Something happened in the last month that every LP with meaningful AI exposure should be sitting with, and most UK institutional allocators haven't yet processed properly.

Leopold Aschenbrenner's Situational Awareness hedge fund - founded in 2024, backed by the Collison brothers, Nat Friedman, and Daniel Gross, built entirely around the thesis that AI would require a massive build-out of chips, memory, data centres and energy - lost roughly $35 billion in assets in a single month. Peak AUM at the start of July was $45 billion. By the end of July it was around $10 billion, after Ken Griffin's Citadel stepped in and bought the entire public equity book at a discount following margin calls from prime brokers at Bank of America, Goldman Sachs and JPMorgan.

Six weeks earlier, the fund was up approximately 450% at the end of June. The wunderkind 25-year-old former OpenAI researcher who had published the definitive AI investment thesis of the last two years, and who had been described by the trade press as "scarily smart," had built one of the most closely-watched AI-focused portfolios in the world. And within four weeks, it was gone.

I want to write about this, because most of the coverage has treated it as a personality story - the wunderkind, the leverage, the FTX ties, the swagger, the "liquor, ladies and leverage" jokes. That framing misses the actual lesson, which is not about Aschenbrenner at all. It's about the shape of every LP portfolio in the UK right now.

See below to continue…

Your network is scattered across inboxes, calendars, messages and apps. SoPhy pulls it all together - and lets you access it the way you already communicate.

Ask it anything about anyone, straight from iMessage, WhatsApp or Telegram. It knows who you met, when you last spoke, and when to follow up.

The first integrated version launches late August. Waitlist members get first access.

Join the waitlist

Continue…

Here's the uncomfortable part. Aschenbrenner was right.

His 2024 essay "Situational Awareness" argued that scaling AI would require enormous expansion in chips, memory, data centres, and electricity generation. Two years on, that thesis has been fully validated. Nvidia is still the most valuable company on earth. The hyperscalers are collectively spending more on AI infrastructure than the entire US highway system cost to build. Data centre capacity is the single most binding constraint on AI progress. Every serious institutional allocator has been repositioning around exactly the layer Aschenbrenner identified.

His portfolio held SK Hynix, CoreWeave, Nebius Group, Sandisk, Bloom Energy, and Micron - all of which were the correct picks for the thesis he was betting on. The thesis was right. The picks were right. He still lost $35 billion.

The reason is the specific gap between having the right AI thesis and being able to survive holding it.

Two things broke simultaneously.

First, the memory stocks - the specific layer Aschenbrenner was most concentrated in - corrected sharply as public markets grew concerned that hyperscaler capex was not translating into near-term revenue. SK Hynix, Sandisk, Micron and the others fell more than 30% in a month.

Second, the leverage he was using to amplify his conviction - reportedly up to 400% - turned that correction into a fund-ending event. The prime brokers called margin. The book had to be sold. Citadel bought it at a discount.

The letter to investors dated July 24, sent days before the collapse, called the sell-off "one of the best buying opportunities since early last year" and invited fresh capital commitments starting August 1. The additional capital didn't materialise in the size needed. That timing gap is the entire story.

Here's what most UK institutional allocators haven't yet worked through carefully.

The reason this story matters for LPs is not that most portfolios are levered 400% on memory stocks. They aren't. The reason it matters is that most portfolios are structurally concentrated in exactly the same AI investment thesis, executed through different instruments.

Look at the UK VC data. AI startups took approximately 74% of all UK VC in H1 2026. Full-year 2025 AI startup funding hit over £6bn - the highest share on record. The specific companies being funded are, in aggregate, expressing more or less the same thesis as Situational Awareness - that AI compute, infrastructure, memory, data, and applications will absorb enormous capital over the next decade. Different instruments, same underlying bet.

Which means when the memory stock correction hit Aschenbrenner, it was a live signal to every LP holding AI-heavy exposure elsewhere in their book. Not because the same margin call is coming for their private positions. Because the same underlying concern - that hyperscaler capex is running ahead of near-term revenue, and public markets are starting to price that gap - applies directly to how those private positions will eventually get marked and exited.

If you're an LP right now allocating to UK VCs whose portfolios are 74% AI-tilted, the Situational Awareness story is a preview of what happens when public market sentiment on AI infrastructure shifts. Public markets moved first. Private markets always move second, with a lag. The question is not whether the lag closes. It's whether your portfolio is positioned for what happens when it does.

If I were running an investment committee this month, there are three specific questions I would want the GPs I'm backing to answer clearly.

One - what is your exposure to positions that depend on hyperscaler capex continuing at current levels?

This is the specific place where Aschenbrenner's thesis broke first. Companies whose business model requires Meta, Alphabet, Microsoft, and Amazon to continue spending 96% of free cash flow on AI infrastructure are exposed to a single point of failure that has now started to move. Some GPs will have almost no exposure to this. Others will have a lot. Most haven't been asked the question directly.

Two - what happens to your unit economics if inference costs fall 50%?

Situational Awareness was long memory and infrastructure precisely because the assumption was that inference costs would stay high and demand for compute would keep expanding. If the current model release cycle (Opus 5 at half the price of Fable) continues compressing prices, some AI infrastructure positions look very different very quickly. The founders and funds who've stress-tested this scenario are in a completely different position from the ones who haven't.

Three - what's your actual liquidity in a scenario where private market valuations correct 30% over 18 months?

Public markets have already given us a preview of what that correction looks like. The prime brokers called margin on Aschenbrenner within days of the memory stock decline. Private LPs don't face margin calls - but they face NAV write-downs, delayed distributions, and pressure from their own institutional stakeholders to explain marks that no longer look defensible.

None of these questions are theoretical. Situational Awareness is the fully-documented, publicly-verifiable case study of what happens when the answers are wrong.

The most instructive detail in the whole story is who ended up owning the positions.

Ken Griffin didn't build Citadel by having a better AI thesis than anyone else. He built it by being one of the most sophisticated risk managers in the industry. When Aschenbrenner's book came onto the market at a discount, Griffin bought it - not because he suddenly agreed with Aschenbrenner's thesis, but because he could hold the same positions without the leverage that had killed the fund.

The picks were the same. The thesis was the same. The only difference was risk management. And that difference was the entire $35 billion.

For LPs, the parallel is direct. The GPs who survive the next 24 months are going to be the ones who held approximately the same theses as everyone else but managed the risk around those theses differently. Reserves for follow-on. Realistic marks. Exposure limits by layer of the AI stack. Genuine model-agnostic architecture in their portfolio companies. Not spectacular new theses. Better plumbing around the ones everyone shares.

The LPs who ask the right questions of their GPs this quarter will end up in the Citadel position. The ones who don't will find themselves adjacent to the Aschenbrenner position, without the leverage but with the same underlying exposure.

The AI trade is not over. The Situational Awareness collapse is not a sign that the underlying thesis was wrong. Aschenbrenner was right about the shape of what AI would need. He was wrong only about how to hold that view through a temporary market correction.

But the collapse is a signal - the first genuinely visible one - that the AI trade has now entered the phase where being right about the thesis stops being enough. Execution, risk management, portfolio construction, leverage discipline, and realistic time horizons are now the variables that separate the winners from the losers.

For LPs allocating fresh capital right now, the smart move is not to reduce AI exposure. It's to interrogate the specific quality of the risk management underneath every AI position in your book. The next twelve months are going to reward that interrogation more than any other single piece of due diligence work you could do.

Because Situational Awareness was not the anomaly. It was the first case study.

The next ones will not be as dramatic. They will be quieter. Slower. Spread across portfolios that don't get CNBC coverage. And the LPs who don't work through the questions above this quarter will find themselves holding those quieter positions when the marks eventually catch up with the public market signal that arrived in July.

Right thesis is no longer enough. Right trade is the whole game now.

✅ Know an LP still comfortable with an AI-heavy portfolio because "the thesis is right"? Forward this their way. The sooner more allocators separate having the right thesis from being able to survive holding it, the better positioned they'll be when the next round of marks lands.

✅ We also share this list in our founder-focused newsletter. Want to be featured next time? Send me a note.

We've built the ultimate aggregator of all UK startup events.
All the UK's best startup events in one place.

Access here and please share with others!

Tech brands are flooding F1 (34 sponsors in 2026, up 70% from last year). Motorsport is where elite connections happen.

Now enter the room: A private BTCC weekend at Silverstone, 26-27 Sept 2026. Thirty hand-picked family offices, fund GPs, and senior execs only.

What's included:

Not for sale. Invite-only. The right thirty get in.

Register interest here - selected for fit & value to the room.

If you found this roundup useful, share it with another investor who should be in the loop. The bigger our community, the more opportunities we can create together.