The real economics of AI compute: power, wafers, and the illusion of infinite scale

Over the last 18 months, the semiconductor industry has been living through a generative-AI gold rush. Capital expenditure commitments, GPU shipments, and data-center power projections are growing at rates the market hasn’t seen since the dot-com or crypto-mining booms.

The latest Nvidia GTC keynote made the scale explicit, and it’s time to separate what’s physically possible from what’s speculative.

Nvidia’s GTC Washington D.C. 2025 keynote confirmed what many suspected: the company’s GPU engine is running at industrial scale.

At the same time, Nvidia teased a total of 20 million “GPU dies” for 2025-2026, but this number is misleading. It refers to dies, not packages. Since each data-center GPU package integrates multiple dies, the physical shipment count in packages is roughly half that figure. In other words, the 20 million figure makes the ramp-up sound even larger than it is.

When we translate Nvidia’s statements into system-level terms, the scale becomes easier to grasp.

For context, OpenAI is for now targeting about 26 GW of total AI compute over the next four years, roughly 6.5 GW per year, split between 10 GW of Nvidia GPUs, 10 GW of Broadcom-custom AI ASICs, and 6 GW of AMD GPUs.  Some of these announcements likely overlap with the Stargate project: the 10 GW, $500 billion AI infrastructure program with Oracle and Softbank. According to an official letter from OpenAI’s Chief Global Affairs Officer, roughly 7 GW of that capacity is already planned for deployment over the next three years. But of course, it doesn’t stop there… OpenAI has also committed to renting $250 billion worth of compute from Microsoft Azure and $38 billion from AWS, and has signed multi-billion-dollar deals with CoreWeave.

And that’s just OpenAI. The rest of the hyperscaler, neocloud, and model-training ecosystem, including xAI, Anthropic, Mistral, CoreWeave, Google, AWS, Microsoft, Meta, Oracle, and others, is stacked on top of this demand.

OpenAI suggests that the US should double its annual power deployment from roughly 50 GW today to around 100 GW per year to sustain the next wave of AI infrastructure growth.

The numbers are unprecedented, and the physics are starting to bite.

The semiconductor industry loves volume ramps, but can Logic wafers, CoWoS packaging, HBM memory, and Power really scale as quickly as the latest announcements suggest?