Top 25 fastest-growing software startups of summer 2026, according to Brex spend data

Fourteen of the 25 fastest-growing software vendors on the Brex Benchmark report this summer — more than half the list — sell infrastructure for building artificial intelligence products, not AI products themselves. Six more are AI products built on those stacks.

Behind No. 1 — Together AI, which rents access to open models — sit the databases, graphics processing units (GPUs), caches, and sandboxes that agent products get built on. The last time a Benchmark list looked like this, the products on top were mobile apps, and the infrastructure was Amazon Web Services (AWS). The fastest-growing line items on startup card statements are the plumbing of the agent economy, and how startups buy that plumbing shows what they’re building.

The summer Benchmark goes deep on the two categories driving the list, AI compute and databases, whose purchase patterns answer questions nobody outside the vendors can measure. When do startups start running open models? Does open compute replace Big Lab bills or grow alongside them? And what happens in the 90 days after a founder prompts an app into existence?

These rankings capture growth acceleration and are the companies pulling away right now, not ones already large. Vendors that are publicly traded, valued above $30 billion, or past $1 billion in annual recurring revenue (ARR) graduate off the list, along with their subsidiaries. The data is real credit card and bill pay activity from tens of thousands of the finance platform’s customers, recency-weighted.

The highlighted vendors in the ranking below marks all 14 startups as the building blocks for shipping AI products.

Together AI sells access to open-source models. The model is freely available but everything required to turn those files into a working product is not, and a whole category of vendors now sells that work preassembled.

An open-weight model is free to download: tens to hundreds of gigabytes of numbers that do nothing on their own. Turning them into a product takes GPUs big enough to hold them, an inference engine like vLLM, and a serving layer that batches web requests into efficient GPU work. A whole category now sells that work preassembled. GPU clouds rent the raw machines. Serverless platforms like Modal and Runpod add containers, auto scaling, and per-second billing. Then at the top, Together AI, DeepInfra, Baseten, and Fireworks serve popular open models behind compatible end points. Change one line of configuration, and the model becomes DeepSeek or Kimi, at whatever price the hosts are currently competing down to. The newest rung, Thinking Machines Lab’s Tinker application programming interface (API), sells managed fine-tuning, where you shape an open model to your data without touching a GPU.

A Nvidia H100 rents for around $3.40 an hour, down from roughly $8 in 2023, and batched well it produces tokens from midsize open models for pennies per million. Closed frontier APIs list at up to $5 per million input tokens and $25 to $30 per million output. For the workloads that dominate a token bill (tagging and routing data, extracting structure from documents, speech-to-text, image and video generation), an open model is usually good enough, and the gap runs five to 20 times over. A whole merchant category grew up to arbitrage GPU-hours against token prices. Scope note: Card and bill-pay data sees the self-serve layer, not invoiced enterprise GPU contracts. That’s a census of what startups do, and startups are where new stacks get chosen.

After building a product-market fit (PMF) AI app with a Big Lab API, the costs can balloon as usage scales. Moving to an open source model (OSM) hosted in the cloud can cut costs 80% overnight. Companies are adding their first open-compute vendor within five months for their first Big Lab API bill, twice as fast as in 2024, according to Brex data.

Measured within each company’s first 12 months on Brex, Supabase adoption grew by roughly 17 times between the 2022 and 2025 startup cohorts; nearly 1 in 10 companies that joined in 2025 paid for it in their first year. Vendor telemetry explains the new default: Neon, a Postgres company (part of Databricks, ineligible here), reports that AI agents, not people, create more than 80% of its new databases, and Supabase independently reports more than 60% launched by AI tools. The majority customer of a database company is now software.

A classic cloud database couples compute and storage on one machine: minutes to provision, billed while idle, nerve-racking to copy. The new generation splits it into disposable compute over durable storage built on a replicated log. A new database becomes a seconds-long metadata operation, an idle one suspends to near zero, and a copy is a pointer, not a transfer.

None of it was designed for agents; it fits them perfectly. Agents work at machine speed, so a minutes-long provision is a blocking failure. Agents are prolific: A coding agent or app builder like Lovable, Bolt, or Replit provisions a database for each app it generates, thousands a day, most of which sit idle. Near-zero idle cost makes that pattern economically boring instead of absurd, and it’s already the norm: Replit’s agent runs its backend on Neon, and Retool manages more than 300,000 Postgres instances there with a single engineer. And because the training corpus is saturated with Postgres, large language models (LLMs) speak it fluently, turning the category’s decade-old Postgres bet into an advantage nobody planned.

Brex data shows what happens when this machinery meets startups:

Entry-tier bills are flat or falling: The median is tens of dollars a month (Supabase’s Pro plan is $25; Neon meters usage with no minimum), and Railway’s median bill halved even as its paying customers grew by nearly eight times. The revenue engine is count, not contract size: Millions of tiny databases created by software, priced like a utility, growing with every app an agent generates. The database industry spent 30 years selling to database administrators. The companies winning the entry point in 2026 sell to other people’s software.

Spring’s trend was abstraction: value accruing to whoever sits in front of a fragmenting market. Summer’s trend is what those abstractions run on. Every voice agent, coding agent, and generative media product on this list resolves into the same two purchases, compute to run models and databases to hold state. Startups now make both in their first year, at small dollar amounts, by the tens of thousands. Those two purchases are the earliest reliable signal for which agent-era companies are becoming real businesses.

All analysis conducted for this report that uses Brex internal customer data is anonymized and aggregated for privacy.

This story was produced by Brex and reviewed and distributed by Stacker.