I Squared Capital launched a U.S.-based AI inference and edge colocation data center platform with a $1 billion commitment, targeting the bottleneck hyperscalers won't solve themselves. The platform focuses on inference workloads and edge compute, not the training centers absorbing most institutional capital since 2022. I Squared, a $43 billion global infrastructure manager with positions in energy transition and digital assets, structured this as a standalone platform rather than bolt-on acquisitions to existing data center holdings.
The timing follows twelve months of inference capacity shortages at cloud providers. Anthropic, OpenAI, and Perplexity faced user throttling during peak periods in Q4 2024. Meta's Llama deployments encountered latency spikes on AWS and Azure infrastructure not optimized for sub-100ms response requirements. Inference accounts for roughly 80% of AI compute costs at scale, but hyperscaler architectures prioritize training density over distributed edge responsiveness. I Squared's thesis: enterprises running customer-facing AI need local compute nodes closer to population centers, not another Midwest GPU farm.
The platform model matters for three groups. First, enterprise customers needing compliant, low-latency inference capacity without hyperscaler lock-in now have negotiating leverage. Second, semiconductor designers—Taiwan Semi included—gain a reference customer for inference-optimized chips as the market splits between training ASICs and inference accelerators. TSMC's 8.4% rally this week reflects anticipation that inference buildouts require different node economics than H100 successors. Third, institutional allocators chasing AI infrastructure exposure without direct exposure to Nvidia's margin compression or Microsoft's CapEx treadmill now have a private-market alternative with yield characteristics closer to tower infrastructure than venture bets.
Operators should watch three signals over the next six months. I Squared will announce anchor tenants by mid-Q2 2025—enterprise names signal demand validation, AI labs signal desperation. Site selection for the first 4-6 facilities will reveal whether the thesis prioritizes coastal enterprise density or secondary markets with power availability. Debt financing terms, expected in Q3, will show whether infrastructure lenders treat this as data center paper or specialty compute with technology risk premia.
TSMC's share price moved on inference capacity expectations, not this deal specifically, but the correlation is the market recognizing that inference economics differ from training. I Squared placed $1 billion where the compute actually runs, not where the models train.