Sunday, August 16, 2026

TL;DR: Google is taking over Crusoe’s Cheyenne campus — 1.8 GW, four halls, a 42-turbine gas fleet Google does not own.

The piece walks a 716-acre, 3-million-square-foot site with more than $7 billion in power investment and a gas fleet still on someone else’s books. The bind is familiar: the campus is real, the electrons are contracted, and ownership of generation is deliberately not Google’s.

TL;DR: Automating AI research could compress years of progress into months, making compute bottlenecks and alignment suddenly decisive.

Greenblatt and Patel debate whether compute, experiment throughput, or scarce human-generated insights will cap the feedback loop once models can perform AI R&D. Even a modest research multiplier could turn gradual progress into a discontinuous jump, leaving alignment and governance little adaptation time.

TL;DR: Ramez Naam on the grid falling behind AI load, sodium batteries an order of magnitude cheaper, and wave-powered data centers as a path to energy abundance.

Saturday's Moonshots (#280) is a long sit-down on the exact bind the stack is in: compute is ready, electrons are not. Naam walks the 200 GW hiding in today's grid, why sodium chemistry changes the storage math, and whether wave power is a real siting option rather than a slide-deck idea.

TL;DR: The GPU is no longer the scarce resource — inference runs on accumulated state, and that state has to live somewhere.

Zeoli maps the tier ladder from HBM down through flash to object storage and argues the hot path is now data movement, not FLOPs. Agent workloads and long KV caches turn “enough HBM next to the GPU” into a storage-architecture problem.

TL;DR: The memory question is no longer “who can secure more HBM” — it is how long AI can scale around a single memory architecture.

FMS 2026 is used as the tell that inference, agents, and persistent KV cache break the HBM-only stack. Tiered memory, zHBM, and packaging capacity become the next allocation fight once HBM4 supply is no longer the only constraint.

TL;DR: Base Power’s real bet is that manufacturing home batteries can create a scalable, distributed power utility.

The profile follows Base from a clever Texas battery offering into the harder business of manufacturing, installation, fleet control, and utility economics. Its tension is whether industrial execution—not software alone—can turn thousands of household assets into dependable grid capacity.