Can Nvidia sustain growth as hyperscalers ramp custom silicon and inference overtakes training demand?

Based on my time at Meta, I think custom chips are the way to go for making Magic Hour's video tools cheaper for creators. We've hit growth walls because of GPU shortages before. These new chips could let us lower prices, but our software is what will actually make us special long-term. So watch the silicon trends, but invest just as much in the creative tools that nobody else has.

In our health-tech work, we run live diagnostics for patients, and speed is everything. We're watching big tech companies build their own chips because they handle these real-time models faster than traditional GPUs. Nvidia should partner with us to design chips specifically for this kind of inference work, or they risk losing the precision health market to custom solutions.

There here is plenty of growth left for Nvidia in innovation and a robust software ecosystem. Its CUDA platform and AI tools are a hard act for competitors to follow. Nvidia's GPUs are optimized for both training and inference, and demand for AI inference is only going to increase. The company is branching into new markets like automotive, health care and edge computing on track for shrinking dependence on hyperscalers. If Nvidia wants to remain competitive, the company should continue building better technology, develop good partnerships and respond to customer needs. And these moves can help it continue growing as the competition heats up.

The strong growth in Nvidia's valuation and performance continues, though this growth story is one which could prove hard to maintain as hyperscalers build their own custom silicon and as inference workloads eclipse training workloads in total AI compute. Nvidia's growth story in training workloads has been largely built around three pillars - a ubiquitous CUDA ecosystem and associated software moat and simply having the best-performing GPUs for AI workloads. Cloud native chips cannot easily emulate the former and hyperscalers are increasingly building in-house chips to bridge the performance gap. As inference workloads account for a greater proportion of AI compute overall, customers will likely demand more cost-effective, energy-optimised hardware that is more suited to persistent deployment as opposed to big model training. Nvidia will need to adjust its product mix to reflect these changes with inference-optimised GPUs and edge AI products, as well as an integrated software stack which increases customer switching costs. But that's not the whole story. Nvidia's long-term durability will be derived from the control of the platform versus the raw performance of the silicon itself. The company's entire software stack, from TensorRT to DGX Cloud, gives it a guaranteed revenue stream and guarantees interoperability that a custom chip can't offer. It also helps that Nvidia is opening up new verticals like automotive AI, robotics and digital twins, which also help Nvidia to avoid some of the hyperscaler cycle while opening up industries that are less likely to build custom silicon. Over time, the company will have to pivot from the "enabler" of AI to the "infrastructure backbone" of intelligent systems. Nvidia's cross-cutting differentiation of compute, software, and stickiness will help to compound the company's already existing dominance, even as the product/market balance shifts toward inference.

Nvidia has had a clear lead with the gaming industry as one of its major stronghold without a doubt. But as data centers and hyperscalers invest more and more on custom silicon for HPC and AI workloads, demand may dwindle for traditional GPUs. As a result, Nvidia has been looking to build its presence into the inference market - offering both hardware and software solutions that allow large data sets to be processed in real time. This diversifies their sources of revenue and sets them up nicely for when inferencing becomes more widespread in industry.

The bigger Nvidia becomes, the more it will face headwinds from the growing share of custom silicon that hyperscalers are designing and of inference over training. To lessen dependency on Nvidia, with which they claim cost advantages and control advantages, companies including Google, Amazon, and Microsoft are investing in proprietary chips (TPUs, Inferentia, Maia AI Accelerators, etc.) These custom solutions have been tuned for specific workloads and are particularly for inference, which is more cost-sensitive, and will benefit more from energy-efficient ASICs. The response from Nvidia is to build out their ecosystem (CUDA, NVLink networking etc) and to develop semi-custom AI systems to keep leading designers as partners. Meanwhile, AMD and Broadcom have been leveraging their low-cost alternatives and custom solutions to try and steal some market share away from NVIDIA.

We only have to look at Nvidia's 20% Wall Street slide in January to understand how vulnerable the stock is to new competitors. While NVDA's blip at the beginning of the year can be attributed to the arrival of China's DeepSeek large language model, new pressures are likely to emerge closer to home in the months and years ahead. It's reasonable to expect Nvidia's market share in chips to erode over time, but it's worth noting that the company is already actively preparing for this scenario, creating integrated hardware and software ecosystems to drive a long-serving network of clients. Major cloud providers like Google, Amazon, and Microsoft are investing billions of dollars into developing their own custom chips, or Application-Specific Integrated Circuits (ASICs), for their workloads. These are set to directly compete with Nvidia's general-purpose GPUs, with their ability to support specific workloads helping for inference purposes. The key issue that Nvidia faces for its longevity is that once AI models are trained using the firm's GPUs, they can be shifted to inference rather than requiring more training. This potentially creates a new market where custom ASICs offer a low-cost alternative to Nvidia's products.

Whether Nvidia can continue to grow as hyperscalers create custom silicon and inference pulls ahead of training will be a test of its innovative mettle. Although hyperscalers such as Google and Amazon are creating custom chips to reduce third-party GPU reliance, Nvidia remains competitive in AI due to its hardware, software ecosystems, such as CUDA, and an end-to-end approach, as stated by Neumann. With workloads for inference increasing, Nvidia has been placing its bets on power-efficient GPUs and AI-tuned hardware such as the H100 to seize this demand. Keeping that lead means continually innovating, holding prices down into the commodity price range, and meeting the challenge of the move-away from general-purpose ships for silicon.

NVIDIA's future depends on whether it can adapt its hardware to emerging materials and shifting compute demands. As hyperscalers ramp up custom silicon and inference begins to outpace training, the company's moat will depend less on CUDA dominance and more on efficiency, thermals, and cost per watt. This is where next-generation materials like graphene could redefine the playing field. Companies such as Avadain are developing graphene-based shielding and conductive layers that can dramatically improve heat dissipation and electrical efficiency—both critical bottlenecks for data centers and GPUs. If NVIDIA embraces these future technologies and integrates advanced materials into its architecture, it can sustain and even accelerate growth. But if it clings to legacy designs while others innovate at the substrate level, it risks losing the very edge that made it dominant. The next frontier isn't just AI—it's the physics that enable AI to scale sustainably. —Pouyan Golshani, MD | Interventional Radiologist & Founder, GigHz and Guide.MD | https://gighz.com

While hyperscalers chase efficiency through custom silicon, Nvidia's real power lies in how fast it adapts its hardware to new workloads. The company treats chip architecture like a living organism, evolving it every few quarters to meet new AI demands. That speed, paired with its massive developer base, gives it an edge that custom silicon rarely achieves. As inference scales, Nvidia can use its flexible architecture and software to stay embedded in every stage of the AI pipeline — from data prep to edge deployment. Growth will depend less on exclusivity and more on ubiquity. Nvidia's play is not to outbuild the hyperscalers, but to make itself indispensable to them.

Nvidia's growth may soon depend less on silicon and more on the knowledge its chips have helped create. Every model trained on Nvidia hardware feeds a global feedback loop of optimization data — insight into architectures, energy efficiency, and performance scaling. That accumulated intelligence gives Nvidia a head start in designing chips that already know what tomorrow's workloads will look like. Hyperscalers can copy hardware designs, but they can't replicate years of learning from millions of AI experiments. Nvidia can turn that knowledge into predictive engineering, where each new generation of GPU anticipates where computing is headed before the market fully shifts. That foresight could be its most valuable asset.