above with a Grafana panel per site, alert below 15% headroom, and wire FastLLM Proxy budgets to defer or shed batch workloads when a site crosses that threshold.
- Ask your data centre or utility partner for written equipment lead times and energisation dates for 2027 and 2028 capacity, and price a gas or HVO-fuelled generation bridge for the gap. Require the vendor's output curve at 45C ambient before any commitment.
- Before signing a 2027 GPU order, run a power-aware platform assessment with Azrty. Our AI infrastructure work starts with platform assessment and design, and our team in the UAE does the deployment and support.