Navigating Vertical AI Startups

We brought together three speakers over the past month: a partner from one of the Nordics' most active early-stage funds, a semiconductor investor watching the infrastructure boom from up close, and two founders building inside the fastest-moving corner of B2B software.

This is what we took away.

I. Business Models for Vertical AI Startups

with Pirkka Palomäki, Founding Partner at Maki.vc

Pirkka opened with a warning: the categories that look most obvious right now are the ones founders should avoid. A light UI on top of an LLM is a weekend project, not a company. Sales and marketing AI is already brutally crowded. Co-pilots — the category every founder pitched in 2024 — have been quietly absorbed by Claude and ChatGPT for most knowledge workers. The opportunities that remain are harder, slower, and more deeply embedded in real workflows than the average pitch deck suggests.

This matches what the broader market is saying. Andreessen Horowitz argued earlier this year that the two-decade SaaS playbook — streamline a human task into software, charge per user — is no longer valid. The new model is software replacing labor entirely, which Bessemer estimates is a 10x larger opportunity than vertical SaaS ever was. Their data shows LLM-native vertical AI companies growing roughly 400% year over year at 65% gross margins — numbers traditional SaaS took a decade to reach.

But Pirkka pushed back on the easy version of this story. He's skeptical of moats that depend on "sustainable competitive advantage" in any classical sense. The models change weekly. Speed and marketing matter more than defensibility. Being first to market is a real barrier to entry in a way that traditional strategy frameworks no longer capture.

The exception is data — but only if it's actionable. His litmus test: does having this data genuinely make your solution better than anyone else's, or is it just an artifact? The strongest plays he sees are services that capture proprietary data as a byproduct of delivering value. Lovable, he noted, is a useful case study. They won not by building a deeper moat, but by reaching a new customer segment with a simpler entry point — attracting people who had never coded before.

On pricing, Pirkka was direct: outcome-based pricing is the most coherent model for AI-enabled services, but the pitch of "we lower your operating costs" only goes so far. Services with simple, transparent pricing are the most interesting structures right now. And on regulated industries like healthcare, finance, and pension funds, he was honest about the tradeoff. The sales cycles are long enough to kill an undercapitalized startup. They are also long enough to keep undisciplined competitors out.

II. Is the AI Market in a Bubble?

with René Kromhof, Founding Partner at Cloudberry VC

René came at this from a different angle: decades in semiconductors and consumer electronics, including the Heptagon exit to ams AG in 2017. He's seen one bubble inflate and burst. He's not certain we're in another one, but the math is making him nervous.

The headline number: the five largest US hyperscalers — Microsoft, Amazon, Alphabet, Meta, and Oracle — have collectively committed to roughly $660–690 billion in capital expenditure for 2026, nearly double 2025 levels. To put that in perspective, building one gigawatt of AI-ready data center capacity now costs roughly $60 billion by credible estimates. And one gigawatt currently generates around $10 billion of revenue a year — a number that has been flat for three years, through two chip generations both sold partly on fixing it. The capex numbers are staggering, and the hyperscalers themselves disagree on how long this hardware will actually be economically useful: Meta recently extended its AI depreciation schedule to six years while Amazon shortened theirs and took a write-down. Depreciate $60 billion over six years and you need roughly $13 billion of annual revenue per gigawatt just to break even — closer to $17 billion to generate cash. Revenue is at $10 billion. There simply isn't a business there yet; the entire bet is that the number breaks upward.

The dotcom comparison is imperfect. In 1999, infrastructure providers were loaning money to startups so those startups could buy their products — circular financing that collapsed when growth didn't come. Today, the buyers are profitable, mature companies. Microsoft and Google would need data centers regardless of AI. That's a real difference, René pointed out. But it doesn't fully resolve the question Sequoia posed eighteen months ago and has only grown louder since: where is the revenue to justify this build-out? Their estimate puts the gap at $600 billion annually — triple the figure from a year earlier.

The valuations are where René gets most cautious. OpenAI closed its latest round at $852B. Anthropic crossed $965B last week. Neither company is profitable. The large majority of their users pay nothing, and ChatGPT is now rolling out ads to address this. The OpenAI IPO, whenever it comes, will be enormous — and built on a business model that is still being invented. In René's framing, the model labs and the IPOs are the gold rush: exuberant, dangerously financed, and exposed if revenue doesn't arrive. The chips, packaging, photonics and data center infrastructure underneath are the picks and shovels — the more rational part of the trade, where the cash flows are real. Both could be mispriced, but for very different reasons.

And one closing observation René kept returning to: humankind has never once had a surplus of compute or bandwidth. Give people more of either, and they find a use for it. So the demand is the surest thing in the whole debate. The real question was never whether the compute gets used — it's who gets paid for building it.

His advice to founders building anything compute-heavy was unusually specific: show the margin you can deliver with current compute costs, then stress-test it with a 50% swing in token prices in either direction. If the model survives that, you have a business. If it doesn't, you have a bet on token deflation continuing at its current pace — which is not a strategy, it's a hope.

III. GTM Strategies for Vertical AI Startups

with Mikko Alasaarela (Agion) and Oona Suominen (Inven)

Mikko opened with a provocation: the software tool, as we've known it, is dead. The moat that SaaS spent twenty years building — sticky workflows, switching costs, seat-based revenue — is being dismantled by agents. What's left, in his framing, is service as a software: a service so deeply powered by AI that it behaves like software. Faster, cheaper, more scalable than any human team, but with judgment and context that automation alone can't provide.

The bar he sets for AI-nativeness is concrete. At least 50% of your product should be shipped by AI to qualify. For early-stage startups, he argues the number should be above 90%. A 4x productivity multiplier is becoming the industry floor across many sectors, not the ceiling. The teams that figure out how to stay aligned with their agents — rather than just directing them at human pace — will build things the rest of the market can't touch. As Mikko put it: humans become the bottleneck the moment you force AI to operate at human timeframes.

Oona's perspective from Inven added the GTM reality. They were first to launch an MCP connector not because they had more resources, but because they watched where the market was heading and moved before it was obvious. In a landscape where models evolve weekly, speed is the strategy. It's also the reason a small, sharp team can still outmaneuver a better-funded competitor: startups can pivot fast, and that pivot speed is the asset.

On winning enterprise trust, both speakers converged on the same uncomfortable point: personal relationships remain the stickiest asset in enterprise sales, and small teams have less time to build them. The implication isn't to deprioritize relationships. It's to let AI handle more of the background work so the human moments carry more weight. Even Anthropic, Mikko noted, runs lean. Teams are getting smaller across the industry, communication is getting harder, and the relationships you do build matter more, not less.

On data as a moat, both speakers agreed it's real — but only if it becomes actionable. Mikko's framing: data has to turn into intelligence that empowers the organization, not just sit in a system of record. Oona added the nuance: data is one angle, not the only one. On pricing, Mikko pointed to Intercom's introduction of outcome-based pricing for its Fin AI agent — $0.99 per resolution — as the template worth studying. His warning: stay out of the token business entirely. If your margins don't survive a swing in compute costs, you're not running a company; you're running an arbitrage that the next model release will close.

What We're Watching

Three nights, one underlying thesis. The companies that win vertical AI over the next few years won't be the ones with the most sophisticated technology, the most defensible moat, or the deepest pockets. They will be the ones that move faster, build trust faster, and learn faster — while staying honest about their unit economics and brutally clear-eyed about what the bubble means for their cost base.

The old playbooks don't transfer cleanly. The founders treating that as an opportunity rather than a problem are the ones worth watching.