ROOT / CPU // DECODED ACCELERATORS: 12 · STATUS: LIVE
CPU // Decoded

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.

$400B
2025 AI capex
90%
Accelerator share: NVIDIA
3-5 nm
Leading process node
1.2 GW
Largest campuses
// Compute telemetry

The physical layer, live

25 data graphs cycling through the silicon, power and economics behind the AI build-out. Auto-advances; hover to pause, click any chart for the detail, tap a dot to jump.

// The silicon

AI accelerator index

The chips training and serving the frontier — filter by maker.

AcceleratorMakerProcessMemoryPower (TDP)What it's for
// The physical layer

What compute actually needs

Map it: chip fabs → · data centers → · supercomputers →

Silicon

Leading-edge nodes

Nearly every frontier accelerator is fabbed by TSMC on 3-5 nm class nodes. The supply of leading-edge wafers is the true bottleneck under the whole AI build-out.

Power

Gigawatt campuses

The largest AI data-center campuses now draw over a gigawatt — the scale of a mid-sized city — reshaping grid planning and driving new on-site generation.

Cooling

From air to liquid

Blackwell-class racks are too dense for air. Direct-to-chip liquid cooling and immersion are becoming standard, tracked by PUE, WUE and TUE efficiency ratios.

// Frontier dispatch · newest in compute

Rubin arrives

What's actually shipping at the leading edge of AI silicon right now.

NEXT-GEN COMPUTE

NVIDIA's next platform is a four-tile module with 288 GB of memory.

The compute frontier's next leap is NVIDIA's Rubin generation, launching in the second half of 2026. Built by TSMC on a 3-nanometre process with next-generation HBM4 memory, the flagship R200 is a true multi-chip module — two compute dies plus two dedicated I/O dies in one package — carrying 288 GB of HBM4 at roughly 22 TB/s of bandwidth.

Rubin is rated at about 50 petaflops of FP4 inference — some 2.5× its Blackwell predecessor — with a "Rubin Ultra" refresh set to double that again and move to TSMC's 2 nm (N2) node around 2027. Early supply goes to hyperscalers and frontier AI labs; broad cloud availability isn't expected until late 2026.

> Sources: NVIDIA GTC 2026 · TrendForce · HPCwire. Specs are the manufacturer's stated figures.