How Together AI Bet on Open-Source AI When Everyone Said Closed Was Better

By 2023, the AI industry was polarized: closed-source companies (OpenAI, Anthropic, Google) argued that powerful AI should be restricted for safety. Open-source advocates (Meta with Llama, Mistral) argued for open access. Both sides built models — but neither focused on making open-source models easy to deploy and scale.

Developers who wanted to use open-source AI models faced enormous friction: finding GPUs, configuring infrastructure, optimizing inference, and managing costs. The gap between "model is available" and "model is production-ready" was vast.

Wardley Mapping revealed a structural gap:

• AI models were rapidly commoditizing (Llama, Mistral, Falcon — all free)
• AI infrastructure (GPU access, inference optimization, fine-tuning) was still custom/expensive
• The value was shifting from "having a model" to "running models efficiently"

Together AI positioned at the infrastructure layer where value was accumulating.

Second-order Thinking validated the open-source bet:

• First order: "Closed models are currently more capable"
• Second order: "But open-source models improve faster because thousands of researchers contribute"
• Third order: "As open-source models reach parity, enterprises will prefer them for data sovereignty, customization, and cost control"
• Fourth order: "The infrastructure provider for open-source AI will capture enormous value — like AWS captured value from open-source web technologies"

Reinforcing Feedback Loop: More open-source model users → more demand for inference infrastructure → Together optimizes for these models → better performance/cost → more users choose open-source → more demand for Together.

Together AI's bet on open-source infrastructure paid off:

• Valued at $3B+ — one of the fastest-growing AI infrastructure companies
• Processes billions of inference requests for open-source models monthly
• Became the default platform for running Llama, Mistral, and other open models at scale
• Offers inference costs 50-80% lower than closed-source alternatives for comparable quality
• Research team published influential papers on efficient training and inference

Together AI proved that in platform shifts, the infrastructure layer is often more valuable than any single application — just as AWS proved during the web era.