Nvidia is not forecasting demand for silicon through 2028; it is forecasting that agentic systems will keep failing to run cheaply, and that is a very different claim.
Let me back that up. Bloomberg reports Nvidia predicting an AI-fueled sales surge extending into 2028. Meanwhile, Tech Times notes SanDisk fell 54% as Morningstar projected a finite AI boom before a 2029 downturn. Two forecasts, roughly the same horizon, opposite emotional registers. The gap between them is not really a disagreement about AI adoption. It is a disagreement about what AI workloads will look like architecturally three years out.
Why a multi-year forecast is an architectural claim
Any hardware vendor projecting demand years ahead is making an implicit bet on the shape of the compute graph. Training runs are lumpy, capital-intensive, and concentrated in a handful of labs. If Nvidia’s confidence rested purely on frontier model training, the forecast would be fragile, because that spending depends on a small number of decisions by a small number of buyers.
A forecast that extends into 2028 reads more naturally as a bet on inference volume, and specifically on the kind of inference that agent systems generate. That distinction matters technically. A chatbot turn is one forward pass. An agent completing a task is a loop: plan, call a tool, observe the result, re-plan, retry when the tool returns something unexpected, summarize, verify. Every one of those steps is a forward pass, and most of them carry a growing context window along for the ride.
The multiplier nobody prices in
Here is what makes agent workloads structurally different from what came before. In a single-turn system, compute scales roughly with user count. In an agentic system, compute scales with user count multiplied by task complexity multiplied by retry depth. That last factor is the one that behaves badly. When a tool call fails or a plan step produces an ambiguous result, the system does not consume a little more compute. It consumes another full cycle, often with a longer context than the last one.
So the honest version of a bullish multi-year forecast sounds something like this: we expect agents to be widely deployed, and we expect them to remain inefficient enough that the compute bill keeps climbing. Both halves are required. Wide deployment with sharply improving efficiency produces a very different revenue curve.
What the bear case is actually arguing
Morningstar’s finite-boom-before-2029 view, as reported, is often read as skepticism about AI demand. I read it as a bet on the efficiency curve instead. If you believe inference cost per useful task falls fast enough, then demand can keep growing while hardware revenue flattens or dips. The two forecasts can both be describing real dynamics on different variables.
The things that would push in the bears’ favor are all architectural, not commercial:
- Model routing that sends easy steps to small models and reserves large models for genuinely hard reasoning
- Caching and memory layers that stop agents from re-deriving the same context on every loop iteration
- Better tool interfaces that reduce retry depth, because most retries are failures of specification, not failures of intelligence
- Structured planning that replaces open-ended reasoning loops with bounded state machines where the problem allows it
- Distillation of task-specific behavior into smaller models once the task pattern stabilizes
None of these are speculative. They are all things teams building agents are already reaching for, mostly because their own cloud bills forced the issue. The open question is aggregate pace, not feasibility.
The infrastructure signal
JLL’s 2026 Global Data Center Outlook sits in this same picture from a different angle. Data center capacity is committed years before it is used, on leases and power agreements that do not flex quickly. That creates a real asymmetry: the physical layer bets on a demand curve, and if agent efficiency improves faster than expected, the buildings still exist. Capacity commitments are a slower-moving, higher-conviction version of the same forecast Nvidia is making verbally.
For anyone designing agent systems, this is the practical read. The infrastructure being built assumes your agents will be expensive to run. Your architecture decisions determine whether that assumption holds for you specifically. Teams that treat inference cost as a first-class design constraint, rather than a line item to optimize later, end up in a structurally different position than teams that scale loop-heavy agents on frontier models and hope prices fall.
What I would watch
Not chip revenue. Watch cost per completed task in production agent deployments, and watch whether retry depth trends down as tool ecosystems mature. Those two metrics sit underneath both forecasts. Nvidia’s 2028 view holds if they stay stubborn. Morningstar’s 2029 view holds if they bend.
The forecast fight looks like a disagreement about AI’s future. It is closer to a disagreement about how quickly we get good at building agents that do not waste compute. That is a solid engineering question, and the answer is being written right now in codebases, not in earnings calls.
🕒 Published: