Aramco's AI Bet: Together AI Raises $800M as Open-Source Inference Crosses $1B

A cloud company's raise, and the thesis underneath it

On July 1, Together AI said it had closed an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures — the venture arm of Saudi oil giant Saudi Aramco — with participation from Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, March Capital, Pegatron and SentinelOne's S Ventures. The raise brings the San Francisco company's total funding to roughly $1.3 billion, according to reports from Business Wire, The Next Web and Yahoo Finance.

On its own, another nine-figure AI round in mid-2026 barely registers. What makes this one worth reading is the company doing the raising. Together AI has never built a proprietary frontier model. It is a cloud platform for running other people's open-weight models — and it just got an eight-billion-dollar vote of confidence, from an oil-money sovereign investor, on the bet that open beats closed on the one axis enterprises actually feel: cost.

What Together AI actually sells

Together AI, co-founded by CEO Vipul Ved Prakash alongside Stanford's Percy Liang and researcher Ce Zhang, sits one layer below the model. Its business is hosting and optimizing open-weight models — the sources name DeepSeek, Nemotron, MiniMax and Kimi — on a unified GPU cloud, bundled with the company's own inference-optimization software.

The pitch is arithmetic. Together claims customers can run comparable workloads at "between six and 60 times lower inference costs than comparable closed-model deployments," per the Business Wire announcement; Yahoo Finance frames the same claim more conservatively as one-fifth to one-seventh of closed-model pricing. Either way, the direction is consistent: take a good-enough open model, host it efficiently, and undercut the metered APIs from OpenAI and Anthropic.

The reported traction is the part investors are presumably underwriting. Together says it has more than a million developers on the platform and annual bookings exceeding $1.15 billion — a bookings figure, note, not recognized revenue, and worth reading with that caveat. Named customers include Cursor, Cognition (the maker of Devin) and the AI-support startup Decagon. Prakash's framing, quoted by Yahoo Finance, leans into the utility metaphor: "Intelligence is becoming a foundational resource for the modern economy, every bit as essential as electricity, bandwidth or capital."

The valuation jump, in context

The headline number climbed fast. Together's February 2025 Series B was a $305 million round at roughly a $3.3 billion valuation; this Series C more than doubles that to $8.3 billion in about seventeen months. Sources differ slightly on the prior mark — some cite ~$3.3 billion, others imply ~$4 billion — but the trajectory is not in dispute.

That kind of step-up is unremarkable by 2025–26 AI standards, and the surrounding comparables show how crowded the "inference layer" has become. Tech Funding News notes Fireworks AI raised $250 million at a $4 billion valuation in October 2025, and that Baseten raised $1.5 billion at a $13 billion valuation the previous month. The infrastructure-and-serving tier is now attracting capital at valuations that would have looked like frontier-lab money a year ago.

Why Aramco leading it matters

The lead investor is the story within the story. Aramco Ventures — associated with the Prosperity7 vehicle that co-led Together's earlier round — is Gulf sovereign capital, and its presence at the top of the cap table continues a now-familiar pattern: petro-state wealth rotating into AI compute as a strategic asset class. Reading it as pure financial appetite would miss the point. For a state whose balance sheet is tied to hydrocarbons, owning a stake in the "electricity of intelligence" is a hedge and a positioning move as much as an investment.

Nvidia's participation cuts the other way and is just as telling. When your GPU supplier is also on your cap table, the alignment is obvious — Nvidia has an interest in a healthy ecosystem of neocloud buyers that aren't the three or four hyperscalers. The investor list, in other words, encodes two bets at once: sovereign diversification and silicon-vendor ecosystem-building.

The plan, and the reasons for caution

Together says the money funds a roughly 50-fold expansion of its compute infrastructure over five years, plus continued work on its inference engine and new products. That is an enormous build-out commitment, and it points straight at the strategic risk.

Together's moat is optimization software layered on rented GPUs — Tech Funding News cites claims of up to 80% lower operating costs from that stack. But the customers it wants to undercut also have suppliers with effectively bottomless balance sheets. The same sources flag the obvious pressure: hyperscalers deploying hundreds of billions in annual capex can build competing inference capacity at a scale a startup cannot match, and can afford to price aggressively to keep workloads on their own clouds. An independent neocloud is a real business today; whether it stays independent, or becomes an acquisition target, is the open question the $8.3 billion doesn't answer.

There's also a demand-side caveat worth stating plainly. Open-model usage on Together's platform reportedly tripled year-over-year, which is a strong signal — but "bookings" and platform-usage growth are not the same as durable, high-margin revenue, and the cost claims (6x to 60x) are the company's own, spanning a range wide enough to demand a benchmark before anyone treats it as settled fact.

Hype versus signal

Strip away the round size and what remains is a genuine data point about where enterprise AI spend is heading. The bull case for open weights was always that most production workloads don't need the single best model — they need a good-enough model at a predictable, lower price, without vendor lock-in. Together crossing a billion dollars in annual bookings, with names like Cursor and Cognition attached, is the most concrete evidence yet that the argument is translating into contracts, not just conference talks.

The hype to discount is the framing of this as an "open beats closed" verdict. It isn't one. The frontier labs still set the capability ceiling that open models chase, and much of Together's inventory — DeepSeek, MiniMax, Kimi — reflects how much of the open-weight frontier now originates outside the U.S., a supply-chain fact with its own policy implications. What the round confirms is narrower and more useful: there is a large, real market for serving open models cheaply, and serious money now believes that market is worth $8 billion.

The takeaway

Together AI's $800 million Series C is less a bet on a company than on a thesis — that inference, not training, is where enterprise AI money will concentrate, and that open weights served at a fraction of closed-API prices are how most of it gets spent. Aramco leading and Nvidia participating tell you who wants that future to arrive. The reasons for caution — self-reported cost multiples, bookings-not-revenue metrics, and hyperscalers who can outspend any neocloud — tell you why it isn't guaranteed. Watch the next two signals: whether Together's 50x compute build-out actually materializes, and whether its cost advantage survives contact with clouds that can afford to give inference away.