The AI Industry Is Burning Fortunes On A Problem We Already Own The Solution To

The AI industry is spending more than half a trillion dollars to build something the world already has. The hardware exists. The power flows to it. The cooling is paid for. It sits in two billion homes doing nothing.

OpenAI announced a $500 billion buildout called Stargate on January 21, 2025. Ten gigawatts. Nine million American homes worth of electricity, dedicated to one company's compute. By September 2025 the company had announced five additional Stargate sites across Texas, New Mexico, Ohio, and Wisconsin, plus a Michigan project at 1 gigawatt and $7 billion. Oracle delivered the first NVIDIA GB200 racks to the Abilene, Texas flagship in June 2025. Six thousand four hundred construction workers were on site by year-end. Total announced Stargate capacity: roughly 7 gigawatts. Total committed capital: north of $400 billion.

Anthropic signed a tens-of-billions deal with Google Cloud on October 23, 2025 for up to one million seventh-generation TPUs, over one gigawatt online in 2026. The deal expanded again in April 2026 through Broadcom for an additional 3.5 gigawatts of next-generation TPU capacity. The Information reported a broader framework totaling 5 gigawatts of committed compute and approximately $200 billion over five years. Microsoft signed a 20-year power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 at a cost of $1.6 billion. Google signed a 25-year PPA with NextEra to restart the 615 megawatt Duane Arnold reactor in Iowa. The 8-K is on file at the SEC. Two retired nuclear plants are coming back online for two corporate customers.

This is the largest infrastructure project in the history of the technology industry. The capex follows the bet. The bet follows the belief that compute is destiny.

What The Industry Refuses To See

Every prior moment of infrastructure scarcity produced the same response. Centralized industry concluded the answer was to build more of the centralized thing. They commission new construction. They lock in supply contracts. They negotiate energy deals. They concentrate. And every time, there is another solution sitting in plain sight that nobody is using because nobody figured out how to coordinate it.

There are roughly 2 billion personal computers in the world. There are over a billion gaming consoles still in homes across multiple generations. There are more than 7.3 billion active smartphone subscriptions on the planet today. IDC reported 1.25 billion new smartphones shipped in 2025 alone. Gartner reported 270 million PCs shipped worldwide in 2025, a 9.1 percent jump from the prior year. The vast majority of these devices contain capable GPUs and NPUs. The vast majority sit idle most of the day.

Nobody is harvesting any of it at meaningful scale. The companies who could most benefit have no incentive to build a coordination layer for hardware they do not own. The users who could most benefit have no way to plug into the larger compute economy. The infrastructure exists as a handful of academic projects. It does not exist as the thing it should be, which is the obvious complement to the centralized buildout consuming hundreds of billions of dollars and stressing the power grid in five states.

The substitution is not perfect. NVIDIA's Vera Rubin NVL144 ships in the second half of 2026 with 144 Rubin GPU dies and 36 Vera CPUs, rated 3.6 exaflops of NVFP4 inference. The 2027 Rubin Ultra NVL576 will draw 600 kilowatts per rack. Consumer hardware does not match data center accelerators on raw frontier-training throughput. The hyperscalers are not paying $100 billion for these systems because they failed to consider alternatives. They are paying because for frontier model training, there is no real alternative yet.

Inference Is The Whole Game Now

Most of the compute the world actually consumes from AI is inference, not training. Inference workloads are smaller, more parallel, and far more tolerant of varied hardware. Apple's M5 chip announced October 15, 2025 delivers more than 4x peak GPU compute for AI versus the M4 through neural accelerators in every GPU core. Qualcomm's Snapdragon X2 Elite Extreme launched September 25, 2025 with an 80 TOPS NPU. Intel Panther Lake at CES 2026 hit 50 TOPS in the NPU alone, around 180 TOPS at the platform level. AMD Ryzen AI PRO 300 ships with 55 TOPS. Microsoft's Copilot+ PC threshold is 40 TOPS sustained. Every major silicon vendor now ships consumer chips that exceed the threshold for usable on-device inference. A quantized 7 billion parameter model runs on any of them. The same workloads that fill racks of GB200s can run, slower but functionally, on hardware already deployed in homes across every country on Earth.

The grid is the part of this that should make every reader stop and think. According to the Lawrence Berkeley National Laboratory, data center demand will grow from 176 terawatt-hours in 2023, which is 4.4 percent of total US electricity consumption, to between 325 and 580 terawatt-hours by 2028, or 6.7 to 12 percent. PJM Interconnection cleared three consecutive capacity auctions at or above the FERC-approved cap. Virginia data centers already consume more than a quarter of the state's electricity. ERCOT's large-load interconnection queue grew from 63 gigawatts in December 2024 to 226 gigawatts by November 2025. Power plants that were supposed to come offline are being kept online. Power plants that were already offline are being brought back. The strategy is colliding with the physical limits of the American electrical grid.

What The Grid Is Telling Us

The defenders of the centralized approach will say that this is what scale requires. That the alternative is hobbyist infrastructure that will never deliver real workloads. That serious AI requires serious capital. They are not wrong about the limits. They are wrong about which limits matter.

Centralized data centers should be built. The argument is not against them. The argument is that there is a parallel resource pool nobody is talking about, and that pool already exists at a scale the hyperscalers cannot match. The pool is already paid for. Already powered. Already cooled. Already distributed across every country on the planet. The hardware refresh cycle is funded by consumer purchases that happen every year without a single capex line on any corporate balance sheet.

The willingness gap is the real bottleneck. The technology to coordinate distributed compute has existed since BitTorrent. The protocols improve every year. Hardware in homes gets more capable every year. Bandwidth gets faster every year. What is missing is the coordination layer that lets the two billion machines do work for the global compute economy and get paid for it.

A coal plant in Ohio is being kept open to power a data center that runs models a consumer GPU could run, in parallel, for free, in a kitchen in the same town. That is the picture that should be hanging on the wall of every infrastructure investment committee in the country. The infrastructure being demanded is also the infrastructure being ignored.

The hardware already exists. Half a trillion dollars is being spent to build a second copy of something we have already paid for. The grid is buckling under the cost of the second copy. The first copy sits idle, two billion machines deep, waiting for somebody to figure out how to use what is already there.