Power, Not Compute, Is Winning | Anthropic's $19B Lock-In Upends the Neocloud Ladder
Anthropic's $19 billion, 20-year lease with TeraWulf and Bloom Energy's $25 billion fuel-cell partnership signal a structural shift: power is no longer a datacenter byproduct but the primary competitive moat. Anthropic is essentially buying optionality on TeraWulf's future capacity at a fixed cost for two decades—a capital move that locks out competitors for the best long-term power contracts. The scale alone ($19B for Kentucky) mirrors betting that power undersupply will persist and that being first to secure it wins the AI lab race. This is not supply-chain negotiation; it is strategic capture of the grid constraint itself.
The market is consolidating around tier-1 players with capital and power. Crusoe Energy's $30 billion valuation (tripled) and Together AI's $8.3 billion Series C validate the neocloud category as real, investable businesses. But that validation came with a trap door: Meta's announcement of internal AI cloud services crashed Nebius by $120 billion in market cap, CoreWeave by 14%, and IREN by 15% in a single trading session. Meta proved that hyperscalers can monetize spare GPU capacity more efficiently than purpose-built competitors. The market was pricing neoclouds as the only path to AI compute; it is now repricing them as commodity suppliers competing against entrenched, margin-advantaged incumbents.
Anthropic and OpenAI are winning because they have both capital and power lock-in. OpenAI is leasing 10 gigawatts in Ohio—another future-capacity bet with the same duration horizon. Crusoe and Together can scale compute, but without long-term power contracts, they are margin-trapped: buying power at spot prices while hyperscalers buy at preferred rates. CoreWeave's 133 MW from Galaxy Energy in Texas is real progress, but it is tactical supply, not strategic moat. In 12 months, the question will not be who has the most GPUs but who secured the most stable, long-term power contracts and at what cost.
China's Meituan fielding a 1.6 trillion parameter LLM on 50,000 domestic chips reshapes supply geography in real time. The US assumption—that AI labs need Western GPUs or will face capability ceilings—no longer holds. Domestic chips are reaching scale and performance, backed by industrial datasets. This fractures the global compute market: neoclouds anchored to US supply now face competition from state-backed capacity independent of US export controls. Combined with Nvidia's 16-layer HBM push (memory supply competition), the infrastructure stack is splintering into US, China, and EU tiers.
Watch for OpenAI, Google, or Amazon to announce their own power-anchored expansions in the next quarter. Whoever secures the second wave of long-term capacity will lock in the tier-2 labs. Mergers and acqui-hires in neoclouds will accelerate—consolidation toward one or two open-model clouds (Together is the leading candidate) and pure compute utilities. CoreWeave and Nebius face a choice: merge with a lab, get acquired by a hyperscaler for balance-sheet strength, or accept margin compression. The power game has a finite number of winners.