TOGETHER AI

The founding team is almost comically credentialed. Percy Liang runs the Center for Research on Foundation Models at Stanford.

Chris Ré is a MacArthur 'genius grant' winner and Stanford professor. Matei Zaharia co-created Apache Spark.

Ce Zhang leads systems research at ETH Zurich. Vipul Ved Prakash previously founded Clearbit.

These are not first-time founders experimenting with a side project — they're the people who literally wrote the research papers everyone else is building on.

The idea crystallized in 2022 as the open-source AI ecosystem was exploding. Models like LLaMA, Mistral, and Falcon were dropping publicly — sometimes matching proprietary models — but running them at production scale was a nightmare.

You needed serious GPU infrastructure, engineering expertise, and ongoing maintenance. Most companies didn't have any of that.

They had a problem to pay rent on a product they could actually build.

Together AI launched with a clear thesis: open-source models are the future, and the bottleneck is infrastructure, not intelligence. If they could make running open-source models as easy as calling the OpenAI API, every developer who wanted control, customization, or lower costs had a reason to switch.

They quietly raised a seed round, built out their inference platform, and by late 2023 were already processing billions of tokens a month.

Together AI is an AI cloud platform. The plain English version: they host open-source AI models on their GPU infrastructure and charge developers per token — every word in, every word out, costs a tiny fraction of a cent.

That's the core business.

Developers and companies that want to run models like Llama 3, Mistral, or Qwen have two options: pay OpenAI and use their proprietary models, or run the open-source equivalents somewhere else. Together AI is that somewhere else.

They handle the GPUs, the serving infrastructure, the latency optimization, and the API — developers just plug in and ship.

Beyond hosted inference, they also offer fine-tuning. Companies that want to train a model on their own proprietary data — to create a version of Llama that knows their codebase or their customer service history — can pay Together AI to handle that compute job.

It's a one-time fee for a job that might cost thousands of dollars in GPU time.

The third leg is dedicated endpoints — enterprise customers who want a private, reserved instance of a model rather than shared infrastructure. That's where the higher-margin recurring revenue lives.

It's the same playbook as AWS: start with commodity usage, move customers up to reserved capacity, keep the stickiness compounding.

The flagship product is the Together Inference API. It's the fastest way to run open-source models in production — you get a REST API that mirrors OpenAI's format almost exactly, which means swapping is often just a URL change and an API key.

They support over 100 models including every major Llama variant, Mistral, Mixtral, Qwen, Falcon, and Code Llama. Latency is competitive, uptime is solid, and pricing is typically 30-70% cheaper than equivalent proprietary options.

Together Fine-tuning lets companies train a custom version of an open-source model on their own data. The workflow is simple: upload your dataset, pick a base model, configure hyperparameters if you care to, hit go.

Together AI handles the multi-GPU training job. You get back a fine-tuned model you can deploy via your own private endpoint.

This is huge for companies that need domain-specific performance — legal, medical, finance — without the cost of training from scratch.

Together Custom Models goes deeper — it's their enterprise offering for companies that want bespoke model development, dedicated infrastructure, and SLAs. Think financial institutions or healthcare companies that need both the performance of a frontier model and the control of on-premise deployment.

They also built Together Embeddings — vector representations of text used for search and retrieval applications — and are expanding into image generation models. The goal is to be the one-stop shop for open-source AI compute, not just the LLM API provider.

The counterintuitive move was positioning against OpenAI not as a competitor but as an alternative infrastructure layer. Instead of building their own proprietary model and competing head-to-head, Together AI made a bet that the best models would eventually be open-source — and that the real value was in making those models accessible.

That timing paid off spectacularly when Meta released Llama 2 in July 2023 and then Llama 3 in 2024. Suddenly, developers had access to genuinely excellent models — but needed somewhere to run them.

Together AI was already there with an API that looked almost identical to OpenAI's. Switching was almost frictionless for anyone who wanted to try.

They also moved aggressively on price. By running optimized inference infrastructure across thousands of GPUs, they could undercut OpenAI's pricing significantly on equivalent open-source models.

For high-volume startups burning $50K a month on API costs, that math matters a lot.

The developer-first go-to-market was pure bottoms-up growth. No enterprise sales team at first — just great documentation, fast APIs, generous free tier, and a product that worked.

Word spread through the same AI developer communities that were already obsessing over open-source models. By the time enterprise customers came knocking, Together AI already had the credibility and the case studies.

The GPU infrastructure business is brutal. Margins depend heavily on how efficiently you're utilizing expensive hardware, and demand for AI compute is lumpy and unpredictable.

When a major model drops — say, Llama 3 — demand spikes overnight and you either have capacity or you lose customers to whoever does. Being perpetually under-provisioned kills growth.

Being over-provisioned kills margins.

The deeper threat is competition from people with unlimited capital. Microsoft, Google, and Amazon all want a piece of the open-source AI infrastructure market.

AWS already offers open-source models through Bedrock. Azure has its own model garden.

Google has Vertex AI. These platforms have existing enterprise relationships, pre-negotiated GPU contracts, and the ability to bundle AI compute with everything else a company already buys.

Together AI has to win on developer experience, latency, price, and breadth of models — and stay ahead on all four simultaneously.

There's also an existential question about where the value accrues in the open-source AI stack. If models keep getting commoditized — better, cheaper, more open — does the infrastructure layer become a race to the bottom?

The counterargument is that Together AI's moat is operational expertise and switching costs, not the models themselves. But that's a thinner moat than owning the model.

The next two years will tell.

Together AI raised a $20 million seed round in 2022 led by Andreessen Horowitz — which is notable because a16z had already backed OpenAI and clearly wanted exposure to the open-source counter-thesis too. Then in November 2023 they raised a $102.5 million Series A at a $1.25 billion valuation, led by Kleiner Perkins.

That round came with an all-star co-investor list: Nvidia, Salesforce Ventures, Emergence Capital, and a handful of strategic angels from the AI research community.

Nvidia's participation is the most interesting signal. When the company that makes the GPUs decides to invest in your GPU utilization platform, they're essentially endorsing your thesis about where the market is going.

It also suggests Together AI gets favorable access to hardware allocation — which in a world of GPU scarcity matters enormously.

The Kleiner Perkins lead brings decades of infrastructure investment experience. They backed Compaq, Sun Microsystems, Amazon, and Google in their respective eras — each time identifying the infrastructure layer that the next wave of computing would run on.

Backing Together AI fits the same pattern: bet on the platform, not the application.

In total, Together AI has raised approximately $228 million. For a two-year-old infrastructure company with no proprietary model and a product that competes directly with big tech giants, that's a remarkable vote of confidence from people who know how these fights usually go.