Together AI raised $800 million on July 1, 2026, at an $8.3 billion valuation — led by Aramco Ventures, with Nvidia, Vista Equity Partners, General Catalyst, Emergence, March Capital, Pegatron, and SentinelOne's S Ventures in the round. Sixteen months ago the company was worth $3.3 billion. The headline is the 2.5x.
The headline is not the interesting number.
The interesting number is $1.15 billion — Together's reported annual bookings — because of what Together is. It is a neocloud: it rents Nvidia GPUs and serves other people's models. DeepSeek. Nemotron. MiniMax. Kimi. Two hundred-plus sets of open weights it did not train and does not own. A company that ships none of the frontier models just crossed a billion dollars a year serving them.
For two years the consensus story was that value in AI concentrates in the labs — that whoever trains the best model captures the economics, and everyone downstream rents at a markup. Together's raise is a crack in that story, and the crack is agent-shaped.
Agents don't buy answers, they buy tokens by the ton#
Here is the mechanism nobody priced in.
A chatbot makes one call per user turn. The user asks, the model answers, the interaction is a single premium transaction and it makes sense to pay frontier prices for the best single answer. That was the workload the closed-model pricing was built for: one shot, high stakes, willingness to pay.
An agent does not work that way. An agent runs a loop. It plans, selects a tool, reads the result, reflects, re-plans, calls another tool, checks its work, summarizes, retries. A single user request can fan out into dozens of model calls, most of them small and none of them glamorous — classify this, extract that, is this done, which tool next. The economics invert. When you make one call, you pay for quality. When you make forty, you pay for the forty.
The unit an agent consumes stopped being the answer and became the token — and tokens are a commodity, priced at the margin.
And on the sub-tasks that make up the bulk of that loop, open weights are no longer the compromise they were in 2024. Routing, extraction, tool selection, short summarization — a well-served open model clears the bar, at a third to half the price per token. (The Groq vs Together vs Fireworks question — who serves those open weights fastest and cheapest — is exactly the fight this raise is funding.) Multiply a modest per-call saving by the call count of an agentic workload and the frontier premium stops looking like a quality tax and starts looking like pure margin you're handing someone else.
Open-model usage roughly tripled over twelve months on the public gateways — a figure Together cites from OpenRouter. That curve is not hobbyists discovering DeepSeek. That is production agents doing arithmetic.
The margin is moving from the model to the meter#
Once a good is a commodity, the money stops accruing to the brand on the label and starts accruing to whoever produces it cheapest at scale while staying neutral across suppliers. Nobody pays a premium for which barrel of crude; they pay for delivery, refining, and reliability. Inference is arriving at the same place. The agent builder does not care whether the routing call ran on DeepSeek or Nemotron — they care that it was fast, cheap, and up. That indifference is the whole thesis. Indifference to the model is a margin transfer to the cloud.
This is why the neutral neocloud is a structurally better place to stand than it looks. A frontier lab has to be right about the next model, every generation, forever, or the premium evaporates. A cloud that serves all the open weights doesn't have to pick the winner — it just has to be the cheapest neutral place to run whichever one wins this quarter. It monetizes the category, not a bet inside it. When the models are substitutes, being the switch beats being one of the things it switches between.
Read the lead investor#
The tell in this round is who led it. Not a software fund chasing the next framework — Aramco Ventures, the venture arm of an oil major. Capital that spent a century learning that the durable money in an extractive business is rarely in the thing everyone photographs. It's in the pipeline, the refinery, the meter on the way out. Building GPU plants near cheap power and metering tokens to everyone is a bet those investors understand in their sleep, and it's a different bet than "our model will be smartest in 2027."
None of this means the labs lose. The hardest single-shot reasoning still belongs to the best closed models, and the frontier is still where the science happens. But the frontier is not where the volume is going. The volume is going into loops — into agents that burn tokens by the thousand on unglamorous steps — and volume is exactly the thing that commoditizes. Together didn't raise $800 million because it built a better brain. It raised it because agents turned inference into a commodity, and someone always gets rich selling the commodity.
The question every AI company should be asking after this round isn't "can we build a better model." It's the older, colder one: are we the toll road, or the car?
