Remember 2023, when half the AI tools you wanted to test had a waitlist and the other half kept throwing “we’re experiencing unusually high demand” at you mid-prompt? That wasn’t a software problem. It was a GPU allocation problem that trickled down to anyone trying to build something with an API key. Vendors got the hardware they could get, and users got the throughput that was left over.
We may be heading into a similar squeeze from a different direction. Morgan Stanley put out a note arguing that Nvidia and Broadcom are relatively insulated from a worsening US data-center power crunch. The brokerage estimates a 34% net power shortfall through 2028 — about 32 GW — after accounting for workarounds like behind-the-meter generation and fuel cells. That’s the shortfall after the clever fixes.
Who the note protects and who it doesn’t
Morgan Stanley doesn’t see these bottlenecks threatening Nvidia or Broadcom’s 2027 forecasts. The reasoning is about information, not luck: both companies have visibility into where their chips land, they’re expanding across geographies, and they coordinate between data centers, semiconductor suppliers, and the power supply chain. When you can see the whole board, you can route around a dark square.
The exposure sits further down. Any delay in AI deployments hits makers of memory, optical, and other secondary chip components. Morgan Stanley flags memory, optical, power-management and analog suppliers as facing greater inventory risk if data-center buildouts slip. These are the companies that build to someone else’s schedule. When the schedule moves, they’re holding parts nobody needs this quarter.
Morgan Stanley isn’t alone in flagging constraints. Goldman Sachs has also pointed at mounting limits on the US data-center buildout, though GS expects limited near-term impact from political pushback. Morgan Stanley’s list runs wider, naming labor and power among the pressures.
Why a tool reviewer cares about megawatts
I spend my days testing agent frameworks, coding assistants, and whatever orchestration layer launched last Tuesday. None of that work involves a substation. But everything I review runs on inference capacity someone else is buying, and inference capacity is a power story wearing a software costume.
Here’s what a 32 GW gap plausibly looks like from a user’s seat, based on how the last crunch played out rather than any forecast in this note:
- Tiering gets stingier. The generous free tier is the first thing to go when capacity is the constraint. Expect more “available on Pro” on features that used to be open.
- Context windows get rationed. Advertised maximums stay on the pricing page. Actual sustained throughput at those maximums is where the quiet trimming happens.
- Model deprecation accelerates. Running three generations of models in parallel costs power. Vendors under pressure consolidate, and your carefully tuned prompts inherit someone else’s migration timeline.
- Regional availability gets weird. Geographic expansion is listed as a strength for the chip vendors. For users, it can mean your workload lands somewhere with different latency than last month.
I want to be careful here. Morgan Stanley’s note is about chip suppliers and equity risk, not about your API quota. The connection I’m drawing is inference, not reporting. But the mechanism isn’t exotic. Power constrains data centers, data centers constrain inference, inference constrains the tools, and the tools constrain what you can ship.
What to actually do about it
Nothing dramatic. The practical move is the same one that was right before this note existed: don’t marry a single provider.
Build your stack so model swaps are a config change, not a rewrite. If your agent framework hard-codes one vendor’s function-calling format, that’s a dependency with a power bill attached. Measure your own baselines — tokens per second, time to first token, error rates — and log them over time. Vendors rarely announce a quiet capacity change. Your own numbers will tell you before the changelog does.
Pay attention to which tools are honest about their limits. The ones that publish real rate limits, explain their fallback behavior, and tell you which model is serving your request are the ones you can plan around. The ones that hide model routing behind a friendly brand name are fine until the day capacity tightens and you can’t tell what changed.
The headline reads as reassuring if you own Nvidia or Broadcom stock. Read as a toolkit user, it says something less comfortable: the parts of the chain most likely to stumble are the unglamorous ones, and problems in unglamorous places have a way of showing up as a slower response and a higher invoice. Worth watching the memory and optical suppliers, not just the names everyone already follows.
🕒 Published: