Together AI raises $800 million to democratize open source AI

Together AI raises $800 million to democratize open source AI

🔎 $8.3 billion for a bet that AI giants refuse to make

On July 1, 2026, Together AI announced an $800 million Series C funding round, boosting its valuation to $8.3 billion. That is a 2.5x jump compared to its $3.3 billion valuation in February 2025, when it closed a round led by General Catalyst.

The detail that changes everything: it is Aramco Ventures, the venture arm of Saudi oil giant Saudi Aramco, that is leading the round. Nvidia, Vista Equity Partners, General Catalyst and March Capital are also participating. The total funds raised since the company's creation now reaches $1.3 billion.

According to Reuters, this round is the strongest signal to date that open source infrastructure has become an institutional market. Not a bet by idealistic foundations. A market that oil capital, tech capital, and private capital deem profitable enough to inject nearly a billion in a single transaction.

Together AI's argument is simple and radical: the economics of closed models do not scale. And the numbers seem to prove them right.

The key points

• $800 million raised in Series C, post-money valuation of $8.3 billion (Reuters, July 2026).
• Aramco Ventures lead investor: first massive entry of Middle Eastern capital into open source AI infra.
• Annual bookings > $1.15 billion in the last quarter, with open source model usage having tripled in 12 months (Together AI Blog).
• Compute infrastructure plans for a ~50x multiplication over 5 years (DataCenterDynamics).
• Cumulated total funding: $1.3 billion.

Recommended tools

The numbers that explain everything — Bookings, tripling, and a symbolic threshold

Together AI doesn't survive on grants. The company generates real revenue, and the BusinessWire press release clears up the ambiguity: annual bookings exceeded $1.15 billion in the last quarter.

This isn't recognized revenue, it's bookings — signed contracts. But the order of magnitude speaks for itself. For a company valued at $8.3 billion, a bookings run rate of over a billion yields a reasonable multiple, not a metaverse-like bubble.

The second key figure: the usage of open source models on the platform has tripled in 12 months. TechTimes frames it clearly: open source inference has crossed the billion-dollar threshold while closed models stagnate.

This tripling isn't a base effect. It's a change in corporate behavior. When an organization signs an AI infrastructure contract, it now systematically compares the cost per million tokens between GPT-5.5 from OpenAI and an open weight model like DeepSeek V4 Pro (Max) hosted on Together AI. And increasingly, the latter is winning.

Why closed models don't scale — The economic argument

Together AI's central argument, detailed in its official announcement, is based on a simple mathematical equation. Closed models charge a margin on two layers: the model itself and the compute infrastructure.

OpenAI, Anthropic, Google — they charge for the right to use their proprietary model AND the cost of the infrastructure that runs it. It's a double taxation. On a small scale, it's invisible. At enterprise scale with millions of requests per day, it becomes unsustainable.

Open source models reverse this logic. The model is free (or almost free via permissive licenses). You only pay for compute. And on compute, competition between providers drives prices down.

The result: for a high-volume text classification use case, the cost per million tokens on a model like DeepSeek V4 Pro (High) via Together AI can be 5 to 10 times lower than that of GPT-5.5 via the OpenAI API. Is the quality identical? No. DeepSeek V4 Pro (High) scores 84 on the general benchmark compared to 91 for GPT-5.5. But for 80% of enterprise use cases, the quality difference does not justify a 500% cost gap.

This is the trade-off that companies have been making en masse for the past 12 months.

Aramco Ventures in the lead — Middle Eastern capital enters AI infra

This is the most important geopolitical signal of this funding round. Quartz points out: the round led by Aramco Ventures marks the entry of Middle Eastern capital into AI infrastructure.

Saudi Arabia is no longer just buying GPUs. It is investing in the layer that controls access to models. This is strategically different. Owning compute means having capacity. Financing the infrastructure that distributes open source models globally means having influence.

The Saudi context sheds light on this move. The kingdom has launched massive initiatives around AI, with the goal of becoming a regional AI hub. Investing in Together AI gives it direct access to the tech stack that serves the most powerful open source models, including Moonshot AI's Kimi K2.6 which reaches 88.1 in agentic score in self-host.

Nvidia is also participating in the round, which is not surprising but significant. The chipmaker is thus validating Together AI's business model: the more open source is democratized, the more GPUs are sold. It is a virtuous circle for Nvidia, which has no interest in seeing only the closed models of OpenAI and Anthropic dominate the market.

Neocloud — Together AI is not a traditional cloud

TechCrunch describes Together AI as a "neocloud" — a new generation of cloud providers dedicated exclusively to the hosting and execution of AI models.

The difference from AWS, Azure or GCP is fundamental. Hyperscalers are generalists. They sell storage, compute, databases, networking. AI is just one revenue stream among others. Their optimization is not done at the token level, but at the instance level.

A neocloud like Together AI optimizes every layer for a single objective: running language models as efficiently as possible. This includes hardware optimizations (precise selection of GPU configurations), software optimizations (custom kernels, optimized scheduling) and architectural optimizations (intelligent routing between models based on request complexity).

PYMNTS emphasizes this aspect: Together AI is not selling cloud, it is selling the execution of open source models at a lower cost. The positioning is precise and differentiated.

This specialization enables significantly higher performance per compute dollar. When you deploy a model on AWS, you pay Amazon's general margin. When you deploy it on Together AI, you pay the margin of a company whose entire architecture is designed to minimize the cost per token.

50x of compute in 5 years — What this means concretely

DataCenterDynamics reports that Together AI plans to multiply its compute infrastructure by about 50x over the next 5 years. This is a staggering figure that deserves to be broken down.

50x of compute does not mean 50x more data centers. It means successive generational leaps in hardware (H200 → B200 →下一代), continuous software optimizations, and geographic expansion into new regions where the cost of energy is low.

This is where the partnership with Aramco Ventures makes perfect sense. Saudi Arabia has low-cost energy and massive infrastructure projects. A neocloud that wants to scale 50x needs two things: capital and cheap energy. Aramco potentially brings both.

For users, this expansion will translate into continuously decreasing per-token prices and increased geographic availability. Today, latency is a real barrier to the adoption of open source in Europe. Closer compute regions will partially solve this problem.

The inference war — Open source vs closed models

The battle between open source and closed models has changed in nature. It is no longer a philosophical debate about openness. It is an economic debate about the marginal cost of intelligence.

Let's take a concrete example. An e-commerce company wants to classify 10 million customer queries per day and generate automated responses. With OpenAI's GPT-5.4 Pro (score 91), the monthly inference cost would be significant. With Anthropic's Claude Sonnet 4.6 (score 83), slightly lower but still high.

By switching to DeepSeek V4 Pro (High) (score 84) or Kimi K2.6 (score 84) hosted on Together AI, the cost drops drastically. The quality is comparable to Claude Sonnet 4.6 for a fraction of the price. And for cases where maximum quality is required, the company can route only those queries to a closed model — a hybrid routing strategy that neoclouds make trivial to implement.

EnterpriseDNA notes that Together AI's explicit focus with this funding round is to reduce inference costs for open source models in enterprise. It's the right message at the right time. Companies have understood that AI is not a project, it's an infrastructure. And like any infrastructure, recurring cost is the deciding factor.

Technical optimizations — OSCAR and memory reduction

Together AI's advantage does not rely solely on the volume of compute. It also relies on software innovations that maximize the efficiency of this compute. A recent and significant example: OSCAR, the 2-bit KV cache quantization open-sourced by Together AI which reduces memory by 8.

The KV cache is one of the main bottlenecks in language model inference. The longer the context, the larger the cache grows, and the more the required memory explodes. OSCAR solves this problem by quantizing the cache to 2 bits, making it possible to serve long-context requests on the same hardware that previously could only handle short contexts.

This is the type of innovation that makes the difference between a generic compute provider and a specialized neocloud. AWS is not going to optimize the KV cache of a specific open source model. Together AI is. Because that is its business.

These optimizations compound. Less memory per request means more simultaneous requests per GPU. More simultaneous requests means better hardware amortization. Better amortization means lower prices. The virtuous circle is in place.

AI Sovereignty — Why open source is a geopolitical issue

Beyond the economy, Together AI's funding round touches on a deeper issue: technological sovereignty. When a European or Asian company relies exclusively on the APIs of OpenAI (American) or Anthropic (American), it takes a geopolitical risk.

Open source models offer an alternative. You download the weights, you host them wherever you want, you control them. The Sentient foundation, which raised 42 million dollars for open source AGI, pursues a similar goal: ensuring that cutting-edge artificial intelligence is not the exclusive property of two or three California companies.

Together AI, by making the execution of these open source models affordable and scalable, is the infrastructure of this sovereignty. Without an efficient neocloud, open source remains a researcher's toy. With Together AI, it becomes a viable alternative for companies of all sizes.

This is also why players like Moonshot AI, which raised 2 billion dollars with its Kimi K2.6 model, are betting on open weight. Their strategy: develop the best possible model, publish it in open weight, and let the ecosystem (including Together AI) handle the distribution. The model gains in adoption, the company gains in influence.

Ollama and local — The open source ecosystem is complete

Together AI's cloud infrastructure is only the visible part of the open source ecosystem. In addition, tools like those allowing you to create open source AI agents locally with Ollama offer a completely decentralized alternative.

The spectrum is now complete. You can run a model locally on your machine with Ollama for development and prototyping. Then, when you move to production, deploy the same model on Together AI to benefit from scalability and low latency. Same model family, same weight format, but two execution infrastructures tailored to two different needs.

This is a flexibility that closed models simply cannot offer. You cannot run GPT-5.5 locally. You cannot move it from one provider to another. You are locked in.

Open source breaks this lock-in. And Together AI monetizes this break.

Quick comparison — Relevant open source models on Together AI

This table illustrates the key point: open source models like DeepSeek V4 Pro (Max) at 88 points are now a hair's breadth away from closed models like GPT-5.4 at 89 points. The quality gap no longer systematically justifies the extra cost.

❌ Common mistakes

Mistake 1: Confusing open source with total free-of-charge

Open source does not mean that everything is free. The model weights are free, but the compute to run them has a cost. Together AI precisely monetizes this compute layer. Thinking that open source eliminates all costs means ignoring that infrastructure is the real expense item in AI.

Mistake 2: Believing that open source is inferior in quality

In 2026, DeepSeek V4 Pro (Max) scores 88, just two points behind GPT-5.5 (91). Kimi K2.6 reaches 88.1 in agentic self-host, surpassing GPT-5.4 (87.6) in this configuration. The quality gap between open and closed has virtually disappeared for the majority of use cases.

Mistake 3: Ignoring geographical latency

Deploying on a neocloud without checking the location of data centers is a mistake. If your users are in Europe and the compute is in the United States, latency will degrade the experience. Check the available regions before migrating.

Mistake 4: Underestimating the cost of migration

Migrating from a closed model to an open source model is not just a simple API URL change. Response formats, tokenizers, and system prompt behaviors differ. Plan a budget for post-migration fine-tuning.

❓ Frequently Asked Questions

Who is Aramco Ventures and why is this choice surprising?

Aramco Ventures is the venture capital arm of Saudi Aramco, the Saudi oil giant. This choice is surprising because it is the first major Middle Eastern investment in Western open-source AI infrastructure. It signals a geopolitical diversification strategy beyond oil.

Is Together AI profitable?

Annual bookings exceed $1.15 billion, but bookings are not recognized revenue. The company is likely still loss-making due to massive infrastructure investments, but the revenue trajectory justifies the valuation.

What is the best open-source model to use on Together AI?

For general use, DeepSeek V4 Pro (Max) offers the best quality/price ratio with a score of 88. For self-host agentic tasks, Kimi K2.6 dominates with 88.1. The choice depends on your specific use case.

Does this funding round directly threaten OpenAI and Anthropic?

Not in the short term. OpenAI and Anthropic still dominate the premium segment. But Together AI is targeting volume, meaning the base of the pyramid. If companies massively migrate their "good enough" use cases to open source, closed-model revenue will be squeezed from the bottom.

Can I use Together AI from Europe?

Yes, via API. The relevant question is rather: where are the data centers closest to your users located? Together AI is expanding its infrastructure, but check the regional coverage for your specific use case before committing.

✅ Conclusion

Together AI just raised $800 million not because open source is an ideal, but because its economics crush those of closed models at the enterprise scale. With $1.15 billion in annual bookings, a tripling of open source usage in 12 months, and planned infrastructure growing 50x, the neocloud has passed the point of no return. Aramco Ventures' entry adds a geopolitical dimension that makes this fundraise strategically irreversible. Open source is no longer the cheap alternative — it is becoming the default infrastructure.