Amazon’s 7.65 GW Cloud Gets a Texas Address

Amazon is investing in 7.65 gigawatts of gas generation for an off-grid Texas AI campus, a proposed plant that The New York Times reported on August 8 could become the largest single U.S. emissions source. By contrast, in 2015 AWS backed a $400 million, 208-megawatt North Carolina wind farm. In California, its $2 billion Gilroy facility spent years in local negotiation after AWS applied to build it in 2020. AWS still sells a cloud without an address while Amazon builds it around power plants and permits tied to specific places.

The scale is clearer than the deal. The reporting cited here identifies Amazon as an investor but does not name the plant’s developer or operator, locate the site within Texas, provide the investment terms or construction date, or explain the electrical arrangement behind “off-grid.” The 7.65 gigawatts describe proposed generation attached to an AI plan, not operating AWS capacity.

AWS now competes by coordinating the industrial chain beneath the API. Each model response creates a recurring cost under inference economics, making electricity, chip supply, land, interconnection rights, permits and financing production inputs. AWS can keep capacity feeling elastic to customers only by making fixed, local commitments years earlier.

Elasticity stops at the substation

AWS and its rivals built the public cloud to separate customers from machines. They pooled hardware, rented virtual capacity and moved software between servers without forcing each buyer to negotiate for a building, cooling system or transmission line. They could schedule work inside installed capacity; they could not summon a substation on demand.

By 2023, cloud providers were already confronting a mismatch: traditional cloud infrastructure had not been designed for large-scale AI. OpenAI and Anthropic later told investors that inference costs exceed half of revenue, placing the recurring expense of serving models alongside the cost of developing them.

AWS can assign an available accelerator far faster than a utility can approve, build and connect generation. In January 2026, PJM Interconnection proposed requiring large data centers to supply their own generation or curtail demand to reduce outage risk. But Bloomberg reported on January 14 that PJM had cut its summer 2027 peak-demand forecast from about 164 gigawatts to 160 gigawatts because some proposed projects, including data centers, lacked firm service or construction commitments. PJM’s two actions draw the useful line: proposed loads had inflated the forecast, while sufficiently firm large loads still prompted a reliability plan.

At AWS, silicon divides the queue by urgency

AWS has spent years developing Trainium, giving it an alternative to the merchant accelerator market. A technical examination of Trainium2 found that the chip had a credible path toward competing in large-language-model inference. Trainium gives AWS more control over cost, supply and system design, while Amazon also buys outside hardware.

AWS’s 2026 agreement with Cerebras makes that approach visible. AWS plans to offer Cerebras hardware for fast inference while retaining Trainium for slower, cheaper computing, routing workloads across different cost and performance envelopes.

AWS coordinates accelerators, networking, storage, cooling and software to produce each completed response. A more expensive accelerator can lower total cost when it finishes latency-sensitive work quickly; a slower proprietary chip can win when price matters more than response time. Customers can ignore those choices because AWS makes the hidden stack heterogeneous.

OpenAI engineers reportedly found an approach that could more than halve inference costs. The report does not say AWS adopted it. It establishes the other side of the allocation problem: software efficiency can change which accelerator is cheapest for a given response.

Gilroy put a city on AWS’s critical path

Amazon and Gilroy, California, negotiated a $2 billion data-center project for years after AWS applied to build it in 2020. The Wall Street Journal reported on August 8 that the project proceeded without public meetings or votes, leaving most residents unaware of it until work began.

A processor ordered for Gilroy still needs an approved site, cooling, power and a completed building before AWS can expose it through a service endpoint. The interval between the 2020 application and construction kept that prospective capacity outside the cloud.

The cited report does not identify which approval consumed most of that interval or quantify a permit-driven delay. Its sharper finding is procedural: work began on a $2 billion project before most residents knew about it.

Lenders decide which campuses become capacity

In November 2025, Barclays tallied announced hyperscaler AI data-center capacity.

At full utilization, the Financial Times reported, that capacity would consume as much energy as roughly 44.2 million U.S. households. The tally measures announcements; each campus still needs financing, power and construction before it can operate.

Barclays also described how Meta moved $30 billion of debt for AI data-center construction off its balance sheet through special-purpose vehicles. Those vehicles separated project financing from the corporate entity while preserving Meta’s access to the resulting infrastructure.

Before a campus runs, lenders decide which construction, power and technology risks they will fund. Debt terms assign construction risk, power contracts determine available load, and chip agreements determine which workloads a campus can serve economically. Infrastructure investors apply the same logic to long-duration, contracted megawatts, underwriting committed access to powered capacity rather than an undifferentiated promise of future demand.

Meta’s $30 billion structure and Amazon’s 7.65-gigawatt plan use different forms of control, but both turn demand forecasts into long-lived obligations before customers consume the first token.

Private generation leaves climate and grid risk behind

Whoever operates the Texas plant would have to secure fuel, maintain reliability and manage emissions. The New York Times reported on August 8 that the proposed plant would be at odds with Amazon’s confirmed commitment to reach net-zero emissions by 2040 and could become the largest single U.S. emissions source. The cited evidence gives no projected annual emissions total, so it supports a potential ranking without quantifying the plant’s footprint.

An off-grid campus can still affect the grid. Data-center operators sometimes switch from utility electricity to local generators without advance notice, and grid operators have warned that such unannounced disconnections can contribute to cascading outages. A large campus changes system balance even when it supplies part of its own load.

That warning concerns data centers generally; it does not establish that Amazon’s Texas design would disconnect in this way. Without connection details, the project’s effect on the grid remains unsettled.

OpenAI can change the cost of a response through engineering. Amazon cannot shrink a gas plant or unwind a municipal agreement at the same speed. AWS protected customers from stranded hardware by pooling it; an integrated AI campus leaves Amazon holding a larger, slower-moving version of that risk.

The token still reaches the screen without a location field; Amazon’s 7.65-gigawatt Texas plan gives the fuel, emissions, financing and permitting risk behind it a fixed address.