The AI apocalypse is happening upstairs. We work downstairs.

Depending on which headline you read this week, AI is either about to crash the stock market or about to put every software company out of business. Sometimes both, in the same article.

They're two different apocalypse stories, and it's worth pulling them apart. One says AI is a bubble that's about to burst. The other says AI works so well it will eat the software industry alive. They can't both be the whole truth.

At our Annual Investor Summit in October, Jeremy Luzinda, Partner at Haatch, took both on. His answer fits in one line: the drama is upstairs. The work is down here. Here's what he meant, and why we think the doom is aimed at the wrong floor.

Yes, there's probably a bubble. It's just on another floor

Let's not pretend the numbers upstairs are normal. The big US tech companies are on course to spend around $646bn on capex in 2026, roughly 2% of US GDP (Apollo). Alphabet alone has signalled more than $300bn for 2027. The last time infrastructure spending ran this hot was the telecoms build-out of the late 1990s, and that ended in a glut.

Around $300bn of venture money was raised globally in the first quarter of 2026, and 80% of it went to AI. OpenAI, Anthropic and xAI took $172bn of that between them (Crunchbase). Meanwhile, each new frontier model is overtaken within months. That's an expensive treadmill.

So who pays the rent on all that compute? Jeremy's answer: that's a great question for Redmond and Mountain View. It's not a question for a seed fund backing UK software companies. We don't build data centres. We're their customers.

Why a bust upstairs would be good news downstairs

Here's the bit the doom stories skip. When the telecoms bubble burst in the early 2000s, the fibre didn't disappear. By early 2002, an estimated 95% of the cable laid in the late-1990s build-out was still dark (Seattle Times). It sat in the ground, suddenly cheap, and a generation of internet companies built on top of it. The investors who paid for the cables lost money. The businesses that rented the bandwidth did very nicely.

AI is already following the same pattern. Epoch AI found that the price of reaching a given level of AI performance has been falling by anywhere from 9x to 900x a year, depending on the task. Only the very newest frontier models still charge a premium, and that premium lasts until the next model ships.

Jeremy calls the arms race a subsidy to the application layer, and that's exactly where we invest: small software companies using AI to do a specific job for a paying customer. Every round of price cuts upstairs makes their product cheaper to run and better to use. If the infrastructure players overbuild, our founders get an even bigger discount.

The "SaaS apocalypse" is a pricing story, not a software story

Now the second apocalypse. In February 2026, $285bn was wiped off listed software companies in 48 hours. The press called it the SaaS apocalypse, and the logic went like this: if an AI agent can do the work, why would anyone pay for software?

Look closer and the fear is about seats. If you sell a licence per user, and a customer can now get the same work done with 15 people and some agents instead of 100 people, you've just lost 85 seats. That's a real problem, but it's a problem with a business model.

So the question to ask is "who was selling seats, and who was selling outcomes?"

We've seen this film before. In 1999, the Economist and the Telegraph were supposedly finished. They're still here. What died was the classified ad that paid their bills. The product survived; the pricing model didn't.

And it gets better for us. The listed incumbent has a thousand engineers maintaining a product built for people clicking buttons, charges per seat, grows slowly, and was bought by someone at 30 times revenue. The companies we back have 5 to 15 people, price on results, and often have revenue in year one. We didn't pay 30 times revenue for them. We paid a seed price.

When incumbents feel the squeeze, they buy what they can't build. Ask why HubSpot bought Trigify, a Haatch portfolio company, in September 2026. That deal delivered a 3.53x return on Haatch's investment, including escrow. As Jeremy put it: we don't own the incumbents, we own the companies built to replace them.

"Won't OpenAI just build everything?"

We hear this one a lot, and it's a fair question. OpenAI, Anthropic and xAI raised $172bn between them in the first three months of 2026, ten times what every UK startup raised in the whole first half (Crunchbase, Dealroom). How does a ten-person UK startup compete with that?

It doesn't need to. The frontier labs are building engines. Jeremy's focus is three words: agents, verticals, boring.

• Agents doing one specific job for one specific customer.
• Vertical software that becomes the system of record for an industry still running on spreadsheets.
• Boring problems with real budgets behind them.

The labs are not going to spend their next six months building software for freight yards, lettings agents or compliance teams. The companies that know those customers, own the workflow and hold the data are the ones that win. A better engine just makes their car faster.

What it looks like on the ground

Forget the headlines for a second. Here's how AI shows up in the companies we've backed:

• Teams of five doing what used to take fifteen. We have companies at £1m of annual recurring revenue with single-digit headcount.
• Charging for outcomes, not logins. Agent-based companies price on the job done, so more work means more revenue.
• Revenue before we invest. The first paying customer used to turn up around month 12. Now founders often arrive with one.

For startups, that's the opposite of an apocalypse. AI makes small teams dangerous. It lowers the cost of building, which is exactly what a seed investor wants.

It's not all smooth. Some founders are learning that an agent working inside a customer's live systems is much harder than a slick demo. But the failure mode we see is "not ready yet", not "nobody wants this". That's a timing problem, not a demand problem.

One line from Jeremy worth underlining: AI is often in the product. It's never the reason we invest. We back a founder who has found one customer with a real problem and a real budget, then help them find the next fifty.

The UK is a good place to do it. UK startups raised $17bn in the first half of 2026, more than double the year before. That's 39% of all venture money in Europe, more than France, Germany and Sweden combined, and 74% of it went into AI (Dealroom). A senior AI engineer in London costs roughly half what they would in San Francisco, and the talent pipeline runs through DeepMind, Oxford, Cambridge and Imperial.

So what would change our minds?

We're not in the business of blind optimism, so here are the tests. Jeremy sees three paths for the wider market: takeoff, grind or drawdown. What decides it upstairs is whether the AI labs' revenues keep compounding fast enough to pay for all that capex. In every one of those paths, inference keeps getting cheaper.

Downstairs, our tests are simpler. Are founders reaching first revenue faster with each cohort? Are strategic buyers still acquiring? Is the UK still the best place in Europe to do this? So far: yes, yes and yes.

The market got a live demo of all this two weeks after the summit. At the end of October, big tech reported and investors flinched at the bill. Meta spent $34.9bn on capex in the quarter, more than the cash its business generated, so free cash flow went negative. Amazon's free cash flow over the last 12 months was minus $42.4bn, against plus $14.7bn a year earlier.

But look at what was growing: Google Cloud up 103%, AWS up 39%. Demand for AI isn't the problem. The market has stopped rewarding spending for its own sake and started asking for returns. That's the infrastructure layer being asked who pays the rent, exactly as Jeremy said it would be.

The drama is upstairs. The work is down here. And down here, it's going rather well.

Important information

Risk warning: Don't invest unless you're prepared to lose all the money you invest. This is a high-risk investment and you are unlikely to be protected if something goes wrong. Investments in early-stage companies are illiquid, and tax reliefs depend on individual circumstances and the companies maintaining qualifying status, and may change. Past performance is not a reliable indicator of future results.

This article is for information only and is not advice or a recommendation. Views are Jeremy Luzinda's as presented at the Haatch Annual Investor Summit, October 2026. Portfolio figures are Haatch's internal marks as at October 2026. Issued by Haatch Ventures LLP, authorised and regulated by the Financial Conduct Authority.

Sources: Haatch Annual Investor Summit 2026 (State of the Market); Dealroom, UK H1 2026 Innovation Update (July 2026); Crunchbase News, global venture funding Q1 2026 (1 April 2026); Apollo Academy, hyperscaler capex (February 2026); Epoch AI, LLM inference price trends (March 2025); Seattle Times, 13 January 2002; WION, February 2026; Alphabet, Meta and Amazon Q3 2026 results; Hurren and Hope (October 2026).