MIT Technology Review · AIApr 8, 2026, 22:00 UTC+8

Mustafa Suleyman: AI development won’t hit a wall anytime soon—here’s why

Mustafa Suleyman argues frontier AI training compute rose from about 10^14 to over 10^26 FLOPs since 2010, a 1 trillion-fold increase, so AI development is not near a wall. He cites a 7x Nvidia chip gain in six years, 3x more HBM3 bandwidth, and Epoch AI estimates that compute needed for fixed performance halves every eight months. The piece is commentary from Microsoft AI’s CEO, not an independent study; the post does not disclose a reproducible basis for the 200GW-by-2030 claim.

Why it matters: HKR-H/K/R all pass: Suleyman takes a hard line in the scaling-wall debate and cites 10^26 flops, 7x chip gains, 3x bandwidth, and 8-month efficiency halving. Held at 82 because this is executive commentary, not independent research, and the 2030 200GW math is not disclosed.

Read the original ↗Export Markdown

Take

Suleyman’s no-wall argument reads less like neutral forecasting and more like Microsoft defending the right to keep spending. The hard evidence is real: frontier training grew from roughly 10^14 FLOPs in 2010 to over 10^26 today; Nvidia moved from 312 TFLOPS on A100 to 2,250 TFLOPS on Blackwell; the piece says one training run fell from 167 minutes to under four on modern hardware.

That proves supply is still climbing. It does not prove capability returns keep climbing at the same slope. Epoch AI’s claim that compute for fixed performance halves every eight months actually weakens the brute-force framing: software efficiency is doing part of the work. The 200GW-by-2030 line is the tell. No reproducible calculation is given, and the author is Microsoft AI’s CEO, with Azure, Maia 200, and data-center capex sitting directly behind the argument.

Source: MIT Technology Review · AI · technologyreview.com
Published: Apr 8, 2026, 22:00 UTC+8 · Apr 8, 2026