GPU Clouds Become a Hyperscaler Utility

Vertical integration and inter-hyperscaler leasing are pricing third-party GPU vendors out of strategic importance.

Hyperscalers have stopped competing for GPU capacity and started building it. Meta, OpenAI, Anthropic, and Google announced $50+ billion in owned infrastructure this month alone: Meta's Louisiana megafacility, OpenAI's $20–30 billion Georgia campus, Anthropic's $19 billion TeraWulf contract, and a $10 billion inter-hyperscaler lease between Meta and Anthropic. Simultaneously, $40 billion in institutional capital (MGX, AIP, BlackRock) closed the Aligned Data Centers acquisition, while Japan announced its sovereign AI infrastructure consortium with NVIDIA. The pattern is unmistakable: hyperscalers and governments are consolidating control of AI infrastructure. Third-party GPU vendors are no longer platforms—they're becoming utilities.

The neocloud market is pricing this correctly. CoreWeave, Nebius, and IREN have lost 30–61% market cap this month as investors recognize the structural problem: when hyperscalers control their own supply, third-party vendors lose leverage. NVIDIA's 9.3% stake in Nebius (shares up 19%) reads as triage, not confidence—the chip giant is placing a bet that at least one independent distributor survives. Hut 8 and TeraWulf have adapted by locking into long-term fixed contracts with hyperscalers, not competing as generalist cloud providers. CoreWeave's $249 billion debt burden is now under systemic scrutiny. GPU supply pluralism (AMD's $5 billion Anthropic partnership, NVIDIA's Vera production) validates chip alternatives, but pluralism only matters if third-party vendors can still compete.

Efficiency is eroding the capex moat. SpaceX AI's gigawatt facility claims $28 billion annualized revenue on inference workloads—a 3–5x premium over traditional cloud—by optimizing software and hardware for a narrower use case. More structurally: Moonshot AI's Kimi K3 reportedly matches Anthropic's Opus 4.8, built by a 300-person team under US export controls. If software efficiency, lean teams, and open-weight models close performance gaps faster than hyperscalers can deploy $20–30 billion megafacilities, the race is no longer about raw GPU accumulation. It's about intelligence-per-GPU and control of the distribution layer.

Three signals to watch: (1) Neocloud M&A—distressed fire sales confirm structural decline; strategic premiums suggest hyperscalers view capacity as complementary. (2) AMD's actual training workload share at Anthropic—if >25%, pluralism is real; if <10%, it's theater. (3) Brownfield scalability—TeraWulf's $19 billion Anthropic deal proves repurposed fossil fuel plants can deliver rapid capex. If five more such contracts close at $15–25 billion each, owned infrastructure + long-term contracts + fixed power becomes the capex standard, replacing cloud tenancy for the next decade.