The silicon
behind it all.
AI isn't a software story — it's power, cooling and silicon. This is the compute index: the accelerators training and serving frontier models, who makes them, and the physical infrastructure they run on. Real specs, primary sources, zero hype.
Who's ahead,
by the numbers.
The frontier accelerators, compared on the spec that actually bottlenecks frontier models. The leader is lit in green. Pick a metric.
A chip a year,
memory doubling.
HBM per GPU across NVIDIA's data-center line. The near-annual cadence — Hopper → Blackwell → Rubin — is why frontier labs buy whole racks.
AI accelerator index
The chips training and serving the frontier. Search, filter by maker, or sort any column. This table is the database on display.
| Accelerator | Maker | Process | Memory (GB) | BW (TB/s) | Year | What it's for |
|---|
Blackwell → Vera Rubin
A new AI platform roughly every year, each about doubling low-precision throughput and memory.
What compute
actually needs
Chips are only the visible tip — the real constraints are wafers, watts and heat.
Leading-edge nodes
Nearly every frontier accelerator is fabbed by TSMC on 3–5 nm-class nodes. Leading-edge wafer supply is the true bottleneck under the whole build-out.
Gigawatt campuses
The largest AI campuses now draw over a gigawatt — the scale of a mid-sized city — reshaping grid planning and driving new on-site generation.
Air → liquid
Blackwell-class racks are too dense for air. Direct-to-chip liquid and immersion are becoming standard, tracked by PUE, WUE and TUE ratios.
Rubin arrives
What's shipping at the leading edge of AI silicon right now. all stories →