Editing Openai/6925f847-0600-800d-8a2e-57670d2a3feb (section)

==== 2. 10 GW build-out (accelerator power only) ====

Now apply your power assumptions to a 10 GW accelerator budget (ignoring CPUs, networking, cooling overhead, etc.).
* Total power budget: 10,000 MW = 10,000,000,000 W

Number of accelerators you can power:
* Ironwood: 10,000,000,000 W / 1,000 W ≈ 10.0 M chips
* MI450X: 10,000,000,000 W / 1,400 W ≈ 7.14 M GPUs
* Rubin: 10,000,000,000 W / 2,300 W ≈ 4.35 M GPUs

(Values rounded to three sig figs below.)

===== 10 GW cluster comparison =====

Units:
* Total HBM = EB (exabytes, 1 EB = 10⁹ GB)
* Total BW = PB/s
* Compute = EFLOP/s (1 EF = 10³ PF)

| Metric @ 10 GW accel power | TPU v7 Ironwood | AMD MI450X | NVIDIA Rubin |
| -------------------------- | --------------- | ---------- | ------------ |
| Accelerators (count)       | 1.00×10⁷        | 7.14×10⁶   | 4.35×10⁶     |
| Total HBM (EB)             | 1.92 EB         | 3.09 EB    | 1.25 EB      |
| Total HBM BW (PB/s)        | 7.4×10⁴         | 1.4×10⁵    | 8.7×10⁴      |
| Total FP8 (EF)             | 4.6×10⁴         | 1.39×10⁵   | 7.25×10⁴     |
| Total FP4 (EF)             | 9.2×10⁴*        | 2.88×10⁵   | 2.17×10⁵     |

*Again, Ironwood FP4 is an inferred 2× FP8, so its FP4 figures are less solid than MI450X/Rubin, which are tied to specific system or chip disclosures. The Next Platform<ref>{{cite web|title=The Next Platform|url=https://www.jonpeddie.com/news/nvidia-prepares-vera-rubin-for-next-ai-wave/|publisher=The Next Platform|access-date=2025-11-29}}</ref>