Small labs can match frontier AI research: $1,000 per experiment and 10k B300s may suffice

yacineMTB · x · 2026-10-15

• ar0cket1 argues AI research breakthroughs need far less compute than assumed: training a 1B-active-param model to show meaningful baseline-vs-method differentiation costs $500–1,000 per run.
• Many recent architectural gains (mHC, KDA, attention residuals, latent MoE) are demonstrable at this scale; optimizer research too. RL is most scale-dependent but only a few times costlier.
• A neolab with 10k B300s is essentially tied with frontier labs — experiments run at half scale or less, but above 100B params scaling laws likely don't change experimental conclusions, so signal loss is minimal.
• What matters is researcher density in a good direction: with more strong researchers on one direction than OpenAI/Anthropic, you can out-research them regardless of raw compute.

Original post →

More from Infra

• CoreWeave executives unload shares for months as stock slides, insider filings show — SumitGup · 2026-10-15
• Musk amplifies a16z compute-scarcity chart: old GPUs becoming appreciating assets — elonmusk · 2026-10-15
• SK Hynix seeks US approval to ship chipmaking equipment to China plants as US factory condition — teortaxesTex · 2026-10-15
• CoreWeave co-founder courts insurers' 'unlimited' capital as private credit hits 35% of US insurer assets — SumitGup · 2026-10-15
• Go 1.27 makes HTTP/1 connection reuse ~5x faster: 100 requests drop from 710ms to 150ms — rakyll · 2026-10-15
• Open textbook "Model-Driven DevOps for Big Data Applications" now freely available — burkov · 2026-10-15

Infra channel →