NVIDIA’s next major data center architecture, Feynman, is already starting to send a warning even before hitting the market: in the era of artificial intelligence, designing the most ambitious chip isn’t enough; you also need the capacity to produce it. Over the past few hours, a widely circulated idea from Asia suggests that the lack of capacity at TSMC’s A16 node is forcing NVIDIA to reconsider parts of its future GPU designs, planned for 2028. There has been no official confirmation from NVIDIA or TSMC on this alleged redesign, so it’s best to treat it as what it is: a supply chain rumor, relevant but not yet publicly verified.
What is confirmed is the context that makes this hypothesis plausible. TSMC introduced A16 as a process specially aimed at HPC and AI products, featuring nanosheet transistors and Super Power Rail technology, with volume production expected to begin in the second half of 2026. Compared to N2P, TSMC claims that A16 offers between 8% and 10% higher speed at the same power, 15% to 20% lower energy consumption for equal performance, and 7% to 10% higher density. On paper, this is precisely the type of node that fits well with a generation like Feynman, which NVIDIA officially projected for 2028.
The reality is that paper can show anything, but manufacturing capacity is what truly matters. In January, TSMC acknowledged in its earnings conference that the demand for advanced silicon driven by AI remained very high, with major clients maintaining positive forecasts, and even their customers—primarily large cloud providers—were directly requesting capacity. This pressure isn’t limited to a single node or client. Reuters reported on March 24, 2026 that Broadcom also views TSMC’s capacity as a bottleneck, and that the Taiwanese foundry itself is reaching production limits as AI chip demand skyrockets.
A plausible rumor, but yet unconfirmed
The origin of the version indicating changes in Feynman stems from reports and feedback from the Asian supply chain suggesting that NVIDIA initially planned this family based entirely on A16 but now might need to combine that node with N3P for certain less-critical blocks. The interpretation from these sources is that the more sensitive dies or chiplets would remain on A16, while other parts might migrate to a more mature process with greater availability. This information hasn’t been confirmed by the involved companies, and this nuance is significant because, in semiconductors, the difference between “design target,” “engineering test,” and “final product” can radically alter the interpretation.
Nevertheless, the hypothesis doesn’t seem far-fetched. TSMC has clearly stated that A16 isn’t a general-purpose node but a technology designed for high-performance chips with complex signal and power requirements. The more specific and advanced a node, the more difficult it is to scale up quickly. Moreover, the industry has been months seeing how AI pushes not only advanced wafers but also packaging, HBM memory, interposers, lasers, and even less visible segments of the supply chain to their limits. If a company like Broadcom is openly talking about constraints at TSMC, it’s not surprising that NVIDIA might also have to adjust schedules, mix nodes, or internally allocate blocks across future generations.
What’s really at stake for NVIDIA
Beyond the specific rumor, what it reveals about the market is significant. NVIDIA continues to dominate AI infrastructure, but even that dominance doesn’t guarantee unlimited access to the most advanced process at the world’s largest foundry. This has several implications. First, from a technical standpoint: if Feynman ends up combining nodes, NVIDIA will need to further optimize the distribution between logic blocks, memory, and interconnects to avoid compromising efficiency, performance, or power consumption. Second, from an industrial perspective: the competitive advantage in AI is no longer solely dependent on architecture, software, and ecosystem but also on the ability to secure supply for multiple years ahead.
The third and broader implication affects the entire industry. If TSMC has to manage A16 with a close eye for top-tier customers, the margin for smaller companies will be even narrower. This reinforces an ongoing trend: longer contracts, more aggressive planning, and partial diversification where possible. In other words, the AI era has not only multiplied demand for computation but also transformed advanced manufacturing capacity into a scarce strategic asset.
For now, caution remains essential. NVIDIA has indicated that Feynman is part of its 2028 generation, and TSMC maintains that A16 will come into volume production in the second half of 2026. The rest—whether Feynman will be entirely on A16, whether it will mix A16 and N3P, or whether it will adjust its design due to capacity shortages—remains in the realm of industrial speculation. Still, the underlying message is quite clear: the next AI battle isn’t just fought in design labs, but also in the ability to turn those designs into real wafers, on time and at sufficient volume.