Nvidia’s AI chip dominance faces a “weakness” tied to what comes after training, CEO of rival chip firm says

Nvidia has built its business into the central plumbing of the artificial intelligence boom, but a Silicon Valley CEO says the company’s lead may be more fragile than investors assume. In a recent interview on the All-In Podcast, Cerebras Systems CEO Frank Bruno argued that shifts in how computers are designed to run AI workloads have historically punished incumbents, even when they dominate their current era.

Bruno’s core claim is that Nvidia’s current computing architecture was engineered to excel at training, the computationally intensive process of teaching large language models how to perform tasks. Training has been the market driver behind explosive demand for Nvidia’s GPUs and the software ecosystem that runs alongside them, the interview content suggests.

The CEO said Nvidia’s “greatest strength” could become its “biggest weakness” if the next phase of AI spending leans more heavily toward inference. Inference is the step where trained models are used in real time to generate answers, recommendations, or decisions. Bruno’s argument is that if inference becomes the larger share of demand, competing approaches optimized for that workflow could win share.

According to the 24/7 Wall St write-up of the interview, Nvidia’s exposure to AI infrastructure spending is unusually concentrated, with data center revenue representing the vast majority of the company’s total revenue in fiscal 2026 results, and data center revenue cited as rising from $15 billion in fiscal 2023 to nearly $194 billion in fiscal 2026. The article also frames Nvidia’s current position as unprecedented, while warning that no technology leader stays on top indefinitely.

The underlying point, as presented by Bruno, is less about Nvidia “falling behind” than about timing and architecture. He used examples of other chip and platform leaders that lost relevance when the industry’s dominant computing model changed, such as IBM after mainframes gave way to personal computers, Intel after the rise of smartphones favored lower-power mobile designs, and Cisco failing to fully capitalize on the cloud transition, according to the summary of the interview.

Bruno’s comments arrive as the AI chip market shows both dominance and intensifying competition. CNBC reported in 2024 that Nvidia’s AI accelerators held an estimated 70% to 95% share range for AI chip sales, but also that competition is rising as more vendors pursue specialized designs. The same theme has played out with broader semiconductor leaders: CNBC has also described Intel as struggling to stay relevant after missing major technology transitions over the past 15-plus years.

Still, the interview account does not claim Nvidia is destined to lose. It suggests the risk is structural: if the industry’s center of gravity moves from training to inference, the architectures and software stacks that best fit inference could become more important than the ones built for training workloads.

What Nvidia did not disclose in the materials cited here is any formal response to the “training-to-inference” critique, any roadmap change to prioritize inference hardware, or any quantified assessment of how much of its future demand depends on training versus inference. The argument also does not identify specific competitors that would gain from such a shift, beyond the general category of firms building inference-leaning systems.