The AI industry is quietly building an exit from its own data centers

In the paperwork that decides who keeps their lights on, the data center has become the load the grid sacrifices first. PJM Interconnection, which runs the grid serving 65 million people, has proposed a rule: new data centers drawing 50 megawatts or more that don't bring their own generation would be the first cut when power runs short, and compensated at half the rate for cutting their demand ^[1].

The unprecedented addition of large loads, most notably data centers, has given rise to resource adequacy shortfalls — and associated real-time operational issues that Interim Resource Adequacy Service is intended to address — PJM Interconnection

A decade of economic machinery rearranged itself around this asset — and now, in grid filings, it's the expendable one. The wall behind that paperwork is physical. Of 3,969 data centers announced across the United States, only 802 are under construction ^[2]. By Goldman Sachs' reading, developers no longer site where the customers are; they site where the electricity is, and grid bottlenecks have already pushed them out of Virginia toward Texas, Georgia and California ^[3]. The constraint has climbed high enough that even the International Monetary Fund now names an energy supply shock as the drag that AI investment must offset to keep global growth at 3 percent ^[4]. But the wall that moves buyers is not the megawatt. It's the meter. Between June and August the price of a token — the slivers of text AI bills are counted in — fell roughly 40 percent, and enterprise AI spending per employee at the biggest firms rose nearly 50 percent in a single month, by Ramp's transaction data ^[5]. The old economist's paradox — Jevons by name — in one breath: cheaper units, more use, a bigger bill. Billing moved to consumption ^[6], and under consumption billing the more a company's agents get used, the faster the cost piles up ^[7]. Buyers have answered by routing routine work to cheap open-weight models and saving the expensive closed ones for hard reasoning ^[6]. Bank of America's read on all this is one sentence.

Basic inference is increasingly commoditized, while frontier capabilities remain differentiated. — Bank of America

And everyone is building the exits, in the open. HP's pitch to move inference — the everyday work of asking a trained model a question — onto local PCs is priced, not just technical ^[8].

the future of inferencing will be “distributed or hybrid.” — Rishi Patel

The company's own arithmetic puts the shift at $650,000 to $2 million saved per 1,000 employees ^[8]. Nvidia, the biggest seller of the expensive hardware, is building a cheaper inference chip specifically "to reduce the high cost of hardware for corporate clients" ^[9]. Google — an operator of some of the world's largest AI data centers — shipped a free app in 2025 that runs generative models entirely on a phone ^[10]. Models under a billion parameters account for 83 percent of all-time downloads on Hugging Face; the work developers actually deploy is exactly the kind small enough to run on a device, not in a rack ^[11]. And the labs have moved too: OpenAI bought tens of thousands of Mac minis and Mac Studios to train its computer-operating agents, because the interfaces an agent must learn sit on a desktop, not in a rack ^[12]. The marketing register never says the word. Perplexity's chief frames the move onto local chips as the mainframe-to-PC transition, argued in the language of privacy, latency and battery ^[13].

The biggest threat to a data center is if the intelligence can be packed locally on a chip that’s running on the device and then there’s no need to inference all of it on like one centralized data center. — Aravind Srinivas

Microchip's edge chief sells the same shift on latency, privacy and lower cloud dependency ^[14]. No brochure says cheaper. The pricing sheets say it for them. The cloud, meanwhile, cannot turn back. Hyperscalers sit on $1 trillion in lease commitments that haven't even begun construction, up from $725 billion in under a month ^[15]. Demand still outruns supply — a 12-gigawatt deficit last year, around $2 trillion in backlogs ^[16]. AWS grew 28 percent and threw off 59 percent of Amazon's profit ^[17]. The industry's grand bargain with the grid is to accept the throttle: Google, Nvidia, Anthropic and the utilities will make data centers "flexible grid assets" that cut their own power during shortages in exchange for faster hookups ^[18]. The same Bank of America research that calls basic inference increasingly commoditized names the large cloud platforms as the profit winners of that very commoditization ^[19]. What the edge migration puts at risk is the math underneath the whole buildout. Lenders demand the chips recover their cost on 3-to-4-year schedules ^[20] — and the shift of inference off the cloud is already reported as a direct threat to those depreciation schedules ^[21]. The honesty beat, before the close: the escape hatch is repricing itself. Nvidia just raised the price of its Shield TV Pro edge device from $199.99 to $299.99 — a 50 percent hike it attributes to the DRAM and memory costs caused by the data center rush ^[22]. The buildout everyone is hedging against is also what makes the hedge more expensive. So the picture is not a stampede. It's a hedge built in plain sight: the supply side holds the cloud contract and cannot turn around, while everyone else — the device makers, the open-weight ecosystem, even the grid that rations the cloud first — is holding a priced route out at the same time. No one is through the door yet. The question that will answer itself is which gives first: the exits getting used, or the short depreciation clocks running out.