Nvidia’s message, as relayed by Bloomberg this week, is that the AI-fueled sales surge isn’t a 2026 story that fades out. The company expects it to run into 2028. That’s a three-year runway spoken out loud by the vendor with the clearest view of who is buying accelerators and how many.
My first reaction wasn’t about the stock. It was about the six-month-old inference code sitting in my repo. If compute keeps arriving on that schedule, a lot of the workarounds I wrote to survive scarce, expensive GPU time are going to look like relics. That’s not a complaint. It’s a planning problem, and it’s the kind I’d rather have.
What a 2028 forecast actually means for builders
Chip roadmaps and bot architecture are further apart than they sound, but not by much. Every design decision I make in a bot pipeline is a bet on the price and availability of tokens. Batching strategy, caching layers, when to route to a small local model versus a hosted frontier model, how aggressively to prune context — all of that is compute economics wearing an engineering costume.
When supply looks tight, you build defensively. You cache hard, you quantize, you keep a fallback model warm, you treat every API call like it might get rate-limited at the worst moment. When supply looks abundant for years, different choices start making sense:
- Agent loops that make several model calls per user turn instead of squeezing everything into one prompt
- Retrieval that re-ranks with a model rather than pure vector similarity
- Evaluation suites you actually run on every commit, because running a few hundred graded completions stops feeling extravagant
- Longer context windows used deliberately instead of as a last resort
None of that is new capability. It’s the same capability with the cost anxiety turned down. I’ve killed features in my own projects purely because the per-conversation cost math didn’t work. A longer runway means revisiting that list.
The counterargument is already on the board
Same week, a different signal: SanDisk fell 54%, with Morningstar reportedly seeing a finite AI boom followed by a downturn before 2029. Put those two headlines side by side and you get something more useful than either alone. Nvidia’s window and Morningstar’s window nearly touch. One says growth through 2028, the other says the correction shows up around 2029.
Read charitably, both parties might be describing the same curve from opposite ends. Growth now, digestion later. That’s a normal infrastructure cycle, and it’s roughly what the physical side suggests too — JLL’s 2026 Global Data Center Outlook exists because capacity planning at that scale runs on multi-year horizons, not quarterly vibes. Buildings get financed years before the racks land.
For anyone shipping bots, the practical read is: you probably have a few good years of improving unit economics, and you should not architect as if that lasts forever. Those aren’t contradictory instructions. They’re the same instruction with a timer on it.
How I’m actually adjusting
Three changes I’m making to my own work, none of them dramatic:
Keep the abstraction layer, even when it’s annoying
Every project of mine has a thin adapter between application logic and whatever model provider is behind it. It adds friction. It also means a price change, a deprecation, or a capacity crunch turns into a config edit instead of a rewrite. If a downturn does arrive late in the decade, portability is what keeps you shipping.
Measure cost per outcome, not cost per token
Token prices are the noisy metric. What matters is cost per resolved support ticket, per generated report, per completed workflow. Instrument that now, while compute is getting cheaper, so you have a real baseline if the direction reverses.
Treat local models as insurance, not ideology
I run small models locally for a subset of tasks. Not because it’s cheaper today in every case, but because it’s a hedge that keeps working when hosted capacity gets contested.
The demand side is louder than the supply side
One more item from the same news cycle worth sitting with: Bloomberg covered AI influencers entering Brazil’s election campaigns. That’s not a data center story or a chip story. It’s a demand story, and it’s the kind that doesn’t show up in any vendor’s guidance until it’s already happened.
The chips get built because software found something to do with them. Political campaigns generating synthetic personas is one answer, and an uncomfortable one. Customer service bots, coding agents, and internal automation are less dramatic answers. All of them consume the same accelerators Nvidia is forecasting demand for through 2028.
Which is the part I keep coming back to. A three-year forecast from a chip vendor is really a bet that builders keep finding uses. If you’re writing bot code right now, you’re not a spectator to that forecast. You’re an input to it. Build accordingly, and keep the exits marked.
🕒 Published: