tomtunguz.com

The frontier AI market is segmenting from both ends of its supply chain. Rationing programs & export limits decide who may run the strongest models. Downstream, enterprises standardize on one or two named vendors, & products ship with a default model ...

Three months into running r2, an agentic AI harness, four lessons held : an inbox beats a task list, routing between local & cloud models is essential, self-healing works but surfaces more errors before it surfaces fewer, and even a no-huddle offense ...

AI model companies buy wholesale electricity by the megawatt and resell it as cognitive work. The unit economics of turning power into intelligence.

NVIDIA's Q2 FY27 revenue reached $96b with a $108b Q3 guide, but hyperscale revenue grew only 13% sequentially against 25% for everyone else. To fund the buyers filling that gap, NVIDIA extended payment terms -- DSO rose from 45 to 60 days & ...

When designing an agent, should its session run for your entire five-year company tenure or reset every day? Perpetual sessions suffer from context rot and security exposure. The winning architecture is a daily coordinator that resets every 24 hours, ...

AI hardware bottlenecks cascade in sequential multi-year waves from GPUs to memory, SSDs, CPUs, HDDs, and physical data center shells, driving facility buildout costs to $20b per gigawatt through the classic Bullwhip Effect.

Local models now answer 89% of everyday chat & reasoning queries as well as frontier models, & their efficiency per watt has improved 5.3x in two years.

Every software product has its own grammar: Figma has frames and components, Salesforce has accounts, contacts, and opportunities, CAD has tools and constraints. Capability and required expertise rise together, so the most powerful tools are the ones ...

Qwen3.6-35B-A3B generates 2.2x faster than Qwen3.8-27B, yet finishes slower because it thinks 3.1x longer. Across 25 tasks, quality is tied. Measure time to answer, not token speed.

Explains test-time training through the analogy of a GPS learning a persistent shortcut around daily traffic rather than a one-time reroute: the model takes a gradient step on the prompt it's answering, so its weights change as it works. Traces three ...