Capital has not retreated from AI. It has become more selective. After the broad enthusiasm of 2023–2024, investors in 2025 and 2026 have concentrated funding in a smaller set of categories and demanded clearer paths to durable businesses. The patterns are legible if you look at where the large rounds are landing.

Three layers that keep attracting capital

1. Foundation and specialized model companies

Training and serving frontier or highly specialized models remains capital-intensive. Investors continue to fund companies that can differentiate on model quality, cost, domain focus, or deployment flexibility. The bar is high: differentiated data, novel training methods, or clear advantages in latency and price are expected, not optional.

2. Infrastructure and tooling

Everything required to build, evaluate, deploy, and monitor AI systems at scale — evaluation platforms, observability, fine-tuning and distillation tools, inference optimization, data pipelines, and agent frameworks — has seen sustained investment. These companies sell to other AI builders, which creates a clearer early revenue story than pure research bets.

3. Vertical applications

Startups that embed AI deeply into a specific industry workflow (legal, healthcare, finance, software development, customer support, education) are raising significant rounds when they can show retention, expansion, and measurable ROI. Generic “AI for X” pitches struggle; deep domain integration and proprietary process data win.

What is harder to fund

Thin wrappers around a single frontier API, without proprietary data or workflow lock-in, have become difficult to raise on. So have pure research projects without a plausible path to product or partnership. Investors have seen enough demos; they want evidence of usage, willingness to pay, and defensibility that survives model improvements from the large labs.

Round dynamics

Several trends stand out in deal structure and timing:

Signals for the next 12 months

Watch for concentration continuing in the three layers above, with particular attention to:

  1. Agentic and multi-step workflow products that cross from pilot to paid production use.
  2. Evaluation, safety, and governance tooling as enterprises scale AI deployments.
  3. Open-weight and specialized model companies that can undercut frontier API costs for high-volume workloads.
  4. Vertical platforms that own the data loop inside a regulated or high-stakes domain.

Outside the pure venture cycle, open research and smaller experimental projects still shape what gets funded later. Techniques and prototypes that surface publicly — including work from independent efforts such as labs.maxpri.workers.dev — often become the seeds of the next product categories investors chase.

Bottom line

AI funding has matured from a broad bet on “AI will change everything” to a more disciplined allocation toward teams that can turn model capability into durable product advantage. Capital is still available in large amounts. It is simply less patient with companies that cannot explain why they will still matter when the next model generation arrives.