4 min read

Happy Wednesday!

Anthropic's IPO paperwork leaked, and the loss is so big it made our own bank balances feel almost tidy. SpaceX learned that renting out computers pays better than launching rockets. Meanwhile Claude spent 64 hours designing proteins to fight scorpion venom, which is more focus than we managed all week.

Anthropic made $16.3 billion in revenue in the first half of 2026, up from $1.1 billion a year earlier. That's according to a draft of its IPO filing seen by The Wall Street Journal, which broke the story on Tuesday evening. The same draft shows a net loss of $80.7 billion.

About $72 billion of that loss is a non-cash accounting charge tied to convertible notes and other financing instruments. Real spending is climbing fast too. Anthropic put nearly $9 billion into computing and infrastructure in the first six months of the year, more than the $7.3 billion it spent in all of 2025.

The draft lists roughly $450 billion in cloud compute deals over the next decade. That includes $111 billion with Google Cloud, $110 billion with Amazon Web Services and $31 billion with Microsoft. About 20% of first-half revenue came from just two customers, and Anthropic is expected to list in November.

Ignore the $72 billion charge. No cash left the building, and it mostly shows how fast the company got valuable. The figure I'd look at is that two customers make up about a fifth of all revenue. That's a lot of trust to put in two relationships. Then there's the compute. Anthropic has signed about $450 billion of cloud deals and spent nearly $9 billion on compute in six months. Half-year revenue went from $1.1 billion to $16.3 billion in a year, so the demand is clearly there. I still think the IPO does well. I just wouldn't call it a safe bet.

Lucas

SpaceX reported third-quarter revenue of $14.6 billion on Tuesday, up 161% from a year earlier, according to The Verge. It made $4.4 billion in net income, against a $1.1 billion loss in the same quarter last year.

The big mover was its AI segment, which SpaceX calls Super Intelligence. Revenue there rose 830% to $7.56 billion, which makes it larger than the Starlink business. Analysts had expected $6.47 billion, according to CNBC. In its earnings release, SpaceX credits new cloud services deals, its purchase of Cursor and more Grok use by businesses and consumers.

The segment made $431 million in operating income after about $1.2 billion of stock compensation tied to acquisitions. It also spent $16.4 billion on capital projects in the quarter. SpaceX says its AI arm now has 1.7 gigawatts of nameplate compute.

SpaceX made more from renting out computers than from launching rockets. Its AI unit brought in $7.56 billion this quarter, up 830%, and Wall Street had expected $6.47 billion. It also spent $16.4 billion on AI in those same three months to make $431 million in operating profit. That only adds up if the chip shortage lasts for years. Musk is betting it will. Right now he's probably right. When chips stop being scarce, it gets a lot less fun to be the landlord.

Tom

Anthropic asked an internal research model to design tiny proteins that latch onto scorpion neurotoxins, and the model mostly ran the project itself. Anthropic wrote the prompts with Citrini Research and gave the model one prompt plus four follow-ups. Over 64 hours it generated 318,484 designs and picked 180 to test. The team added 31 more with Claude's help.

Adaptyv Bio and Twist Bioscience tested all 211 in the lab, and 17 met Anthropic's bar for a binder. Each one grabs a single toxin. Anthropic hasn't checked whether they neutralize venom in cells, and it says plainly that this is not an antivenom. Scorpions sting at least 1.2 million people a year and kill more than 3,250, mostly children. Anthropic has released all 211 designs and their data.

Claude came up with 318,484 protein designs. The lab tested 211, and 17 stuck to a toxin. That isn't an antivenom yet, and Anthropic says so. Still, I like the choice. Drug companies mostly skip scorpion stings because there's no money in them. If AI makes the first attempt that cheap, neglected problems finally get one. My guess is a nonprofit takes these designs the rest of the way, slowly.

Lucas