OCP's Silicon Photonics Vision Roasted as Per-Token Energy Nears 1 Joule
jwt0625 · x · 2026-08-18
- OCP released an architecture vision for open silicon photonics in AI systems, mixing units like pJ per token and pJ per training step, which the author calls a high-entropy blend of based and unhinged AI slop.
- The striking number: inference energy per token is now close to 1 joule — think lifting a basket of apples at 1 m/s, or a row of tennis ball machines firing continuously at 6 m/s.
Related event: OCP's Open Silicon Photonics Vision Draws Criticism(2 posts)→
More from Infra
- Parlor: Open-source, on-device real-time multimodal AI similar to GPT-Live — tom_doerr · 2026-08-18
- Unsloth Studio desktop implodes on Windows; what GUI app can run MoE-quantized models? — IngwiePhoenix · 2026-08-18
- Opinion: transformer will eventually be replaced — can Nvidia disrupt itself and stay ahead? — yangyi · 2026-08-18
- File Systems Emerge as Core Paradigm for AI Data Interaction — blaizedsouza · 2026-08-18
- Merge Partners with Mastra to Provide Unified API Gateway for AI Agents — shensi · 2026-08-18
- Does High Concurrency Make MoE Serving Load Nearly All Weights Per Token? — LocalLLaMa_reader · 2026-08-18