Together AI signs 250 MW HUMAIN capacity deal for open models

• The deal combines a planned 250 MW Saudi facility with a pledge of more than 100,000 chips of capacity over the next 12 months, as the official announcement and vipulved's capacity claim describe it.
• More than $5 billion is an expected first-year gross annualized revenue figure, a qualification abbreviated in togethercompute's announcement as "$5B+ in annualized revenue."
• Together positions the capacity for open-model inference and training, while togethercompute's announcement frames the deal as open source gaining infrastructure comparable to closed-model labs.
• The public hardware description reaches "multi-chip" infrastructure and a 100,000-plus-chip claim in vipulved's post, without naming an accelerator generation, vendor, or network design.

A RuntimeWire report puts the plan at 120,000 chips, a more precise figure than the company's 100,000-plus claim. The official release calls the project new data-center construction, says workloads will run from Saudi facilities, and promises low-latency access to Europe, the Middle East, and Africa.

The formal announcement says the partnership includes construction of a new 250 MW data center in Saudi Arabia. Together's social post describes capacity offered over the next 12 months, so the public wording does not specify which portion comes from a completed build versus the broader capacity plan.

The 100,000-plus number appears in vipulved's post, while RuntimeWire reports a 120,000-chip path. Neither the company release nor the posts identify the exact chip count by accelerator type.

The published material does not identify:

• accelerator supplier or generation;
• GPU, CPU, and networking mix;
• rack topology, interconnect, or storage design;
• power-usage effectiveness or usable IT load;
• construction and service-availability milestones.

The official announcement divides the partnership into physical infrastructure from HUMAIN and cloud-platform distribution from Together AI:

• HUMAIN contributes large-scale, multi-chip infrastructure and capital deployment.
• Together contributes its global customer ecosystem and AI-native inference platform.
• The stated service covers high-performance infrastructure for inference and training workloads in Saudi data centers.

Together's public framing is explicitly about open models. Closed-model companies had hyperscalers bankrolling their compute for years, togethercompute's follow-up argues, while this agreement supplies a capacity route without saying which open-model families, API products, training configurations, or access terms will be offered.

Together and HUMAIN project more than $5 billion in gross annualized revenue during the first year. Against the announced 250 MW ceiling, the two headline thresholds imply more than $20 million of annualized gross revenue per stated megawatt.

That number is a projected gross run rate, not a stated facility budget or signed revenue total. A Magica analysis makes the same distinction from the published wording; the announcement does not disclose pricing, customer commitments, a revenue split, or construction financing.

The release identifies Saudi power capacity as an advantage and says the location provides low-latency access to Europe, the Middle East, and Africa. Together calls power the immediate AI bottleneck in togethercompute's power post.

HUMAIN describes its wider stack as data centers, AI infrastructure, cloud platforms, foundation models, and AI applications. Together frames this particular arrangement as a partnership that unlocks compute rather than requiring every company to raise billions itself in togethercompute's partnership post.