Anthropic vs Meta: Room to Be Wrong
The author compares Anthropic and Meta, which are both valued near $1.5 trillion but got there differently. The piece asks how much of each price rests on a core business that already earns money versus a speculative bet on future products and model leads.
- About 60% of Meta's 2026 capex, and roughly two-thirds of the ~$212 billion expected next year, goes to Meta Superintelligence Labs and other AI work whose products haven't yet proven their economics.
- Anthropic's compute share serving paying customers fell from 44% to about 25% between early 2025 and 3Q'26, a bet that research and training will pay off more than near-term serving, though revenue per R&D dollar has started slipping.
- The author estimates Meta's core ads business, excluding AI and Reality Labs costs, implies only about a 10x EV/EBIT multiple, suggesting the market prices in a fair amount of conservatism.
Anthropic and Meta have both landed near a $1.5 trillion valuation, but they got there by different routes. Meta's core is advertising, and the bet is Meta Superintelligence Labs (MSL) and its Muse models. Anthropic's core is a fast-growing API business, and its bet is a set of first-party products that mostly haven't launched. The author's main question is how much of each price rests on the core business and how much rides on the bet. Some investors are already pushing Anthropic's IPO target toward $1 trillion after a recent AI selloff.
On Meta, the author estimates roughly 60% of this year's $135 to $145 billion capex goes to MSL and other AI work, and about two-thirds of the roughly $212 billion consensus for next year. The line is blurry, since MSL models already improve ad ranking and creative tools, but much of the return still depends on Muse or Meta's enterprise platform, and Zuckerberg has said Muse needs to get more than 10 times cheaper to run before it's profitable at scale. Anthropic's version of the same gamble runs through its income statement. The share of its compute serving paying customers fell from 44% in the first half of 2025 to about 25% in the third quarter of 2026, with the rest going to research, training and the free tier. That is a bet that research will make serving cheap enough to pay off, and so far the numbers support it: revenue grew about 29 times while serving compute grew about five times. Revenue per R&D dollar rose from $0.74 to $1.91 before slipping to $1.78 as R&D outpaced revenue.
The author's most interesting material is on efficiency. METR's outside assessment found Anthropic's models gaining capability about 1.9 times faster since October 2025, with roughly 80% of that attributable to AI itself rather than more training compute. Anthropic's R&D spending rose from about $1.9 billion to $9.7 billion over a year, though, and benchmark progress per R&D dollar fell sharply. John Hempton's estimate that Anthropic has spent under $20 billion on compute in its whole history, against far larger figures at OpenAI and Google, suggests Claude is unusually compute-efficient, though the author flags that these comparisons are shaky because OpenAI doesn't file with the SEC and Google's AI spending is buried in a larger business. On a like-for-like basis of compute that hits the income statement, Anthropic has spent about $26 billion since early 2025 and Meta about $16.5 billion. Meta's advantage is owning its fleet, which avoids a cloud provider's margin, but it carries the capital and the risk that Anthropic's rented model avoids.