Brex’s summer 2026 Benchmark report highlights a shift in the software startup landscape. Fourteen of the 25 fastest‑growing companies are not selling finished artificial intelligence products; instead, they provide the essential infrastructure—databases, graphics processing units (GPUs), caches, and sandbox environments—that developers need to build AI agents.
Infrastructure Takes Center Stage
The top‑ranked firm, Together AI, rents access to open‑source models, allowing startups to avoid the high costs of proprietary APIs. Historically, similar lists were dominated by mobile‑app developers and cloud providers like Amazon Web Services. This year’s data shows that the “plumbing” of the AI economy—compute power and data storage—is the primary growth driver.
Open‑weight models are freely downloadable but require substantial resources to run. Startups must provision GPUs large enough to host the models, an inference engine such as vLLM, and a serving layer that batches web requests. Companies like Modal and Runpod provide serverless platforms that add containers, auto‑scaling, and per‑second billing. At the top of the stack, firms such as Together AI, DeepInfra, Baseten, and Fireworks deliver popular open models through ready‑to‑use endpoints, letting developers switch between models like DeepSeek or Kimi with a single configuration change.
Cost Savings Drive Adoption
A single Nvidia H100 GPU now rents for about $3.40 per hour, down from roughly $8 in 2023. When efficiently batched, these GPUs can generate tokens for open models at pennies per million, compared with closed‑source APIs that charge $5‑$30 per million tokens. For common workloads—data tagging, document extraction, speech‑to‑text, image and video generation—open models often meet performance needs at a fraction of the cost.
Brex data shows startups are adding an open‑compute vendor within five months of their first large‑scale API bill, twice as fast as in 2024. This rapid transition can cut operating expenses by up to 80 percent, allowing young companies to achieve product‑market fit without the financial strain of proprietary AI services.
Database Innovation Fuels Agent Development
Database usage patterns have also evolved. Supabase adoption grew roughly 17‑fold between the 2022 and 2025 startup cohorts, with nearly one‑in‑ten 2025 entrants paying for the service in their first year. Neon, a Postgres‑compatible database, reports that artificial intelligence agents now create more than 80 percent of its new databases, while Supabase says over 60 percent of its launches are initiated by AI tools.
Modern cloud databases separate compute from durable storage, enabling instant provisioning and near‑zero idle costs. This architecture aligns perfectly with AI agents that spin up thousands of databases daily for generated applications. Companies like Replit and Retool already rely on such databases, running massive numbers of instances with minimal engineering overhead.
What the Spending Data Reveals
Entry‑level spending remains modest. Median monthly bills for services like Supabase’s $25 Pro plan or Neon’s usage‑based pricing sit in the tens of dollars range. Railway’s median bill halved even as its paying customer base grew eightfold. The revenue model is volume‑based: millions of tiny databases are created and billed like utilities, scaling with each new AI‑generated app.
The summer trend emphasizes the two core purchases that underpin the agent economy—compute power to run models and databases to store state. Startups are making both purchases within their first year, often spending only a few thousand dollars total. These early signals help identify which AI‑enabled companies are transitioning from experimental projects to sustainable businesses.
All analysis in this report is based on anonymized, aggregated Brex customer data, ensuring privacy while providing insight into emerging technology trends.
Original reporting: KRDO (Colorado Springs metro) — read the source article.
