Core Event: Negotiations for 1GW+ Data Center Lease Underway
Anthropic is in early-stage negotiations to lease up to 1 gigawatt (GW) of compute capacity, with plans to入驻 Stream Data Centers-developed datacenter campuses and deploy TPU chips co-designed by Broadcom and Google. The company is pursuing two parallel paths: direct datacenter infrastructure leases and chip-level compute rentals. Construction of 1 GW capacity is estimated to require at least $4 billion in capital expenditure (approximately 26.9 billion RMB), while Anthropic currently relies primarily on cloud providers including AWS, Google, and SpaceX.
Key facts:
- Lease scale: Up to 1 GW compute capacity
- Datacenter partner: Stream Data Centers (27-year history, nationwide U.S. coverage)
- Primary chip solution: Broadcom-Google co-designed TPU (with flexibility to deploy NVIDIA GPU)
- Capital expenditure: Minimum $4 billion
- Google’s role: Exploring credit guarantee provision
- Existing compute sources: AWS, Google, SpaceX
Operational Details: Dual-Track Compute Acquisition Strategy
Anthropic’s compute acquisition strategy shows clear divergence. On one path, the company plans to enter campuses as direct tenant—securing both physical space and power infrastructure rather than just virtual servers. On the other, it maintains hardware flexibility by permitting NVIDIA GPU deployment alongside TPU, reflecting strategic diversification away from single-vendor dependency.
The counterintuitive fact: While the approach centers on TPU, NVIDIA has been injecting massive capital into datacenter firms to secure GPU inclusion in AI infrastructure—meaning Anthropic’s decision will influence chip ecosystem dynamics.
Notably, Anthropic has already secured compute via multiple channels:
- AMD agreement: Locked 2 GW chip supply
- TPU lease: $3.5 billion deal through Apollo-Blackstone special-purpose vehicle (approx. 23.5 billion RMB)
The $3.5 billion agreement runs parallel to the proposed $4 billion infrastructure investment, revealing a multi-tier, multi-vendor compute reliance system.
Vendor Resource Comparison
| Lease/Partnership Type | Scale/Amount | Partner | Chip Solution | Financial Model |
|---|---|---|---|---|
| New datacenter lease negotiation | Up to 1 GW | Stream Data Centers | TPU primary, GPU supported | Direct tenancy + Google credit guarantee under discussion |
| TPU lease agreement | $3.5 billion | Apollo + Blackstone | TPU exclusively | Special-purpose vehicle dedicated funding |
| Chip supply agreement | 2 GW | AMD | CPU/GPU | Direct chip capacity locking |
| Existing compute sources | Not disclosed | AWS, Google, SpaceX | Hybrid architecture | Conventional cloud service procurement |
Practical Recommendations
For developers: Anthropic’s multi-chip approach suggests its model framework must support heterogeneous hardware ecosystems. If inference services open publicly, users will benefit from greater hardware flexibility and reduced vendor lock-in risk.
For datacenter operators: Stream Data Centers, as physical infrastructure provider, would become a strategic long-term partner if the deal closes. This direct-tenant model may be replicated by other AI companies aiming for power and space resource bargaining power beyond cloud provider ecosystems.
If you’re evaluating Anthropic API access, monitor AMD collaboration progress closely: 2 GW chip scale implies significant training/inference capacity growth; however, if TPU-optimized inference dominates access patterns, short-term API availability may remain tied to Google’s infrastructure.
Final Word
Anthropic’s compute sourcing is evolving from “cloud rental” toward “infrastructure direct control + chip customization.” This dual-track strategy enhances supply chain resilience while reflecting AI firms gaining end-to-end control from physical layer to chip layer—an operational shift that signals profound restructuring of AI infrastructure value chains.