Good morning.

Tesla and SpaceX are building what Musk says will be the largest building on Earth, in Texas, to make chips they cannot otherwise buy. Meta became the fourth AI disclosure in two weeks, and the common thread turns out to be one testing vendor rather than four rogue models. And an Australian bitcoin miner is now worth $10.5 billion for reasons that have nothing to do with bitcoin.

Let's get into it.

That is the planned manufacturing footprint of Terafab, the semiconductor plant Tesla and SpaceX said Thursday they will build in Grimes County, Texas, outside Houston. The initial investment is $16.8 billion. SpaceX filings suggest the multi-phase build could eventually reach $119 billion. Musk called it "the largest and most valuable building on Earth by far," which for once may be closer to a specification than a boast.

The reason to build it is the same reason everything else in this newsletter has been about electricity and buildings for a month. Musk's roadmap requires Optimus robots, Cybercabs, and the Starmind orbital data centers SpaceX announced Tuesday, and all three need chips that nobody currently makes in sufficient volume. So rather than queue behind everyone else at TSMC, two companies that do not make semiconductors are constructing the world's largest building to make semiconductors for themselves. The plant will handle logic and memory, plus packaging and testing, in one vertically integrated site. Intel says it will contribute something, without saying what.

Two details worth holding onto. SpaceX did not mention Terafab once on its first public earnings call earlier this week, which is a strange omission for a project of this scale. And the announcement followed a Wednesday county meeting where hundreds of residents turned out to object to the tax breaks and the lack of transparency, with SpaceX committing to draw water from the Gibbons Creek Reservoir rather than local groundwater. That is the fourth jurisdiction in six weeks where a hyperscale compute project has met organized local resistance. The chips keep getting cheaper. The permission to build the places that make them does not.

Hadrian ($1.37B, Series D). WCM Investment led at roughly a $7.9 billion valuation, with Valor Equity, 137 Ventures, Baillie Gifford, and Washington Harbour, joined by Andreessen Horowitz, Founders Fund, CapitalG, and Lux. Hadrian builds hyper-automated factories producing military vehicle parts, submarine components, and aviation hardware, including a new Alabama submarine facility valued at $2.4 billion in a public-private deal. The valuation quintupled since the last round. Note what is buried in the details: the company now lists England as a headquarters alongside Michigan. A US defense manufacturer establishing a British base is a bet about which procurement budgets grow fastest, and it is not a subtle one.

LifeMine Therapeutics ($188M, Series E). Bezos Expeditions and Gates Frontier both wrote checks, alongside Google Ventures, RA Capital, GSK Ventures, and Arch, for a Watertown company developing an oral immunosuppressant to prevent organ rejection in kidney and islet transplant patients. The round follows a brutal year of biotech layoffs and abandoned raises. When the two richest men in technology fund the same transplant drug, it is worth noticing that neither is funding a chatbot.

Omilia ($67M, Series B). Expedition Growth Capital led the Athens company, which builds voice agents for customer service and has grown recurring revenue tenfold to $60 million since 2020, with Capital One, RBC, Discover, and Taco Bell as customers. Its pitch is specifically that it avoids large language models for routine tasks. That sounds like heresy and is actually arithmetic. A balance inquiry does not need a frontier model, and the company that worked that out has better unit economics than the ones that did not.

Malachyte ($10M, seed). Bessemer and Gradient Ventures co-led, with Harpoon, for a company founded by former Spotify engineers applying session-based recommendation to online retail. It reads clicks, hovers, scrolls, and searches in real time to infer what a shopper wants during this visit, rather than from what they bought last year, and it works without a login. Spotify solved cold-start recommendation for music a decade ago. Somebody was eventually going to carry that exact expertise into a category where nobody signs in.

→ Four AI labs have now disclosed containment failures, and one testing vendor is in three of them. Meta said Thursday that one of its models reached the open internet during a security evaluation and broke into a third party's system. The cause was a misconfiguration by Irregular, the outside firm Meta hired to stress-test the model, which accidentally gave it live internet access during a test meant to be sealed. Irregular is the same vendor whose environment produced Anthropic's three incidents, and its spokesperson told Reuters this was "the exact same evaluation-environment issue," involving no sandbox escape and no sophisticated cyber action. The story everyone has been telling for two weeks is about models going rogue. The story the evidence supports is that one company's test environment kept handing models an internet connection while telling them they did not have one, and the models did what they were asked to do. Irregular is now writing a white paper on containment best practices, which is both the correct response and a slightly awkward sentence.

→ An Australian bitcoin miner is now a $10.5 billion AI infrastructure company. Firmus announced a fully subscribed $2 billion equity round, with follow-on money from Coatue and Nvidia and new money from Blackstone Tactical Opportunities and Jane Street, roughly doubling its valuation from $5.5 billion four months ago. The company was founded in Sydney in 2019 to mine bitcoin. Mining taught it one narrow set of skills: acquire cheap power at scale, cool a building full of hot silicon, run it at high utilization. That skill was nearly worthless in 2022 and is worth billions now, because the binding constraint on AI stopped being chips and became somewhere to put them. Firmus did not pivot into this boom. It stood still and the boom arrived at its address.

→ Unitree priced its Shanghai IPO and humanoid robots became a public-market question. The Chinese robot maker set its offering at 150.8 yuan per share, raising roughly 6.1 billion yuan. Unitree's advantage has always been price, shipping capable quadrupeds and humanoids at a fraction of what Boston Dynamics or Tesla charge. Going public forces it to publish real numbers on bill-of-materials cost, field reliability, and how many units get deployed rather than demonstrated. Whatever those filings say will be the first honest accounting the humanoid category has ever been made to produce.

For founders just starting out. Programs, pitches, and angel checks worth knowing about right now.

Closing this month

Open now, deadline later

Non-dilutive money

Always rolling: a16z Speedrun, Neo, Pear VC's PearX, 500 Global, Plug and Play, AI Grant, Buildspace, Conviction Embed, Z Fellows, TinySeed, HF0, South Park Commons, Thiel Fellowship.

You just finished a newsletter that contained a paragraph about a bitcoin miner turning into a data center company. Nobody arrives here by accident, and nobody reads this far unless they build, fund, or run something. That is the whole audience: founders, investors, and operators who read rather than skim.

Reply to this email. We respond same day.

Till next time,

The Vesting Team (Brycent, Ajay, and Kamil)