Leopold Aschenbrenner's fund was up more than 400% earlier in 2026. Situational Awareness had become the purest available bet that AI would keep going, and for a long stretch that bet paid in a way almost nothing else did.
Then, over roughly a month, it went from $45 billion in assets to $10 billion. Leveraged positions in AI and tech soured, the fund was forced into a fire sale, and Citadel bought it out inside a twenty-four hour window.
That happened days after Meta reported a quarter in which revenue rose 28% to $60.8 billion and free cash flow fell 91%, from $8.55 billion a year earlier down to $784 million, on the back of $31.1 billion in quarterly capital expenditure. The stock dropped 10% in extended trading. Two very different events, one uncomfortable question, and a great deal of confident commentary about a bubble.
What actually happened
It is worth separating these carefully, because they are not the same kind of event and conflating them produces the wrong conclusion.
Meta's quarter was not a bad quarter. Revenue beat estimates and grew 28%. What happened is that Meta chose to spend an enormous amount of money, $31.1 billion in three months, on AI infrastructure, and narrowed its full-year capital expenditure guidance to between $130 and $145 billion. Free cash flow collapsed because cash went into data centres rather than because business deteriorated. Mark Zuckerberg also hinted at selling or leasing excess compute, which is an interesting signal about how much capacity is being built relative to current demand.
The hedge fund collapse is a genuinely different animal. That is what happens when leveraged bets on a concentrated theme move against you: the leverage that produced the 400% gain produces the reverse with equal efficiency. It says something about the fragility of financial positions built on AI enthusiasm. It says very little about whether AI tools work.
One is a company spending heavily on a long bet. The other is a trader losing a leveraged one. The commentary that bundles them into a single narrative about the bubble popping is doing something rhetorically convenient and analytically sloppy.
Why this is not evidence of a crash
The strain is real and it is worth taking seriously. What it is not is evidence that the underlying technology is failing, and the distinction matters for how you should act.
Note what is happening on the product side during the same weeks. OpenAI cut its budget model price by 80%. MiniMax released a competitive video model and published the weights. Anthropic shipped a materially more capable model at an unchanged price. Those are not the actions of an industry in retreat. They are the actions of an industry in a brutal fight for market share, which is a different thing and produces different symptoms.
The financial strain and the capability progress are both real simultaneously, and that combination is historically normal. It describes the railways in the 1840s, the telecoms buildout of the late 1990s, and a dozen episodes in between. Enormous capital goes into infrastructure faster than revenue can justify. Some investors lose a great deal of money. The infrastructure remains and gets used by everyone who comes afterward, often at prices well below what it cost to build.
If that is the pattern here, the users of the technology come out ahead and the financiers of it do not. As a small business you are firmly in the first category, and the correct emotional response to overbuilt capacity that you did not pay for is not anxiety.
The risk that actually applies to you
Here is where it gets genuinely relevant, and it is not about markets at all.
The real exposure for a small business is that a service you have come to depend on is currently priced below what it costs to provide, because a well-funded company is buying market share, and that pricing is not permanent. When the money gets more expensive, subsidised pricing is the first thing that changes.
We have already seen both directions of this in a single summer. GPT-5.6 Luna fell 80% because competitive pressure at the budget tier is severe. Claude Sonnet 5 rises 50% on 1 September as promotional pricing ends. The second of those is exactly the mechanism to watch: adopt at an attractive rate, build a dependency, absorb the real rate later. We went through that in detail in what AI models actually cost now.
The other version of the same exposure is the small AI vendor rather than the large one. A tool built by a fifteen-person startup running on venture money, priced attractively to acquire customers, is a genuinely useful tool and also a business that needs to raise again. If capital tightens, the outcomes are a price rise, an acquisition, or a shutdown notice, and all three arrive as an email you were not expecting. This is the vendor availability question we covered in what happens when your AI tool disappears.
How to tell if you are being subsidised
There is a rough test for this and it does not require reading anyone's accounts.
Ask what the compute underneath a tool plausibly costs, and compare it to what you pay. If a service processes long documents, generates video, or runs multi-step agent work for a flat €29 a month with generous limits, somebody is absorbing a difference. That is not an accusation, it is an ordinary growth strategy and it may continue for years. But you should know you are inside it, because the alternative is finding out through a pricing email on a Thursday.
The second signal is funding stage relative to price. A tool from a recently funded startup priced clearly below larger competitors is running the market-share play, which is fine. A tool from an established profitable company at a similar price is more likely to be sustainable, because it does not need the price to change to make the business work.
The third and most practical signal is simply how much of your operation runs through it. A subsidised tool doing something peripheral is not a risk worth managing. A subsidised tool that your quoting process cannot function without is a different situation, and the difference is not the tool, it is the dependency you built on top of it.
Building for the boring outcome
The response that works here is not defensive and it is not expensive. It is designing so that any single vendor changing its mind is an inconvenience rather than an emergency.
The most valuable habit is knowing, for each tool that matters, what you would do if it doubled in price or disappeared next month. Not having a migration plan written up, just being able to answer the question. For most tools the answer is that an alternative exists and switching would take a day, which means there is nothing to worry about. For one or two tools the answer will be that you genuinely do not know, and those are the only ones that deserve attention.
Keep your data somewhere you control, or at least somewhere exportable. The difference between a vendor change being annoying and being catastrophic is almost always whether the accumulated history lives inside their system with no way out. Check that the export exists before you need it, because discovering that it does not is a bad thing to discover during a wind-down notice.
And prefer standard interfaces where the choice is available. A workflow that talks to a model through a common interface can move providers with a configuration change. A workflow built around one vendor's proprietary features cannot. That is not an argument against ever using proprietary features, some of them are genuinely worth the lock-in. It is an argument for knowing which of your dependencies are the sticky kind.
The honest perspective
A hedge fund losing $35 billion in a month is a dramatic story and it is not your story. Meta spending $31 billion in a quarter is a bet on a scale that has no bearing on how you run a business with nine employees.
What both events genuinely tell you is that the money behind AI is under more pressure than it was six months ago, and that pressure eventually reaches pricing. That is worth knowing and it is not worth panicking about. The tools work. They keep improving. The prices at the low end keep falling. None of that is contingent on any particular investor being right about anything.
The businesses that come out of this period well are not the ones that predicted the market correctly. They are the ones that got real value from these tools while remaining able to change their minds when a vendor changed theirs. That is an unglamorous position and it is available to anyone willing to spend an hour asking what they would do if their most important tool sent an awkward email next Thursday.