Situational Awareness sold its stocks, kept Anthropic. Why?
The Situational Awareness hedge fund dumped its public AI stocks but held its $5B Anthropic stake. Here's what that split signals for anyone building on AI APIs from Sri Lanka.
The Situational Awareness hedge fund just sold most of its public stock portfolio and kept its Anthropic shares, and that ordering tells you more about the AI trade than any of the headlines about the loss did.
TechCrunch reported the unwind on 30 July 2026 in this piece on the fund's forced sale. I'm not commenting on the trading. I'm commenting on what a small builder should read into it.
🔍 The numbers behind the unwind
Leopold Aschenbrenner, a 25-year-old former OpenAI researcher, started the fund in 2024 with no prior trading experience. Per TechCrunch, the arc looks like this:
| Marker | Figure |
|---|---|
| Return through June 2026 | 439% |
| Peak assets | ~$45 billion |
| Assets before the sale | ~$20 billion |
| Assets after selling to Citadel | ~$10 billion |
| AI infrastructure equities, one month | down 30%+ |
| Anthropic stake (private, retained) | $5 billion |
The stocks that got hit were the picks-and-shovels names: SK Hynix, Sandisk, Bloom Energy, Nebius Group. Leverage turned a bad month into a forced one. Ken Griffin's Citadel took the book.
Key takeaway: The AI selloff wasn't a verdict on whether models work. It was a verdict on whether the companies spending billions on data centres will see revenue soon enough to justify it. Those are completely different questions, and only one of them affects your API bill.
⚡ Public panic, private conviction
Look at what got sold versus what got kept.
- Sold: memory, storage, power, cloud compute. The physical layer.
- Kept: Anthropic ($5B), MatX (AI chips), Fluidstack (AI data centres, in talks at an $18 billion valuation in April).
Fluidstack is a data centre company, so this isn't a clean "sold hardware, kept software" story. The real split is liquidity. Public positions get marked to market every second and margin-called on the way down. Private positions don't. When your leveraged public book breaks, the private book is the part you're allowed to keep believing in.
That's worth remembering when you read confident takes about "the market losing faith in AI." A forced seller sells what's sellable, not what they're least convinced by.
💰 What it means for what you pay per token
Short version: nothing yet. Slightly worse later, if anything.
The cheap inference you and I use is subsidised by capital that is currently being questioned. If capex financing gets more expensive, the pressure lands on price and on free tiers eventually, not this quarter.
Here's how I'd separate signal from noise:
| Signal | Does it move your API bill? |
|---|---|
| SK Hynix / Nebius share price falls 30% | No |
| A fund gets margin-called | No |
| Labs raise less capital for 3–4 quarters | Yes, with a long lag |
| Free tiers quietly get rate-limited | Yes, immediately |
| A lab publishes real inference gross margins | Yes, it sets the price floor |
If you're building anything with an LLM in it, know your actual cost per request before you worry about macro. Our AI token counter and embedding cost calculator exist for exactly that, and the AI model comparison page lets you check whether a cheaper model would survive your workload.
🛠️ How I'd de-risk an AI side project built from Sri Lanka
We're at the far end of this supply chain. We don't set prices, we don't get early access, and we pay in USD out of LKR income. So the practical moves are defensive:
- Measure cost per user action, not cost per token. "Rs 4 per generated CV" is a number you can defend to a customer. "$0.003 per 1K tokens" is not.
- Keep the model behind an interface. One
callModel()function, one place to swap providers. If a free tier disappears on a Tuesday, this is a config change instead of a rewrite. - Don't build a business that only works at today's price. If your margin dies when inference costs 2× more, you don't have a product, you have an arbitrage.
- Push work to the client where you can. Anything that runs in the browser costs you nothing per request forever. That's why a big share of our tools are client-side.
- Price in your own currency risk. If you're freelancing on top of this, the USD-LKR earnings calculator will show you what a rate change actually does to your take-home.
If a single provider's pricing page can kill your project, that provider is a co-founder you didn't agree to take on.
📊 The October IPO is the number to actually watch
TechCrunch notes Anthropic is expected to go public as soon as October, potentially at a higher valuation than the fund's entry. That matters to us for a boring reason: an IPO comes with financial disclosure.
Right now nobody outside these companies knows what inference actually costs to serve. An S-1 would give us a real read on:
- Gross margin on API revenue
- How much of revenue goes back into compute
- Whether enterprise contracts or consumer subscriptions carry the business
That's the first honest input any of us will have had for cost planning. Every pricing prediction before it, mine included, is inference from the outside.
I'd also flag the timing. Per TechCrunch, on 24 July Aschenbrenner invited investors to commit fresh capital starting 1 August. A fund raising into an expected October listing of its largest private holding is a fund with a plan, not a fund in retreat. Read the loss headline accordingly.
💡 What this means for you
If you're a student, an engineer, or running a two-person team here, the useful conclusions are narrow:
- Nothing about your stack broke this week. The models you use are unaffected by a fund's margin call.
- Cheap inference is a funding decision, not a law of physics. Build as if it could get more expensive, because the capital funding it just got a scare.
- Free tiers are the first thing to tighten. They're a marketing line item. Have a fallback and know what it costs.
- Wait for the S-1 before believing anyone's cost analysis, including this one.
The AI trade wobbling in public markets doesn't change what you can ship from a laptop in Colombo tonight. It just changes how confident you should be that tonight's prices will hold next year.