Blame the Grid, Not the GPU

Morgan Stanley’s read on the US data-center power crunch is, in effect, that the two companies everyone watches most closely will barely feel it. The brokerage’s analysts put the net power shortfall through 2028 at 34%, or roughly 32 GW, even after crediting the industry for behind-the-meter generation, fuel cells, and similar workarounds. And yet they do not see those bottlenecks threatening Nvidia or Broadcom’s 2027 forecasts, pointing to the visibility both firms have into where chips get placed, how deployments spread geographically, and how closely data centers, semiconductor suppliers, and the power supply chain now coordinate.

That second half is the interesting part. A 32 GW hole is not a rounding error, and the claim is not that it disappears. The claim is that it gets absorbed somewhere other than the top of the stack.

A shortfall has to land on someone

Think about what a power shortfall actually does to a buildout. It does not usually cancel the compute order. It moves it. A site slips two quarters because the interconnect queue is long, so the racks arrive later, or arrive somewhere else, or arrive in a different configuration than the one originally specified. For a vendor with line-of-sight into placement decisions across many customers and many regions, that reshuffling is manageable. You know roughly how much silicon is going out the door; you just do not know precisely which building it lands in.

For the companies supplying memory, optics, power management, and analog parts, the same reshuffling looks very different. Those components are built against forecasts handed down the chain, stocked ahead of installation, and tied to specific system configurations. When the rack slips, the inventory does not. Morgan Stanley’s framing — that delays in AI deployment hit makers of memory, optical, and other secondary components — is really a statement about who holds the timing risk in this supply chain. It sits lower down, where visibility is thinner and the buffer is physical stock rather than a forecast you can re-aim.

Why this matters for agent architecture specifically

My interest here is not the trade. It is what the exposure pattern says about the workloads we are building toward.

Agentic systems are not shaped like the training runs that justified the first wave of this buildout. A training job is a dense, scheduled, FLOP-bound thing you can site wherever power is cheapest and run for weeks. An agent is the opposite: long-lived context, many sequential model calls, tool invocations that wait on external systems, retrieval against large stores, and memory that has to persist across turns. The binding constraint is rarely raw arithmetic. It is capacity to hold state, bandwidth to move it, and latency between the components that touch it.

Which means agent workloads lean hardest on exactly the parts of the bill of materials that Morgan Stanley flags as most exposed. High-bandwidth memory determines how much context you can keep resident and how many concurrent sessions a node can serve. Optical interconnect determines whether a multi-step agent loop spread across machines stays inside a tolerable latency budget. Power delivery and analog components determine how densely you can pack any of it into a rack that a constrained site can actually feed.

So the comfortable reading — accelerator demand is intact, therefore the AI capex story is intact — is incomplete for anyone designing agent infrastructure. You can have all the accelerators you ordered and still be short the memory depth and interconnect you need for the workload you actually want to run. Compute availability and agent-serving capacity are not the same quantity.

What a 34% gap does to design choices

If you take the shortfall estimate seriously as a planning input rather than a headline, a few things follow for system design.

• Expect more geographic dispersion of inference capacity, driven by where power is available rather than where latency would prefer it to be. Agent architectures that assume tight co-location of model, memory store, and tool endpoints become fragile.
• Expect state management to carry more weight. If memory parts are the exposed link, context compaction, cache reuse, and persistent summarization stop being optimizations and start being capacity strategy.
• Expect uneven hardware generations serving the same fleet. When sites slip and configurations shift, you end up scheduling agent workloads across heterogeneous nodes, which argues for routing logic that is aware of what each node can actually hold.

None of that is a reason for gloom about agent systems. It is a reason to stop treating power as an infrastructure footnote. The analysis says the shortfall will be absorbed, and it will be — just not by the companies whose quarterly numbers get the most attention. It gets absorbed by the components that determine how much state an agent can carry and how fast it can move that state around. Those are the same constraints that decide whether an agent can hold a long task together or loses the thread halfway through.

The grid, in other words, is quietly becoming an architectural variable. Worth designing around rather than waiting out.

🕒 Published: October 6, 2026