The scale of the Pecos buildout
Amazon has acquired GW Ranch, an 8,000-plus-acre site in Pecos County in West Texas, to build a data center campus fed by a purpose-built natural gas plant capable of generating 7.65 gigawatts across 35 turbines. Construction permits for three data center buildings were filed in early August, and satellite imagery showed active land clearing by late July, indicating Amazon is moving fast on a site previously being developed by Pacifico Energy. The Texas Commission on Environmental Quality has already issued the air permit needed to build.
The numbers here are worth sitting with. 7.65 gigawatts is enough generation capacity to power millions of homes, concentrated on a single site to run AI infrastructure. Amazon is pairing that gas capacity with 1.8 gigawatts of battery storage and 750 megawatts of solar, and the whole campus is designed to operate as what reporting describes as a self-contained energy island, generating and consuming power independent of the regional ERCOT grid, at least until interconnection timelines allow a transition to grid-connected service.
Behind-the-meter power is now core cloud strategy
Behind-the-meter generation, building dedicated power plants next to the data centers they serve rather than waiting years for utility grid connections, has moved from a workaround to a standard hyperscaler strategy in 2026. Industry tracking cited alongside this deal counts at least 59 such projects announced since early 2025, totaling roughly 90 gigawatts of capacity. Microsoft is building its own nearby project in Pecos, a 2-gigawatt data center backed by a 2.67-gigawatt gas plant announced in June, and Amazon is separately exploring another gas-powered facility near Pittsburgh.
We read this as confirmation that the AI capacity race has shifted its center of gravity from chip availability to power availability. Grid interconnection queues in most U.S. regions now run three to five years, far too slow for hyperscalers racing to bring gigawatt-scale AI training and inference capacity online in 2026 and 2027. Building your own power plant is the only way to hit that timeline, and Amazon, Microsoft, Meta, and OpenAI-aligned developers are all making the same bet simultaneously, which is starting to strain turbine supply chains too.
The emissions math CIOs cannot ignore
The Pecos plant is permitted to emit up to 33 million tons of carbon dioxide annually, a figure reporting describes as potentially making it the most carbon-intensive power facility ever permitted in the United States. That number lands squarely on any enterprise that has published net-zero or Scope 3 emissions targets and relies on AWS for compute, because cloud provider power sourcing increasingly shows up in corporate carbon accounting whether or not the enterprise chose that outcome.
This is a live reputational risk today, not a hypothetical one. Sustainability teams at large enterprises are already fielding board-level questions about AI's energy footprint, and a single hyperscaler decision to run gigawatt-scale gas generation reshapes the emissions intensity of every workload hosted on that infrastructure. CIOs should push AWS account teams now for region-specific and workload-specific emissions disclosure rather than relying on company-wide renewable energy percentages that can mask exactly this kind of concentrated fossil buildout. Enterprises that report Scope 3 emissions tied to cloud consumption will need a far more granular data feed from AWS than most contracts currently require, and getting that commitment in writing now is easier than retrofitting it after the Pecos plant is already running.
What this means for AWS capacity planning
Amazon's own spokesperson framed the investment plainly: Amazon believes in paying the full costs of powering its operations, and the Pecos campus does that by building dedicated generation rather than drawing down a constrained public grid. That is a defensible engineering answer to a real bottleneck, but it also means AWS capacity growth is now gated by gas turbine delivery schedules, permitting timelines, and interconnection engineering, not just server and chip procurement.
For enterprise buyers, the practical implication is that AWS capacity commitments tied to new or expanding regions should come with explicit contingency language around power-driven delays. First power at Pecos is targeted for the first quarter of 2027, an aggressive timeline for a facility of this scale, and any slippage there cascades directly into AI training and inference capacity availability. Enterprises planning large GPU reservations or dedicated capacity blocks with AWS should ask which power source underpins that specific region and what the fallback plan is if the generation asset is delayed.
The competitive and cost implications ahead
Behind-the-meter power gives hyperscalers a way to route around grid constraints, but it is not cheap, and someone eventually pays for 35 gas turbines and a battery farm. We expect these capital costs to show up gradually in AWS pricing for GPU-heavy instance types and reserved capacity, even if Amazon frames the investment as protecting rather than raising customer prices. Enterprises with heavy AI training workloads should model that pricing risk into multi-year capacity plans now rather than treating current on-demand and reserved pricing as stable.
There is also a strategic angle worth watching: hyperscalers that control their own power generation gain a durable capacity advantage over cloud providers still dependent on public grid interconnection. That advantage compounds over the next several years as AI demand keeps outpacing grid buildout nationally. CIOs evaluating multi-cloud AI strategies should treat a provider's power self-sufficiency as a genuine differentiator in vendor selection alongside its chip inventory, because it increasingly predicts who can actually deliver committed capacity on schedule. Smaller cloud providers without the balance sheet to fund gigawatt-scale gas plants will find it progressively harder to compete for the largest AI workloads, which concentrates enterprise leverage in negotiations with an ever-smaller set of vendors that can guarantee power at scale.
