Sunday, October 18, 2026

TL;DR: Oilfield service firms running behind-the-meter gas engines have become the de facto utilities of AI buildout.

Trinetti maps how the burden of securing power has shifted from utilities to data center customers, and how companies that once powered drilling rigs now sell multi-year on-site generation contracts. The piece explains the economics of BYOG arrangements and why they are becoming structural rather than a bridge until the grid catches up.

TL;DR: CC Wei's two-year-old promise to Musk is why TSMC is still in the Texas Terafab hunt.

Culpan reconstructs how a chairman's handshake assurance to a then-minor client became the backdrop for a three-way dance between Musk, Intel and TSMC over a Texas mega-fab. The tension is pace: Intel is moving fast to lock in the deal while TSMC plays a slower, relationship-driven game, and the piece shows how personal trust still moves multi-billion-dollar fab decisions.

TL;DR: Export controls on DUV tools, not just EUV, are the real lever on China's chip trajectory.

Saif Khan went from writing CSET deep dives to building the export-control regime inside the NSC and Commerce across two administrations, and now argues through his new think tank's DUV report that tool restrictions matter more than chip bans. The conversation also makes the case for slow, technical policy research in a field that moves weekly — a rare insider view of how posting becomes policy.

TL;DR: AlphaFold predicts structures but not the folding process, and the Bitter Lesson may not apply to biology.

DeepMind's Kohli and Biohub's Candido push back on the triumphalist AlphaFold narrative, separating structure prediction from the still-unsolved dynamics of folding. The debate over whether pure scaling or domain-specific inductive bias wins in biology is a useful counterweight for anyone assuming compute alone cracks every science.

TL;DR: Data center EPC margins have surged on fixed-price work, but schedule risk still sits with owners.

Using Primoris's renewables overruns as a foil, the piece dissects how multi-prime AI campus contracts allocate cost versus timing risk and why builders nearest the data center have seen margins rise fastest. The investor lens is historical: the contractors that survived the last downturn were the ones paid from regulated budgets with repeat work and cash.

TL;DR: Microsoft designed Maia 300 around moving data, not FLOPs, for agentic inference at multi-gigawatt scale.

Azure's hardware chief explains the co-design choices behind a 96-accelerator, 167kW rack, 18 TB/s of HBM4 and a scale-up fabric for 12,288 chips on TSMC N2P. It's a detailed look at how a hyperscaler thinks about inference economics when it intends to run most of its own workloads off NVIDIA.