I want to write something uncomfortable this week because I've been sitting on it for a couple of months and it's genuinely starting to worry me.

I've been in a lot of investor rooms in the UK recently. Senior investors. Top-tier names. People whose LinkedIn bios are impressive and whose track records are real. And I keep noticing the same thing, quietly, every single time: most of them do not understand how AI actually works.

Not in a hand-wavy "I'm not a nerd" way. In a fundamental way. They understand the label. They understand the hype. They can recite the headline claims. But when you drill down into the mechanics of what a large language model is actually doing under the hood, and why that determines what a company built on top of it can or cannot defensibly own, the conversation collapses.

And I'm sitting there thinking - this is the person deciding which companies get funded in the UK AI ecosystem. This is the person our whole capital allocation depends on. This is the person whose LPs are trusting them to distinguish signal from noise.

That's a problem. And it's a much bigger problem than most people are willing to name out loud.

See below to continue…

The founders who close rounds fastest aren't just the ones with the best pitch; they're the ones who show up prepared. Carta Launch gives early-stage founders a clean, investor-ready cap table for free, so when the conversation gets serious, you're already serious. 

Manage equity, issue shares, and track dilution, then instantly share it with a built-in data room that gives investors exactly what they need, exactly when they ask.

The best time to get your cap table right is before you need to.

Learn more about Carta Launch and get started for free.

Continue…

I want to put the stakes in context, because I don't think most people realise how much money is now flowing through hands that don't fully understand what they're pricing.

According to HSBC Innovation Banking and Dealroom, UK AI startups raised $5.8 billion in Q1 2026 alone - roughly 74% of all UK venture capital. That's a fivefold increase in AI's share of UK VC since 2022. Natwest's Future of UK Innovation report put full-year 2025 AI startup funding at over £6 billion, more than a third of all UK venture capital and the highest share on record. The British Business Bank's 2026 Small Business Equity Tracker shows AI companies took a record 44% share of UK smaller-business equity investment in 2025, with AI-related deal value up 48% year on year.

For every £3 a UK VC deployed into a startup in H1 2026, roughly £2.20 of it went into something with "AI" somewhere in the pitch deck.

That is not a tilt. That is not a preference. That is the near-total capture of an entire national capital market by a single narrative. And the honest question I want to ask is: how much of that capital is being allocated by people who could pass a basic technical literacy test on the thing they're buying?

Because based on the conversations I've been having, my genuine estimate is: not enough.

Let me give you the simplest possible example of what I mean, because it's become a running joke on the internet and almost everyone has missed the actual point.

If you ask ChatGPT how many Rs are in the word "strawberry," it will often confidently tell you two. The answer is three. The internet spent six months mocking this as proof that AI is stupid.

Here's what's actually happening. Large language models don't see letters. They see tokens - chunks of text, typically about four characters long, that the model has been trained to recognise as units. When you type "strawberry," the model doesn't see s-t-r-a-w-b-e-r-r-y. It sees something like "straw" plus "berry" - two tokens - and those tokens are what it processes. The individual letters inside those chunks are, in a very real sense, invisible to it. When you ask how many Rs are in the word, you're asking the model to answer a question about characters it literally cannot see.

Once you understand that, the "AI is stupid" reading collapses. What you're actually looking at is a specific, well-documented architectural constraint that determines what LLMs are and aren't reliable at. And the same architecture is behind almost every strange, wrong, or off-feeling AI output you've ever seen.

Now here's the reason I'm telling you this in a piece about investors: that is the level of technical literacy every AI investor should have. Not deep learning theory. Not the maths. Just the basic mechanics of what the thing is actually doing when it produces an output, and therefore where its edges are.

I've asked variations of this question to a lot of UK investors this year. Not the strawberry one specifically - things like "what makes this AI company defensible against a competitor with the same LLM API and a weekend of engineering time" or "what's the actual proprietary layer here that a competitor can't replicate" or "what does this company own that the frontier labs can't just build in the next model release." And I've watched the answers slide into vibes almost every time.

That is not because these are bad investors. They are, mostly, good ones. It's because the entire industry has moved faster than the people writing the cheques could keep up with.

If I were an LP right now, allocating to UK VCs whose portfolios are 74% AI-tilted, I would want to know that the GPs I'm backing can answer three specific questions about every AI company in their portfolio.

One: what would break if the underlying model changed tomorrow? If the answer is "nothing much, we'd just swap providers," this is a wrapper company. Wrappers can be great businesses. But they don't have moats, and they should be priced accordingly. If the answer involves a specific integration, a proprietary fine-tune, a workflow-specific evaluation layer, or a genuine data feedback loop - now you're looking at something more interesting.

Two: what does this company own that a well-funded competitor with the same API access couldn't replicate in six weeks? This is the actual moat question. Most AI companies right now have three answers to this: brand, distribution, or a very specific customer workflow embedded into their product. If the answer is "our AI is better," that's not a moat - the frontier labs will have caught up by the next model release. If the answer is "we have three years of proprietary user interaction data that our AI improves on," that might be one.

Three: what happens to margins when inference costs double? Because they will. The Anthropic and OpenAI pricing structures we all built businesses on this year are not the structures we're going to be looking at in two years. The companies that can absorb that shift have unit economics that don't depend on subsidised inference. The companies that can't are being priced right now as if the current cost curve is permanent.

If a founder can't answer these questions in a first meeting, that tells you something. If an investor doesn't ask them, that tells you something too.

I know this all sounds like a criticism of investors. I want to be direct about why it's a much bigger deal than the mirror-image criticism of founders.

Founders who don't understand AI fundamentals lose their own time and money. That's bad but it's contained. It's their runway, their equity, their decision.

Investors who don't understand AI fundamentals lose everyone's money. LP capital that could have backed a real deep-tech company goes instead into a wrapper. A defensible enterprise AI company gets passed on because it doesn't hype-signal in the language a non-technical investor recognises. Capital gets concentrated at valuations that don't match the underlying business quality, because everyone in the room is working from vibes.

The distortion this creates is real. UK founders with genuinely differentiated AI products are getting less capital than they should. UK founders with a Cursor licence and a good pitch deck are getting more. LPs are subsidising the difference. And the whole system is doing this while telling itself it's making rational, informed decisions about the future of British technology.

If this sounds harsh, I want to name what I'm not saying. I'm not saying every UK AI investor is out of their depth. I know some who are technically sharp, ask the right questions, and are quietly making genuinely great bets on companies with real defensibility. But they are the exception, not the rule, and they'll be the first to tell you the same.

Two things.

If you're an investor active in AI, do the reading. Not a whitepaper. Not a McKinsey deck. The actual technical fundamentals of how these models work. Spend an hour on tokenisation. Spend another hour on what fine-tuning actually does and doesn't do. Understand why retrieval-augmented generation is a moat for some companies and a cost centre for others. Read Andrej Karpathy's YouTube tutorials - they're free, they're pitched at exactly the level a non-technical person needs, and they'll take you from vibes to actual understanding in about ten hours of viewing.

That is not a big ask against the size of the cheques you're writing.

If you're a founder building in AI, learn to explain your moat in ten seconds to a non-technical investor. Because you're going to be pitching to a lot of them, and the ones who don't understand the fundamentals will fund you or pass on you based on whether your pitch matches the shape of pitches they've seen work before. That shape is mostly hype. If you can find a way to make the actual mechanical defensibility of your product legible to someone who doesn't have the technical grounding, you'll close rounds that better companies with worse pitchers won't.

That's not a fair situation. It just happens to be the situation.

We are, right now, allocating billions of pounds of British capital into a category the majority of the allocators don't fully understand. Some of that capital is going to genuinely brilliant companies that will define the next decade of UK technology. A lot of it is going to companies that will look, in retrospect, like they were funded on the strength of the word "AI" appearing on their landing page.

The correction is coming. It always does. And when it arrives, the investors who'll survive it are the ones who did the work to understand what they were buying. The founders who'll survive it are the ones who built something with real defensibility rather than the appearance of it.

I'm writing this because I want more of you to be in both categories when the music stops.

The strawberry example is not just about AI being weird. It's about how quickly the surface of something impressive-looking can hide a completely different mechanical reality underneath. And in a market where 74% of British venture capital is now flowing into a single technology category, the question of how many people writing those cheques could correctly explain what's actually inside the box has stopped being academic.

It's the whole game.

✅ Know an investor confidently backing AI companies without ever having asked a founder what would break if the underlying model changed tomorrow? Forward this their way. The sooner more allocators do the reading, the sooner UK capital starts finding the founders who actually deserve it.

(👉 Vote now - we’ll share the results in next week’s issue. All votes are anonymous.)

(Last week’s poll results)

Access here and please share it with others.