Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarWired→ original

Huang Steps Down at Nvidia as AMD's Share of AI Chips Grows from 2% to 7–10%

Nvidia sells about 80% of AI accelerators, the chips used to train and run neural networks, but CEO Jensen Huang has stepped down, and CFO Colette Kress is temporarily running the company. AMD's share grew from 2% in 2022 to 7–10% in 2026, while Google, Amazon and Microsoft are making their own chips. Previously, moving to different hardware meant rewriting code for CUDA. Now AI can do it in a matter of hours. That is why it's worth testing your models on AMD in advance and comparing launch costs.

AI-processed from Wired; edited by Hamidun News
Huang Steps Down at Nvidia as AMD's Share of AI Chips Grows from 2% to 7–10%
Source: Wired. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceWired ↗
Editor-in-Chief
Key metric #1According to Wafer's CEO, new models run on AMD within hours at no less than 80% efficiency
Key metric #2Nvidia's China revenue share fell from 20% in 2024 to 8%
Nvidia revenue$26.91 billion in January 2022 and $215.94 billion in January 2026

Nvidia remains the most influential company of the AI era. According to Silicon Analysis, it accounts for about 80% of AI accelerator sales, the chips used to train and run neural networks. But, as Wired reports, the company now faces several problems at once. CEO Jensen Huang, who was among the first to realize that graphics chips suit AI, has stepped down. Chief Financial Officer Colette Kress has taken over his duties. Meanwhile, competitors are closing in from several directions, and one of Nvidia's main moats, the CUDA software platform, has proven less durable than assumed.

Chips are designed over years, but markets change in months

An anonymous head of a major AI company explained the challenge to Wired: everything sold today was designed two to three years ago. That means Kress, or whoever leads Nvidia permanently, must already guess what chips will look like two years from now.

For now, the numbers favor the company. Revenue grew from $26.91 billion in January 2022 to $215.94 billion in January 2026. But AMD's share of the AI accelerator market rose from 2% to 7–10% over the same period. Google, Amazon and Microsoft are developing their own chips for training and running models to cut costs and depend less on Nvidia. Cerebras sells giant chips for giant models. AMD bought Taalas, a company that burns model weights directly into silicon so computations run very efficiently. The drawback of all these solutions is the same: running modern models on them requires rewriting a lot of code.

Analyst Patrick Moorhead of MoorInStrat adds two more factors. First, because of AI agents, inference, meaning running already trained models, is becoming more important, and Nvidia has fewer advantages there than in training. Second, demand is growing for conventional processors (CPUs), where Nvidia barely competes. In December 2025, the company bought a startup with specialized chips, but Moorhead does not consider this a complete answer.

The CUDA moat has become permeable

For years, Nvidia kept developers with CUDA, a set of software tools used to write code and build models for its chips. Imagine a company whose entire production stack is built around CUDA: switching to AMD would mean months of rewriting code, so almost nobody made that move.

Now AI models read and write code so well that they can port other models to different hardware themselves. The startup Wafer uses AI to adapt models for AMD and then sells their deployment at a lower price. Its CEO, Emilio Andere, says that after each major open-weight model release, his team runs it on AMD within hours at no less than 80% of Nvidia's efficiency. According to him, engineers barely need to edit or check the code. Wafer's revenue, the founder claims, has been growing rapidly in recent months. This is Andere's own claim, and the material contains no independent confirmation.

Moorhead notes that Nvidia also has access to AI, so automating chip design won't necessarily weaken it. But the ability of neural networks to quickly rewrite code for other hardware could undermine the company's main advantage.

China and the question of management

The second front is China. Washington restricts shipments of the most powerful chips, while Beijing is developing its own industry. Sales of permitted previous-generation and non-AI chips doubled over the year, but China's share of Nvidia's revenue fell from 20% in 2024 to 8%. Huawei builds supercomputers from very large numbers of chips, and some Chinese companies, such as Z.ai, claim they trained their models entirely on Huawei hardware. If Chinese open models take hold in other countries, a market for Chinese chips could emerge there too.

China is also linked to Huang's departure. According to data cited by Wired, he had a relationship with Huang Le, head of the Singapore-based Megaspeed, which US authorities suspect of ties to Chinese intelligence. Megaspeed reportedly imported Nvidia chips worth $4.6–7.5 billion into Southeast Asia and may have given Chinese AI companies access to them. Babak Minovi, an analyst at Eurasia Group, believes Huang was right on the main point: a complete ban on sales to China strengthens local competitors.

What this means in practice

Industry veteran Dan Hutchinson compares the situation to the "Swan moment" at Intel: in 2019, engineer-CEO Bob Swan, a finance man, replaced the engineering-focused leader, the focus shifted to sales, bureaucracy expanded, and revenue declined. "Nvidia is now the Intel that is about to happen," he says. This is one analyst's opinion, not a fact. Wired notes that Nvidia's problems are not necessarily fatal, but they are accumulating.

The practical takeaway for teams building AI products: dependence on a single accelerator supplier is becoming a manageable risk. If code can be moved quickly, it makes sense to test your models on AMD or cloud chips in advance and compare launch costs. And Nvidia's strategy for the next two years is worth watching closely.

Story

Key Questions Answered

What are the key details on Chips are designed over years, but markets change in months?

An anonymous head of a major AI company explained the challenge to Wired: everything sold today was designed two to three years ago. That means Kress, or whoever leads Nvidia permanently, must already guess what chips will look like two years from now. For now, the numbers favor the company. Revenue grew from $26.91 billion in January 2022 to $215.94 billion in January 2026. But AMD's share of the AI accelerator market rose from 2% to 7–10% over the same period.

What are the key details on The CUDA moat has become permeable?

For years, Nvidia kept developers with CUDA, a set of software tools used to write code and build models for its chips. Imagine a company whose entire production stack is built around CUDA: switching to AMD would mean months of rewriting code, so almost nobody made that move. Now AI models read and write code so well that they can port other models to different hardware themselves. The startup Wafer uses AI to adapt models for AMD and then sells their deployment at a lower price.

What are the key details on China and the question of management?

The second front is China. Washington restricts shipments of the most powerful chips, while Beijing is developing its own industry. Sales of permitted previous-generation and non-AI chips doubled over the year, but China's share of Nvidia's revenue fell from 20% in 2024 to 8%. Huawei builds supercomputers from very large numbers of chips, and some Chinese companies, such as Z.ai, claim they trained their models entirely on Huawei hardware.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…