A friend of mine runs inference for a small startup on a rack of Nvidia GPUs that are old enough to be embarrassing. Two hardware generations behind. He keeps a spreadsheet of what it would cost to move to newer silicon, and every quarter he opens it, looks at the throughput his aging cards are still delivering, and closes it again. The upgrade never happens because the old stuff refuses to stop working.
That’s a small moment, but it explains something bigger about why Nvidia’s position in 2026 keeps confounding people who called the top two years ago. The bear case was never really about performance. It was about the assumption that hardware this expensive would depreciate on a normal schedule, that competitors would catch up on the next node, and that customers would defect the moment a cheaper option appeared. Some of that is happening. Most of it isn’t happening fast enough to matter.
The durability angle nobody priced in
Jensen Huang has been making a claim that sounded like sales talk when he first said it: Nvidia’s aging chips hold their value and their usefulness far longer than the market expects. From where I sit, testing tools that run on this hardware, he’s mostly right, and it’s less about the silicon than the software stack wrapped around it.
When a chip generation stays supported, stays optimized, and keeps picking up performance from driver and library updates, its practical lifespan stretches. That has a compounding effect on the economics. Every year an older card stays productive is a year the buyer’s original spend looks smarter, which makes the next purchase easier to justify. It’s a quiet flywheel and it doesn’t show up in benchmark comparisons.
Nvidia’s platform work in 2026 reinforces this. At GTC, Marco Pavone, senior director of autonomous vehicle research, walked through updates to Alpamayo, a family of open AI models, simulation tools and datasets for autonomous driving development. Note what that is: not a chip announcement. It’s an ecosystem play. Models, sim tooling, data. The kind of thing that makes a developer’s existing hardware more useful rather than obsolete.
The competition is real, though
I’d be a bad reviewer if I pretended the challenge was imaginary. Reports that Meta might shift billions in compute spend from Nvidia GPUs to Google TPUs knocked the stock, and for good reason. That’s not a startup hedging its bets. That’s one of the largest compute buyers on earth signaling that a second architecture is viable at scale.
The framing I keep seeing, and I think it’s accurate, is that AI compute is becoming a multi-architecture fight rather than a single-vendor market. GPUs won’t own everything. TPUs won’t stay a niche curiosity. Anyone building tooling right now should assume they’ll eventually need to support more than one target.
Nvidia has also had a self-inflicted stumble. The company reportedly delayed its next AI chip over a design flaw tied to the COWOS-L fabrication process, with packaging, thermal coefficient differences and warpage all named as factors. Read that list and you’ll notice none of it is about compute design. It’s advanced packaging physics, the hard, unglamorous part of modern chipmaking where things warp under heat and stop lining up. Delays there are a reminder that Nvidia is bound by the same manufacturing reality as everyone else.
What this means if you’re picking tools
Here’s where I’ll get practical, because that’s the whole point of this site.
- Don’t upgrade on vibes. If your current cards are hitting your latency and throughput targets, the case for new silicon is weaker than the marketing suggests. Measure first.
- Assume portability will matter. If you’re writing anything that touches the metal, budget time for a second backend. The multi-architecture fight means today’s single-vendor shortcut is tomorrow’s rewrite.
- Watch the software cadence, not the launch events. A vendor that keeps shipping optimizations for hardware you already own is doing more for your budget than one announcing a faster part you can’t buy.
- Factor in supply. Capacity constraints are a real planning input, not a footnote. The best chip you can’t get is worse than the adequate chip sitting in your rack.
The uncomfortable read
Nvidia’s chips continue to outperform competitors in 2026, and the company’s market dominance is holding despite a delayed product, a serious rival architecture, and a very large customer publicly shopping around. Doubters weren’t wrong about the pressure. They were wrong about how quickly pressure converts into lost position.
What’s actually keeping Nvidia ahead is the least exciting thing about it. Hardware that stays useful. Platforms that keep getting fed. A stack that makes last year’s purchase feel fine. My friend’s embarrassing old GPU rack is the whole argument, sitting in a closet, still churning through requests.
đź•’ Published: