@FinanceYF5: 3/ The speed at which startups are switching to open models is clearly accelerating. From the first payment for OpenAI or Anthropic API to the first purchase of open model compute, the median time has shortened from 23 months in 2022 to 5 months in 2025. The core reason is straightforward: some workloads can see cost reductions of up to 80%.

3/ The speed at which startups are shifting to open models is clearly accelerating.

The median time from first payment for OpenAI or Anthropic API to first purchasing open model compute power has shortened from 23 months in 2022 to just 5 months in 2025.

The core reason is straightforward: certain workloads can reduce costs by up to 80%. https://t.co/1fXx2PKG1T

Where the AI Boom is Really Making Money is Shifting from Applications to “Pipelines”

1/ Brex compiled a list of the 25 fastest-growing software companies as of summer 2026.

14 of them sell AI infrastructure, while another 6 are AI products built upon that infrastructure. Topping the list is Together AI, which provides open model services.

In the AI boom, the fastest growth is now occurring in compute, databases, caching, and sandboxes.

2/ Open models are free to download, but there’s an entire technology stack between them and being truly usable.

GPU clouds rent machines; Modal and RunPod handle containers and auto-scaling; Together AI, Deepinfra, and others package models into APIs; Tinker further up the stack offers managed fine-tuning.

Each layer sells “a little less to worry about.”

3/ The speed at which startups are shifting to open models is clearly accelerating.

From the first payment for OpenAI or Anthropic APIs to the first purchase of open model computing.

4/ However, open models haven’t replaced closed-source ones. Startups are increasingly taking a dual approach.

Complex reasoning and customer-facing conversations continue using closed-source APIs; large-scale tasks like classification, extraction, transcription, and generation are shifting to open models that are 5–20 times cheaper.

98% of open compute buyers are still paying OpenAI or Anthropic.

5/ The second-fastest growing infrastructure category is databases.

Neon states that over 80% of new databases are created by AI Agents; Supabase reports a similar figure exceeding 60%.

Agents can generate thousands of applications per day, requiring databases that provision in seconds, have near-zero idle costs, and can be instantly duplicated. Traditional architectures struggle to adapt to this pace.

6/ Startups’ “first database” has also changed.

Supabase’s share in initial standalone database purchases has risen from approximately 15% in 2023 to 38% in 2026; Pinecone’s share has fallen from 22% to under 2%.

Vector search hasn’t disappeared; it’s gradually being integrated into Postgres.

7/ Among startups that joined Brex in 2025, nearly one in ten purchased Supabase within their first year.

The median time from generating an app with Lovable, Replit, or Bolt to starting to pay for Supabase or Neon is just 106 days.

The earliest real signal of an Agent economy might be these two bills: compute and databases.

Read the original: