Gary Marcus questions Astra's $1B training bill as scaling data goes dark
GaryMarcus · x · 2026-09-07
Nina Schick claims Astra was trained on 100k Blackwell GPUs with an estimated $1B compute bill and 200 MW of power, arguing scaling still holds.
Gary Marcus suspects the real figure is even higher, and makes a broader point: base-model scaling data is no longer published—each new release bundles a new base model with undisclosed harnesses and augmented data, making it impossible to attribute improvements to any single factor. "Zero science is being reported," he writes, while asking what happens when the first gigawatt cluster comes online.
More from Infra
- Compute financing risk will fall to hedge funds and commodity traders, not private credit — AccBalanced · 2026-09-08
- Solo dev ships Jenny, an MIT-licensed local LLM desktop app after 1.5 years — TangySword · 2026-09-08
- Cacheon launches GLM-5.3 kernel arena, paying up to 33 TAO daily to beat sglang — JosephJacks_ · 2026-09-08
- Johns Hopkins report urges strategic power islanding to cut blackout risk — philvenables · 2026-09-08
- HBF math: matching H200 bandwidth needs ~4,900 concurrent NAND planes — lauriewired · 2026-09-08
- BMO: record US power demand isn't lifting gas use — renewables and batteries are filling the gap — aronchick · 2026-09-08