Startups in AI are competing in the fastest-growing software category on record. Enterprise AI spend hit $37 billion in 2025, up from $11.5 billion the year before, according to Menlo Ventures. But the money is not spread evenly. Some layers reward capital and incumbents. Others reward speed, focus, and small teams.
Here's a map for founders deciding where to play. You'll see which categories are saturated, where the real openings sit, and what separates the startups in AI that pull ahead.
Where the money flows today
More than half of enterprise AI spend now lands at the application layer. Menlo Ventures puts 2025 application spend at $19 billion, against $18 billion for infrastructure. Buyers want immediate productivity, not long-term infrastructure bets.
That's a shift. Earlier in the cycle, a16z observed that infrastructure vendors were "likely the biggest winners," capturing most of the dollars. Apps grew fast but fought thin margins. App companies spend roughly 20 to 40% of revenue on inference and per-customer fine-tuning, by a16z's estimate.
The more useful finding: a16z had a "hard time finding structural defensibility anywhere in the stack." The usual defenses, like scale, supply chains, distribution, and data pipelines, tend to fade over time. For a new founder, that's an opening, not a warning.
Which categories are saturated for small teams
"Saturated" here means capital-intensive and leader-concentrated. It doesn't mean closed. Three areas are hard right now.
Start with foundation models. Three providers control 88% of enterprise LLM API spend: Anthropic at 40%, OpenAI at 27%, and Google at 21%, per Menlo Ventures. Enterprise open-source share fell from 19% to 11%. Competing here means out-spending labs with nine-figure training budgets. a16z notes training can cost "up to tens or hundreds of millions of dollars."
General-purpose copilots are just as crowded. Horizontal AI is the largest app category at $8.4 billion, and copilots hold 86% of it. ChatGPT Enterprise, Claude for Work, and Microsoft Copilot lead the pack. A thin wrapper over the same models won't displace them.
Core infrastructure is the third tough spot. Incumbents hold 56% of the AI infrastructure market. Challengers are growing, but they face hyperscalers that collectively spend more than $100 billion a year on capex, a16z estimates.
One caveat: these numbers describe concentration, not permanence. a16z found no layer structurally defensible. For startups in AI, saturation is a high cost of entry, not a wall.
Where the real openings are
The application layer is the clearest opening. Startups in AI now earn nearly $2 at the app layer for every $1 incumbents earn. That's 63% of the market, up from 36% a year ago, Menlo Ventures reports. Below the app layer the story flips: incumbents keep 56% of infrastructure.
Departmental workflows are next. Departmental AI reached $7.3 billion, up 4.1x in a year. Startup share runs highest in product and engineering (71%), sales (78%), and finance and operations (91%). Coding alone accounts for $4 billion, the category's first killer use case.
Vertical AI is growing even faster. It hit $3.5 billion, nearly triple 2024, and healthcare leads at roughly $1.5 billion. Adoption clusters in industries with manual, unstructured workflows that software historically skipped.
AI agents keep drawing founders too. Of the 144 startups in Y Combinator's Spring 2025 batch, 67 (about 46%) were AI agent companies, PitchBook reports. The prior batch had 58 of 163.
Then there's the buyer side. 76% of enterprise AI solutions are purchased rather than built, up from 53% in 2024. Enterprises want finished products, not frameworks.
Five patterns among winning startups in AI
These patterns recur across the winners named above. Treat them as hypotheses to test in your own market, not laws.
Own a niche workflow, not a category. Clay wins in sales by attacking workflows Salesforce doesn't own: research, enrichment, and personalization from signals outside the CRM (Menlo Ventures). Narrow surface, deep fit.
Ship faster than incumbents can copy. Cursor beat GitHub Copilot to repo-level context, multi-file editing, and diff approvals. GitHub had the distribution. Cursor had the velocity.
Stay model-agnostic. Cursor adopted Claude Sonnet 3.5 the moment it launched instead of waiting on one vendor's roadmap. a16z's view is that long-term model differentiation is unclear. Flexibility beats lock-in.
Move bottom-up. Cursor reached $200 million in revenue before hiring a single enterprise sales rep. Product-led motions drive 27% of AI application spend, versus 7% in traditional software. Individual users now carry these tools into the enterprise themselves.
Compound proprietary context. a16z names data pipelines as a defense but calls it rarely durable. The stronger play is usage that generates data no one else has. In finance, startups like Rillet and Numeric build AI-first systems where every transaction feeds the automation. Test whether your product gets better with each cycle of use.
How startups in AI compete with funded giants
Incumbents have distribution, data, and balance sheets. Even so, Menlo's data shows exactly where lean teams win. Startups in AI thrive in agile departments: market research, sales, marketing, and product. Incumbents hold IT and data science, where reliability and deep integrations outweigh iteration speed.
Compete where speed matters more than certified integrations. Avoid head-on fights where the incumbent's slowness is a feature.
There's a second edge. Incumbents in regulated domains face accuracy demands that slow AI-native shipping. In finance, that paralysis opened room for AI-first ERPs. Where a giant can't move fast enough, a vacuum forms.
A positioning rubric for founders
Score each question as a starting point, not a verdict. Missing data isn't a bad score. It's a reason to run a small test.
- Does the workflow sit outside the incumbent's system of record?
- Can you ship weekly while the incumbent ships quarterly?
- Can you swap model providers without rewriting the product?
- Can one user get value without a procurement cycle?
- Does usage produce context that improves the next interaction?
One hard blocker overrides the rest. If a giant can bundle your feature next quarter, the aggregate score doesn't matter. Narrow the niche, or add depth they can't copy. Usually that means owning the workflow itself and the context data behind it.
The lean operator's version of this thesis
Zrevick's founding thesis is that AI-native operators can outrun larger incumbents. Technical founders already ship products at AI speed with tools like Claude Code and Codex. The business side lags behind, stuck in five to eight disconnected tools: website builders, CRMs, lead databases, sequencers, support desks.
Zrevick consolidates those functions into one connected workspace. A shared Identity Graph, the Company Brain, keeps every teammate and AI agent working from the same picture. MCP and REST APIs let agents run go-to-market work straight from the development environment. It takes the compounding power founders already have in code and applies it to sales, marketing, and support. See the platform for details, or read more on the Zrevick blog.
The bottom line
Enterprise AI spend more than tripled in a single year. The market rewards capital at the bottom of the stack and speed at the top. Startups in AI win most at the application layer, in narrow workflows, with products that compound context over time. Pick a surface the incumbent doesn't own, ship faster than they can copy, and let usage build a defense they can't buy. That's the whole game right now. And if you want your go-to-market moving at the same speed as your code, see what the Zrevick platform does.