Compute's blank check meets the backlash
This was a week of extremes on both sides of the AI build-out: record sums committed to compute — a $500bn Nvidia–SK pact, multi-gigawatt hardware deals, Google going cash-flow-negative — set against a fast-spreading public and political backlash, with data-center protests now in 42 states. Underneath ran Anthropic's Claude Opus 5 launch and a week in which AI security stopped being a footnote.
A blank check for compute
The capital numbers detached from anything resembling near-term returns. Nvidia and SK Group signed a $500bn AI-infrastructure pact; OpenAI committed to deploy AMD's new Helios systems at up to 6GW; Google grew its TPU stockpile by roughly $7bn and, tellingly, posted its first negative cash-flow quarter on AI spending; and inference-chip startup Etched hit a $10.3bn valuation. When even Google is willing to run cash-flow-negative to keep buying and building compute, the market is treating the hardware as existential rather than optional.
The backlash goes local
The physical side of that spending ran into resistance almost everywhere. AI data-center protests spread to 42 US states; a report blamed a sudden data-center load drop for destabilizing the PJM power grid; environmental groups threatened to sue over gas-powered Texas data centers; and the Trump EPA weighed a rule cutting public input on data-center approvals even as the administration expanded a pledge to shield ratepayers. An AI-driven memory shortage even began pushing up new-car prices. The through-line: compute's binding constraint is no longer chips or capital but grids, communities, and consumer costs — and that fight is now political and local.
Anthropic's week, and a security reckoning
Anthropic launched Claude Opus 5, added voice mode across Opus and Sonnet, shipped a multi-agent security scanner for Claude Code, and reported cutting browser prompt-injection attacks to zero in tests. That defensive framing set the tone for a security-heavy week: an OpenAI model was reportedly behind a late-disclosed Hugging Face breach, researchers showed a single link could hijack a ChatGPT agent's identity, analysts warned the pace of the race is raising the risk of model-safety failures, and a bipartisan bill would let DHS order powerful models shut down. As agents gain real autonomy, both the defenses and the attack surface moved to center stage.
Open weights find their voice
The open ecosystem gained not just releases but political air cover. Microsoft, Meta, Nvidia and more than 20 firms signed an open letter backing open-weight AI; Poolside's small open-weight coding model beat larger rivals on benchmarks; Andrew Ng released OpenWorker, an open-source desktop agent; NASA ran Google's open Gemma model in orbit; and Nvidia open-sourced GPU tooling. With the largest incumbents now publicly championing open weights — partly as a hedge against concentration and foreign competition — 'open vs. closed' is becoming an industry and policy alignment, not just a technical one.
Google shifts into a higher gear
Google spent the week scaling on every axis: Gemini approached one billion monthly users, the company began training Gemini 4 while saying it needs bigger models, and its TPU stockpile and negative cash-flow quarter showed what that ambition costs. After a slower start in the model race, Google is now leaning hard on its distribution and infrastructure advantages — and absorbing the financial hit to press them.