Enterprise Adoption Is Slower and Stickier Than the Demos Suggest

The chip supply chain is quietly consolidating around a handful of chokepoints: advanced packaging, high-bandwidth memory, and leading-edge fabrication. Any one of these can gate the whole system, which is why capacity announcements from the packaging and memory vendors deserve as much attention as the accelerator launches themselves.

It is worth remembering that the enterprises paying for all of this are not buying models — they are buying outcomes. The vendor that can credibly tie its product to a line item a CFO already understands will win a budget fight against a dozen technically superior tools that cannot.

When a lab ships a new frontier model, the interesting question is rarely whether the benchmark went up. It is whether the price-performance curve shifted enough to unlock a category of application that was previously uneconomical. Watch the pricing page, not the leaderboard.

The headline number everyone fixates on is training compute, but the margin story is increasingly about inference. As model providers push cheaper, faster variants, the cost of serving a query has collapsed by roughly an order of magnitude in eighteen months — and that changes which products are viable to build on top of them.

Open-weight models keep closing the distance to the closed frontier, and each release compresses the premium that proprietary providers can charge. That does not erase the moat — the frontier still leads on the hardest tasks — but it caps how much of the market the leaders can defend at the low and middle tiers.

Memory bandwidth has quietly become the constraint that dictates real-world throughput. You can stack more accelerators, but if the model cannot be fed fast enough, the extra compute sits idle. This is why the high-bandwidth memory roadmap is worth tracking as closely as the flagship chip roadmap.

There is a persistent gap between what a model can do in a demo and what an enterprise will actually deploy. Procurement cycles are long, security reviews are longer, and the switching costs — once a workflow is embedded — cut both ways. The lesson is that adoption is slow to arrive and slow to leave.

Sovereign and regional buildouts are becoming a demand source in their own right, driven less by economics than by the desire not to depend on someone else's cloud. That demand is price-insensitive and politically durable, which makes it a floor under accelerator orders that a pure ROI model would miss entirely.

The bull case for the capex supercycle rests on durable demand and expanding use cases; the bear case rests on the possibility that a great deal of this spending is defensive, undertaken because no incumbent can afford to be the one that under-invested. Both can be true at once, and the timing of the reckoning is the whole game.

The headline number everyone fixates on is training compute, but the margin story is increasingly about inference. As model providers push cheaper, faster variants, the cost of serving a query has collapsed by roughly an order of magnitude in eighteen months — and that changes which products are viable to build on top of them.

There is a persistent gap between what a model can do in a demo and what an enterprise will actually deploy. Procurement cycles are long, security reviews are longer, and the switching costs — once a workflow is embedded — cut both ways. The lesson is that adoption is slow to arrive and slow to leave.

The bull case for the capex supercycle rests on durable demand and expanding use cases; the bear case rests on the possibility that a great deal of this spending is defensive, undertaken because no incumbent can afford to be the one that under-invested. Both can be true at once, and the timing of the reckoning is the whole game.