Huang Renxun's resignation: AI's central banker is now a conservative
Nvidia's true risk isn't the scandal itself, but losing the only founder capable of driving transformation while providing industry-wide liquidity with a permissive style—replaced by a notoriously cautious ex-CFO.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Major transformation requires founders, not professional managers
Tod argues strategic shifts demand founders, not managers maintaining operations. In 33 years, Huang drove 4–5 major transformations: 2006 CUDA ecosystem against board resistance, 2013 deep learning bet, 2019 Mellanox acquisition for datacenter shift. A professional manager wouldn't have made these calls. Nvidia has no clear number-two; Huang directly manages 60+ executives, creating governance single-point-of-failure risk upon departure.
— TodAI industry's central banker replaced by a conservative
Liu Yilun metaphorically frames Huang as AI's central bank chief, aggressively providing industry liquidity—including backing $250B to OpenAI and 8G watts of infrastructure. His replacement, the ex-CFO, is historically cautious and previously opposed such massive guarantees that Huang overruled. Leadership change risks shifting industry liquidity from accommodative to restrictive, an impact deeper than any stock price move.
— Liu YilunNvidia faced a founder-or-bust moment even without the scandal
Tod outlines three structural challenges meeting founder departure: (1) Nvidia extracts 70–75% of industry gross margin; everyone else subsidizes it, sparking eventual reversal; (2) like GPU-to-ASIC transition in Bitcoin mining history, once Transformer architecture and parameter scale stabilize, everyone shifts to custom ASIC chips, eroding GPU's generalist premium; (3) CUDA is fundamentally software, and AI translation tech reduces cross-architecture code pain, weakening CUDA's moat. These converge at the worst moment.
— TodFounder premium is evaporating as Nvidia flat-lined over a year
The rally commenced late 2022 on monetary loosening and GPT-3.5 establishing AI infrastructure narrative. Since mid-2025, stock barely moved despite earnings delivery—textbook value-trap dynamics of valuation compression alongside rising earnings. Critical point: long-term holders weathered 20–30% swings unchanged because they believed in Huang's industry leadership. That ‘founder premium’ dissolves with him; justifying long-term valuation ceiling becomes impossible.
— TodAMD and Broadcom caught Nvidia's losses but can't solve long-term problems
Nvidia stock declines saw AMD and Broadcom rally in tandem—Taiwan Semiconductor reallocating advanced packaging capacity. AMD's TSMC advanced packaging share rose from 9% in 2026 to 20% in 2027; freed capacity likely went to AMD and Broadcom (TPU manufacturer). This represents short-term trading reallocation, not structural shift. CapEx growth peaking, Democratic administration potentially constraining US datacenter builds, China compute business off-limits—structural headwinds hit AMD and Broadcom equally. Fundamental beta upside exhausted.
— TodCrushing competition backfires: blocking domestic players forces genuine innovation
Tod offers counterintuitively: if Nvidia's void can't be filled by AMD and Google TPU, only domestic GPU steps in—‘赌不如输’. True defeat of rivals requires market flooding with cheap goods leaving domestic players zero development space; instead, access blocks create exactly that space. Liu adds: domestic card makers' constraint isn't demand but upstream capacity (SMIC, North Chip). DeepSeek already adapts using open-source CUDA alternatives like TileLang; mainland GPU training transition may accelerate.
— Tod/Liu YilunB300 servers jumped 4M to 20M yuan; channels panic-hoarding
Liu reports frontline channel intelligence: the GPU hoarding window is explicit and narrow—buy everything available, no hesitation. A B300 server cost ~4M yuan domestically before Spring Festival; by October, 20M yuan. Meanwhile BAT collectively ordered over 3M compliant Pro5500 servers for compute-intensive preprocessing, but actual delivery sits below 5% of expected volume.
— Liu YilunHuang must prove he didn't know, or knew but opposed it
Tod judges Huang's short-term return nearly impossible. This scandal fundamentally targets Trump, not Huang directly—standard playbook: damage reputation, reassign from key roles, construct charges, publish reputation reversal. Huang's immediate task: prove he didn't know about high-performance chip shipments reaching unauthorized destinations, or knew but disagreed and subordinates acted without authority. Failing to distance himself risks criminal exposure.
— TodUS chip export controls now extend four layers deep
Liu systematically maps America's export control infrastructure: (1) destination controls—advanced chips to mainland China and Macau require license; (2) ultimate parent penetration—entities registered in Singapore or Cayman face restrictions if HQ is mainland; (3) end-user and end-use audits—even foreign buyers trigger controls if sellers know chips will transfer to restricted entities; (4) foreign direct product rule—overseas foundries like TSMC and Samsung face controls on chips using designated US technology even if produced abroad. This architecture continues tightening.
— Liu YilunTech giants face hidden conflicts allocating compute resources internally
Liu reveals Alibaba Cloud's bind: to maximize revenue, it must rent capacity to external model vendors like Kimi and Zhipu, but Tongyi Qianwen uses Alibaba Cloud without internal billing, removing sales incentive to reserve cards for internal use. Balancing internal and external demand remains unresolved. Tencent takes the compliance path, using H20-class chips internally while offering third parties only Kunlun and Haiguang chips. Independent model companies chronically shortage capacity, exploring token-split: capacity providers earn 70%, model providers 30%.
— Liu YilunIn their own words · checked verbatim
I find it deeply sad. I really don't think this is targeting Huang at all. In reality, it's ultimately targeting Trump.
我觉得非常可悲 我觉得这事根本就不是冲黄来的 实际上最后背后就是川普
Tod0:00
The timeline is actually very clear: in coming days there will be frenzied GPU hoarding—place all the orders you can, buy whatever you can get. Don't harbor any illusions. American controls are only getting stricter.
这个时间范围其实说得非常明确 就是最近几天会疯狂地去囤GPU 就下GT2订单 有多少买多少 大家不要抱有侥幸心理 这个美国的管控已经是会越来越严格的
Liu Yilun0:00
That's why people sometimes say: after it crashes enough, they just pump themselves up saying ‘just believe in Lao Huang, he knows better than I do’.
所以大家有时候说 这个跌麻了之后 就跟自己打气说 相信老黄就完了 他比我懂
Tod13:09
It's like what the Democratic party did—sometimes losing is better than competing. If you really want to crush it, you'd flood the market with cheap goods so domestic players have zero development space. But by blocking instead, you forced them to innovate themselves.
就像民主党这种做法 其实赌不如输 就如果你想真正击溃它 你就比如大规模的倾销 你就不让它国产 有任何发展的空间 但你反而堵起来以后 它只能被迫就是自己做
Tod21:14
That October 2025 report documented close ties between Nvidia and Megaspeed, and suggested Megaspeed was selling Nvidia's high-performance chips to unauthorized destinations.
25年10月那个报道 他是报道了英伟达和Megaspeed之间 有比较密切的关系 以及Megaspeed 有可能就是在向一些县职地 去卖英伟达的高性能的芯片
Liu Yilun42:24
On November 5th, three Congressional committees requested that Huang Renxun appear for closed-door questioning on November 18th.
11月5号 国会的三个委员会 就是要求黄仁群 11月18号那一天 要去接受闭门质询
Liu Yilun47:25
Figures
| Nvidia stock dipped to | 148.8 USD | 11:08 |
| B300 server domestic price (pre-Spring Festival) | ~4M yuan RMB | 26:09 |
| BAT combined Pro5500 orders | 3M+ units | 27:15 |
| BAT combined Pro5500 orders value | over 35B USD | 27:15 |
| Hong Kong H200 inventory | 600K units | 55:30 |
| Token split ratio | Compute provider 70%, model provider 30% | 1:04:43 |
Glossary
- ASIC
- Application-specific integrated circuit custom-designed for particular algorithms; higher performance but less flexible than GPU.
- CUDA
- Nvidia's GPU programming ecosystem; the driver layer connecting hardware circuits to upper-layer software and a core part of Nvidia's competitive moat.
- TileLang
- Open-source operator-writing tool from Peking University researchers; analogous to Triton in CUDA ecosystem, helping escape Nvidia's software stack.
- Foreign direct product rule
- US export control mechanism treating any overseas chip using designated US technology as controlled even if produced outside America.
How to listen
Secondary market investors tracking Nvidia stock, practitioners following domestic GPU substitution opportunities, and those requiring understanding of US-China chip export control mechanisms for international business and compliance.
14:11–16:12 section covering detailed US equity terminology like ‘Davis double’ and value traps; skippable for those not trading American stocks.