Together AI's $800 million raise is a bet that customers will choose cheaper AI infrastructure when the output is good enough to ship.
Together AI closed an $800 million Series C on July 1 at an $8.3 billion valuation, and the lead investor is the part that should make you look twice. Aramco Ventures, Saudi Aramco's venture arm, led the round. Vista Equity Partners, General Catalyst, Nvidia and Pegatron joined it.
You don't usually see an oil major's venture arm anchoring an AI infrastructure round. This one did. As Axios also noted in its July 1 venture roundup, Together was one of two companies that day to announce an $800 million raise, which tells you how hot this corner of AI infrastructure still is even after two years of crowded funding headlines.
The bet is blunt. Aramco Ventures manages roughly $7.5 billion, and AI is now its largest portfolio slice. It has already backed inference hardware maker Groq and Qualcomm's edge AI push. Together fits that pattern neatly: it sells access to the pipes that AI compute runs through, not another chatbot with a nicer landing page.
What Together Actually Sells
The San Francisco company was founded in 2022 by Vipul Ved Prakash, Percy Liang and Ce Zhang. It doesn't build its own frontier model. It runs open-weight models including DeepSeek, Nvidia's Nemotron, MiniMax and Kimi at scale, then rents that compute to companies that don't want to pay OpenAI or Anthropic prices for every task.
That pitch now has numbers behind it. Together says annual bookings reached $1.15 billion last quarter. Customers including Cursor, Cognition and Decagon have reported cost cuts ranging from 6x to 60x compared with closed-model pricing. Decagon, which builds AI customer service agents, says it cut its own inference bill sixfold after moving workloads onto Together's platform.
That isn't a rounding error. If you're building AI into a product people actually use all day, inference cost becomes the business model very quickly. A support agent that works at demo scale can still fail as a company if every conversation carries too much model cost. You don't need a grand theory to understand that. You need a bill.
Frankly, this is the clearest sign yet that the open source is catching up argument has moved from conference talk into procurement. Together says usage of open-weight models across the industry roughly tripled over the past twelve months. Every time a developer swaps a GPT class call for DeepSeek or Kimi and the output holds up, that is one more vote against the idea that only a few closed labs can serve production traffic.
The Money Behind The Bet
Together isn't just banking the cash. It plans to grow infrastructure capacity roughly 50x over the next five years, and it says it has secured commitments for more than 500 megawatts of compute capacity. That is the part of the story that sounds least like software and most like heavy industry. AI infrastructure now has land, power, supply chains and manufacturers sitting right in the middle of the pitch.
Nvidia's participation is the tell here. The company that makes the chips is now backing a neocloud that rents them out. Pegatron matters too, because the Taiwanese manufacturer builds the physical servers this industry runs on. When a supplier invests in a customer instead of simply selling to one, it is telling you something about expected demand.
This isn't Together's first big round. It raised $305 million at a $3.3 billion valuation before this financing. Sixteen months later, the valuation has more than doubled. Few infrastructure startups outside the foundation model labs themselves are moving that fast.
None of this guarantees Together wins. Cheaper isn't the same as better. Closed labs like OpenAI and Anthropic still ship models that lead on the hardest reasoning tasks, and enterprise buyers are not going to swap providers just because a benchmark chart looks friendly. Reliability matters. So does security. So does support, once the model is sitting inside a customer's workflow.
But Together doesn't need to win every frontier contest. It needs enough companies to decide that chatbots, coding assistants, support agents and internal tools do not always need the absolute best model on the market. They need something good enough, running reliably, at a price that does not wreck the unit economics.
That is the bet Aramco Ventures backed with $800 million in real capital. Cursor, Cognition, Decagon: all three are making the same bet already, one API call at a time.
Also read: Delaware Weighs a New Legal Entity Built for Companies Run by AI Agents • SBI and Solana Foundation Team Up to Build Japan's Onchain Financial Market • A Shanghai Chip Startup Is Betting 3D Stacking Can Break Nvidia's Grip
