Meta raises CapEx by $10B — and its stock drops 8.6%

For a software company, lifting the annual investment budget is usually read as confidence. For a hyperscaler, the same move is now read as a receipt. The market no longer rewards the spending. It checks the composition.

CapEx means two different things. For a software firm it is almost zero: laptops, leases, cloud subscriptions. For a datacenter operator it is concrete, steel, GPUs, and 20-year power purchase agreements. When Meta raised its 2026 CapEx guidance from $115–135B to $125–145B on its Q1 2026 call, the extra $10B was not a rounding error. It was a down payment on physical scarcity. The stock fell 8.6% that day, according to Trefis. The market did not punish growth. It questioned where the marginal dollar was going.

Alphabet followed the same path. After guiding 2026 CapEx up to $185B in Q4 2025, it raised the ceiling again to "up to $190B" in Q1 2026 and warned of a "significant" increase in 2027. Google Cloud revenue grew 63%, yet the conversation is shifting from top-line growth to what that growth costs.

Consensus estimates for the four hyperscalers in 2026 range from roughly $474B to $750B depending on the source and date. The dispersion itself is the signal: nobody knows exactly how high spending will go, but everyone agrees it is rising fast. What now matters is the share of that envelope reserved for power, land, and interconnection rather than silicon.

The useful conversion is not CapEx to revenue. It is CapEx to megawatts. Building datacenter capacity costs roughly $9–15M per MW, per Alpha-Matica. At that range, Meta's extra $10B implies about 0.7–1.1 GW of new power — roughly the scale of its Prometheus cluster. That is not a chip order. It is an order for electrons, land, and grid access.

Scale this up and the picture becomes physical. US AI datacenters are expected to need 20–30 GW by the end of 2027, according to Epoch AI. That is roughly 5% of US generation capacity, but about half of France's entire grid. PJM, the largest US grid operator, now attributes 94% of its peak-load growth to datacenters. The constraint is no longer the ability to buy GPUs. It is the ability to plug them in.

The contracts show the transmission channel. Meta has signed more than 6 GW of nuclear agreements, including a 20-year PPA with Constellation (CEG) at Clinton. Google has a 25-year PPA with NextEra (NEE) tied to the restart of Duane Arnold and a 615 MW plant. Vistra (VST) keeps appearing as the utility counterparty in hyperscale deals. These are the names inside the NBC Compute Index that collect the marginal dollar as it migrates from the hyperscaler to the electron.

Meta, Alphabet, Microsoft and Amazon are not in the index. They are the upstream signal. The investable consequence sits one layer down, in L1 energy.

This is not a proven law. It is a corroborated thesis built on published guidances, disclosed PPAs, and grid data. If the Q2 calls confirm that the CapEx uplift is landing in power contracts and grid access, the read strengthens. If the calls emphasize GPU volume instead, the chip story stays intact and energy remains a secondary play. Either way, the market has already voted: it will not cheer topline spending. It will price the composition.