UK AI Company Formations Hit Record 86% Surge While Class 42 Trademark Filings Drop 25%: The IP Gap Widening at the Core of Britain's Software Boom

Something quietly extraordinary happened inside the UK's company registry data this summer. While fourteen sectors contracted in Q1 2026, software and AI company formations surged 86% in a single quarter, with business and domestic software development registrations jumping from 4,035 to 7,506 in three months. At the same time, AIBD analysis of UK Intellectual Property Office data shows Class 42 trademark filings fell to 5,530 in Q3 2026, a drop of 25.3% against the prior period. Two curves moving in sharply opposite directions. That divergence is the story.

Two Classification Systems, One Fault Line

To understand why it matters, you need to hold two taxonomic changes in your head simultaneously. The first is taxonomic in the industrial sense: the ONS published UK SIC 2026 in late July, formally separating general-purpose AI software and AI model development from broader software categories at the 5-digit level, the first time AI-related activities have been explicitly recognised in Britain's statistical classification of economic activities. The second is taxonomic in the IP sense: the Nice Classification 13th Edition entered force on 1 January 2026, explicitly adding "Artificial Intelligence as a Service [AIaaS]" as a recognised service in Class 42, covering SaaS, cloud computing, research, and scientific technology services.

Think of it as two engineers finally labelling the pipes in a building that has been running on unmarked plumbing for two decades. The pipes were always there. Now they have names. The problem is that the builders are laying new pipe faster than anyone is putting labels on.

The Formation Surge Is Structurally Different

The 86% growth in software and AI formations is not noise. Research from Your Company Formations published 28 July found the surge was sustained across all three months of Q1 2026, including February, the month Companies House doubled the cost of digital incorporation from £50 to £100. While fourteen other sectors declined that month, software development grew 10.8%. A demand signal that strong, against a doubling of friction, is rare. The full-year 2025 growth rate for the same SIC category was 38.4%, according to Beauhurst's New Startup Index 2026. Q1 2026 has already more than doubled that annual trajectory in a single quarter.

The ONS's own July 2026 report on AI in UK businesses adds context: around 36% of UK workers now report using AI at work, placing Britain among Europe's higher adopters. Only 10% of businesses using at least one AI technology report using it extensively. Adoption is broad but shallow, the classic early-infrastructure phase. You see this pattern in computing history too: the minicomputer era of the 1970s generated enormous formation activity in SIC-equivalent categories long before the architecture settled.

So what does the formation data actually tell us about the engineering layer? It is a proxy, because there is still no dedicated SIC code for artificial intelligence companies in the UK. Businesses working in AI register primarily under business and domestic software development. SIC 2026 begins to fix this at the statistical level, separating AI model development from generic software publishing, but Companies House adoption is not expected to be mandatory until the 2031 Blue Book cycle. Founders building AI inference pipelines and founders building tax-filing SaaS are still, for now, the same code.

The IP Exposure Problem

Here is where the trademark drop becomes architecturally interesting. Class 42 covers precisely the territory where the AI formation surge is happening: software as a service, cloud computing, AI-based research services, and, from 1 January 2026, AIaaS as an explicitly codified category. Filing volume in that class falling 25.3% in Q3 2026, as measured by AIBD's analysis of IPO data, suggests one of three things: companies are deferring filings due to cost pressure; they are unaware the Nice 13th Edition created cleaner, more defensible terminology for their actual service offering; or they are simply moving too fast to think about it.

All three explanations point to the same structural gap. The Temporal State of Development Report, published in August 2026 and based on surveys of over 550 engineers in the US and UK, found daily AI agent use among engineers rose 33 percentage points in a year, with the majority describing agents as either helpful or transformative in production workflows. Andreessen Horowitz, analysing OpenRouter traffic data, estimated that autonomous AI agent workloads consume approximately five times as many LLM tokens as human-driven usage. Infrastructure is being built, deployed and scaled at velocity. The brand names sitting on top of that infrastructure, the names engineers will remember, enterprise buyers will put in procurement forms, and courts will litigate over, are being left without formal protection.

The NHS SBS £900m Healthcare AI Solutions framework, published in May 2026 with a tender deadline of 23 June, illustrates the stakes precisely. Organised across eight lots spanning radiology, predictive analytics, robotics and operational efficiency, it is the largest public-sector AI procurement vehicle in the UK. Contract awards expected in March 2027 will route substantial government spend through supplier brand identities. Vendors without registered Class 42 marks covering AIaaS, diagnostic platforms and clinical decision-support software face a situation analogous to bidding on a long-term infrastructure contract without a registered company name: technically possible, strategically exposed.

The Architecture of the Gap

The engineering analogy worth reaching for is TCP/IP adoption in the early 1990s. The protocol stack was being deployed everywhere, across universities, telcos and government labs, but the address space management was chaotic, the naming conventions were informal, and the assumption was that formalisation would catch up later. It did, eventually, but the lag produced two decades of domain-name disputes, squatting, and adversarial registrations that cost the industry vastly more than early registration would have. The Class 42 trademark gap has the same shape.

SIC 2026's explicit recognition of AI model development as a distinct economic activity is useful signal here. When a classification framework built for national accounts starts treating "general-purpose AI software" differently from "business and domestic software development," it is because the economic and operational distinction has become too large to ignore. The IP framework, via Nice 13th Edition, made the same move six months earlier. The market, measured in trademark filings, has not followed yet.

That lag is not fatal, but it has a ticking clock. Open frameworks like NHS SBS compress it: once award decisions are made in March 2027, the downstream licensing, white-labelling and sub-contracting activity that follows will be far harder to protect retrospectively.

6-12 Month Implications

Three near-term pressure points are worth watching. First, the ONS's annual review of SIC 2026 explanatory notes, expected end of March 2027, will further refine AI activity classifications, and any change to how Companies House eventually mandates the new codes will flush a large volume of re-registration activity that could also trigger trademark priority disputes. Second, the DRCF's AI consumer risk consultation, which closed 2 September 2026, may produce guidance affecting how AIaaS products are described in regulated contexts, with direct downstream implications for how Class 42 specifications are drafted. Third, NHS SBS contract awards in Q1 2027 will create the first real public test of whether AI vendors have their IP house in order before public money moves.

The 86% formation surge is genuinely impressive. Britain is building. But in infrastructure terms, the foundations precede the load-bearing work. Right now, the load is arriving faster than the foundations are being registered.