Nvidia’s AI Chip Lead Faces Pressure From Cloud Chips, AMD and Easier Porting
- Cargo
- Machine Learning
- Shipped
- Shipper
- Will Knight
- Readers
- 614
Nvidia still leads AI accelerators, but cloud chips, China’s growing alternatives and AI-assisted code porting are testing its advantages.

Nvidia’s position in AI chips rests on more than fast processors. Its software, developer adoption and early investment in AI helped make its hardware the default choice for many model builders. Those advantages still matter, but competitors are gaining ground in several parts of the market at once. With CFO Colette Kress now leading the company, the choices made today could shape its position years from now.
That timing matters because chip development runs on a long cycle. One AI company CEO, speaking anonymously to avoid affecting business relationships, said products shipping today were designed two to three years ago. The person leading Nvidia now must make decisions for a market whose needs may look quite different by the time those products arrive.
More routes around Nvidia
Nvidia still holds about 80 percent of AI accelerator sales, according to Silicon Analysis. AMD, its longtime rival, has increased its share from 2 percent in 2022 to between 7 and 10 percent in 2026. Those figures leave Nvidia far ahead, but they show that customers have alternatives—and that AMD is working to make its chips a more serious option.
Google, Amazon and Microsoft are developing their own specialized processors for training and running AI models. Designing silicon in-house can help cloud providers reduce costs and rely less on Nvidia. It can also give them more control over how their infrastructure serves their own systems and customers. These efforts don’t need to displace Nvidia across the market to matter: even shifting some workloads to in-house chips can change how much hardware cloud companies buy from it.
Startups are pursuing narrower designs. Cerebras sells large chips intended for very large AI models, while Taalas has developed an approach that encodes model weights directly in silicon. AMD recently bought Taalas. Designs like these may improve efficiency for particular workloads, but adopting them can require substantial code changes, especially for newer models. That engineering work remains a practical hurdle for customers considering specialized hardware.
The mix of workloads is changing, too. Patrick Moorhead, chief analyst at MoorInStrat, points to AI agents as a reason demand is growing for inference: running a model after training. He says Nvidia has less of an advantage in this area against cloud providers and startups. Nvidia acquired Groq, a startup making specialized chips, in December 2025, a move that puts it into this part of the market.
AI assistants and agents may also increase demand for general-purpose central processing units, or CPUs. Moorhead says Nvidia has little presence in that market. If these systems rely on more kinds of computing than model training alone, the company may need to compete for a broader set of hardware purchases.
China is a smaller market for Nvidia
Nvidia’s growth has been tied to the boom in AI. Jensen Huang cofounded the company in 1993 to make graphics chips for personal computers. By the mid-2010s, he was backing AI research, supporting software for AI engineering and providing hardware to labs and startups. After ChatGPT brought AI to a wider audience in 2022, Nvidia’s revenue rose from $26.91 billion in January 2022 to $215.94 billion in January 2026.
China, once another major market for AI chips, now accounts for a smaller share of Nvidia’s business. The company cannot sell its most powerful chips there, though it can sell earlier generations. Sales of those chips, alongside non-AI chips, doubled over the past year. Even so, China’s share of Nvidia revenue fell from 20 percent in 2024 to 8 percent, according to company filings. It’s also unclear how many chips shipped to other countries might reach Chinese customers or be accessed remotely by Chinese firms.
Domestic suppliers offer a potential alternative. Huawei has built supercomputers that use larger numbers of chips to train and run AI models. Z.ai, a maker of open-weight models, has said it trained its models entirely on Huawei chips. If those models attract users outside China, demand for Chinese hardware could follow them into other markets.
Software may weaken Nvidia’s lock-in
Nvidia’s software has helped keep developers on its hardware. Its CUDA platform gives engineers tools for writing programs and building models for Nvidia chips, and moving established systems to different processors has often required extra work. That switching cost has been one of the company’s strongest advantages.
AI coding tools may reduce that burden. Models can now help rewrite code for different hardware, making it quicker to move workloads to competing chips. Wafer, a startup that ports models to AMD hardware and serves them at lower cost, says it can get a major new open-weight model running on AMD chips within hours at at least 80 percent efficiency. CEO Emilio Andere says his engineers now need to edit or review little of the generated code.
AI could also assist with the chip design process itself. Chip firms already use specialized models to arrange processor components, a task that has traditionally depended on deep engineering experience. OpenAI and Anthropic have described newer models capable of designing electronics. Moorhead argues that Nvidia can benefit from these tools as well as its rivals. The more immediate concern may be that AI makes Nvidia’s software advantage easier to overcome by helping developers adapt code for competing hardware.
Veteran chip analyst Dan Hutchinson warns that Nvidia could follow a pattern he associates with Intel after its engineering-focused CEO Brian Krzanich was replaced by CFO Bob Swan in 2019: more attention to sales, slower technical progress and growing bureaucracy. That comparison is a warning, not a forecast. Nvidia’s lead remains substantial, and the company is investing in new areas such as inference chips. The test for its next leadership will be whether it keeps pace as customers gain more hardware choices and software tools make switching less difficult.