Meta's $20B capex revision points at memory, not GPUs

The post-call coverage framed the print as a spending shock: AI capex going up, returns uncertain, stock down about 7% in extended trading. That framing is true at the strategy layer. It is not the most useful framing at the infrastructure layer.

Three numbers do the work, and they cross-reference cleanly across one quarter.

First, Meta's own disclosure. The capex revision is +$10B at the midpoint and +$10B at the top end of the range. The press release names "higher component pricing" generically. The earnings call goes one step further: Zuckerberg names memory specifically. That is the first time in this cycle a top-five hyperscaler has put memory in front of GPUs in named-component prepared-remarks commentary.

Second, the supplier side. SK hynix reported a 198% year-over-year revenue print at a 72% operating margin. On the call, the company said HBM4 demand over the next three years exceeds its production capacity, and the DRAM marketing VP framed the buyer behavior as "prioritizing volume over price." Read together, supplier and buyer are describing the same constraint from opposite sides.

Third, the spot indicators. TrendForce reports DRAM contract prices rising 55–60% QoQ in Q1 2026. HBM3E module supply is reported as fully booked through year-end 2026, with HBM4 ramping at SK hynix's M15X line in Cheongju and Samsung's P4 in Pyeongtaek. The 2.5x-off-2025 framing in earlier secondary coverage is not directly sourced; the verifiable shape is "Q1 2026 alone +55–60% QoQ on conventional DRAM, with server DRAM compounding faster on top."

The verified facts, restated: a top-five hyperscaler raised its 2026 capex by roughly $10B at the midpoint and named memory specifically as a cost driver in CEO prepared remarks. The dominant memory supplier reported a 198% year-over-year revenue print, said HBM4 demand exceeds three-year supply, and runs a 72% operating margin. Advanced packaging at TSMC is reported as fully booked.

The read: the cost mix of an AI training and inference build is migrating from logic to memory. The 2023–2024 story, where one GPU SKU bottlenecked everything, is being replaced by a memory-bandwidth story, where the binding inputs are HBM stacks, advanced packaging slots at TSMC's CoWoS lines, and the back-end capacity at memory fabs in Korea. A hyperscaler that wants to bring inference capacity online in 2027 cannot solve that with a GPU order alone. The order has to clear three queues at once: silicon, memory, and packaging. Meta's revision is what it looks like when the memory queue starts pricing.

What this would mean if it holds: the right way to read hyperscaler capex over the next four quarters is not "AI spending up" but "memory share of capex up." Models that backed into capex from GPU unit volumes will under-call total spend because they will under-weight HBM cost per accelerator and over-weight logic cost per accelerator. Forecasts that anchor on GPU shipment counts alone will miss the same way that 2022 hyperscaler capex models missed by anchoring on data-center square footage and ignoring transformer lead times.

There is also a structural read on the supplier side. SK hynix's stated three-year HBM4 imbalance is a planning input, not a marketing line. Long-dated HBM pricing power is now a binding factor in hyperscaler ROI calculations on AI infrastructure. Either hyperscalers absorb it through capex (which is what Meta just did), or they pass it through to enterprise AI compute pricing (which lifts the floor for inference cost-per-token and reshapes which workloads are economic to deploy at the edge versus in the cloud).

Reading: this is the first quarter where the published infrastructure-cost arithmetic for large-model inference has memory, not logic, as the dominant marginal cost. That is the layer-beneath-the-label change in this cycle.