What next for the compute crunch?

Demand surge

It has been reported that GitHub saw an approximately 14x annualised increase in commits over three months — a crude but telling proxy for inference demand as coding agents and "vibecoding" hit the mainstream. OpenAI and Anthropic are publicly saying they’re compute-starved, and the domino effect is already visible: when one service tightens limits, traffic floods competitors. Panic? Not yet. But the suddenness is uncomfortable. Who knew developer tools could become a canary in the coal mine for datacenter capacity?

Supply and rollout headaches

Big headline deals don’t conjure silicon overnight. Concrete must be poured, power hooked up, turbines and racks ordered, and GPUs fabricated, racked and networked. It has been reported that Nvidia’s GB200 rollout — a generation that’s largely liquid-cooled rather than air-cooled — has been unusually painful, with shortages of skilled labor and specialty plumbing components slowing deployments. The result: promised capacity lands late, and the short-term squeeze tightens.

Geopolitics and memory limits

It has been reported that Stargate’s 1GW datacenter under construction in the UAE has become a geopolitical chess piece amid recent U.S.–Iran tensions, after footage of the site circulated. Meanwhile, the deeper, slower bottleneck is DRAM: memory fabs don’t scale fast. SK Hynix’s roughly $8bn order of EUV kit from ASML will take years to meaningfully increase supply, and it has been reported that executives including Sundar Pichai have called out memory as a major constraint.

What next?

Short-term pain looks likely — throttling, outages, and rationing of premium inference cycles. But there’s another side: pressure breeds efficiency. Expect more aggressive model distillation, software-hardware co-design, and architectural tweaks to stretch scarce memory and teraflops further. Will this choke AI’s march forward, or force smarter engineering and cheaper innovation? For now, teams are literally wrestling power cords and pipelines — and the industry is learning, the hard way, that raw compute is not just rare silicon but an ecosystem of bricks, coolant and geopolitics.

Sources: martinalderson.com, Lobsters