Cerebras/Groq chaining thousands of chips enables 100T scaling
scaling01 · x · 2026-08-25
User observes that Cerebras and Groq are essentially chaining together dozens, hundreds, or even thousands of chips. Based on this, they argue that if we want to scale to 100T parameter models, we could theoretically do it today.
More from Infra
- Neon deep dive: WAL+S3 storage architecture for the era of agents — matei_zaharia · 2026-08-25
- Hot Chips reveals Waymo's sensor-heavy self-driving stack — firstadopter · 2026-08-25
- Darkbloom goes viral, faces Qwen latency issues — gajesh · 2026-08-25
- 3D die stacking is the new norm at Hot Chips 2026 — appenz · 2026-08-25
- 1Password Targets Standing-Access Gap Left Open by AI Agents — shashib · 2026-08-25
- AI Capex to Reach $765B in 2026, Overtaking Oil and Gas as Compute Futures Emerge — rohanpaul_ai · 2026-08-25