Anthropic's 10GW Infrastructure Buildout — And Why Korea Needs to Be Inside
Scale of power demand that makes proprietary silicon — and supply chain ownership — essential
Power Demand Stack
AWS Exclusive Contract
5 GW
Dedicated data center capacity for Claude inference. Equivalent to 5 nuclear reactors running 24/7 exclusively for Anthropic.
+
Google + Broadcom TPU Network
5 GW
Distributed computing network co-developed with Google and Broadcom. Custom TPU-based accelerator infrastructure.
10 GW Total
Anthropic's committed infrastructure capacity — for a single AI company
Scale Reference: Nuclear Equivalents
⚛️
1 nuclear plant ≈ 1 GW
Standard large-scale nuclear power station average output
🏙️
10 GW = Seoul metro area
Approximately equal to the peak electricity demand of greater Seoul (10M+ people)
🌍
10 GW = Top-20 country
Total AI compute demand exceeds entire national electricity capacity of many mid-size nations
💰
$65B fuel for this
Series H capital primarily directed at building and operating this physical infrastructure
Why this scale drives the Korean semiconductor partnership
- At 10 GW scale, per-chip energy efficiency becomes the single largest operating cost variable — driving demand for custom ASIC over general-purpose GPU
- Custom ASIC needs a foundry partner → Samsung is the only strategic partner with logic fab capability
- Each new AI accelerator rack requires multiple HBM4 stacks → direct pull for SK hynix's highest-margin product
- Supply chain disruption at this scale costs millions per hour → makes equity-level partnerships with suppliers strategically necessary, not optional
Source: BloombergNEF AI Data Center Energy Report 2026; AWS-Anthropic agreement disclosures; Google Cloud + Anthropic partnership filings. Nuclear equivalent estimates based on average 1 GW nameplate capacity.