Epoch AI released its AI Chip Sales database on January 8th, and the headline number is almost too clean: 15 million H100-equivalents of global computing capacity. That's the combined output of every major chip designer, normalized to the performance of Nvidia's H100, the workhorse of the last two years of AI training.
But the more interesting story is buried in the breakdown. The H100, which felt ubiquitous approximately five minutes ago, now accounts for less than 10% of Nvidia's AI revenue. The B300, announced at GTC 2025 and shipping in the second half of the year, already generates the majority.
That's a faster transition than most people expected.
The generational churn
Hardware generations in AI move fast, but this is something else. The H100 was the chip that trained GPT-4, the chip that every startup spent 2023 and 2024 desperately trying to get allocations for. CoreWeave built a business around it. Now it's becoming legacy inventory.
The Epoch AI data covers Nvidia, Google (TPUs), Amazon (Trainium), AMD (Instinct), and Huawei (Ascend). Together these account for what the researchers estimate is the large majority of global AI compute capacity. They don't track Microsoft's Maia or Meta's MTIA, noting these aren't yet deployed at significant scale for training.
One thing I couldn't find in their documentation: any accounting for chip retirements. They acknowledge this is an imperfect measure, noting that "AI hardware generally lasts for several years" and that the last two years of shipments would dominate total stock anyway even with infinite lifetimes. Make of that what you will.
The power situation
Here's where it gets uncomfortable. Epoch AI estimates this hardware collectively draws over 10 gigawatts of power. For context, New York City's peak summer demand tops out around 10 to 11 gigawatts. The city's average load sits closer to 5 or 6 gigawatts.
So we're now running, globally, something like twice New York City's worth of electricity just on AI accelerators. And that's probably conservative, because the researchers note that data center facility power typically runs about 1.3x higher than the chips alone, once you factor in cooling and other overhead.
The doubling time on this capacity is around 7 months, according to their trend analysis. Computing capacity has been growing at roughly 3.3x per year since 2022.
What they're actually measuring
Epoch AI is careful about what they claim. The database tracks chips "delivered and ready for installation," not necessarily online in a data center. For Nvidia and AMD, their estimates come from revenue figures, which recognize sales upon delivery. For Amazon Trainium, they lean more heavily on their separate data center tracking project.
The "H100-equivalent" metric itself is a simplification. They're comparing peak 8-bit operations per second. Real-world performance depends on memory bandwidth, software optimization, networking, and a dozen other factors their FAQ admits they're ignoring. But it's a reasonable approximation for getting a sense of scale, and citing "15 million H100-equivalents" is certainly more intuitive than whatever the raw operations-per-second figure would be.
Confidence intervals on most estimates span roughly 2x in either direction. They're most confident about Nvidia numbers (which benefit from quarterly earnings disclosures and extensive analyst coverage) and least confident about Amazon Trainium (limited public information, though their data center tracking provides a floor).
Why this matters
The dataset is free and open, which is the whole point. AI infrastructure has been weirdly opaque given how much money is flowing through it. Chip vendors don't publish unit sales. Cloud providers don't disclose how many GPUs they're actually running. Training run compute requirements are often estimated after the fact by outside researchers.
Epoch AI has been filling this gap for years, most notably with their training compute estimates for frontier models. This chip sales database is the supply-side counterpart.
The timing is interesting. Blackwell chips are sold out through mid-2026 according to some reports. Meta is projecting over $100 billion in 2026 capex. The infrastructure buildout continues even as questions about when it pays off get louder. Having better data on the denominator, how much compute actually exists, seems useful.
Next update from Epoch AI is expected in the coming weeks as Q4 2025 earnings roll in.
