A Surprising Return of the Chip Startup
One of the more striking developments in technology markets is the resurgence of initial public offerings, and nowhere is this more notable than in semiconductors. The biggest IPO of the year — and the largest semiconductor IPO ever — signals a real shift. For years, the conventional wisdom held that starting a chip company was simply too hard. The field appeared locked up by giants: Intel as the historic incumbent, and later Nvidia and TSMC dominating design and manufacturing respectively. Yet we are now witnessing genuine innovation in a space that was assumed to be closed to newcomers. In an age defined by semiconductors, the arrival of a major new chip company is itself a meaningful signal that the market structure is more fluid than it appeared.
From Training to Inference
The most important question facing the dominant player is not whether its current business is healthy — by every measure it is performing extraordinarily well — but what happens as the industry transitions from training to inference. Training large models was the phase that built Nvidia's dominance. Inference, the act of running those models in production, is a different competitive landscape, and it is here that Nvidia has allowed rivals into the market.
The reason is straightforward: prices and gross margins are exceptionally high, with gross margins exceeding 70%. Such margins are wonderful for the company itself, but they create economic space for competitors. Nvidia could have dominated inference from the very beginning. Had its prices been half of what they are today, it could still have operated as a profitable business while making it pointless for anyone to look elsewhere. It chose not to do that — a perfectly defensible decision given how well it is doing — but the consequence is that room has opened up in the inference market for others to build viable solutions.
Cerebras is a clear example. Its largest customer is OpenAI, and much of its revenue and promised future revenue flows from that single relationship. Positron, a private company, is pursuing a similar path, focused on inference and on better power efficiency. These companies exist in a competitive gap that Nvidia's pricing effectively created. Nor should the established alternatives be forgotten: Intel and AMD are both actively arguing that they have a place in a world increasingly oriented around inference rather than training.
The Customer Revolt Against High Prices
There is a fundamental difference between raising prices in a one-off transaction and doing so within a long-term business relationship. When a supplier knows a customer desperately needs a scarce product and raises the price accordingly, that may be acceptable in the short term. But when that supplier dramatically increases prices on partners who depend on it year after year, those partners begin to look for alternatives — and that is precisely what is happening across Nvidia's customer base.
The pattern is consistent among the hyperscalers, the handful of companies — Microsoft, Amazon, Google Alphabet, and Meta — that each spend tens of billions of dollars annually on AI infrastructure. Amazon, facing high costs, is designing its own Trainium chip. Google, also charged heavily, has developed its own TPU. Meta, paying substantial sums where it operates data centers, is working on its own silicon. Oracle has not gone the in-house route, but it is increasingly turning to AMD chips instead of Nvidia in many instances.
The underlying preference has not changed: every one of these companies would rather use an Nvidia chip wherever it could. The chips are the best available. But high prices are pushing even the most committed customers to seek alternatives. There is a certain irony here, given that OpenAI helped create Nvidia's dominance in the first place; now the same ecosystem of large buyers is actively encouraging diversification away from a single supplier.
This is, to be fair, a deliberate business model. The strategy is to extract as much value as possible while the opportunity exists, to acquire competitors and emerging technologies, and to buy a way into desirable adjacent markets. It is the model of a fantastic, highly profitable company. The open question is whether it can last forever — and there is good reason to doubt that it can.
China as the Critical Wild Card
China represents one of the most important uncertainties in this story. The decision to permit ten Chinese companies to acquire H200 chips has not translated into eager demand; those companies have not clearly signaled that they want the chips. China has instead grown more self-sufficient, leaning on domestic players such as Huawei. The stakes are large — were the China market to open fully, it could represent more than $50 billion in revenue for Nvidia — but the trajectory of self-reliance complicates that prospect.
What makes China especially significant is talent: arguably some of the most capable engineers working on non-Nvidia AI chips are based there. A historical analogy is instructive. During the Cold War, when the Soviet Union could not obtain the best computing hardware, it compensated by producing some of the world's finest software programmers. Constrained by inferior machines, Soviet engineers learned to do more with less by writing exceptionally efficient software.
The same dynamic is visible in China today. DeepSeek is the clearest illustration: it likely used few or no Nvidia chips and nevertheless delivered remarkable AI results. China is demonstrating an ability to build impressive systems on subpar semiconductors. For Nvidia, this matters profoundly. The company would prefer to be the single global standard rather than see multiple competing standards develop — and constrained but resourceful adversaries are exactly the conditions under which alternative standards take root.
The Ecosystem Beyond the GPU
Focusing only on the GPU misses the larger picture. Nvidia is no longer simply a maker of graphics processors; it is an entire ecosystem, and its expansion is increasingly visible in areas surrounding the core chip. The company is talking far more about networking, about GPU usage in inference rather than training, and about power management — including significant changes coming over the next year in how power is delivered directly to the chip itself.
Networking deserves particular attention. Optical networking has been an important theme for years, encompassing coherent optics and the companies operating in this space, including AOI and Fabrinet. These firms occupy an unusual position: they both enable Nvidia and, in some respects, compete with it. As a result, the moves Nvidia makes in networking could move a great many other stocks — potentially more than they move Nvidia itself — precisely because the ecosystem extends so far outward from the central player.
This is why the substance of an earnings report matters more than its effect on any single share price. A stock price is ultimately just a reflection of the underlying business. The meaningful signals are the developments around the GPU: networking strategy, the inference roadmap, and power management. Demand for the core product is not in question — the chips are sold out, and the company is selling every unit it can possibly manufacture. It would gladly sell more into China, but at present it is already selling everything it can make.
Looking Ahead
Given a sold-out product, expanding ecosystem ambitions, and a business of unprecedented scale, the near-term outlook points toward strong results — numbers large enough to make observers' jaws drop yet again, simply because no one has ever seen a business of this size operate this way. The deeper story, however, is not the next set of headline figures. It is the slow structural shift underneath them: a pricing strategy so aggressive that it is funding its own competition, a transition to inference that rewards efficiency over raw power, and a geopolitical rival learning, as others have before, to do extraordinary things with constrained resources. The current dominance is real and impressive. Whether it is permanent is a far less certain proposition.