The Nvidia Corporation logo, representing the American multinational corporation and technology company that designs and supplies graphics processing units, application programming interfaces, and high-performance computing, is exhibited during the Mobile World Congress 2025 in Barcelona, Spain, on March 5, 2025. (Photo by Joan Cros/NurPhoto via Getty Images)
NurPhoto via Getty Images
Nvidia (NASDAQ: NVDA) has indisputably led the AI boom. Its GPUs are considered the gold standard for training extensive AI models, resulting in sales growth from $27 billion in FY’23 to an anticipated $200 billion this fiscal year. In addition to selling the highest performance chips, the company’s CUDA software ecosystem has effectively retained customers. While Nvidia is expected to continue its reign as the leader in AI hardware for years ahead, its stock valuation of nearly 40x forward earnings signifies not only its leadership but also expectations of ongoing, multi-year growth.
This renders it vulnerable: even a slight decrease in demand or a structural shift in the AI lifecycle – from focusing on training workloads to inference – could undermine investor confidence and cause significant declines. Historical events highlight this risk – following the Covid-era surge in GPU demand for gaming and cryptocurrency, inflation and reduced demand led Nvidia shares to plummet almost 66% from peak to trough, in contrast to a mere 25% decline for the S&P 500. The stock’s volatility suggests a similar drop could occur if the current AI growth phase slows. Read NVDA Dip Buyer Analyses to examine how the stock has bounced back from sharp dips in the past.
Training vs. Inference Shift
Over the past two years, companies have invested vast resources into developing AI models. Training these enormous models usually requires a concentrated effort that demands significant computing power, with Nvidia being the largest beneficiary, as its GPUs are widely seen as the fastest and most efficient for these operations.
However, the AI landscape may be evolving. Incremental performance improvements are waning as models become larger, while access to high-quality training data is becoming a limiting factor – much of what is readily available online has already been utilized in current models. Collectively, these factors imply that the most demanding phase of AI training might start to plateau. Adding to the uncertainty, the economics of the GPU market remains challenging, as numerous Nvidia customers continue to struggle to yield significant returns on their substantial AI investments.
In contrast, inference involves applying trained models to new data in real-time and at scale. It is less intensive per task but occurs continually across millions of users and applications. As AI evolves, a larger portion of value creation could shift from training to inference. The challenge for Nvidia is that its growth has been primarily linked to training, where its high-end GPUs have a stronghold. Inference presents an opportunity for more mid-performance and cost-effective chip alternatives, as well as specialized offerings. Here’s a glimpse at some of the main competitors in the inference sector.
MORE FOR YOU
What The Inference Landscape Looks Like
AMD has notably trailed Nvidia during the initial phase of AI development, but it could become a significant competitor to Nvidia in inference. Its chips are becoming increasingly competitive in performance while providing cost and memory advantages. Not all organizations require or can afford Nvidia’s premier GPUs. Many are likely to choose older Nvidia models or budget-friendly alternatives such as AMD’s (NASDAQ:AMD) MI series, which delivers solid performance for inference and fine-tuning of models. See How AMD stock surges to $330.
ASICs or Application-Specific Integrated Circuits are also gaining momentum. Unlike GPUs, which are flexible and programmable, ASICs are designed for a specific task, making them more cost- and power-efficient for inference workloads. The cryptocurrency sector offers a precedent: Bitcoin mining began with GPUs but swiftly transitioned to ASICs once scale and efficiency became vital. A similar trend could emerge in AI inference. Two companies likely to benefit from this shift are Marvell and Broadcom, both of which possess expertise in creating custom silicon for hyperscalers.
U.S. Big Tech players such as Amazon (NASDAQ:AMZN), Alphabet, and Meta are all designing AI chips. Amazon has focused on training-oriented chips, Meta began with inference and is expanding into training, while Google accommodates both with its TPU (tensor processing unit) lineup. For these hyperscalers, the goal is not necessarily to outdo Nvidia in the marketplace but to reduce costs, enhance bargaining power, and manage supply for their expansive cloud ecosystems. Over time, this translates to diminished incremental demand for Nvidia’s GPUs. In Q2, Nvidia reported that just two of its clients accounted for approximately 39% of total revenue, and it is highly plausible that these were major U.S. tech firms. This concentration makes Nvidia significantly more susceptible. If hyperscalers increasingly turn to proprietary silicon, even minor changes in purchasing behavior could lead to substantial revenue implications.
Chinese Players In China, firms like Alibaba (NYSE:BABA), Baidu, and Huawei are enhancing their AI chip initiatives. Reports last week indicated that Alibaba plans to introduce a new inference chip for its cloud division. The strategy is twofold: to facilitate inference at scale using its own technology stack, and to ensure a reliable supply of semiconductors in light of U.S. export restrictions. Currently, Nvidia’s GPUs are still anticipated to serve as the backbone of Alibaba’s AI training operations, but inference is poised to become the long-term growth driver for the firm.
Nvidia Remains Dominant, but Risks Are Increasing
Nvidia’s position is still robust due to its established ecosystem, substantial R&D investments, and supremacy in training. However, inference is expected to become the next growth engine for AI hardware, and the competitive landscape is significantly more crowded. Even a slight dip in growth could heavily impact the stock, considering how much future performance is already factored in. For investors, the critical question is whether Nvidia’s growth trajectory can align with the elevated expectations set by the market. If the economics of inference prove less advantageous than those of training, the stock could still face a “valuation reset” despite retaining its technological leadership.
The Trefis High Quality (HQ) Portfolio, which comprises 30 stocks, has a history of consistently outperforming its benchmark, inclusive of the S&P 500, Russell, and S&P midcap. What accounts for this? As a collective, HQ Portfolio stocks have delivered superior returns with reduced risk compared to the benchmark index; a smoother ride, as evident in HQ Portfolio performance metrics.