Analysis
Data center developers are increasingly turning to purpose-built natural-gas power plants to solve a problem the regulated grid can't move fast enough to fix: getting enough electricity to a new AI data center on a timeline measured in months, not the years a standard utility interconnection typically takes. BloombergNEF tracked 99 such projects planned across 22 states, which could emit roughly 318 million metric tons of CO2 annually if run at industry-standard utilization, Bloomberg reported. The US electric power industry as a whole emitted about 1,485 million metric tons of carbon last year, per Energy Information Administration data cited in the same reporting -- meaning this one category of infrastructure could lift total US power-sector emissions by 20%, and as much as a third if the plants run at full capacity.
Why "behind-the-meter" is the mechanism that matters
The structural detail that makes this possible is that these are privately operated, behind-the-meter facilities -- gas plants built to serve a single data center directly, permitted and constructed without requiring approval from a utility or an independent system operator. That's the entire point, from a data-center developer's perspective: standard grid interconnection queues in many parts of the US now run years long because of the sheer volume of new generation and demand requests flooding regional grid operators. A behind-the-meter gas plant sidesteps that queue entirely, trading a longer-term climate cost for a much faster path to power a facility that needs to come online on an investor-driven timeline.
More than a third of the tracked projects are in Texas, which has both an independent, lightly-regulated grid (ERCOT) and permissive permitting for this kind of self-supplied generation -- the same regulatory environment that made Texas the location of choice for Tesla's Cybercab manufacturing and multiple other fast-moving infrastructure projects this year.
The companies whose climate commitments this complicates
Amazon, Microsoft, OpenAI and Anthropic are all named among the backers of these projects, per Bloomberg's reporting. Each of these companies has made public net-zero or carbon-reduction commitments on a multi-year timeline; each is also, per this reporting, financing or hosting operations at gas-fired plants built specifically to bypass the slower, cleaner path of grid-supplied and renewable power. That's not necessarily hypocrisy -- a company facing a genuine choice between missing a compute deployment timeline or temporarily burning more gas to hit it is making a real tradeoff, not a cynical one -- but it is a direct, quantifiable tension between AI infrastructure speed and climate commitments that these companies will have to reconcile publicly as the emissions math becomes harder to avoid.
The 20%-to-33% range itself carries real uncertainty: it assumes all 99 tracked projects get built and run at either industry-standard or full utilization, and pipeline power projects routinely get delayed, cancelled, or run below assumed capacity. But even the low end of that range, applied against a base of 1,485 million tons, represents one of the largest single new sources of US power-sector emissions growth identified this year -- and it's a direct, traceable consequence of the AI infrastructure buildout, not an indirect or contested one.
