Together AI raised $106 million as investors backed a commercial platform built around openly available and custom artificial-intelligence models. The transaction illustrates an emerging business model: distribute model choices widely, then charge for scarce compute, optimisation, deployment and enterprise support. Openness may reduce lock-in, but it does not remove infrastructure cost, licensing questions or operational responsibility.

Open models attract another large investment

In its 22 March 2024 newspaper edition, The Wall Street Journal reported on open-model startups challenging proprietary AI providers. Belle Lin described companies using free model distribution to attract developers while selling computing power, subscriptions, support and specialised tools.

Together AI, based in the United States, was one example. Rather than relying on payment for access to a single model, it offers infrastructure for training, fine-tuning and inference across open and custom models. That positions the company between developers and the expensive hardware needed to operate generative AI at scale.

The $106 million round funds the platform

Together AI announced the $106 million financing on 13 March 2024. Salesforce Ventures led the round, with new and existing investors participating. The company said it would expand enterprise features and add international compute capacity.

Together reported more than 45,000 registered developers and traffic growing threefold month over month. Those are company-supplied adoption indicators rather than audited revenue figures. Registered users do not necessarily become paying, retained customers, but developer participation can create demand for hosted inference and deployment services.

The commercial stack contains several paid layers

  • Access to scarce accelerator capacity for training and inference.
  • Optimisation that improves throughput and hardware utilisation.
  • Fine-tuning and model-shaping tools for specialised tasks.
  • Secure deployment, monitoring and integration for enterprises.
  • Technical support and predictable service levels in production.

Free distribution is not free operation

An organisation may download model weights without paying a proprietary API fee and still face major costs. It needs suitable hardware, engineering, data preparation, evaluation, security, monitoring and energy. Larger models can demand specialised accelerators and distributed systems that few buyers want to build alone.

This creates room for a platform business. The model can remain openly available while a customer pays for convenient, reliable operation. The analogy is familiar from open software: access to code is separate from the economics of hosting, maintenance, integration and accountable support.

A text-free physical chart shows wide developer adoption feeding five coloured revenue tracks for compute and services while large GPU blocks and dark cost stacks remain
Model access can be widely shared even when training, inference, optimisation and reliable production support remain paid services.

The meaning of open remains contested

Traditional open-source licences were designed for software code. AI systems add model weights, training data, architecture, evaluation and usage restrictions. A model described as open may expose some components while withholding others or limiting commercial use. Buyers must therefore examine the actual licence rather than rely on a category label.

Control over deployment can still be valuable. A company may want to fine-tune a model on private information, run it inside a chosen environment or move workloads among providers. Yet portability depends on compatible tools, data rights and operational skill. Open weights reduce one form of dependency without eliminating every switching cost.

Capital must become efficient usage

The investment lets Together AI buy capacity, improve its systems and pursue enterprise customers. It also creates pressure to turn developer interest into durable revenue while competing with hyperscale clouds, specialist AI infrastructure companies and proprietary model vendors. Infrastructure expansion ahead of demand can be costly.

Customers should compare price per useful result, latency, model quality, data controls, availability and the effort required to migrate. Together AI's round shows that investors see a business around open models. The lasting test is whether the platform can make varied models reliably easier and cheaper to operate than customers could achieve elsewhere.