Tera signs LOIs for 15 MW of low-cost data-center power

United States, June 25, 2026. Tera has signed letters of intent (LOIs) for 15 megawatts of hosting and power across two independent U.S. data-center operators. The capacity is the foundation of Tera's plan to serve open-model inference on hardware it owns, at a structurally lower cost than renting from cloud providers.

Tera's thesis is that the lowest-cost producer of open-source inference wins on price. Open-model tokens are fungible: a Llama or Qwen token is the same regardless of who serves it, so cost of production is the durable advantage. Owning GPUs and running them on low-cost power is how Tera intends to drive that cost down.

A staged path to owned infrastructure

Tera is following a staged approach. It serves first on cloud and rented hardware to validate demand and utilization, then deploys owned GPUs into low-cost sites as volume grows. The 15 MW under LOI reserves the power and hosting for that owned stage.

Both operators run sites in the United States with access to low-cost power. Tera reserved capacity at each under signed LOIs, with phased deployment beginning later in 2026.

Power and hosting are the real constraint in inference economics. Locking in low-cost capacity now means that when we deploy owned hardware, our cost per token is set by physics, not by a rental markup.

Tera serves open models including Llama, Qwen, DeepSeek, Mistral, and Kimi through an OpenAI-compatible API and leading LLM gateways. The company will scale owned capacity as demand on those channels grows.