Nvidia is telling investors the AI-fueled sales surge runs through 2028. Morningstar is telling investors the AI boom is finite and a downturn arrives before 2029. Both of those things can be true at the same time, and that gap — roughly twelve months wide — is where a lot of AI toolkit buyers are about to get caught.
I review tools for a living. I don’t trade semiconductor stocks. So my interest in Nvidia’s forecast isn’t about the share price. It’s about what a three-year demand runway means for the software sitting on top of all that silicon, and what happens to your stack when the runway ends.
Why a chip forecast is a toolkit story
Every AI tool you pay for is a wrapper around compute someone else is renting. That’s not a criticism, it’s just the architecture. When the underlying compute is abundant and subsidized by growth expectations, your tools get generous free tiers, cheap per-seat pricing, and vendors who eat inference costs to buy market share.
Nvidia projecting demand into 2028 is a signal that the money keeping that arrangement alive is still flowing. Vendors will read the same headlines you did. Expect more aggressive land-grab pricing, more products shipped before they’re ready, more “AI-powered” features bolted onto tools that were doing fine without them.
That’s the part I’d flag for anyone building a stack right now. A long demand forecast is not a quality signal. It’s a permission slip for vendors to keep spending on acquisition instead of durability.
The Morningstar half of the picture
SanDisk falling 54% while Morningstar calls the AI boom finite with a downturn before 2029 is the counterweight, and it’s the more useful number for planning purposes. Not because I think Morningstar has better foresight than Nvidia, but because it puts a shape on the thing.
A finite boom means the current pricing environment has an expiration date. Whatever you’re paying today for that agent orchestration platform or that code assistant seat, the version of the market that made that price possible is temporary. When compute stops being cheap relative to expectations, the cost gets pushed downstream. To you.
The JLL 2026 Global Data Center Outlook sits in the middle of these two views, and data center planning cycles are long. Buildings and power contracts don’t flex on a quarterly basis. Whatever capacity gets committed now shows up years later, whether or not demand held.
What I’d actually do with this
Nothing dramatic. But a few habits get more valuable when the ground under your tools is this uncertain:
- Test the export path before you commit. Not the documented one — the actual one. Export your data, then try to load it somewhere else. If it takes an afternoon, that tool has you.
- Treat annual contracts as a bet on vendor survival. The discount is real. So is the risk that you’re prepaying a company whose unit economics only work in the current environment.
- Prefer tools where AI is a feature, not the product. Software with a working business model before the boom will still have one after. Software that exists because inference was cheap is a different bet.
- Watch for silent tier changes. Free tiers shrinking, rate limits tightening, “fair use” language appearing in terms of service. These are early indicators, and they show up before price increases do.
- Keep one boring fallback per critical workflow. Slower, dumber, self-hosted if possible. You may never use it. That’s fine.
The honest uncertainty
I don’t know which forecast is closer to right. Nvidia has visibility into order books I’ll never see. Morningstar has an incentive to be the sober voice in the room and a track record that’s mixed like everyone else’s. The 54% drop in SanDisk shows the market already prices in doubt somewhere in the supply chain, but a single stock move isn’t a thesis.
What I’m fairly confident about is narrower: the tools I review are cheaper right now than their long-run cost, and the people building them know it. Some are using the runway to build something that survives. Others are using it to grow numbers for a funding round. From the outside, those two look nearly identical, which is exactly why the export test matters more than the feature list.
Three years is a long time in this space. Long enough to build something real, and long enough to forget that the pricing you got used to was never permanent. Plan for the second one and you’ll be fine either way.
🕒 Published: