Together AI raises $800M Series B as open-source models hit enterprise-grade performance

Together AI raises $800M Series B

Together AI has grown its bookings run rate from $100M to $1.2B in a year and closed an $800M Series B, with Aramco Ventures named as a notable strategic investor. Vipul Ved Prakash, co-founder and CEO, attributes the growth to two converging forces: open-weights models have closed the performance gap with closed models, and enterprise demand for AI tokens has surged as companies move from prototypes to production-scale deployments.

“Our business, which really serves open source models and powerful AI at dramatically cheaper cost, has been growing tremendously. We've grown from $100M in bookings to $1.2B in the year, and we raised $800M in our Series B. Open weights models are now able to really address the largest workloads of agentic long range, long horizon tasks that have been deployed in enterprises over the last six months.”

Open source catching up

Prakash argues open-weights models can now handle the long-horizon agentic workloads that enterprises have been deploying over the past six months — tasks that previously required frontier closed models. The gap, he says, has been narrowing on both first and second derivatives. The practical consequence is substitution: companies that prototyped on closed models are hitting cost walls at production scale and finding open models increasingly viable as drop-in replacements.

A second driver is data sovereignty. As models get smarter, more companies are becoming uncomfortable feeding proprietary workflows into closed APIs, on the logic that doing so hands a strategic asset to the model provider. Prakash expects this concern to accelerate the shift toward open-weights deployments.

China's open-source stance

On why Chinese labs release openly while US labs mostly don't, Prakash's read is structural rather than ideological. Chinese AI companies are already generating revenue through APIs and applications, so releasing models openly is just how that market competes. The US market formed around closed models for different reasons. He expects more US companies to release open models as the token economy matures and tooling becomes modular enough to route workloads to either open or closed models interchangeably.

Distillation and the regulatory question

Asked whether US open-source labs are disadvantaged by restrictions on distillation from closed frontier models, Prakash thinks the role of distillation is overstated. Recent gains in open models have come primarily from reinforcement learning and the standardization of the transformer stack — shared architecture, training tools, and benchmarking infrastructure that any lab can now build on. That commoditization of the training machinery, he argues, means labs starting today will outperform those that attempted the same thing two years ago, regardless of distillation access.

On safety regulation, Prakash acknowledges the risks may be real but notes quantification is lacking. His case for open models in a regulatory context is that they can be inspected and benchmarked in ways closed models cannot — making them, in principle, more amenable to rigorous oversight than the alternative.

$1.2B bookings run rate on a business that was at $100M a year ago is the number that anchors everything else here.