AI took three quarters of UK venture funding this year. Its margin problem is still unsolved. — TFN

British AI corporations have simply had their finest six months on report. Evaluation from HSBC Innovation Banking UK and Dealroom places UK AI funding at $12.6bn within the first half of 2026, greater than 4 instances the identical interval a yr earlier and near three-quarters of all enterprise capital invested within the nation. Whole UK enterprise funding reached $17bn, the strongest opening to a yr since 2022, and the UK took 39% of all European enterprise capital.

That capital is being deployed into companies whose unit economics look nothing just like the software program companies that skilled a technology of traders. The hole is just not small; it’s not momentary, and most of it’s compute.

The quantity that broke the mannequin

ICONIQ’s State of AI survey, protecting roughly 300 executives at corporations constructing AI merchandise, has AI product gross margins averaging round 52% in 2026. That’s up from 41% in 2024 and 45% in 2025, so the course is encouraging. It additionally sits 25 to 30 factors under the 75% to 85% that conventional software program has delivered for 20 years.

The distribution issues greater than the common. Bessemer Enterprise Companions’ State of AI work discovered the fastest-scaling corporations, these reaching $100m in annual recurring income in round eighteen months, had been operating at roughly 25% gross margins, successfully shopping for distribution with compute. The extra capital-efficient cohort in the identical evaluation averaged round 60%. Each figures are a great distance from what a software program investor’s mannequin assumes.

Then there may be the discovering that ought to fear anybody underwriting these companies on a software program template. In ICONIQ’s knowledge, mannequin inference rises from 20% to 23% of whole spend as merchandise mature, whereas expertise prices fall from 32% to 26%. Learn that once more. The compute line grows as a share of spend as the corporate scales, quite than shrinking.

Software program economics rested on the other assumption. Price of products bought was largely mounted, so every further buyer arrived at near zero marginal price and margins expanded with quantity. In an AI product, the dominant price is incurred per question. It scales with utilization, typically quicker than utilization, and it doesn’t amortise. A enterprise that’s underwater at 100 prospects doesn’t get rescued by reaching a thousand.

This flows straight by way of to how these corporations are valued. The Rule of 40, which asks that development plus profitability clears 40, quietly assumed COGS was largely mounted. Strip 25 factors out of gross margin and development has to hold a far heavier share of the load to clear the identical bar. Traders are adjusting, however the benchmarks in large circulation nonetheless largely date from a distinct price construction.

From an IT choice to a capital allocation choice

The consequence for founders and boards is that the place compute bodily runs has stopped being an engineering desire and turn out to be a query concerning the form of the P&L.

For many corporations more often than not, the reply continues to be to name an API. In case your volumes are modest, your utilization is spiky, or you might be nonetheless altering the product each fortnight, paying a supplier for elasticity is straightforwardly appropriate. You might be shopping for optionality, and optionality is price paying for when you nonetheless want it.

The query arises on the level the place an organization is operating its personal fashions, by itself GPUs, at sustained and predictable utilisation. That could be a actual threshold and a rising variety of British AI corporations are crossing it, significantly these coaching or fine-tuning their very own fashions quite than wrapping another person’s.

Right here I’ve to be straight concerning the state of the proof, as a result of the web is stuffed with assured solutions and most of them are nugatory. Seek for the break-even level at which self-hosting beats renting, and you will see that figures starting from two million tokens a day to 100 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 printed by corporations that promote one facet of the commerce. Anybody quoting you a common crossover quantity is promoting one thing. The sincere reply is that it is determined by your utilisation, your mannequin dimension, your site visitors profile and your engineering capability, and the one dependable calculation is the one finished by yourself invoice.

The case that’s really documented

One firm has printed its numbers in sufficient element to be helpful, and it’s price inspecting exactly as a result of it’s not an AI enterprise.

37signals, which makes Basecamp and HEY, spent $3,201,564 on cloud providers in 2022 and determined to depart. The Register reported that the corporate purchased round $700,000 of Dell {hardware} to interchange its cloud situations, and that this outlay was fully recouped throughout 2023 as contract commitments expired. Writing on his personal weblog, co-founder David Heinemeier Hansson put the projected saving at well over $10m across five years, with the {hardware} operating in colocation quite than in a facility the corporate constructed itself.

Three caveats belong with these figures, and Hansson provides most of them himself. The comparability is rarely absolutely like for like. An organization with no present racks pays to lease them, although he argues that price is trivial subsequent to a cloud invoice. The financial savings goal retains transferring because the enterprise grows and desires extra {hardware}. And Data Centre Dynamics, reporting the identical story, famous that the figures exclude {hardware} refresh cycles and the extra operations employees that proudly owning infrastructure requires.

The important caveat is the one that almost all retellings skip: it is a steady, predictable SaaS workload, not AI inference. It doesn’t switch routinely. What transfers is the precept, which Hansson states extra bluntly than any analyst would: as soon as the payments turn out to be substantial, you owe it to your self and your traders to a minimum of do the maths.

That precept applies with extra power to AI, not much less, as a result of the compute line is bigger and since ICONIQ’s knowledge says it grows quite than shrinks with scale.

Why the British model has a lead time

There’s a wrinkle particular to corporations working right here, and it’s one that doesn’t seem in any American evaluation of this query.

If the maths does level in the direction of operating your personal {hardware}, the capability to place it in is rationed. Britain’s demand-connection queue for the transmission community reached 125GW towards a nationwide peak demand of round 45GW, and the Nationwide Vitality System Operator pushed by way of its largest ever set of connection reforms in December 2025 to clear speculative initiatives out of it. Excessive-voltage transformers carry lead instances of two to 4 years. Knowledge centre emptiness throughout the UK has fallen from 27% in 2016 to eight% within the first quarter of 2026.

The sensible implication for a board is that this choice can’t be left till the compute invoice turns into painful. Securing capability with colocation providers that already maintain energy and might help high-density, liquid-cooled {hardware} takes planning measured in quarters. An organization that runs the numbers, concludes it ought to transfer, after which begins trying is starting a course of quite than finishing one. The evaluation and the procurement have to run in parallel.

The counterweight

Two issues lower towards every little thing above, and a bit that ignored them wouldn’t be price studying.

The primary is that inference prices are collapsing. Equal mannequin efficiency has turn out to be dramatically cheaper yr on yr, and each month that continues weakens the case for purchasing {hardware} to flee an API invoice which may be considerably smaller by the point the equipment is racked. ICONIQ’s personal margin development, from 41% to 52% in two years, is partly this impact. If that curve holds, a few of immediately’s margin drawback solves itself with none infrastructure choice in any respect.

The second is that almost all AI corporations shouldn’t be fascinated by this. In case you are pre-product-market-fit, iterating weekly, or operating variable workloads, the flexibleness premium is cash properly spent and infrastructure possession is a distraction from the factor that truly kills startups, which isn’t constructing one thing individuals need.

What has modified is narrower than the loudest model of the argument, and extra sturdy. Compute is now a everlasting, rising, structural line in an AI firm’s price of income quite than a brief inefficiency to be optimised later. That makes it a board-level query about capital quite than a procurement element. In a yr when AI has absorbed three-quarters of British enterprise funding, the businesses that perceive their very own compute economics exactly, and the traders who ask about them early, are working with a significantly higher image than those nonetheless making use of a 2021 software program mannequin to a enterprise that not resembles one.