Huang's exit leaves a gap, and chip challengers are pitching the memory wall

The opening

Nvidia's near-term business is not at significant risk after Jensen Huang's departure, according to Bernstein Research on the Odd Lots podcast. The long-term roadmap is the part at risk, because Huang was the strategist behind the long-term architecture calls. That gap is exactly where specialised chipmakers are now pitching their case. m:…0452

Fractile chief executive Walter Goodwin told the Twenty Minute VC podcast that the memory wall is the primary constraint on large-model inference: bandwidth has scaled 25,000 times more slowly than raw compute over the last 25 years. His company's answer is to stack memory directly on top of logic, aiming for the speed of SRAM with the capacity of DRAM. The claim is his company's own, and the test is whether the chips ship and win real workloads. m:…0297

Goodwin also argues that vertical integration is now a strategic necessity for frontier labs, which need to protect their margins rather than be commoditised. He points to DeepSeek's inference stack as evidence, saying it delivers higher throughput and better interactivity than standard inference servers. The logic is that the labs with the most to lose from generic hardware will pay for hardware built for their workloads. m:…0297

What Huang held together

Bernstein's Stacey Rasgan, speaking on Odd Lots, described Huang as the driver of Nvidia's role as a financier of the AI ecosystem, the company that funded customers and cut supply deals on personal handshakes. She said the downgrade followed disclosures about Huang's relationships with subordinates and dishonesty with the board. Whatever the board decides next, the personal-deal machinery does not transfer automatically to a successor. m:…0452

The revenue map shows why the succession matters beyond one company. Bernstein puts US customers at over 60% of Nvidia's revenue and Taiwanese customers at roughly 20%, with direct sales to China at about 8% and Malaysia and Singapore at about 1%. The small China share is the point: the fight over where chips end up is about control of the technology, not about current sales. m:…0452

Bloomberg's Surveillance panel discussed reports that Nvidia chips intended for Singapore were diverted to a company in Shanghai with ties to the Chinese government. Odd Lots also raised the prospect of a US crackdown on Chinese users reaching AI capacity remotely. Whoever runs Nvidia next inherits both the supply chain and the scrutiny around it. m:…9685 m:…0452

What to watch

Goodwin's other warning is about the bottleneck nobody pitches around: data-centre capacity. He told Twenty Minute VC the shortage of data-centre capacity is a more acute constraint than the silicon supply chain, and that the build-out is a long-term, positive-sum game rather than a bubble. If he is right, the challengers' window stays open even if Nvidia's chips stay ahead, because the constraint moves to power and buildings. m:…0297

For a venture investor, the question is not whether a startup can beat Nvidia's best chip. It is whether a narrow bet, on memory bandwidth for inference, can win enough frontier workloads to matter while Nvidia sorts out its leadership. The pitch is narrow by design: win the inference workloads where memory, not raw compute, is the limit. That is a smaller prize than replacing Nvidia, but it is a prize a specialist can actually claim. m:…0297