Two threads ran together today. Serving, not training, is where the compute money goes: OpenAI's compute chief puts inference above 80 percent of future demand, Crusoe reports power rather than GPU supply as the binding constraint on new capacity, and the prefill and decode split is the architectural answer to reasoning models that spend tokens by the thousand. At the same time the open-weight tier got more usable, with a 27B vision model running quantized on a laptop and GLM 5.3 measured task by task against the closed frontier. Baseten's argument joins the two threads: past a certain volume, post-training an open model stops being a preference and becomes the cost decision.
Sachin Katti puts inference above 80 percent of all AI compute spend and lays out a road to 30GW, with revenue trailing capacity rather than leading it. Chase Lockmiller supplies the physical half from Crusoe's 2.1GW Abilene campus: at roughly 60 million dollars per megawatt, siting follows available power, and negative West Texas prices are already deciding where the buildout lands.
Qwen 3.8 27B ships Apache 2 licensed and vision capable, quantized to about 17GB and comfortable on a well specced laptop. The catch is the shipped default: reasoning effort is set to xhigh, which produces spectacular over-thinking on consumer hardware. Lower it before forming an opinion of the model.
A dated brief from the vibe-coding frontier. Today’s Intel.
Tuhin Srivastava's case is that renting a frontier API is the cheap option only until you have volume, after which post-training an open model is both the cost play and the durable advantage. A task by task read of GLM 5.3 against closed models is the empirical half of the same question, and it is the read anyone weighing an open-weight backend actually needs.
Reasoning models changed the shape of an inference workload, and disaggregation is the response: separate the prefill and decode phases and schedule each on hardware suited to it. Gerstner and Madra trace what Groq built on that idea, why served token costs fell as far as they did, and what Nvidia bought for 20 billion dollars.
GitHub made Balanced depth generally available for Copilot code review, settable as an organization or repository default and overridable per pull request, so thorough analysis can be reserved for risky changes and Lite handles the rest. On the ecosystem side, Awesome Claude Code pulls the scattered guides, skills, plugins and observability tooling into one maintained index.
Guillermo Rauch describes Vercel moving from elastic compute sized for humans to infrastructure a coding agent can drive on its own, deploying, securing and maintaining software across hours or days without a person in the loop. That is the demand side of every inference number above.
Two readings of where the next bottleneck sits. Ali Ghodsi argues capability is no longer the limit inside a company, missing tacit organizational context is, which is why agentic deployment is harder than a benchmark score suggests. Yash Patel puts it upstream, at continual and data efficient learning from sparse real feedback, and explains why verifiable domains like code and math dominate RL training in the meantime.