Monday, June 29
China Surges, AI Stumbles
China has claimed the top spot on the TOP500 supercomputer list (opens in a new tab) with a machine built entirely on domestic chips, a direct rebuke to US export restrictions designed to kneecap its compute ambitions. The system, unveiled at ISC'26, outpaces its nearest rival and signals that American chip controls may be accelerating Chinese self-sufficiency rather than containing it. Meanwhile, Z.ai is claiming its GLM-52 model can match Mythos on cybersecurity benchmarks (opens in a new tab), adding another data point to a pattern of Chinese labs closing gaps that were supposed to take years. The strategic calculus in Washington just got considerably harder.
Ford's decision to rehire veteran engineers (opens in a new tab) — internally called 'gray beards' — after AI-assisted design processes fell short of expectations is a rare public admission that the current generation of AI tools cannot yet replace decades of domain expertise. The move follows a broader industry pattern of overcorrection: companies automate, hit the limits of what models can reliably do, then quietly walk it back. Separately, prosecutors in the Palisades fire arson case used ChatGPT conversation logs as trial evidence (opens in a new tab), a landmark moment that turns everyday AI interactions into potential legal records. Taken together, both stories expose the same uncomfortable truth — AI outputs carry real-world consequences that neither vendors nor users have fully priced in.
SemiAnalysis projects that behind-the-meter datacenter capacity in the US could exceed 40GW by 2028 (opens in a new tab), a figure that reflects just how severely grid infrastructure is now constraining AI buildout. Hyperscalers are essentially constructing private power ecosystems to bypass a transmission network that was never designed for this load. At the same time, Wall Street is positioning Micron as the next Nvidia (opens in a new tab), betting that memory — not just compute — becomes the critical bottleneck as model inference scales and HBM demand explodes. The infrastructure layer of the AI economy is where the next concentration of power, both literal and financial, is forming.