Meta's AI chief could offer only a 'vibe report' and some charts before a check of up to $200 billion

The $100 billion to $200 billion has no stated period, and Zuckerberg's own phrasing was loose: "We're going to write this check for like, I don't know, $100 billion, $200 billion," he said on the Cheeky Pint podcast ^[2]. Cryptobriefing dates the decision to late 2025 ^[6]. Neither account says whether the sum spans one year or several. The answer changes what the 2026 guide means. At $130 billion to $145 billion ^[8], it is 65% to 72.5% of the $200 billion top of the range ^[17] and already above the $100 billion bottom. Cryptobriefing also reports Meta's infrastructure chief saying more than $100 billion goes to infrastructure in 2026 alone ^[10]. If the range is a multi-year total, one year has used most of it. If it is annual, 2026 sits inside it.

Meta did have models, or rather it had models that failed to impress: Bloomberg, as relayed by Quartz, reported that they drew a weak reception in Silicon Valley last year ^[23]. Wang's account of the evidence at the time is short. The best he could offer executives, he said, was a "vibe report" on research progress and some charts ^[4]. Zuckerberg recalled his own question to Wang: "Alex, are you sure this is going to work?" ^[7]

The outside evidence of payback is a download count. Suppose every one of the Muse app's 5 million downloads paid for a full year (downloads are not subscribers, so this is a ceiling). At the $20 tier that is $1.2 billion a year, and at the $100 tier $6 billion ^[18]. Set against the $130 billion low end of the 2026 guide, those are 0.9% and 4.6% ^[19]. The comparison is crude, because capex buys assets that last years ^[14] and the subscription figure is a single year of revenue.

We think the shares' gain of more than 12% this year ^[12] is the price of an option on Muse Wonder holding up. The support for that option so far is Meta's own claim that the model matches ChatGPT and Claude ^[21] and Wang telling Bloomberg the models are "competitive with the best models out there" ^[22]. There are three ways it goes differently. Independent evaluations could confirm parity, and the check looks cheap. Parity could be real and still not pay, if rivals match it and prices fall. Or the model could lag while the fixed costs keep running. The first is the live case against us. We would call the thesis wrong if independent tests put Muse Wonder level with ChatGPT and Claude and Meta shows subscription revenue growing into the billions.