Nvidia scandal persists as AI chiefs keep queueing for chips

Nvidia scandal persists as AI chiefs keep queueing for chips

On November 16, 2026, The Verge AI reported that despite continuing controversy over Nvidia’s market power, Jensen Huang still controls AI compute supply as CEO, co-founder and president, and AI company chiefs are still lining up for his chips. The article called Nvidia chips the bedrock of advanced AI technology in the United States and said demand keeps outrunning supply. The controversy has not weakened Nvidia’s pricing power; it has exposed how much global AI infrastructure depends on one

On November 16, 2026, The Verge AI published a piece titled The Nvidia scandal isn’t slowing down. The article revisited Huang’s unusually broad influence over AI’s direction as Nvidia’s CEO, co-founder and president. A week earlier, The Verge had described Nvidia’s chips as the bedrock upon which advanced AI technology in the US is being built. This week’s article added that ever-increasing demand has consistently outpaced supply, leaving AI CEOs effectively bowing to Huang for allocation. The report did not disclose a new regulatory penalty, lawsuit figure or financial number; its core fact was the coexistence of a persistent controversy and a buying rush. The signal is blunt: public debate over the limits of Nvidia’s power has not cooled, but customer behavior has not changed. Nvidia GPUs are used for pretraining, post-training and inference clusters, and the latest generation still

Why now: Nvidia has become the biggest beneficiary of AI capital spending over the past two years. Data-center GPUs, NVLink interconnects and the CUDA software stack form the de facto standard for training clusters. Cloud providers and model developers have concentrated compute budgets on Nvidia’s roadmap, while TSMC’s CoWoS advanced packaging and HBM supply limit how fast the industry can add capacity. That has widened the supply-demand gap. Multiple regulators have already opened antitrust inquiries into Nvidia, and the market has repeatedly debated its share, bundling practices and customer lock-in. Export controls have further fragmented regional supply. In an earlier comparable period,

Liu Gong's take: over the next six months, Nvidia is likely to keep more than 80% of AI accelerator revenue, but custom ASICs from cloud providers plus AMD will approach 20% of new compute purchases for the first time. The controversy will not make orders disappear; it will make customers prepare second sources faster.

From a technical standpoint, Nvidia’s moat is not single-chip performance but the system layer: CUDA, NVLink, Spectrum-X networking and rack-scale delivery. B200 carries 192GB of HBM per GPU, 140% more than H100’s 80GB, and GB200 NVL72 places 72 GPUs and 36 Grace CPUs in one interconnect domain. That vertical integration makes it hard for rivals to catch up with chip specs alone. Liu Gong sees a similar pattern in industrial vision inspection: once a production line has calibrated a set of cameras, lighting and algorithm stack, switching costs far exceed the unit-price difference. AI clusters work the same way. Once model code, communication libraries and operations tools are optimized around Nvidia, customers may complain about monopoly in meetings and still write NV on the purchase order. More important, fault recovery and linear scaling efficiency in training clusters depend on accumulated engineering experience. Even if a second source matches paper performance, customers face months of tuning.

On the business side, the controversy actually reinforces Nvidia’s scarcity pricing power. Public filings show Nvidia’s data-center revenue reached about $115 billion in fiscal 2025, up roughly 142% year on year, with gross margin around 75%. When demand exceeds supply, allocation rights are profit. Custom cloud chips can lower inference costs, but they cannot quickly replace the CUDA ecosystem on the training side. Liu Gong predicts that within 12 months, leading cloud providers will prioritize custom chips for recommendation, search and some inference workloads, while training clusters remain anchored on Nvidia. The result is a two-track structure: Nvidia for training, multiple sources for inference. That structure will slow Nvidia’s growth rate, not reverse its absolute revenue, because total AI demand is still expanding and inference diversification enlarges the overall compute pool.

For the industry, regulation will raise Nvidia’s compliance costs but is unlikely to change the supply landscape within six months. Antitrust investigations usually take years, and customers will still lock in orders for 2027 cluster expansions while probes proceed. Export controls split regional markets into different price bands, encouraging gray trade and local substitutes. The real medium-term variable is not litigation but whether AMD and custom ASICs can close the software gap after TSMC’s CoWoS capacity expands. Liu Gong believes that if second-source cluster interconnect and compiler tools remain one generation behind for another 12 months, Nvidia’s share will fall from its peak but not collapse. If second sources cut unit costs for inference by more than 30%, cloud procurement structures could see a real inflection in 2027.

  • Nvidia’s data-center revenue reached about $115 billion in fiscal 2025, up roughly 142% year on year; total revenue was about $130.5 billion, with data center close to 88%.
  • Public market research generally puts Nvidia’s AI accelerator share at 80%-90%, AMD at single digits to 10%, and cloud custom ASICs combined below 20%.
  • B200 offers 192GB of HBM per GPU, 140% more than H100’s 80GB; GB200 NVL72 puts 72 GPUs and 36 Grace CPUs in one interconnect domain.
  • Nvidia’s gross margin is around 75%, while TSMC’s CoWoS advanced packaging capacity remains the expansion bottleneck; an eight-GPU H100 server system typically costs hundreds of thousands of dollars, and system premiums widen when

For the industry: over the next three months, regulatory questions and public pressure will continue, but AI cloud capex is already committed and GPU allocation will stay tight. The near-term pain falls on smaller AI companies facing queue costs and cloud GPU rents; large customers are more likely to secure quotas. Within a year, Nvidia will remain hard to replace for training, while AMD, Google TPU, AWS Trainium and other custom routes will accelerate diversion on inference, pushing customers from single sourcing to dual sourcing. For ordinary users: within three months, AI application prices are unlikely to fall because of the controversy, and popular models will still face compute

Across public discussion, AMD markets large HBM capacity and an open software stack as differentiators, Intel emphasizes Gaudi’s price-performance, and cloud providers prefer to build custom ASICs rather than attack Nvidia publicly. Analysts generally argue that antitrust probes take years and are unlikely to change GPU allocation in the short term. Some enterprise customers support multi-sourcing but acknowledge high CUDA migration costs. Developer communities are more direct: they complain about GPU rents and quotas while continuing to write kernels and operators for Nvidia platforms. Academics and policy groups focus on allocation power: if a handful of cloud providers and Nvidia jointly

Questions worth exploring:

  • Why Nvidia’s CUDA ecosystem is harder to replace than raw chip performance
  • The pace at which cloud custom ASICs divert inference workloads
  • How antitrust and export controls reshape global compute pricing

Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.

This item comes from The Verge AI (published 2026-11-16 13:00:00). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private

Leave a Reply

Your email address will not be published. Required fields are marked *

中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn