From SemiAnalysis, September 10:
Dumb Science Experiments vs. Money Printing Machines
Last year we were the first to call out Onsite Gas Generation as the primary method adopted by AI Labs and Hyperscalers to solve power constraints. Our positive view was far from being consensus: behind-the-meter primary power solutions have been called all sorts of names, such as “science experiments”, “Dark Gigawatts”, and “literally the dumbest thing that human beings have ever attempted to do”!
But since then, that supply chain has witnessed a massive acceleration. Our Energy Model now tracks 75GW of firm, binding orders in the supply chain only for for behind-the-meter AI compute - of which ~20GW alone ordered in Q2 2026. What started as an Elon Musk experiment is now mainstream for every single AI Lab and hyperscaler. To be clear, this data does not include the hundreds of GWs of speculative, baseless announcements that many other analysts track in their numbers - we only focus on binding orders received by OEMs specifically serving BTM AI compute, tracked at the project-level.
Sources: SemiAnalysis Energy Model; sales@semianalysis.com
The path from firm equipment order to delivered project is still long and challenging. There is substantial execution risk and that’s what we’ll focus on in this report. But the industry is more experienced than you’d think: by the end of the year, ~3GW of operational US datacenter IT capacity will be powered behind-the-meter, and that number will experience multiple straight years of triple-digit growth. Our Energy and Datacenter models account for all potential delays, as we’ve explained in depth in our piece Stop Saying Half of 2026 US Datacenter Capacity is Canceled.
Source: SemiAnalysis Energy Model, sales@semianalysis.com
Some of the most strategic projects developed by leading AI labs and hyperscalers are relying on behind-the-meter, supporting hundreds of billions of future revenue. Adoption has never been more broad-based. A few examples:
In 2026 year-to-date, Microsoft has signed over 5GW of behind-the-meter nameplate capacity, of which 2.7GW with Joulent & Chevron, and well over 2GW through turnkey datacenter leases with companies like Crusoe. That 5GW encompasses a broad range of different types of power equipment; full breakdown available to our Energy Model subscribers.
Google, historically the most reluctant to onsite gas, is deploying 930MW of off-grid aeroderivative turbines in a flagship campus in Armstrong County. In addition, the Search Giant will deploy 900MW of Bloom Energy Fuel Cells in Wyoming - as we called out back in February 2026 as a huge positive for Bloom Energy. They’ll be paired with >1GW of Mitsubishi J-class turbines.
Both Anthropic and Meta have signed 300-500MW deals with Enchanted Rock, a supplier of 0.5MW gensets built around a 21.9-liter V12 gas engine. Separately, Anthropic’s flagship campus in Texas, backstopped by Google, will also deploy over 1.5GW of off-grid generation; full breakdown of Anthropic’s exact datacenter facilities available to our Datacenter Model subscribers.
OpenAI will imminently start operations in its flagship off-grid 1.4GW (IT capacity) campus in Shackelford County, TX, using over five hundred 4.25MW Jenbacher J624 engines. We show below a portion of the campus. Combined with their 1.3GW IT site in New Mexico, that represents over $150B of contracted spending that OpenAI signed with Oracle relying on behind-the-meter power.
***
Why is this happening, and why have many energy experts been so wrong? It comes down to understanding AI economics. We’ve discussed this at length in many other articles and in our Tokenomics Model. As a quick reminder, the value of megawatts for end-users is skyrocketing. A power plant supporting an islanded 1GW IT datacenter typically costs ~$5B. In today’s environment, inference API revenue can yield $100B per GW per year, at 90%+ gross margins. Paying 2x more money or accepting 30% lower efficiency for faster speed of deployment becomes a no-brainer. Said differently, Anthropic and its peers can pay back the value of a power plant in 20 days of inference revenue. As we discussed six months ago, most of the value in the AI infrastructure stack is shifting to frontier model developers....
....MUCH MORE
