British AI companies have just had their best six months on record. Analysis from HSBC Innovation Banking UK and Dealroom puts UK AI funding at $12.6bn in the first half of 2026, more than four times the same period a year earlier and close to three-quarters of all venture capital invested in the country. Total UK venture funding reached $17bn, the strongest opening to a year since 2022, and the UK took 39% of all European venture capital.
That capital is being deployed into businesses whose unit economics look nothing like the software businesses that trained a generation of investors. The gap is not small; it is not temporary, and most of it is compute.
The number that broke the model
ICONIQ’s State of AI survey, covering roughly 300 executives at companies building AI products, has AI product gross margins averaging around 52% in 2026. That is up from 41% in 2024 and 45% in 2025, so the direction is encouraging. It also sits 25 to 30 points below the 75% to 85% that traditional software has delivered for two decades.
The distribution matters more than the average. Bessemer Venture Partners’ State of AI work found the fastest-scaling companies, those reaching $100m in annual recurring revenue in around eighteen months, were running at roughly 25% gross margins, effectively buying distribution with compute. The more capital-efficient cohort in the same analysis averaged around 60%. Both figures are a long way from what a software investor’s model assumes.
Then there is the finding that should worry anyone underwriting these businesses on a software template. In ICONIQ’s data, model inference rises from 20% to 23% of total spend as products mature, while talent costs fall from 32% to 26%. Read that again. The compute line grows as a share of spend as the company scales, rather than shrinking.
Software economics rested on the opposite assumption. Cost of goods sold was largely fixed, so each additional customer arrived at close to zero marginal cost and margins expanded with volume. In an AI product, the dominant cost is incurred per query. It scales with usage, sometimes faster than usage, and it does not amortise. A business that is underwater at a hundred customers does not get rescued by reaching a thousand.
This flows straight through to how these companies are valued. The Rule of 40, which asks that growth plus profitability clears 40, quietly assumed COGS was mostly fixed. Strip 25 points out of gross margin and growth has to carry a far heavier share of the load to clear the same bar. Investors are adjusting, but the benchmarks in wide circulation still largely date from a different cost structure.
From an IT decision to a capital allocation decision
The consequence for founders and boards is that where compute physically runs has stopped being an engineering preference and become a question about the shape of the P&L.
For most companies most of the time, the answer is still to call an API. If your volumes are modest, your usage is spiky, or you are still changing the product every fortnight, paying a provider for elasticity is straightforwardly correct. You are buying optionality, and optionality is worth paying for while you still need it.
The question arises at the point where a company is running its own models, on its own GPUs, at sustained and predictable utilisation. That is a real threshold and a growing number of British AI companies are crossing it, particularly those training or fine-tuning their own models rather than wrapping someone else’s.
Here I have to be straight about the state of the evidence, because the internet is full of confident answers and most of them are worthless. Search for the break-even point at which self-hosting beats renting, and you will find figures ranging from two million tokens a day to a hundred million tokens a month, from 8,000 conversations a day to eighteen months of near-total utilisation. They disagree by orders of magnitude, and the overwhelming majority are published by companies that sell one side of the trade. Anyone quoting you a universal crossover number is selling something. The honest answer is that it depends on your utilisation, your model size, your traffic profile and your engineering capacity, and the only reliable calculation is the one done on your own bill.
The case that is actually documented
One company has published its numbers in enough detail to be useful, and it is worth examining precisely because it is not an AI business.
37signals, which makes Basecamp and HEY, spent $3,201,564 on cloud services in 2022 and decided to leave. The Register reported that the company bought around $700,000 of Dell hardware to replace its cloud instances, and that this outlay was entirely recouped during 2023 as contract commitments expired. Writing on his own blog, co-founder David Heinemeier Hansson put the projected saving at well over $10m across five years, with the hardware running in colocation rather than in a facility the company built itself.
Three caveats belong with those figures, and Hansson supplies most of them himself. The comparison is never fully like for like. A company with no existing racks pays to rent them, though he argues that cost is trivial next to a cloud bill. The savings target keeps moving as the business grows and needs more hardware. And Data Centre Dynamics, reporting the same story, noted that the figures exclude hardware refresh cycles and the additional operations staff that owning infrastructure requires.
The critical caveat is the one that most retellings skip: this is a stable, predictable SaaS workload, not AI inference. It does not transfer automatically. What transfers is the principle, which Hansson states more bluntly than any analyst would: once the bills become substantial, you owe it to yourself and your investors to at least do the maths.
That principle applies with more force to AI, not less, because the compute line is larger and because ICONIQ’s data says it grows rather than shrinks with scale.
Why the British version has a lead time
There is a wrinkle specific to companies operating here, and it is one that does not appear in any American analysis of this question.
If the maths does point towards running your own hardware, the capacity to put it in is rationed. Britain’s demand-connection queue for the transmission network reached 125GW against a national peak demand of around 45GW, and the National Energy System Operator pushed through its largest ever set of connection reforms in December 2025 to clear speculative projects out of it. High-voltage transformers carry lead times of two to four years. Data centre vacancy across the UK has fallen from 27% in 2016 to 8% in the first quarter of 2026.
The practical implication for a board is that this decision cannot be left until the compute bill becomes painful. Securing capacity with colocation providers that already hold power and can support high-density, liquid-cooled hardware takes planning measured in quarters. A company that runs the numbers, concludes it should move, and then starts looking is beginning a process rather than completing one. The analysis and the procurement need to run in parallel.
The counterweight
Two things cut against everything above, and a piece that ignored them would not be worth reading.
The first is that inference costs are collapsing. Equivalent model performance has become dramatically cheaper year on year, and every month that continues weakens the case for buying hardware to escape an API bill that may be substantially smaller by the time the kit is racked. ICONIQ’s own margin trend, from 41% to 52% in two years, is partly this effect. If that curve holds, some of today’s margin problem solves itself without any infrastructure decision at all.
The second is that most AI companies should not be thinking about this. If you are pre-product-market-fit, iterating weekly, or running variable workloads, the flexibility premium is money well spent and infrastructure ownership is a distraction from the thing that actually kills startups, which is not building something people want.
What has changed is narrower than the loudest version of the argument, and more durable. Compute is now a permanent, growing, structural line in an AI company’s cost of revenue rather than a temporary inefficiency to be optimised later. That makes it a board-level question about capital rather than a procurement detail. In a year when AI has absorbed three-quarters of British venture funding, the companies that understand their own compute economics precisely, and the investors who ask about them early, are working with a considerably better picture than the ones still applying a 2021 software model to a business that no longer resembles one.
Sponsored by Carbon-Z
