Together AI just landed $800 million to prove that the cheapest way to run AI isn't the most obvious one.
The San Francisco company said this week it closed an $800 million Series C at an $8.3 billion post-money valuation, in a round led by Saudi Arabia's Aramco Ventures, with Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, March Capital, Pegatron and S Ventures also writing checks. That figure landed well above the $7.5 billion valuation that Investing.com had reported was under discussion just weeks earlier, and it comes less than eighteen months after Together AI's $305 million Series B closed in February 2025. Across four rounds since its 2023 seed, the company has now raised roughly $1.3 billion.
You don't need to squint to see why an oil-money sovereign fund and the world's most valuable chipmaker both wanted in. Together AI doesn't build its own frontier model. It runs everyone else's, at scale, on infrastructure it controls, for customers who have grown tired of paying frontier-lab prices for tasks that don't need a frontier-lab model.
That bet is paying off in the numbers. Together AI's annual bookings crossed $1.15 billion last quarter, the company said, while open source model usage across the industry has tripled over the past twelve months. Customers including Cursor, Cognition and Decagon are running production workloads on models like DeepSeek, Nemotron, MiniMax and Kimi instead of closed alternatives, and reporting cost savings ranging from 6x to 60x for equal or better performance. Decagon, the customer support AI company, has said it cut its own inference bill sixfold after making the switch.
For most of the last three years, the working assumption in Silicon Valley was that OpenAI and Anthropic would keep enough of a quality lead that enterprises would pay whatever it cost to stay on their APIs. That assumption is cracking. Chinese labs, DeepSeek chief among them, have shipped open-weight models that match or beat closed competitors on plenty of benchmarks, and Together AI's entire business exists to let companies run those models without operating their own GPU clusters or negotiating their own chip supply. Vipul Ved Prakash, who co-founded Together AI in 2022 after building the anti-spam system Vipul's Razor and the social analytics firm Topsy before Apple acquired it, has pitched the company's mission as making frontier-grade AI economically accessible rather than exclusive to whoever can absorb the biggest API bill. The new funding backs that pitch directly. Together AI says it plans to grow its infrastructure footprint roughly 50-fold over the next five years, a scale-up that only makes sense if demand for open models keeps compounding rather than leveling off.
Frankly, the investor list tells its own story. Aramco Ventures leading a Series C for an AI infrastructure company has nothing to do with oil and everything to do with a sovereign wealth arm betting that compute capacity, not any single model, is where the durable value sits. Nvidia's participation is more direct. Every workload Together AI runs, no matter which open model sits on top of it, still needs Nvidia chips underneath.
None of this means the closed labs are in trouble. OpenAI and Anthropic still set the pace on the hardest reasoning and coding tasks, and plenty of enterprises will keep paying a premium for that ceiling. But Together AI's bookings growth points to a real migration underway for the much larger share of enterprise AI workloads, chatbots, support agents, coding assistants, that don't need the absolute frontier, just something close to it at a fraction of the price.
That's the wedge Cursor and Cognition already found when they moved workloads onto Together's infrastructure. It's the same wedge Together AI just raised $800 million to widen.
Also read: Venice AI became a unicorn by promising to forget everything you tell it • Cisco is betting its whole workforce on AI agents while it cuts jobs • Meta Is Quietly Building a Business to Sell Its Leftover AI Computing Power