Artificial Intelligence

Aramco Just Bet 800 Million on Cheaper AI

Together AI raised 800 million dollars at an 8.3 billion valuation, led by Aramco Ventures, to run open-weight models cheaply. What the inference-layer bet means for owners weighing AI costs.

Artificial IntelligenceBy Servola Tech Desk2026-07-054 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Aramco Just Bet 800 Million on Cheaper AI

Key takeaways

Together AI raised 800 million dollars at an 8.3 billion valuation on 1.15 billion in annual bookings, which means the money is now moving to the layer that runs open models cheaply rather than to the closed model itself.
The company claims six to sixty times lower inference cost than comparable closed-model deployments, so the per-token bill you treat as fixed is now a contestable line item.
Aramco Ventures led the round with Nvidia participating, a signal that capital and compute both see the open-weight inference market as durable, not a passing arbitrage.
The raise reprices your build-versus-rent question: running DeepSeek, Kimi, or a MiniMax model on a specialist platform is now a supported path, not a science project.
For a European operator, this widens the field of vendors that sit outside the two largest US closed-model labs, which matters for both cost and sovereignty.

What Together AI actually raised, and why the backers matter

On 1 July 2026 Together AI announced an 800 million dollar Series C at an 8.3 billion dollar post-money valuation, roughly doubling its worth from the 3.3 billion it carried in early 2025. Aramco Ventures led the round through its Prosperity7 program, with Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, March Capital, Pegatron, and SE Ventures joining. Total capital raised now stands at about 1.3 billion dollars.

The number that carries the story is not the valuation. It is the 1.15 billion dollars in annual bookings the company reported for its most recent quarter, serving over one million developers and thousands of enterprises including Cursor, Cognition, and Decagon. This is not a lab burning capital in search of a model. It is a business selling something companies already pay for at scale.

When both a sovereign oil investor and the dominant chipmaker write cheques into the same company, they are not betting on a single model winning. They are betting that the work of running many models cheaply is a permanent, growing market. That is the signal owners should read.

The bet is on the inference layer, not the model

For two years the default enterprise assumption was simple: pick a closed frontier model from a large US lab, call its API, and treat the per-token bill as a cost of doing business. Together AI is funded on the opposite premise. Its platform runs open-weight models such as DeepSeek, Kimi, and MiniMax, wrapped in its own inference-optimisation software, and it claims customers pay between six and sixty times less than comparable closed-model deployments. Decagon reported cutting inference cost roughly sixfold after switching.

The distinction matters because it separates two things owners tend to conflate. The intelligence you rent and the infrastructure that serves it are now different products, sold by different companies, priced on different curves. The model may be free to download. The efficient serving of it, at low latency and predictable cost, is what 800 million dollars just paid for.

Independent tracking backs the shift. OpenRouter data cited alongside the raise shows open-model usage on Together's platform tripling year over year. This is not a marketing claim about future adoption. It is measured traffic moving from closed endpoints to open ones because the arithmetic changed.

What changes for an owner or operator

The practical consequence is that your AI cost structure has become a decision rather than a given. If a meaningful share of your workload is summarisation, extraction, classification, or routing, you are very likely overpaying by running it against a flagship closed model. The task does not need frontier reasoning; it needs a competent open model served efficiently. That is precisely the workload this market is built to absorb.

The counterweight is real and should not be waved away. Hyperscalers are building their own inference capacity at a scale a single startup cannot match, and a specialist platform is one more vendor to govern, secure, and depend on. The honest question is not whether open-model inference is cheaper. It is whether the saving justifies the added supplier and the engineering to route work between tiers. For a business running tens of thousands of euros a month in AI spend, it usually does.

The European and sovereignty read

For an operator in the EU or UK, this raise quietly widens the menu. Every credible alternative to the two largest US closed-model labs improves your negotiating position and reduces the concentration risk of building a business process on a single vendor's pricing and availability. The models Together serves are open weights, which means in principle they can be run inside a jurisdiction and infrastructure you choose, not only through a US-hosted API.

That does not make Together itself a European sovereignty answer; it is a US company, and Aramco capital carries its own considerations. But the pattern it validates, open weights plus an efficient serving layer, is the same architecture Europe's own sovereign-AI efforts are pursuing. The commercial proof that this model makes money is useful even to operators who will never sign with Together. It tells you the option is real, funded, and unlikely to disappear.

Frequently asked questions

Does this mean we should drop our closed-model vendor?

No. The disciplined move is to tier your workloads. Keep a frontier closed model for the tasks that genuinely need its reasoning, and move high-volume, lower-complexity work to a cheaper open-model path. The point of this raise is that the second option is now well-funded and production-grade, not that it replaces the first.

Is 8.3 billion a bubble valuation?

It sits on 1.15 billion in annual bookings and paying customers such as Cursor and Cognition, so it is anchored to revenue rather than promise. That does not remove risk. Hyperscalers building inference at scale are the real competitive threat, not the valuation itself.

Why does Aramco leading the round matter?

It signals that patient, non-US-tech capital views open-model inference as a durable market worth a nine-figure entry, not a short arbitrage. Combined with Nvidia's participation, it tells you both the money and the compute side expect this layer to persist.

The frontier model you rent and the infrastructure that serves it are now two markets, and this raise funds the cheaper one. Treat your per-token bill as a decision you can revisit, not a fixed cost you inherit.

Together AIAramco VenturesNvidiaOpen WeightsInferenceEnterprise AI

Nvidia Funds the Labs That Buy Its Chips. Is the AI Boom Paying Itself?

Nvidia funds the labs that buy its chips. Why the circular financing behind the AI boom is a balance-sheet risk owners and family offices should price now.

3 min read

Energy

Three privately built microreactors reached criticality inside one month

Three privately built microreactors reached criticality under the DOE pilot program before July 4, and one powered an Nvidia chip. What the milestone changes for compute buyers, and what it does not yet prove.

4 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.

2 min read14 views

← All articles