Together AI: Betting Against Vendor Lock-In in the Open-Source Boom

Every company building on AI eventually confronts an uncomfortable question: what happens if you build your entire product around one model provider's API, and that provider changes its pricing, its policies, or its priorities? Together AI's entire business exists to make that question less scary — a cloud platform purpose-built for the growing world of open-source models, founded by researchers who understood early that not everyone wants to be locked into a single closed provider.

Academic Roots, Infrastructure Ambitions

Together AI was founded in 2022 by a team drawn from Stanford AI Lab, CMU, and MIT — three of the most influential AI research institutions in the world. That academic pedigree matters for a company whose core product is fundamentally about deep technical execution: running inference, fine-tuning, and full training-from-scratch on dedicated GPU clusters is not a trivial engineering problem, and getting the performance and cost economics right requires exactly the kind of systems expertise that background suggests.

Three Forces Driving Demand

The report identifies three distinct growth drivers behind Together AI's rapid expansion. First, what it describes as a "Cambrian explosion" of open-source models — every new release from Meta's Llama family, Mistral, DeepSeek, or Qwen creates a fresh wave of demand from developers who want to deploy and fine-tune that specific model immediately, rather than waiting. Second, the broader rise of agentic applications, which need inference that's both fast and cheap enough to support the rapid, repeated model calls an autonomous agent makes during a single task. Third, and perhaps most strategically important, deliberate anti-positioning against OpenAI: Together AI doesn't use client data to train its own or other models, and in its Zero Data Retention mode, doesn't store prompts or responses at all.

That third point is a genuine differentiator for enterprise and research clients handling sensitive data — a guarantee that your prompts aren't quietly becoming someone else's training data is a meaningfully different value proposition from providers with murkier data usage policies.

Clients Doing Things That Wouldn't Work Elsewhere

Together AI's customer list includes recognizable enterprise names like Salesforce, Zoom, and Pika Labs, but the more technically interesting case is the Arc Institute, a genomics research organization that used the platform to train its Evo genomic model — a project the report notes would have been unachievable on the Institute's own infrastructure. That's the clearest possible demonstration of Together AI's core value proposition: giving research and engineering teams access to training-grade infrastructure they couldn't otherwise build or afford themselves.

Expanding the Stack Through Acquisition

In May 2025, Together AI acquired Refuel.ai, adding a data-preparation layer to its existing infrastructure stack. That acquisition reflects a natural expansion pattern for infrastructure companies — once you own the compute layer, moving up the stack into data preparation, which sits directly upstream of training and fine-tuning, is a logical way to capture more of a customer's total AI workflow rather than just the GPU-cycles portion of it.

Growth in Numbers

Together AI's valuation grew 2.6x in eleven months, from $1.25 billion in March 2024 to $3.3 billion in February 2025.

The Bigger Bet: Open Source Keeps Winning Share

Together AI's entire growth thesis depends on one continuing trend: that open-source models keep improving fast enough, and closely enough behind closed frontier models, that developers keep choosing flexibility and cost control over the raw capability ceiling of a closed provider. Every open-source release in 2025 that narrowed the gap with GPT-class models has effectively been free marketing for Together AI's entire business model — and there's no obvious sign of that open-source momentum slowing down.

FAQ

Source: CrackTheDeck Research.