AI compute is now strategic

Artificial intelligence used to look like software. Now it looks like heavy industry. That is the shift behind the new scramble for “compute.” Training and running frontier AI models no longer depends on code alone. It depends on scarce chips, dense racks, specialized networking, cooling systems, grid connections, and enough electricity to keep all of it running. Domain-b reports that global AI infrastructure spending is set to exceed $2 trillion in 2026, with longer-term investment potentially reaching $3 trillion to $4 trillion by the end of the decade. That number is so large because AI has moved out of the cloud metaphor and into the world of physical bottlenecks. The bottlenecks are not subtle. Nvidia’s GB200 NVL72 system packs 72 Blackwell GPUs and 36 Grace CPUs into a single liquid-cooled rack-scale machine. Nvidia says the rack behaves like one giant GPU for trillion-parameter inference workloads. That is the point. The industry is no longer buying generic servers by the pallet. It is buying tightly integrated AI factories, where the chip, the interconnect, the rack, and the cooling loop all matter at once. That changes what a “capacity decision” means. A company choosing a model provider, a cloud region, or a colocation site is also choosing access to transformers, memory bandwidth, power delivery, and network fabric. The hardware stack has become strategic because each layer constrains the next one. A fast chip without enough power is stranded. A powered site without transformers is delayed. A finished building without GPUs is just an expensive shell. The data center market already shows what happens when demand outruns all of those layers at once. CBRE says vacancy in primary North American data center markets fell to 1.4% at the end of 2025, a record low, while asking rates for wholesale colocation kept rising. In plain English, the empty rooms are gone. Much of the market is now preleased before construction finishes. That pushes buyers into off-market deals, frontier locations, and long lead-time commitments that would have looked absurd a few years ago. And even that is only the real estate side. The deeper constraint is power. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of global consumption. In its base case, electricity supplied to data centers rises from 460 TWh in 2024 to more than 1,000 TWh in 2030. In the United States, the IEA says data centers are on course to account for almost half of electricity demand growth through 2030. AI is not just competing for GPUs anymore. It is competing for megawatts. That is why the biggest AI announcements now sound like utility planning. OpenAI said in January 2025 that Stargate intended to invest $500 billion over four years in U.S. AI infrastructure. By September 2025, OpenAI said new sites put the project ahead of schedule toward a 10-gigawatt build-out. Later updates said the broader effort had moved past 8 gigawatts of planned capacity and more than $450 billion in investment commitments. Those are not software numbers. They are grid numbers. Once compute becomes this physical, control shifts upward. The winners are not just the firms with the best models. They are the firms that can lock in chips, reserve capacity years ahead, finance custom builds, and negotiate directly with utilities. Compute stops being a background utility and starts behaving like oil pipelines, shipping lanes, or semiconductor fabs. The strategic question is no longer who has the smartest model. It is who can keep 2 million chips fed.