# Together AI raises $800 million at an $8.3 billion valuation as enterprises abandon closed AI models > Source: > Published: 2026-07-30 19:19:25+00:00 *Together AI raised $800 million at an $8.3 billion valuation on July 1, but the real story is the money now chasing the inference layer, not one company's funding victory.* The check that matters most in AI infrastructure this month didn't go to OpenAI or Anthropic. Not even close. It went to Together AI, a four-year-old startup built around a blunt idea: you don't need a closed model for every serious AI product. Reuters reported on July 1 that Together AI raised an $800 million Series C led by Aramco Ventures, with NVIDIA, General Catalyst, Vista Equity Partners, Emergence Capital, Salesforce Ventures, March Capital, Pegatron and SentinelOne's S Ventures among the participants. The round valued the company at $8.3 billion, up from $3.3 billion in February 2025. That's a serious re-rating. It isn't only a valuation story. Together AI said annual bookings crossed $1.15 billion last quarter, and it counts Cursor, Cognition, Decagon, Eleven Labs and Suno among its customers. The company runs cloud infrastructure and APIs for open-weight models including DeepSeek, MiniMax and Kimi, letting companies train and serve AI workloads without paying every bill to a closed frontier lab. That is where you should pay attention. Together AI's own announcement says companies using open models routinely see 6x to 20x lower costs while maintaining equal or better performance. Decagon, the customer service automation company, cut its inference costs sixfold after moving to Together AI, according to the same announcement. Those are company-supplied figures, so you shouldn't treat them like audited economics. But you also shouldn't wave them away. AI products with high token volume live or die on exactly this math. According to TechCrunch, Together AI pointed to OpenRouter research showing that industry usage of open-source models has tripled over the past year. That isn't a philosophy seminar anymore. It's a purchasing decision showing up in production traffic. ## The inference layer is getting priced like infrastructure Together AI isn't alone. Fireworks AI announced on July 15 that it raised a $1.505 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures and TCV. The company said it had passed $1 billion in annualized revenue run rate and was serving more than 40 trillion tokens a day. More than 95% of those tokens, Fireworks said, come from models specialized on customers' own data. Baseten has moved even faster than the older draft of this story allowed. It did raise a $300 million Series E at a $5 billion valuation earlier this year, but that is no longer the latest fact. Baseten announced on June 22 that it had raised $1.5 billion across two tranches at $13 billion and $11 billion valuations, and said revenue had grown about 20x year over year while its platform processed more than 1 billion inference calls a day. Put Together AI, Fireworks and Baseten side by side and you get roughly $39 billion of private valuation around a layer of AI infrastructure that barely had a clean name three years ago. The number can look absurd until you follow the usage. These companies sit where model choice, GPU supply, latency and cost all meet. That is a valuable corner of the stack. Frankly, the old way of talking about open versus closed models is getting tired. Enterprises aren't switching workloads to DeepSeek, Llama, Kimi or custom post-trained models because they want to win an argument on GitHub. They're doing it because the performance gap has narrowed enough for the finance team to ask why the expensive API is still the default. ## What this means if you're building on a closed API If your startup is still running every workload through a proprietary model, the question isn't whether closed models are good. They are. The question is whether you need them everywhere. Some work still belongs with frontier systems. Hard reasoning tasks, anything at the frontier of coding, agents where raw capability is genuinely the product - those can justify the bill. But a lot of production AI is less glamorous than that: support tickets, document extraction, routing, summarization, voice workflows, coding assistant features that need speed more than mystique. Those are the workloads where open-weight models have become dangerous to ignore. Reuters quoted Together AI CEO Vipul Ved Prakash saying, "The future of AI won't be owned by a few companies. It will be built by millions of developers and businesses, and open-source models are making that possible." That line is self-interested, of course. He runs the company that benefits if it comes true. But the capital behind this round says investors believe the same direction of travel. The harder constraint now is compute. Together AI said it has secured commitments for more than 500 megawatts of compute capacity to be capitalized independently by its new investors, and expects its infrastructure footprint to grow roughly 50-fold over the next five years. That tells you what this market really is. The winners won't only have nicer APIs. They'll have dependable GPU supply, tuned software, enterprise controls and enough scale to keep token costs falling. For builders, the practical move is simple. Audit your model spend. Work out which workloads genuinely need frontier capability and which ones open-weight models can already handle. Don't turn loyalty to a closed API into an operating expense you never question. **Also read:** [A federal judge says the government's case for banning Anthropic has gotten worse](https://startupfortune.com/a-federal-judge-says-the-governments-case-for-banning-anthropic-has-gotten-worse/), [ChipAgents raises $134 million as semiconductor giants pay AI to design their own next chips faster](https://startupfortune.com/chipagents-raises-134-million-as-semiconductor-giants-pay-ai-to-design-their-own-next-chips-faster/), [Seedance 2.5 Set to Expand Multimodal Input Support to 50 Media Files](https://startupfortune.com/seedance-25-set-to-expand-multimodal-input-support-to-50-media-files/)