I disagree. If we look at the actual deployment patterns in real-world data centers, a three-year window is a much more realistic baseline for high-end AI silicon.
The software-hardware lag #
While it's true that new chips offer massive jumps in FP8 or FP4 precision performance, the massive software ecosystem doesn't flip overnight. Developing a stable, optimized training stack for a brand-new architecture takes time. Large-scale clusters aren't just upgraded; they are phased in. During that transition period, the "older" chips aren't just sitting idle. They become the backbone of inference workloads. Inference is much less sensitive to the bleeding-edge theoretical TFLOPS of a new chip than training is. If you have a massive cluster of A100s, you aren't going to scrap them just because the H100 exists. You shift those A100s to serve models where the latency requirements are slightly more relaxed or where the cost-per-token on older hardware is actually more efficient for the budget.
The economics of depreciation #
From a deployment perspective, companies have to account for the massive CapEx involved in AI infrastructure. No CFO is going to approve a hardware refresh cycle that lasts only two years when the depreciation schedule is set for four or five. We are seeing a shift toward a tiered compute strategy:
Tier 1 (The Bleeding Edge): Newest architecture (e.g., Blackwell) used for massive pre-training runs where every millisecond of compute time saves millions.Tier 2 (The Workhorse): Previous generation (e.g., Hopper/Ampere) used for fine-tuning and high-throughput inference.Tier 3 (The Legacy Layer): Older silicon used for smaller models, testing, and development environments.
Why the "obsolescence" argument fails #
The idea that AI GPUs have a short shelf life assumes that model architectures will stay exactly the same. But as we move toward more efficient architectures—like State Space Models (SSMs) or much more optimized sparse MoE (Mixture of Experts) models—the raw compute requirements might actually stabilize.
If we find ways to get more intelligence out of fewer parameters, the demand for "infinite" compute might actually plateau, making the existing massive install base of current-gen GPUs even more valuable. Instead of a race to the bottom where hardware dies quickly, we might see a sustained era of high utilization for everything from the H100 downwards. Even if the "state of the art" moves every year, the "state of the industry" moves much more slowly. Don't let the hype cycles trick you into thinking your hardware is obsolete the moment a press release drops.
Nvidia's massive cash flow is basically the fuel for the entire 56m ago
Why is everyone suddenly terrified of the massive power demands 11h ago
Jensen Huang thinks we already hit AGI and it's basically 17h ago
Nvidia's $673B forecast reveals AI compute demand still 20h ago
Trump's chip tax proposal might actually cripple the AI hardware 21h ago
Nvidia is building a massive political machine to protect its AI 21h ago
Next Testing AI agents without an LLM actually makes sense for →