Depreciation is the real benchmark.
Not FLOPS. Not memory bandwidth. Not whatever number gets projected onto a keynote screen in forty-foot type. The metric that decides who wins the compute market in 2026 is how slowly a chip becomes worthless, and that is the metric almost every Nvidia skeptic has been ignoring.
The bear case has been consistent for a couple of years now: margins that fat invite competition, hyperscalers will build their own silicon, and a GPU is a commodity waiting to be commoditized. Parts of that case have landed. Reports that Meta might move billions in compute spend from Nvidia GPUs to Google TPUs knocked the stock, and rightly so. That is real money moving to a real alternative. Nvidia has also reportedly delayed its next AI chip over a design flaw tied to the COWOS-L packaging process, involving thermal coefficient differences and warpage. Advanced packaging is where the hard physics lives, and hard physics does not care about roadmaps.
And yet the chips keep outperforming competitors, the market position keeps holding, and the stock is circling record highs. The reason given in most coverage is the surprising durability of the hardware. I want to unpack what durability actually means at the architecture level, because I think it is the most under-analyzed force in this market.
Agents changed the shape of the workload
When the dominant workload was training one large model on one large cluster, hardware value was easy to model. Newest silicon does the frontier run. Older silicon gets shoved toward less glamorous jobs and its economic value falls off a cliff. That is the depreciation curve everyone priced in.
Agent systems break that model. An agent is not one inference call. It is a graph of them, and the nodes are wildly uneven in what they demand:
- Routing and intent classification, which are tiny and latency-sensitive
- Retrieval and reranking, which are memory-bound rather than compute-bound
- Tool-call generation, often handled by a small distilled model
- The reasoning step that genuinely needs the newest, fastest part
- Verification and critique passes, which can run on anything
- Long-horizon state management, where KV cache capacity matters more than raw throughput
Run that graph across a fleet and you discover something operators have been quietly learning all year. A three-generation-old accelerator is not obsolete. It is the right tool for four of those six nodes. It sits behind the same CUDA calls, the same kernels, the same profiling tools, and the same scheduling logic as the newest part in the rack. Nothing needs rewriting to move a workload down a tier.
That is the actual durability story. It is not that the metal lasts longer. It is that a stable software layer gives aging silicon somewhere useful to go. Jensen Huang has been making a bold claim about aging AI chips, and whatever the specific framing, the underlying economics are sound: heterogeneous agent pipelines create demand for a spread of performance tiers, not just the top one.
Why the delay stings less than it should
A packaging-related delay on a next-generation part would be a serious wound for a vendor whose value proposition is peak performance per socket. It matters less for a vendor whose value proposition is a fleet that stays productive for years. If your installed base is still earning its keep, a slipped launch date pushes revenue around a calendar rather than stranding customers.
The same logic explains why platform work matters more than benchmark wins. At GTC 2026, Marco Pavone, senior director of autonomous vehicle research, presented updates to Nvidia Alpamayo, a family of open models, simulation tools and datasets for autonomous driving development. Read that as a chip announcement and it looks like a side quest. Read it as depreciation management and it is central. Every open model, simulator and dataset released into the ecosystem is another workload that will run on hardware customers already own.
What would actually change my mind
I am not arguing the competition is fake. The multi-architecture future is arriving, and TPUs moving out of niche status is the clearest signal we have. But the challenge that would genuinely dent Nvidia is not a competitor posting a better number on one node of the agent graph. It is a competitor whose software layer lets customers reassign old accelerators as fluidly as CUDA does, so that a purchase made in 2026 still has obvious work to do in 2030.
That is a much harder problem than winning a benchmark, which is why the doubters keep getting the timing wrong. They have been modeling a chip company competing on chips. The thing that keeps proving them wrong is a fleet company competing on residual value.
Watch the resale and reassignment behavior of two-generation-old accelerators. That number will tell you more about the next two years than any launch event will.
đź•’ Published: