Together AI has become the fastest-growing software startup this summer, according to Brex's analysis of real spending data from thousands of companies, revealing a fundamental shift in how startups build AI products. The company rents access to open-source AI models, and its rise to the top of the list signals something bigger: startups are rapidly moving away from expensive proprietary AI services toward cheaper alternatives, reshaping the entire economics of artificial intelligence.

What Is Together AI Actually Selling?

Together AI doesn't build its own AI models. Instead, it takes freely available open-source models and handles all the complex infrastructure work that makes them usable. An open-source model is essentially a massive collection of numbers, tens to hundreds of gigabytes in size, that does nothing on its own. Turning those numbers into a working product requires specialized graphics processing units (GPUs), software to run the model efficiently, and a system to handle web requests. Together AI packages all of this work and sells it as a service.

This sits at the top of a growing stack of AI infrastructure companies. Below Together AI are serverless platforms that add containers and automatic scaling. Below those are raw GPU cloud providers that rent the machines themselves. What makes Together AI different is that it handles the final layer, serving popular open models behind compatible endpoints so customers can swap between different models by changing a single line of configuration.

Why Are Startups Abandoning Expensive AI APIs?

The math is stark. A Nvidia H100 graphics processor, one of the most powerful chips available for AI work, rents for around $3.40 per hour in 2026, down from roughly $8 in 2023. When batched efficiently, these machines produce AI-generated text from mid-size open models for pennies per million words. By contrast, closed proprietary APIs from major AI labs charge up to $5 per million input words and $25 to $30 per million output words. For common tasks like tagging and routing data, extracting information from documents, and generating images, open models are usually good enough, and the cost difference runs five to 20 times in favor of open models.

This cost advantage is driving rapid adoption. According to Brex data, startups are now adding their first open-compute vendor within five months of their first bill from a major proprietary AI lab, twice as fast as they were in 2024. Many startups report cutting costs by 80% overnight when they move from proprietary APIs to open models hosted on compute clouds.

How to Evaluate Your AI Infrastructure Choices

• Cost Comparison: Calculate your actual token costs with proprietary APIs versus the hourly GPU rental cost for open models at your expected usage volume, accounting for batching efficiency.
• Model Capability Match: Assess whether open models perform adequately for your specific use case, such as document extraction or data tagging, rather than assuming you need the most advanced proprietary model.
• Timeline to Adoption: Plan to evaluate open-compute vendors within your first five months of using proprietary AI services, when cost pressures typically become acute and switching is most feasible.

What Does the Brex Data Actually Measure?

Brex analyzed credit card and bill-pay activity from tens of thousands of companies using its financial platform, focusing on real spending patterns rather than marketing claims. The data is weighted toward recent activity, capturing which vendors are pulling away right now rather than which ones are already large. The rankings exclude publicly traded companies, those valued above $30 billion, and those generating more than $1 billion in annual recurring revenue, so the list reflects emerging trends rather than established giants.

What Brex found is striking: fourteen of the top 25 fastest-growing software vendors sell infrastructure for building AI products, not AI products themselves. Six more are AI products built on those infrastructure stacks. This mirrors a historical pattern. The last time a Brex benchmark list looked this way, the products on top were mobile apps, and the infrastructure was Amazon Web Services (AWS). Today, the fastest-growing line items on startup spending are the foundational components of what Brex calls the "agent economy," and how startups buy that infrastructure reveals what they're actually building.

Why Does This Matter Beyond Just Saving Money?

The shift from proprietary APIs to open models represents a fundamental restructuring of AI economics. When startups can cut costs by 80% by switching infrastructure providers, it changes which AI products become economically viable. It also means that the competitive advantage in AI is shifting from who builds the best model to who builds the best infrastructure for running models efficiently. Together AI's position at the top of the fastest-growing list suggests that this infrastructure layer is where the real growth is happening right now.

The broader implication is that open-source AI models, once seen as inferior alternatives to proprietary systems, are becoming the default choice for startups building real products. This doesn't mean proprietary models will disappear, but it does mean the economics of AI are becoming more competitive and more accessible to smaller companies. The infrastructure companies that make open models easy to use and cost-effective to run are capturing the most growth, and Together AI's rise to the top of the fastest-growing startups list is the clearest signal yet that this shift is accelerating.