Announcing our $800M Series C to accelerate the shift to open-source AI

Together AI's Series C isn't just a big number — $800 million from a roster that includes NVIDIA, Aramco Ventures, Vista Equity, and Salesforce Ventures — it's a statement about where the real money in AI is heading. Alongside the equity, the company secured commitments for more than 500 megawatts of compute capacity, financed independently by its new backers. That's the kind of figure usually reserved for hyperscalers building data centers, not a four-year-old startup, and it signals just how much Together AI expects inference demand to grow.

The company's pitch centers on a problem a lot of AI-native businesses are quietly wrestling with: the gap between demo-stage AI costs and production-stage AI costs. Firing off a few hundred API calls to a closed frontier model during a pilot feels cheap. Doing that same thing at the scale of an actual product, with agents writing code, answering support tickets, or chewing through documents around the clock, turns into a budget problem fast. Together AI's argument is that open-weight models — DeepSeek, Nemotron, MiniMax, Kimi, GLM among them — have closed enough of the performance gap with closed models that switching now saves real money, not just ideology points. The company claims customers see cost reductions between 6x and 20x, and cites Decagon specifically cutting its inference bill sixfold after migrating.

What's notable is that Together AI isn't just reselling open models with a nicer API. It's been shipping its own infrastructure layer — FlashAttention-4 tuned for NVIDIA's Blackwell chips, a kernel system called Together Megakernel, a compiler tool named together.compile — the kind of low-level plumbing that determines whether an inference request costs a fraction of a cent or several times that. The company says it's now one of the largest producers of AI tokens anywhere, which, whether or not you find that framing a little grandiose, does reflect a real shift: customers like Cursor, Eleven Labs, Suno, and Cognition are routing meaningful production traffic through it rather than treating it as a side experiment.

The bigger backdrop here is a market correction of sorts. For the last couple of years, plenty of companies defaulted to closed frontier APIs because they were the obvious, well-marketed choice. Now that AI features are actual line items in production budgets rather than R&D curiosities, the calculus is changing, and infrastructure providers betting on open models are positioned to benefit if that trend holds. Together AI's founders are framing this raise as early innings, not a victory lap, and given the scale of compute they're lining up, they're clearly expecting the inference bill for the entire industry to keep climbing for a long while yet.