Why Nvidia is reworking the Feynman AI chip platform
Nvidia is being pushed to rethink the design of its next-generation Feynman AI chip platform after it didn’t secure enough production capacity on TSMC’s most advanced manufacturing process. The situation was reported by Taiwanese financial news outlet cnYES.
The key issue is straightforward: the original plan depended on broad access to TSMC’s cutting-edge A16 (1.6-nanometer) process. With that capacity not available in sufficient volume, Nvidia is now looking at a split approach that keeps the highest-priority pieces on the most advanced node while moving other parts to an older process.
The revised manufacturing plan: A16 for critical components, N3P for the rest
What stays on TSMC A16 (1.6nm)
Under the revised plan, only the most critical components of the Feynman chips would be manufactured using TSMC’s A16 (1.6-nanometer) process.
This keeps the most important parts aligned with the most advanced manufacturing capabilities Nvidia had initially aimed to use across the platform.
What moves to TSMC N3P (3nm)
Less essential parts would shift to the older N3P 3-nanometer node.
This kind of mix-and-match approach is a compromise: it’s meant to reduce dependence on a single, capacity-constrained process, but it also introduces practical and strategic consequences for the platform as a whole.
What the compromise could mean for cost, design, and supply
The revised approach could bring:
- Design trade-offs
- Higher costs
- Potential supply constraints
And that matters because Feynman is described as central to Nvidia’s AI ambitions. When a flagship platform gets redesigned around manufacturing availability—rather than purely around ideal technical targets—there’s usually a ripple effect across planning, timing, and how much volume can realistically reach customers.
TSMC’s advanced node capacity crunch and why it’s not a quick fix
Overwhelming demand for 2nm and sub-2nm production
TSMC’s advanced 2-nanometer and sub-2-nanometer processes are facing demand pressure from AI and high-performance computing customers. The result is that capacity is fully booked through 2028 and potentially beyond.
That multi-year saturation is what turns this from a short-term scheduling headache into something that can force architectural and packaging decisions—exactly what Nvidia is now confronting with Feynman.
Expansion plans in Tainan Science Park, with a 2028 target
TSMC is expanding aggressively. It has filed environmental review documents for a new 15.46-hectare fabrication facility in southern Taiwan’s Tainan Science Park.
The plan includes:
- Construction set to begin this year
- Completion targeted for 2028
Even with aggressive expansion, the timeline underscores the reality that leading-edge capacity can’t be ramped instantly. When demand is already booked out years ahead, new fabs and new lines become part of a longer, tightly constrained supply story.
Pricing pressure as supply stays constrained
TSMC is expected to raise prices in response to the constrained supply. That reinforces its pricing power across the semiconductor industry.
In practical terms, that expectation adds another dimension to Nvidia’s redesign decision: it’s not just about whether capacity exists, but also about what it costs to lock it in—and what trade-offs become acceptable when pricing moves upward.
Nvidia’s view on demand and future capacity growth
Nvidia CEO Jensen Huang said earlier this year that the company faces exceptionally strong demand. He also projected that TSMC’s overall capacity could more than double over the next decade.
That projection doesn’t resolve the near-term constraints around A16 capacity and the 2028 booking window, but it frames how Nvidia is thinking about the long arc: demand is strong now, and future capacity growth is anticipated—yet the immediate bottleneck still shapes what can be built, when, and at what scale.
Feynman’s role in Nvidia’s chip roadmap
Previewed at GTC 2026 as the successor to Vera Rubin
Nvidia first previewed the Feynman architecture at its GTC 2026 conference in San Jose on March 15. During that event, Huang outlined the platform as the successor to the Vera Rubin chip family.
That positioning places Feynman as a major step in Nvidia’s forward roadmap—so the manufacturing constraints affecting it aren’t just operational details. They connect directly to Nvidia’s platform continuity and future AI product strategy.
Built on TSMC A16, including backside power delivery technology
Feynman was described as being built on TSMC’s A16 process, which introduces backside power delivery technology.
In the context of the redesign, that’s an important detail: A16 isn’t just “smaller.” It’s tied to specific process capabilities. If only the most critical components remain on A16 while other parts shift to N3P, the platform becomes a hybrid from a manufacturing standpoint—potentially complicating design decisions and production coordination.
Launch and delivery window tied to tight industry capacity
Feynman is targeted for launch in 2028.
Customer deliveries could potentially extend into 2029 or 2030.
That stretched delivery window sits alongside the broader reality described above: TSMC’s most advanced capacity is booked through 2028 and potentially beyond. Put simply, the timeline and the constraint are interlocked, and the redesign is one way to keep the platform on track inside those limits.
Supply chain diversification: expanding beyond a single foundry path
TrendForce notes possible moves beyond Taiwan
Industry analysts at TrendForce have noted that Nvidia may also diversify its Feynman supply chain beyond Taiwan.
This is presented as a potential strategic adjustment, consistent with the idea that Nvidia is looking for ways to reduce dependence on the most constrained slices of capacity.
Intel considered for less complex components like the I/O die
Reports suggest Intel is being considered for producing less complex components such as the chip’s I/O die.
The emphasis here is “less complex.” The idea isn’t that the entire platform moves away from TSMC’s advanced nodes, but that certain components could be sourced differently to relieve pressure on the most in-demand manufacturing resources.
What this situation signals for the semiconductor industry
This episode highlights the strategic leverage TSMC holds as the world’s dominant advanced chipmaker. It also points to a growing tension: surging AI demand is colliding with the physical limits of manufacturing capacity.
For Nvidia, the immediate outcome is a redesign and a more segmented manufacturing plan. For the broader industry, the underlying message is that cutting-edge node access is becoming a defining competitive constraint—one that can reshape product plans even for the biggest AI-focused chip platforms.
Q&A
What is Nvidia changing in the Feynman chip manufacturing plan?
Only the most critical components would use TSMC’s A16 (1.6nm) process, while less essential parts would shift to the N3P 3nm node.
Why is TSMC capacity forcing a redesign now?
Demand for advanced 2nm and sub-2nm processes from AI and high-performance computing customers is overwhelming, with advanced capacity fully booked through 2028 and potentially beyond.
When is Feynman expected to launch and reach customers?
Feynman is targeted for launch in 2028, with customer deliveries potentially extending into 2029 or 2030.