01 Trigger Event
Amazon is investing in a new 7.65GW gas turbine power plant in Pecos County, West Texas, comprising 35 natural gas units. Initially dedicated to its new data center, the facility will not interconnect with the Texas grid (ERCOT). According to Cleanview, which tracks data center electricity, the project (GW Ranch) has received Texas emissions permits, and the emissions from this single plant may rank among the dirtiest gas-fired facilities in the country.
02 What This Really Means
This is not an environmental news story—this is the first hard data point showing the AI compute expansion curve hitting the physical ceiling of the US power grid.
What 7.65GW actually means: roughly equivalent to the installed capacity of 6-7 large nuclear reactors, exceeding the peak load of most American cities. The reason Amazon chose to build behind the meter (BTM, on-site self-supply) rather than connecting to ERCOT lies in two details I'm not entirely certain about but worth investigating:
First, ERCOT interconnection queue. I haven't run the actual numbers internally on Texas interconnection applications, but there's an industry rumor that the current ERCOT large-load interconnection queue is already extending past 2030. When your load is at the GW scale, the interconnection approval cycle may take longer than the data center's civil construction.
Second, 7.65GW represents the load of a single campus. This isn't a phased ramp-up order—it's a one-time drop of an enormous block. Any grid operator will tell you that this kind of step load change either has to be absorbed by BTM or triggers grid stability upgrades—the time and money costs of which hyperscalers are unwilling to pay.
What AWS, Microsoft, and Google are collectively running toward with behind-the-meter solutions represents a fundamental shift: electricity is becoming a co-located input, no longer a commodity purchased from utilities. This is a Stratechery-style mutation of aggregation theory: when demand reaches critical mass, you vertically integrate upstream. Amazon isn't entering the gas turbine business because they want to—the grid cannot scale to match their demand cadence, so they must grow their own generation capacity.
03 Historical Analogy / Structural Comparison
What comes to mind is Apple designing its own chips in the 2010s. The real catalyst for the M-series wasn't "we want to build a better CPU"—it was that Intel's iteration rhythm couldn't fit into the iPhone's power budget and launch timeline.
What Amazon is doing in Pecos is the same structural pattern: when external supply (interconnection capacity, wholesale price stability, grid service quality) cannot evolve fast enough to match your product cadence, you build it yourself.
Deeper layer: this is the same script's v2 as the hyperscaler self-built data center wave around 2008. The first wave was capex on real estate; this wave is capex on electrons. Pushing further, we will likely see capex on transmission lines, and even capex on cooling water rights (Texas has already started).
I also think of a counter-analogy: in the 1980s, energy-intensive giants like Alcoa built their own hydroelectric facilities in the Pacific Northwest. The difference this time is that AI's load growth slope is far steeper than any historical precedent—nuclear power takes 10 years from groundbreaking to grid connection, while a gas peaker like GW Ranch takes only 24-36 months. So gas becomes the transitional answer, but it's an answer carrying a transition tax.
04 What This Means for AI Builders
What I tell opcx.ai's clients: over the next 12 months, watch at least three things closely.
Inference pricing's "power passthrough" will become explicit. In AWS, Azure, and GCP's GPU instance pricing for 2026-2027, an increasing share of marginal cost will come not from H100/H200 depreciation, but from PPA (power purchase agreement) fixed cost pass-through. When gas peaker capital costs are amortized into pricing, you will see a new "regional compute premium"—cheaper in the West and South, more expensive in PJM/NEISO regions.
Colocation site selection becomes an infrastructure decision, not a financial decision. If you're building training clusters or locking in long-term inference capacity, start looking at ERCOT / MISO / PJM interconnection queue depths instead of focusing solely on GPU prices. I may be overestimating the practical impact on smaller builders—most application-layer companies won't self-build GW-scale loads—but as a secondary signal for region selection, it is starting to matter.
Contractualization of carbon costs will become a hidden variable in enterprise AI procurement. EU CSRD and SEC climate disclosure rules are pushing procurement teams to ask: how many grams of CO2e per million inference tokens? This supply chain pressure will force API providers to sign PPAs, purchase RECs, and amortize those costs into token pricing. For pure application layers this is noise; for the infrastructure layer, it's a new source of pricing power moat.
More directly: this week's action item is to review your inference workload geographic distribution and check whether you're over-concentrated in the PJM (US East) region—where the electricity cost upward slope over the next 24 months may run 1.5-2x steeper than in western gas-rich regions. I haven't seen precise public data on this; this is inference based on historical IRP filings, and readers should judge for themselves.
05 Counterarguments / Risks
I may be wrong in three places.
First, GW Ranch may be an exception rather than a trend. This single Amazon project may reflect Pecos County's specific grid bottleneck plus Texas's regulatory permissiveness combination, rather than a nationwide systemic inflection point. If interconnection queues in Mississippi or Virginia aren't actually that bad, the entire behind-the-meter narrative shrinks to a local story. Amazon declined to comment on specific site selection logic, and I haven't obtained their internal interconnection application timelines—I acknowledge this judgment is built on reasoning from publicly available information.
Second, I may be overestimating gas peakers' transitional nature. If grid upgrades remain slow for another decade (entirely possible), gas becomes the "new normal" rather than a transitional solution, and the entire energy transition narrative gets delayed by AI demand. This would in turn make the claim—repeated by many climate-tech VCs—that "AI is a clean tech enabler" turn into a bitter irony.
Third, and most critically: I may be underestimating demand-side elasticity. If model efficiency improvements (MoE architecture, MLA, SSM, speculative decoding, prompt caching) really cut inference compute per task by 50%+ annually, then the GW-scale new load demand assumption collapses. Whether this assumption holds, I'm less certain than I was six months ago. The debate between the scaling law camp and the post-training camp is actually answering this question—how many watts do we actually need to run AI?
The sharpest counterargument: AI's power hunger is hype self-reinforcing, not structurally inevitable. If this is the case, today's gas peaker investments become stranded assets. But I have to be honest here—I can't figure out how to judge this. I can only place stakes on both sides and wait for the next data point (most likely Q2 2026 hyperscaler capex guidance) to falsify one side.