Quick Answer: Nvidia still captures about 80 percent of AI accelerator sales, but its lead is narrowing. AMD's share rose from roughly 2 percent in 2022 to 7 to 10 percent in 2026, cloud giants are building their own chips, and AI tools can now port code to rival silicon in hours. Nvidia's problems are structural, and they predate the CEO change.
Nvidia's revenue went from $26.91 billion in January 2022 to $215.94 billion in January 2026. Those numbers look like a company at its peak. But the real risk to Nvidia isn't on its income statement yet. It's in what customers are building, where they're building it, and how easily they can leave. The WIRED report describes a company that owns the current market while its rivals line up for the next one. Here's what that means for Nvidia AI chip competition, and why the timing of Huang's exit makes it harder.
The Lead Is Real, but It's Narrowing
Nvidia still dominates. Silicon Analysis estimates it captures roughly 80 percent of AI accelerator sales. That's an enormous share, and it's why the company's position looks unshakable from the outside.
The trend underneath is what matters. AMD's share rose from about 2 percent in 2022 to between 7 and 10 percent in 2026. That's a fivefold gain in a few years, and AMD is now pushing hard with its own AI chips.
Here's the part investors often skip. Nvidia's problem isn't that it's losing sales today. Its problem is that the chips it sells now were designed two to three years ago. As one AI company CEO told WIRED, "Everything sold today was designed two-to-three years ago." Whoever runs Nvidia has to make bets about 2028 hardware right now, and those bets are less certain than the current revenue suggests.
Cloud Giants Are Building Their Own Silicon
Google, Amazon, and Microsoft are all designing specialized chips to train and serve AI models. Each has seen its own share grow. Their motive is simple: lower costs and less dependence on a single supplier.
For these companies, every custom chip is a negotiating lever. Even if a Google TPU or an Amazon Trainium chip doesn't replace every Nvidia GPU, it caps what Nvidia can charge for the workloads it does lose. That's a cost-and-leverage effect, and it compounds over time.
Startups add pressure too. Cerebras sells oversized chips built for giant AI models. AMD recently bought Taalas, a company that burns model weights directly into silicon to make computation far more efficient. These designs promise real gains, but they come with a catch. Most cutting-edge models were written for Nvidia's software stack, so using these chips often means rewriting significant amounts of code.
That rewrite cost is the main reason adoption is slower than the hardware specs suggest. It's also the reason the next section matters.
Inference Is Where Nvidia's Edge Thins
Training a model is one job. Running it for millions of users is another, and it's called inference. Patrick Moorhead, chief analyst at MoorInStrat, argues that AI agents make inference more important, and inference is where Nvidia has less of an edge than cloud hyperscalers and startups.
Why does this matter? Inference economics depend on cost per token served, latency, and power draw. A chip that's excellent for training a frontier model may be overbuilt for serving it. Customers who run huge inference fleets have more reason to test alternatives.
Nvidia isn't ignoring this. It acquired the specialized chip startup Groq in December 2025, which signals that the company sees the inference gap. The question is whether one acquisition closes a gap that's opening across the whole category.
Moorhead also points to demand for general-purpose CPUs, driven by AI assistants and agents. That's a market where Nvidia barely competes. If agent workloads pull more spending toward CPUs, Nvidia's share of the total AI compute budget could shrink even if its GPU sales hold steady.
| Player | Position (per source) | Main lever | Main constraint |
|---|---|---|---|
| Nvidia | About 80% of AI accelerator sales | CUDA software ecosystem and an installed base | Current chips designed 2 to 3 years ago; weaker in inference and CPUs |
| AMD | Share up from about 2% (2022) to 7 to 10% (2026) | Lower-cost GPUs; bought Taalas for weight-in-silicon chips | Still a small share; customers must port code |
| Google, Amazon, Microsoft | Each reports share gains from in-house chips | Cost control and independence from Nvidia | Chips serve their own clouds; limited sales to others |
| Cerebras | Sells oversized chips for giant models | Speed on very large models | Requires rewriting code for cutting-edge models |
| Nvidia (Groq deal) | Acquired Groq in December 2025 | Specialized inference hardware | Integration risk; inference gap still widening across the field |
China Is Already Slipping Away
Nvidia's China story is a clear example of how policy can override technical strength. Washington restricts the most advanced chips from being sold there, so Nvidia can sell only previous-generation parts. Beijing is also pushing its own chip industry forward.
The numbers are stark. According to Nvidia's filings, China now accounts for only 8 percent of its overall revenue, down from 20 percent in 2024. Sales of the older chips and non-AI products doubled over the past year, but that growth didn't offset the loss.
Huawei is the company to watch. Despite years of US sanctions, it has built new kinds of supercomputers that link many more chips than usual to train and run AI models. Z.ai, which makes one of the world's strongest open-source models, says it trained its models entirely on Huawei chips. That's the company's own claim, and independent verification is still limited. Still, if Chinese open-weight models gain traction abroad, they could pull demand toward Chinese hardware in other countries.
There's also a murkier issue. Megaspeed, a Singaporean company, reportedly imported between $4.6 and $7.5 billion in Nvidia chips into Southeast Asia. Reports suggest some of those chips may have reached Chinese AI firms. Nothing here is proven in the source, and the company's executives haven't been convicted of anything. But it shows how hard it is to control where advanced chips end up once they leave the US.
Babak Minovi of the Eurasia Group argues that Huang had a point. Cutting off China, he says, pushes Chinese firms toward domestic rivals. Huang's position was pragmatic. The business cost of that position, though, is now visible in Nvidia's revenue mix.
The CUDA Moat Is Easier to Cross Than You Think
Most analysts still describe Nvidia's CUDA platform as a fortress. Developers learned to write for CUDA over more than a decade, and that switching cost has kept AI companies loyal. It's the default explanation for why Nvidia's lead lasts.
Here's where the conventional wisdom breaks down. AI models are now good enough at reading and writing code that porting work, which used to take engineering teams months, can happen far faster. The moat isn't gone, but it's shrinking in exactly the place where AI is strongest.
Consider Wafer, a startup that uses AI to port models to AMD hardware and serve them more cheaply. Its CEO, Emilio Andere, says that every time a major new open-weight model ships, Wafer can have it running on AMD at least 80 percent as efficiently within hours. That's a company claim, not an independent benchmark, so treat it as a signal rather than a proof. Andere also says his engineers barely need to edit or review the generated code.
Moorhead raises the same concern. AI models that rewrite code to run on different hardware can erode the advantage Nvidia built with CUDA. This is the counterintuitive finding: the tool that makes Nvidia's software dominant, AI-assisted coding, is also the tool that dismantles lock-in.
When AI Designs Chips, Nvidia Rides the Same Wave
There's a second irony. Most chip design firms already use specialized AI models to arrange components on processors, which used to demand deep engineering experience. OpenAI and Anthropic have both recently touted their newest models' ability to design electronics from scratch.
So the tools that Nvidia sells to its customers are also reshaping how chips get designed, including Nvidia's own. Moorhead doesn't think this is a serious threat, because Nvidia has access to the same AI. "They have access to AI too," he says.
That's probably true, but access isn't the same as advantage. If AI design becomes cheap, the barrier to entering chip design drops too. A startup with good models and a few engineers can now attempt designs that once required a large team. Nvidia's scale helps, but it no longer guarantees a head start.
The Intel Warning Sign
Dan Hutchinson, a veteran chip analyst, argues Nvidia may be entering its "Intel swan moment." He's referring to Intel's leadership change, when engineering-focused CEO Brian Krzanich was replaced by CFO Bob Swan. Krzanich also resigned after a relationship with an employee, which is why the comparison gets personal. Hutchinson's point is about strategy, not about individual conduct.
His argument is that after the handover, Intel's focus shifted toward sales and away from technical progress. Bureaucracy grew, and revenue suffered. "Nvidia is now an Intel waiting to happen," he says.
This comparison has limits. Nvidia's current financial position is far stronger than Intel's was, and a CFO-led transition doesn't automatically produce the same outcome. Still, the analogy highlights a real risk: a leader who spotted the AI wave early is being replaced by a successor who has to make multi-year bets with less personal insight into the technology. That's the part most investors should watch closely.
Frequently Asked Questions
What is Nvidia AI chip competition?
Nvidia AI chip competition is the rivalry between Nvidia and companies that make alternative AI accelerators, including AMD, cloud providers designing their own chips, and startups like Cerebras. Nvidia still holds about 80 percent of AI accelerator sales. Competitors are gaining ground, especially in inference and in cost-sensitive cloud workloads.
How do I switch AI workloads from Nvidia to AMD chips?
Start by checking whether your models and frameworks support AMD's software stack. Most cutting-edge models were written for CUDA, so you may need porting work or AI-assisted code translation. Test with a representative workload first, because efficiency gains vary by model and serving setup.
Why is Nvidia losing market share in China?
US export rules bar Nvidia from selling its most advanced chips in China, so it can sell only older generations. Meanwhile, Beijing is pushing domestic chipmakers such as Huawei. China fell to 8 percent of Nvidia's revenue from 20 percent in 2024, according to company filings.
When does the CUDA moat stop protecting Nvidia?
The moat weakens when porting code becomes cheap. AI models that can rewrite code quickly reduce the switching cost customers have long faced. The effect is still uneven, though, and large, complex production systems still need careful validation.
Nvidia isn't going to lose its lead overnight, and the company's scale still matters. But the trouble that WIRED describes existed before Huang's exit, and it's rooted in product timing, inference economics, and software lock-in. If you're tracking Nvidia AI chip competition, watch inference pricing, AMD's share gains, and whether AI-assisted porting keeps getting faster. Pick one of those signals and check it against the next quarterly filing. That's the fastest way to see whether the trouble is accelerating.