Who Controls Compute Controls AI: The Multi‑Gigawatt Arms Race

The Moment Everything Changed

On a single week in April, the battleground for the next half‑decade of AI stopped being only model design and started looking very much like power, silicon and rack‑scale lock‑ins. A string of announcements — Broadcom’s long‑term chip deal for Google, Anthropic’s multi‑gigawatt compute commitments, CoreWeave’s multi‑year GPU supply and NVIDIA’s Vera Rubin systems and software moves — suddenly made access to accelerators, HBM and interconnects the scarcest commodity in the stack. The question that follows is simple and brutal: whoever controls the compute fabric will control who can build, run and monetize frontier AI.

Background

The industry’s early years of the large‑model boom were defined by architecture, data and training recipes: better models beat lesser ones. Hardware mattered, but it was largely fungible — buy GPUs or rent TPUs and iterate. That shifted as models ballooned into trillions of parameters and the economics of training and inference came to hinge not just on raw FLOPS but on memory bandwidth, on‑chip SRAM, NVLink fabric and entire rack efficiency. Hyperscalers had quietly invested in custom accelerators for years; now, those investments are being transacted into long‑term supply agreements and co‑designed systems that promise dramatically lower per‑token costs and higher throughput. The result is a visible transition from model‑led competition to infrastructure‑led advantage, with geopolitical and market consequences that are only beginning to ripple outward.

What Happened

This week crystallized the transition. Broadcom disclosed a long‑term agreement to develop and supply future generations of custom AI processors and rack components to Google, tying Broadcom to Google’s TPU roadmap and signaling multi‑year supply assurances CNBC. Anthropic announced expanded agreements for “multiple gigawatts” of next‑generation TPU capacity with Google and Broadcom, saying most of that capacity will be sited in the United States and scheduled to come online starting in 2027 Anthropic. Separately, Anthropic signed a multi‑year arrangement with CoreWeave for NVIDIA GPU capacity to come online in 2026 CoreWeave. At the same time, Reuters reported Anthropic is exploring designing its own ASICs as a hedge against vendor dependency and cost exposure Reuters.

NVIDIA amplified the pressure with Vera Rubin, a rack‑scale platform that fuses new Rubin GPUs, Vera CPUs, DPUs and high‑bandwidth fabrics into turnkey PODs and claims major efficiency gains versus prior generations NVIDIA. NVIDIA’s December acquisition of SchedMD — maker of the Slurm scheduler — and subsequent emphasis on orchestration raised concerns that control over scheduling software could tilt neutrality in multi‑vendor clusters Reuters. Meanwhile, OpenAI’s massive funding round and reported $122 billion of committed capital further concentrates buying power among a handful of deeply capitalized labs, increasing competition for scarce compute and memory resources OpenAI CNBC.

Why It Matters

The immediate technical effect is straightforward: whole‑rack efficiency beats chip‑only gains for modern models. Memory and interconnect shortages, already pressuring DRAM and HBM prices, mean that having guaranteed access to accelerators and HBM at scale is now a strategic asset. Economically, long‑term, multi‑gigawatt commitments and co‑designed racks create a moat that is far harder and slower to replicate than an open‑source model checkpoint. Startups that lack deep pockets will face higher spot pricing or contractual denial, compressing innovation at the margins and shifting power toward well‑capitalized labs and hyperscalers.

There are geopolitical and policy dimensions as well. Locating multi‑gigawatt clusters within particular jurisdictions is a national security decision; governments will want assurance about where models are trained and who controls the underlying hardware. Memory concentration among a few fabs, and supplier leverage in accelerators and DPUs, invites scrutiny from antitrust and national‑security regulators alike. Finally, the energy and environmental calculus changes when the unit of competition becomes site power measured in gigawatts rather than chips measured in teraflops.

Expert Perspectives

Industry leaders framed these moves as necessary at scale. Dario Amodei of Anthropic praised the Vera Rubin platform’s capabilities in NVIDIA’s materials: “NVIDIA’s Vera Rubin platform gives us the compute, networking and system design to keep delivering while advancing the safety and reliability our customers depend on” NVIDIA. Anthropic’s CFO Krishna Rao described the expanded deal as “a continuation of our disciplined approach to scaling infrastructure” and emphasized U.S. siting and enterprise demand Anthropic. Jensen Huang framed Vera Rubin as a “generational leap” that unites chips and systems into one supercomputer NVIDIA.

From the specialist cloud side, CoreWeave’s CEO Michael Intrator positioned his company as an essential bridge between model builders and production systems: “We’re excited to work with Anthropic at the center of where models are put to work and performance in production shows up” CoreWeave. On the skeptical side, analysts and HPC operators have warned that vertical integration — hardware, interconnect, scheduler and software — risks vendor lock‑in and could reduce neutrality in cluster orchestration, a concern that gained traction after NVIDIA’s SchedMD acquisition Reuters. Market analysts like Mizuho projected enormous revenue upside for suppliers that capture these commitments, underscoring how supplier leverage can reshape valuations CNBC.

What to Watch

The decisive signals in the next 12–24 months will be capacity actually delivered, not just contracts signed. Track when Anthropic’s multi‑gigawatt TPU clusters come online (they’ve signaled 2027 starts) and whether CoreWeave’s promised GPU capacity appears in the market later in 2026 Anthropic CoreWeave. Monitor NVIDIA partner OEM shipments for Vera Rubin systems and independent benchmarks (MLPerf and third‑party rack tests) to verify claimed 10x inference throughput and token‑cost reductions NVIDIA. Memory supply is the single choke: watch HBM wafer start dates, Micron/SK Hynix/Samsung fab timelines and industry price indices for DRAM/HBM to see whether shortages persist or ease The Verge.

Policy and market structure matters too. Regulatory scrutiny of acquisitions that touch orchestration layers, and any moves toward “sovereign AI” procurement rules, could reshape where and how compute is sited. Finally, watch Anthropic’s ASIC exploration closely — if it commits to chip design and a foundry path, the industry will have crossed into a second phase where large labs both own models and the silicon that runs them Reuters.

What emerges from this week is not simply a set of deals but the opening salvo of a structural contest: models will still matter, but the ability to cheaply and reliably run those models at scale will decide who can lead. For startups, governments and investors, the strategic imperative is clear — secure compute, or lose the race.