Meta's 2026 Capex Plan More Than Doubles Two Years of Spending; the Product It Funds Is Still on Paper

The person this lands on is whoever keeps the model routing config: a few vendor rows, a default, a fallback, and a note about which version the prompts were tuned against. Adding a Meta row is a two-minute edit. Defending it in a quarterly review is the hard part, and the arithmetic is why.

Take the low end of the plan. If $130bn in 2026 is more than double what Meta spent on capital across 2024 and 2025 together, those two years came to under $65bn combined ^[4][1]. The Louisiana site alone, at $50bn, is roughly 38% of that low-end year ^[5][2]. The plan's midpoint, $137.5bn, is about 8.7 times the $15.8bn of net income Meta reported for the second quarter of 2026 ^[12][3]. The bulk of it goes to computing resources rather than to anything a user opens ^[10]. Fast Company says Meta declined to put executives on the record for its piece ^[16].

Teams like to imagine users running the eval, comparing outputs, and switching to whichever model wins on merit. In practice, most pin one version and shape their prompts around its quirks until something forces a move. Three Muse Spark releases in under four months works out to a new flagship every six weeks or less ^[15][4]. For a pinning team, that cadence is a cost, and it is paid in re-evaluation time, not in licence fees.

Now separate the thing being pitched from the thing being done. The pitch is an individual directing superintelligence, with small-business owners and an 8-year-old named as the beneficiaries ^[1][2]. What Meta is actually building is a frontier lab assembled by purchase: it paid $14bn for 49% of Scale AI, which implies about $28.6bn for the whole supplier without a controlling stake ^[6][5], installed Alexandr Wang atop Meta Superintelligence Labs ^[7], and has reportedly offered first-year packages of $100m or more ^[9]. The manifesto's unit of value is a person holding a capability. Nothing in Fast Company's inventory of what has shipped names the screen where that person meets it.

The bill is paid by advertising, which was 98% of second-quarter revenue, and eMarketer expects Meta to pass Google as the largest digital ad company by the end of 2026 ^[12][13]. My read: while the consumer surface stays unnamed, the surface with a P&L attached is ranking and targeting, which no user directs and none can inspect. That is not the thesis in the manifesto.

So score each vendor row on two axes. First, whether you can name the end-user surface where that vendor's own stated thesis gets tested, because a named surface is something you can watch fail. Second, whether you could swap the row out inside one sprint. Meta currently sits in the quadrant of large capability and unnamed surface, and that quadrant has a legitimate use: interchangeable capacity, priced per token, re-evaluated on their release schedule rather than yours. The trade-off is worth saying out loud. If the distributed version does arrive and you treated Meta as a commodity, you will be behind whoever co-designed with them, and you will have banked every migration you did not have to run in the meantime.