
→ Jensen Huang resigned on Nov 11, and CFO Colette Kress is now interim CEO. Nvidia's product roadmap and its orders for the first half of 2027 are still in place.
→ The silicon isn't what's at risk. What's at risk is the personal relationships Huang used to run Nvidia's Taiwan supply chain.
→ Expect capacity to be allocated by contract, not over dinner. That means faster diversification and more robots built on chips from more than one vendor.

trillion dollar banquet Read more: https://www.tweaktown.com/news/102686/nvidia-ceo-has-trillion-dollar-banquet-with-bosses-of-quanta-asus-acer-gigabyte-msi-more/index.html
1. The Stress Test
On Nov 11, Nvidia's board accepted Huang's resignation as CEO and director, effective immediately. Colette Kress will serve as interim CEO, alongside her CFO role, while the board considers internal and external candidates. Harvey C. Jones became the company's first-ever board chair. The resignation came during a board investigation into Huang's reported ties to an executive at Megaspeed, a Singapore data center company. Those allegations remain unproven, and Beijing has called them a smear.

Physical AI takes this harder than cloud AI does. A robotics team commits to one stack for years: Jetson Thor at the edge, Isaac, Omniverse and Cosmos for simulation, and hardware qualified against that silicon. Switching vendors means qualifying the whole robot again.
The near term is covered. Taiwanese exporters say many orders for the first half of 2027 are already placed. The open question is 2028 and beyond.

2. The Relationship Layer: Will It Change?

In Taiwan, Huang was more than Nvidia's CEO. He was its chief diplomat to suppliers.
→ Jan 2026: He hosted about two dozen top executives at the "trillion-dollar dinner": TSMC's C.C. Wei, Foxconn's Young Liu, Quanta's Barry Lam, plus leaders from MediaTek, Wistron, Inventec and Delta. Earlier in the same trip he had a private dinner with TSMC founder Morris Chang.
→ May 2026: More than 30 supply-chain CEOs joined him at a modest Taipei restaurant, covering wafers, substrates, power and liquid cooling.
→ Long-term commitment: Nvidia committed to Constellation, a new Taiwan headquarters in the Beitou-Shilin Science Park, expected to open around 2030.
Taiwanese media called him the only Western tech CEO who could bring 30 of the island's top executives to a casual dinner.
What stays the same: The working-level ties are built into both companies after decades of joint engineering, new-product ramps and shared test flows. Taiwanese suppliers say orders and capacity haven't changed, and that these partnerships can't be rebuilt overnight.
What changes:
→ The escalation path. When CoWoS, substrates or racks ran short, a call from Huang to a supplier's chairman cleared the jam. A CFO serving as interim CEO won't have that channel on day one, and neither will a successor without Huang's Taiwan roots.
→ Capacity built on trust. Suppliers partly built capacity ahead of Nvidia's orders on Huang's word. Expect firmer volume commitments, prepayments and take-or-pay terms instead.
→ Supplier hedging. With Nvidia's claim to first call weaker, Taiwanese firms have more reason to court AMD, hyperscaler ASIC programs and Tesla.
Watch: who gives the Computex 2027 keynote, whether the supplier dinner still happens, and whether Constellation stays on schedule.3. Where the Chain Lives
→ Silicon & packaging: TSMC (wafers and CoWoS), ASE/SPIL, Amkor, KYEC (test), and Unimicron and Ibiden (substrates)
→ Systems: Foxconn, Quanta, Wistron/Wiwynn, Inventec and Supermicro, with Jabil, Flex and Celestica as non-Taiwan options
→ Motion & perception:
Industrial arms: FANUC, ABB, KUKA, Yaskawa
Humanoids: Figure, Agility, Apptronik, Tesla Optimus
Components: Harmonic Drive Systems, Nabtesco, Hiwin, Sony (image sensors), onsemi
Robot software: Skild AI, Physical Intelligence
4. How Concentrated Is Taiwan?

→ Leading-edge logic: TSMC makes roughly 90% of the world's most advanced chips, including all of Nvidia's flagship chips. Arizona now produces some Blackwell wafers, but advanced packaging is still done in Taiwan.
→ AI servers: Taiwanese ODMs build most of the world's AI servers and racks.
→ CoWoS packaging: Mostly TSMC's own capacity in Taiwan, with ASE/SPIL and KYEC handling overflow and testing. Volume packaging in the U.S. is still years away.
The takeaway: the robot's brain comes from one island. Its muscles (harmonic drives, roller screws, sensors) come from Japan, Germany, China, and Taiwan's Hiwin.

5. Who's Moving In
→ For freed TSMC capacity: Taiwan's Commercial Times reports that Nvidia may slow chip orders. If it does, the freed TSMC capacity could go to AMD and to ASIC designers such as Broadcom, MediaTek, Alchip and GUC.
→ Hyperscaler ASICs: Google's TPU, Amazon's Trainium and Microsoft's Maia are competing for CoWoS capacity and ODM rack slots.
→ Edge compute rivals: Qualcomm, AMD embedded, Intel and MediaTek are each pitching themselves as a second source of chips for robots.
→ Vertical integrators: Tesla, with its AI5 chip made at both TSMC and Samsung, and Optimus competing for the same actuator suppliers.
→ The counterpoint: Taiwanese suppliers say nothing has moved yet, and switching suppliers takes years, not quarters.
6. Winners vs. Losers
Under pressure
Suppliers concentrated on Nvidia, as procurement becomes more formal and competitive
Robot makers built only on Jetson, which are likely to hedge rather than leave
Cross-border chip resellers in Southeast Asia, facing tighter export compliance
Positioned to gain
Capacity outside Taiwan: Amkor, Samsung, Jabil, Flex, Celestica
Robot software that runs on any chip
Owners of actuator and sensor bottlenecks, whose pricing power doesn't depend on who runs Nvidia
7. Bottom Line
Nvidia's CUDA and Omniverse moat and its engineering bench are intact, so Physical AI deployment doesn't stop. What ends is the era when one founder's relationships decided who got capacity.
→ Founders: qualify a second compute path now, while it's cheap.
→ Investors: watch CoWoS allocation, Computex 2027, and actuator lead times, not the headlines.
📊 Poll: Which Physical AI bottleneck worries you most in 2027: Advanced Packaging, Precision Actuators, or Edge Compute?
Reply with your vote, and share this with a builder who's betting on one chip.
Sources