Compute & Infrastructure

The Trump administration has established a functional pre-release review regime for frontier AI models, and this week's events confirm it is operational rather than theoretical. Anthropic's Mythos 5 was taken offline after a Friday ultimatum two weeks ago and has now been partially restored to a select group of 'trusted partners' following White House negotiations, per Bloomberg. Separately, OpenAI limited the rollout of GPT-5.6 — a three-tier suite comprising flagship Sol, mid-tier Terra, and Luna — to a limited preview after government pressure, with a source indicating Washington cautioned OpenAI against release without prior approval, per Tom's Hardware.

The mechanism appears to be a 30-day advance access window tied to the President's executive order, which OpenAI is voluntarily complying with while seeking 'a more sustainable approach for future releases.' From an infrastructure strategy perspective, this is significant: it inserts a regulatory variable into deployment timelines that downstream compute capacity planners — data centre operators, cloud providers, enterprise buyers — must now account for. Model release dates, which drive inference capacity procurement cycles, are no longer solely within the labs' control.

Vantage Data Centers has confirmed structural completion — 'topping out' — of the second building at OpenAI's Lighthouse campus in Kenosha, Wisconsin, per Data Center Dynamics. This is confirmed physical progress, not an announcement: steel is up. The Lighthouse campus is one of the flagship sites in OpenAI's multi-billion-dollar domestic infrastructure push and represents the kind of purpose-built, hyperscale inference and training capacity that the company will need to support GPT-5.6 and successor models at scale.

The broader buildout picture remains mixed. A Dogecoin-origin cryptominer, Z Squared, is acquiring a site in Arkansas for a 150MW immersion-cooled AI/HPC campus powered by behind-the-meter natural gas — announced but unbuilt, per Data Center Dynamics. SuperX has launched an AI inference cloud location in Denver, its first North American deployment — operational but at modest scale, per Data Center Dynamics. These smaller entrants signal that the mid-tier inference market is attracting speculative capital, but the gap between their scale and hyperscaler buildouts remains vast.

Two distinct geothermal signals emerged this week. In the US, I-Pulse CEO Robert Friedland outlined a $250 million investment — described as US government-backed — in semiconductor drilling technology designed to access deep geothermal heat for around-the-clock clean baseload power, per Bloomberg. Friedland's argument is explicit: geothermal, not wind or solar, is the correct answer to AI's power demand because it provides firm, dispatchable capacity independent of weather. In Japan, Japan Semiconductor Co. has signed a 15-year geothermal virtual power purchase agreement with the Waita geothermal project, purchasing environmental certificates over the contract term, per Data Center Dynamics.

The strategic logic is consistent across both cases: AI data centres require power that is continuous and predictable, not just low-carbon. Geothermal's value proposition is its load factor — near 90% capacity utilisation versus 25-35% for solar and wind. The $250M US investment is at early-stage technology development; the Japan vPPA is a financial instrument rather than direct generation ownership. Neither is operational baseload capacity today, but the 15-year contract horizon in Japan and the government-backed capital in the US suggest these are durable commitments rather than speculative announcements.

Onsemi's $7 billion all-stock acquisition of Synaptics, confirmed this week, targets the intersection of smart power management and edge AI hardware — specifically robotics and physical AI applications, per Tom's Hardware. Synaptics brings edge inference silicon and human-machine interface IP; Onsemi contributes wide-bandgap power semiconductors and automotive-grade power management. The combination is architecturally coherent for the robot/autonomous systems market, where efficient power delivery and on-device inference are co-dependencies.

The Semiconductor Engineering weekly review also flags IBM's 7-angstrom chip with 40% more SRAM area, 1nm MoS2 nanotube research, a $250M CHIPS Act award, and a new advanced packaging site entering development, per Semiconductor Engineering. The CHIPS Act award and advanced packaging site are confirmed funding and site announcements respectively; the MoS2 nanotube work is research-stage. The cumulative picture is of a semiconductor industry under pressure to diversify packaging geography and materials science simultaneously.

Bloomberg's explainer on AI distillation this week crystallises a structural risk to the capital-intensive AI infrastructure model: rivals can build competing systems for materially less by distilling from frontier models, undermining the returns on the hundreds of billions being spent on training compute, per Bloomberg. This is not a new dynamic — DeepSeek's R1 demonstrated it in early 2025 — but its persistence as an investor concern shapes the risk appetite for continued hyperscale infrastructure commitment.

Simultaneously, Apple and Microsoft price increases driven by rising component costs triggered a selloff in Asian memory chip stocks, per Bloomberg. The concern is that input cost inflation compresses device demand, which flows back through the stack to dampen the memory chip cycle that has underwritten much of Asia's semiconductor capex expansion. These two pressures — distillation eroding training ROI from below, and component inflation threatening demand from above — create a structural squeeze on the economics of the current AI infrastructure investment thesis.