How Does Together AI Make Money? GPU Rentals, Per-Token Inference, and an $8.3B Valuation

Together AI's annual bookings crossed $1.15 billion in Q2 2026. Almost all of it comes from two mechanics: renting out Nvidia GPU clusters by the hour, and charging per token when a customer calls one of its hosted open-source models.

Together AI doesn't sell a subscription or charge per seat. It's a "neocloud" — infrastructure that sits between raw Nvidia chips and the companies that want to train or run AI models without building their own data centers. Every dollar it makes is tied directly to compute consumed, which is exactly why its revenue has scaled as fast as the AI buildout itself.

How does Together AI make money?

Together AI makes money by renting Nvidia GPU clusters by the hour and by charging per token for serverless inference on open-source models — usage-based pricing rather than a flat license fee. Together AI's pricing page lists GPU Clusters at $3.99/hour on-demand for an H100 up to $9.99/hour for a B300, with cheaper reserved and preemptible tiers for longer commitments. A separate Dedicated Inference tier — a single GPU reserved for one customer's model — runs $5.49/hour for an H100 and $8.99/hour for a B200. Serverless inference is billed per million tokens processed instead of per hour, which is the product most comparable to calling the OpenAI or Anthropic APIs directly.

How we counted: GPU-cluster and dedicated-inference hourly rates, per together.ai/pricing, fetched October 27, 2026. Preemptible and longer-term reserved rates run lower; enterprise capacity is quoted directly.

Together AI's serverless inference: pricing by the token

On the per-token side, Together AI's rates vary widely by model size. Smaller or distilled models like Llama 3 8B Instruct Lite run $0.14 per million input and output tokens, while a larger model like Kimi K3 runs $3.00 per million input tokens and $15.00 per million output tokens. A Batch API, for workloads that don't need an instant response, bills at half the real-time rate.

How we counted: serverless per-token rates listed on together.ai/pricing, fetched October 27, 2026; rates change frequently.

How much revenue does Together AI actually have?

Together AI's own disclosed figure is annual bookings, not revenue: the company said its annual bookings crossed $1.15 billion in the quarter before its July 1, 2026 Series C announcement. Bookings can include multi-year contract commitments that haven't yet been recognized as revenue, so it isn't the same number as ARR. Sacra, a research firm that tracks private-company financials, separately estimates Together AI reached roughly $1 billion in annualized revenue by February 2026, up from about $618 million at the end of 2025 — a 375% year-over-year growth rate — with a gross margin of around 45%, consistent with a business that still pays third-party clouds for a meaningful share of its GPU capacity.

No audited revenue figure has been published for either year. Treat the $1.15 billion bookings number as the one Together AI itself has put on the record, and the ~$1 billion ARR estimate as a third-party research estimate that has not been independently confirmed.

Together AI's funding history: from $3.3B to $8.3B

Together AI raised an $800 million Series C on July 1, 2026, at an $8.3 billion post-money valuation, led by Aramco Ventures — Saudi Aramco's venture arm — with Vista Equity Partners, General Catalyst, Nvidia, and others participating. That's 2.5x the $3.3 billion valuation from its roughly $305 million Series B about 16 months earlier, around February 2025. For the full breakdown of that round and who's backing it, see our deep dive on Together AI's $800M Series C.

Who pays Together AI? Customers and the IBM deal

Together AI says it powers more than a million developers, with Cursor, Salesforce, Decagon, Zoom, Datadog, ElevenLabs, Google DeepMind, and Quora listed among the companies running inference on its platform, according to Together AI's own site. Its largest disclosed enterprise commitment is a $240 million, multi-year deal signed with IBM on August 11, 2026 to build a dedicated inference cluster on IBM Cloud, running on Nvidia HGX B300 systems and targeted for availability in Q1 2027. At the time of that announcement, Together AI said its platform was processing roughly 400 trillion tokens a month.

The $5B Humain bet: Together AI's push into Saudi Arabia

On August 31, 2026, Together AI and HUMAIN, Saudi Arabia's state-backed AI company, announced a strategic partnership to build a 250-megawatt data center in the kingdom. The companies said the facility is expected to generate "over $5 billion in gross annualized revenue in its first year" — a projection for year one of a planned buildout, not revenue that has actually been booked or recognized yet, and separate from the $1.15 billion in bookings Together AI had already disclosed two months earlier.

The deal follows Together AI's own pattern elsewhere: lease or build compute capacity with a partner, then sell access to that capacity to enterprise customers through the same per-GPU-hour and per-token mechanics described above, rather than funding data center buildouts entirely off its own balance sheet.

How Together AI compares to Fireworks AI, Baseten, and Modal

Together AI competes with a small group of AI inference infrastructure companies that all use the same usage-based, GPU-time or per-token pricing mechanics rather than seat-based licensing: Fireworks AI, Baseten, and Modal Labs have each raised at multi-billion-dollar valuations within the same stretch of 2026, on the same core bet: enterprises would rather pay a specialized neocloud for inference than build and operate their own GPU fleets, or pay hyperscaler list prices through AWS Bedrock or Google Vertex. Together AI's $1.15 billion in disclosed annual bookings is the largest top-line figure any of the four has put on the record so far, though the companies report different metrics (bookings versus ARR) on different dates, so the comparison is directional rather than apples-to-apples.

Where we could be wrong