Together AI is now processing over 400 trillion inference tokens per month. A year ago, that number was 30 billion. If your mental math is struggling with that jump, here’s the shorthand: it’s a roughly 13,000x increase in twelve months.

The cloud inference platform has quietly become one of the fastest-scaling AI infrastructure companies in the world, propelled by a simple thesis: enterprises want powerful AI models without the eye-watering costs of proprietary APIs.

The numbers behind the surge

Together AI founder Vipul Ved Prakash has described the growth trajectory in terms that would make most SaaS founders weep with envy. Daily token processing climbed from approximately 1 billion tokens to over 1 trillion, a leap of more than 1,000x.

The company reportedly reached an estimated annualized revenue of approximately $1B by early 2026. That’s not a valuation number or a fundraising figure. That’s revenue, the kind of metric that separates companies with actual traction from those running on vibes and venture capital.

Why open-source models are winning the cost war

The economics here are straightforward. Running inference on proprietary frontier models from companies like OpenAI or Anthropic comes with per-token pricing that adds up fast at enterprise scale. Open-source models, by contrast, offer organizations the ability to run comparable workloads at a fraction of the cost, with the added bonus of customization and fine-tuning.

AI platforms globally are now handling trillions of tokens daily, with open models capturing a growing share of the market. Together AI has positioned itself squarely at the center of this shift, offering infrastructure that makes deploying and scaling open-source models as frictionless as possible. The platform appeals to both scrappy startups watching every dollar and large enterprises looking to avoid vendor lock-in with proprietary model providers.

New infrastructure partnerships signal bigger ambitions

On June 3, 2026, Together AI became the first commercial customer for Vector Core Compute’s new inference cloud, a platform built on a hybrid CPU/GPU/RDU architecture designed specifically for high-throughput AI workloads.

The partnership also reflects the broader reality that inference, not training, is becoming the dominant compute workload. Training a large language model is a one-time (or periodic) expense. Inference, the actual serving of that model to users, runs continuously and scales with adoption.

What this means for the AI market and investors

The 400 trillion token milestone is a signal that the AI inference market is entering a new phase, one where scale and cost efficiency matter more than model novelty. The rise of open-source model adoption also threatens the pricing power of proprietary model providers, as enterprises can run open-source models on platforms like Together AI at significantly lower cost.

The risk for Together AI and similar platforms is concentration. If a handful of open-source models dominate (Meta’s Llama family being the obvious example), the inference layer could become commoditized quickly. The Vector Core Compute partnership suggests Together AI is already thinking about this, locking in next-generation hardware advantages before the market gets crowded.