The AI Infrastructure Arms Race: Google, CoreWeave, and Why Compute Beats Models
Three deals in the same week tell the same story: the frontier AI competition has moved from model benchmarks to gigawatt-scale infrastructure.
Google’s $5B Anthropic Bet
Google committed over $5 billion to finance a Texas data center for Anthropic through Nexus Data Centers. The facility targets 500 megawatts by late 2026, expandable to 7.7 gigawatts — enough to power a mid-sized city. It will run on direct natural gas via on-site turbines, bypassing the public grid entirely.
Google already holds a 14% stake in Anthropic after investing $3B between 2023-2025. The new financing gives Google a second revenue channel — selling compute to a model it partially owns — while ensuring Claude doesn’t become an AWS exclusive. Amazon has invested $8B in Anthropic and is its primary cloud partner for training workloads.
CoreWeave’s $8.5B Debt Facility
CoreWeave closed an $8.5 billion delayed draw term loan on March 31, maturing 2032. The structure is non-recourse, ring-fenced to a dedicated entity, and secured by GPU clusters and contracted revenue — including an estimated $19B backlog from Meta.
The deal represents one of the fastest cost-of-capital compressions in technology infrastructure finance:
| Facility | Year | Size | Rate |
|---|---|---|---|
| DDTL 1.0 | 2023 | $2.3B | ~15% floating |
| DDTL 3.0 | 2025 | $2.6B | SOFR + 4.00% |
| DDTL 4.0 | 2026 | $8.5B | SOFR + 2.25% |
In three years, CoreWeave moved from high-yield equipment financing to investment-grade institutional infrastructure debt. That compression reflects a fundamental reclassification by lenders: GPU clusters with hyperscaler contracts are now treated as infrastructure assets, not technology bets.
The Scorecard
| Player | Commitment |
|---|---|
| Microsoft Stargate | $500B (4 years) |
| Meta US Expansion | $600B (through 2028) |
| OpenAI/Oracle | $300B (5 years, 4.5GW by 2027) |
| Google/Anthropic Texas | $5B+ |
The constraint is no longer chips. It’s power delivery at scale. The behind-the-meter model — on-site generation bypassing the public grid — is becoming the standard approach for frontier AI facilities because grid interconnection queues in the US now run 5-7 years.
What It Means for the Model Race
The infrastructure arms race matters for Stack Futures because it determines which models exist in 2027-2028. A lab that can’t secure gigawatt-scale compute can’t train the next generation of frontier models. The labs with locked-in infrastructure — OpenAI, Google, Anthropic, Meta — have a structural advantage that benchmark scores don’t capture.
The SFX-10 is a snapshot of today’s frontier. The infrastructure being built now determines who’s on it in two years.