Together AI Raises $800 Million at $8.3 Billion Valuation as Open-Source Inference Becomes the Agent Cost Escape Hatch

Together AI has closed an $800 million Series C round at an $8.3 billion valuation, according to the New York Times Dealbook and Reuters. The company, which hosts and optimizes open-source AI models for enterprise customers, is positioning itself as the cost-efficient alternative to per-token pricing from OpenAI and Anthropic as companies scale agent deployments.

The Business Model Bet

The core proposition is straightforward. Closed-model providers charge per token: every query, every reasoning step, every tool call an agent makes hits the meter. At chatbot scale, this is manageable. At agent scale, where a single autonomous workflow can generate thousands of inference calls, per-token costs compound into budgets that force companies to reconsider their architecture.

Together AI offers a different model: open-source models hosted on managed infrastructure with compute-based pricing. Companies pay for capacity rather than consumption, making costs predictable regardless of how many calls their agents make.

Ashwin Sreenivas, CEO of customer service automation company Decagon AI, told the Times that Together AI saved a “massive” amount compared to OpenAI pricing. Decagon runs a mix of closed and open models, using OpenAI for high-stakes reasoning and Together for high-volume, lower-complexity tasks.

The Infrastructure Category

Together AI’s raise is the latest in a wave of infrastructure-layer funding. Groq closed $647 million to expand its inference cloud across 13 data centers. Baseten was valued at $13 billion in its latest raise. Cerebras hit a $50 billion market cap after its May IPO. Each company attacks the same constraint from a different angle: reducing the cost of running AI at production scale.

Together AI CEO Vipul Shukla plans to spend most of the new capital on R&D and compute buildout, according to the Times. The company sees a path to IPO following the Cerebras template: enterprise demand first, public listing second.

The Agent Angle

The timing of Together AI’s raise aligns with a visible shift in how companies evaluate inference costs. Palantir CEO Alex Karp publicly argued this week that per-token pricing forces companies toward open-weight models and efficiency-first architecture. OpenAI’s own engineering team just demonstrated that software optimization can cut inference costs by 50%, validating the thesis that the industry’s GPU-intensive approach has room for dramatic efficiency gains.

For agent builders, the math is simple. An agent making 1,000 inference calls per task at $0.01 per call costs $10 per task. The same agent running on open-source models through Together AI at compute-based pricing can reduce that to a fraction. At enterprise scale, across hundreds of agents running continuously, the difference between per-token and compute-based pricing is the difference between a viable product and an unsustainable burn rate.

The $8.3 billion valuation reflects investor conviction that this cost differential will drive a structural shift in how companies source inference for autonomous systems.