Knowledge base

• Compute density per rack increased by 28X from an H100 rack to a Rubin rack.
• Networking within a rack has grown similarly, hitting the limits of copper cabling
• Rack power moved from roughly 5-10 kW to 100-250 kW to support today’s systems and will increase to 1MW over the next 3 years.
• Data center scale is moving from tens to hundreds of MW and, in some cases, GW-scale campuses.
• Power is increasingly supplied not only from grid-only to grid plus behind-the-meter or captive sources.

Let that sink in.

https://x.com/rajananandan/status/2089566695946109432?s=46&t=ay_h33GThR69kRjbFIQm_w

In one night, two hours apart, DeepSeek V4 Pro and Grok 4.6 both shipped, and both pushed near Claude Fable 5's level. DeepSeek matched Fable on tool benchmarks (Terminal Bench 87.9 to Fable's 88.0) at a fraction of the price: output around $0.87 per million tokens against Fable's $50, though DeepSeek has warned of price hikes. Grok landed close on quality with startling speed: coding tasks that drag past half an hour on GPT-5.6 Sol finished in under 20 minutes, some in three, at $2 in, $6 out.

Google launched Gemini 3.7 Flash on Thursday, three weeks after shipping 3.6 Flash, at half the price of its predecessor. The model targets coding and agent workflows, with real gains in debugging and code generation per Google's own benchmarks.

Meta Superintelligence Labs released Muse Glimmer, a 30B open-weight agent model under a permissive Apache 2.0 license, runnable on a Mac or a single consumer GPU with no cloud, no per-step API calls, working even offline. On its own benchmarks it nearly matches GPT-5.5 at tool use and agent tasks, while scoring roughly half of frontier models on hard general reasoning. Meta shipped the weights, though not the training data, and llama.cpp, Ollama and MLX support lands within days.

Seems even a16z has observed image-gen issues when trying to make it watertight !

Releasing next week qwen3.8

qwen.ai/blog?…

celeris.ai/blog-benchmarks.html

Celeris-1: Claims to be the fastest AI model.

For exact same prompt token consumption.

CPU only model

huggingface.co/jlnsrk/GLM-5.2-colibri-int4