AINEWS
A projected tenfold jump in AI chips by 2028 makes the build out a foundations question, because power, water, and capital, not chip supply, will decide who builds and who falls behind.
About 20 million A.I. chips now sit inside the world's data centers. By the end of 2028, the research firm Epoch AI projects that count will reach roughly 200 million, a tenfold expansion that doubles the global installed base every nine months.
Each new gigawatt of data-center demand now meets a queue of siting, water, and grid-connection limits before it meets a chip order. The substrate decision being made in real time is who can run those constraints at hyperscaler scale, and which regions capture the leverage when they do.
Etched, a chip firm that designs processors specifically for large language models, has raised more than $1 billion in the past year. Its co-founder Rob Wachen has called the build-out the largest infrastructure expansion in human history. Industry analysts at Futurum Group size 2026 AI infrastructure capex at roughly $690 billion, a number that captures the data-center shells, power agreements, and grid interconnects required to run the next chip generation, not the chips themselves.
Amazon has doubled its computing capacity since 2022 and expects to double it again by 2027, according to Peter DeSantis, who leads foundational AI models at AWS. Amazon provides compute to Anthropic, OpenAI, and other model labs. The capacity curve is the operational face of the chip-count model: more compute, deployed on the same cadence, is the bet that the next model will be more capable than the last.
Industry voices call this bet "Scaling Laws": the claim that feeding more compute to an AI system yields a more capable system, with capability gains that hold across orders of magnitude. Held as a thesis by named engineers and executives, it is a reasonable bet. The risk is treating it as a law of nature, which papers over the binding limits actually shaping the build-out.
Department of Energy has begun coordinating federal resources to meet data-center electricity demand, a step that reflects how quickly the grid has become the bottleneck. A 2026 outlook from the law firm Morgan Lewis flags siting, water, and interconnection queue length as the constraints that will determine which projects clear and which stall. Brookings has tracked the same pattern, noting that the regions with available power and willing utilities will set the pace, because the substrate is power, water, and grid capacity, not labor.
The historical comparisons the source industry reaches for are 1800s railroads, FDR's New Deal, and the Manhattan Project. They work as one sentence of scale. The substrate is closer to a controlled semiconductor fab build-out than to a turn-of-the-century rail expansion: a small number of buyers, a small number of builders, and a small number of regions with the power capacity to host the next increment. Each doubling pulls more of the global compute base into that narrow set.
An AI system passed the bar exam in 2023, AI helped identify a suspected cause of Alzheimer's in 2025, an AI system solved a math problem open for roughly 80 years in May 2026, and two AI systems under testing recently hacked a company database. Each of these is a Scaling-Laws receipt from the capability side. The open question is whether the next tenfold of compute produces the same slope, or whether the gains taper as the substrate tightens.
What to watch over the next 24 months:
Twenty million chips today, 200 million by 2028, and a build-out whose binding constraints are measured in megawatts, acre-feet, and balance-sheet capacity. The next two years will show who can actually deliver the doubling, and who is betting on a cadence the substrate cannot hold.