Jensen Huang resigned as Nvidia’s CEO on November 12, effective immediately, after 33 years running the company he co-founded. The board named CFO Colette Kress as interim CEO and opened a search for a permanent replacement. The trigger: an extramarital affairs scandal, one involving a woman U.S. intelligence officials identified as a Chinese operative. For AI developers, the real question is not what happened in Huang’s personal life — it’s whether any of this touches the GPU stack you depend on.
Nothing breaks today
CUDA still compiles. Your training jobs still run. The GPU roadmap Nvidia published is still the roadmap: Rubin entered full production before Huang left, Rubin Ultra is confirmed for H2 2027, and Feynman follows in 2028. All customer commitments remain in force per the board’s statement. Kress, a 13-year Nvidia veteran who helped engineer the Mellanox acquisition and the company’s AI infrastructure pivot, opened with a continuity message. Ian Buck — the engineer who built CUDA — still works at Nvidia and is reportedly on the CEO shortlist. If Buck gets the job, the developer roadmap probably stays exactly where it was.
The CUDA moat just got more interesting
One day before the resignation, Anthropic dropped a detail that matters more than the scandal. The company revealed that its internal Mythos model had ported Claude’s prompt-processing workload to AMD MI355X GPUs in roughly three days, with almost no human involvement. The same task has historically taken six to twelve months of kernel engineering. Anthropic’s director of computing confirmed they now run Claude on four chip vendors: AWS Trainium, Google TPU, Nvidia GPUs, and AMD GPUs.
This is not a fluke. AMD’s ROCm 10 — launched in August 2026 — claims 3.3x inference gains over its previous version and now ships with stable support for vLLM, SGLang, FlashInfer, PyTorch, and TensorFlow. The performance gap between AMD and Nvidia for standard ML workloads has narrowed from a 2-5x deficit to roughly 10-30%. Custom kernels still require explicit rewriting, and HIP conversion alone does not achieve parity. But for inference on standard model architectures, AMD is a real option in a way it was not two years ago.
Huang was the person most identified with Nvidia’s developer ecosystem and its “CUDA or nothing” gravitational pull. His departure does not break CUDA, but it removes the clearest symbol of the platform’s singular authority at exactly the moment a credible alternative is maturing.
The supply chain problem nobody is talking about
Nvidia carried $279 billion in component and manufacturing commitments as of its July 2026 filing. It also held $36 billion in backstops for neocloud operators — the mid-tier GPU cloud companies that raised capital partly because Jensen Huang personally vouched for them. Fortune’s assessment was direct: “That Nvidia guarantee is not worth what it was worth a week ago.” Neocloud stocks fell the day of the resignation announcement.
For individual developers this matters less immediately than it does for infrastructure teams. But if any of those commitments unravel, GPU availability tightens. The companies that absorb the first squeeze are not hyperscalers — they already have long-term capacity contracts. The companies at risk are mid-sized teams renting compute from neoclouds month to month.
What to actually do
Three things worth doing in the next 30 days:
Watch the CEO search. The board’s choice of permanent CEO is a signal. An internal technical hire — Buck is the name to track — means the roadmap stays on its current trajectory. An external or finance-focused hire signals the board is in control mode: shorter planning horizons, more shareholder focus, less tolerance for long-horizon developer programs.
Start a ROCm parallel evaluation. Not a migration — an evaluation. Stand up a small inference workload on AMD hardware alongside your existing stack. SemiAnalysis outlines the full competitive picture. Anthropic’s three-day result is not a reason to migrate today. It is a reason to know your switching cost before you need to know it.
Do not change your production stack. Huang leaving does not break CUDA. The software does not know who the CEO is. If you have stable training and inference pipelines, keep running them. The uncertainty is real but it is medium-term, not this week.
Huang built Nvidia into something genuinely exceptional. Thirty-three years of compounding technical bets produced a GPU platform the entire AI industry runs on. That does not evaporate in a resignation. What changes is the degree of conviction you can place on that platform remaining uncontested for the next five years — and after Anthropic’s three-day migration, that conviction was already eroding.
