Todd | 晚点聊 LateTalk Summary

1. Key Themes

Nvidia as the AI Industry's "Central Bank" — and a Hawkish Governor Takes Over

The most original framing of the episode: Nvidia, via Jensen Huang, functioned as the AI sector's monetary authority, supplying liquidity through credit guarantees and circular financing. Replacing a dovish chair with a conservative CFO tightens the whole system. Todd explains: "我们有时候讲说现在 AI 这个行业的中央银行的行长就是黄仁勋。那这个行长的风格呢相对比较激进，就像比如说你中央银行它不断地在降息或者扩表，就整个行业实际上流动性是比较充沛的。" (Translated: "We sometimes say the central bank governor of the AI industry is Jensen Huang. His style was aggressive, like a central bank constantly cutting rates and expanding its balance sheet, so liquidity across the industry was abundant.") [00:06:27] On the successor: "其实 CFO 呢这个人呢就其实一直就比较保守和谨慎。比如说包括英伟达对比如 OpenAI 2500 亿美金的 8GW 的建设的应该是最终的贷款人，其实本质上就是一种提供信用。现在...马上他就开始又提要反对。" (The CFO has always been conservative and cautious; Nvidia is effectively the lender of last resort for OpenAI's 8GW, $250 billion buildout, and he is already moving to oppose it.) [00:06:50] The consequence: "如果之前是螺旋上升的话，一旦开始收缩的话就变成螺旋下降。" (If it was a spiral up, once contraction starts it becomes a spiral down.) [00:37:21]

The Founder Premium: Key-Man Risk, Governance Gaps, and Taiwan Supply Chain Relationships

Todd argues founders drive strategic pivots that managers can't, and Nvidia has no deputy. Huang pushed at least "4 到 5 次" transformations in 33 years [00:03:00]. Manchi adds the specifics: CUDA in 2006 despite board opposition, the 2013 deep learning bet, and the 2019 Mellanox acquisition [00:03:20, 00:03:42]. On structure: "他 60 多个高管实际上都是他直接对的...如果公司没有比如说非常坚定的二号位或者三号位，这个就一下就类似有可能就这个局面就陷入瘫痪的状态了。" (His 60-plus executives all report directly to him; without a firm number two or three, the company could be paralyzed.) [00:04:40] Huang's Taiwanese roots also formed a profit-sharing community: "大家听黄的都挣到钱了...这个事就变成了一个非常大的利益共同体，这个事情也是我觉得其他人替代不了的。" (Everyone who followed Huang made money...it became a huge community of shared interest that nobody else can replace.) [00:06:02]

"Old-Guard Stock" Valuation: Earnings Up, Multiple Down, Capex Second Derivative Peaked

Todd's investment thesis: the AI infrastructure trade is largely played out in public markets. "AI的大的基建周期，从二级市场上讲啊，就基本上可能也就这样了。因为达大概在 25 年下半年以后，到现在基本就没怎么涨过了，但是你业绩是很好，所以这是典型地进入一个老登股的状态，就是估值在不断地下跌，业绩不断地在上升和兑现。" [00:15:42] Hyperscaler cash is exhausted: "每个季度挣多少，基本上花多少，甚至还要多借一点点的状态。" (They spend about what they earn each quarter, even borrowing a bit more.) [00:15:19] His framing of the cycle's start: late 2022 combined the inflation second derivative turning negative with GPT-3.5, which converted "原来搞不明白的科学问题给它转化成了既定的工程问题" (an unsolved science problem into an established engineering problem). [00:14:55] Nvidia guided $400B revenue in 2026 and $700B in 2027, yet trades at under 10x on 2028 numbers [00:13:56, 00:16:23].

Nvidia's Three Structural Moats Are Each Under Threat

Todd lists: (1) 70-75% gross margins mean "全行业都在给它打工" (the whole industry works for it), creating incentive to defect [00:08:31]; (2) the Bitcoin-mining precedent, where GPUs gave way to ASICs at "成本是 GPU 的 1%" (1% of GPU cost) once the algorithm stabilized, and the same could happen if Transformer architecture and scaling laws plateau [00:09:10, 00:09:40]; (3) CUDA is software, and software is cheap in the AI era: "如果这个东西是软件的话，那 AI 时代里可能软件是最不值钱的...现在用 AI 翻译其实并不那么麻烦，就你其实所谓 CUDA 的壁垒就没那么高企了。" [00:09:57, 00:10:23] Manchi notes the irony that Nvidia's own success is eroding its moats [00:10:50].

Who Wins and Loses on Nvidia's Wobble

Todd sorts the market into three groups. Winners: AMD and Broadcom (the latter via Google TPU), through reallocation of TSMC advanced packaging. "AMD 在 26 年占台积电的先进封装大概 9 个点，然后到 27 年加早 20%...多让出来的 CoWoS 先进封装产能大概率主要会被 AMD 吃掉，然后次要可能会被博通，也就是 Google 的 TPU 吃掉。" [00:17:57] Losers: CoreWeave and Nebius, the "地方小银行" (local small banks) funded by Nvidia credit: "央行一旦开始流动性收缩或者转变...CoreWeave 可能是最先倒霉或者遭殃的。" [00:20:20, 00:20:35]; plus Southeast Asian GPU resellers and Hong Kong stocks tied to GPU diversion [00:20:50]. Potential winners: domestic Chinese compute, because "堵不如疏" (blocking is worse than channeling) creates a vacuum only domestic chips can fill [00:21:32]. His trading call: the AMD/Nvidia rotation likely reverses once Huang's departure is official, and the short-term trade ends [00:19:14].

China's Compute Reality: Supply-Constrained, Not Demand-Constrained

Yilun Liu reports domestic card capacity is already fully booked: "国产卡的产能已经被大厂们包圆了，所以它的问题在于上游供应不足，而不是说下游的需求不足。" [00:23:16] SMIC allocated most capacity to Huawei (roughly 700-800K chips per year), so small GPU firms are bottlenecked [00:26:01]. Software ecosystem work may accelerate: Manchi notes DeepSeek began using TileLang in 2025 as a Triton/CUDA alternative [00:24:20] and expects training on domestic chips to become a key turning point [00:24:36]. A gap remains: a researcher at a major Chinese firm said OpenAI has "100 倍" (100 times) their compute [00:33:26]. Todd adds that Chinese model labs will fall further behind US labs unless domestic chips catch up quickly [00:32:04].

The Chip Smuggling Compliance Stack: Four Layers of US Control

Yilun lays out how US export controls close the offshore loophole: (1) destination controls covering mainland China and Macau; (2) ultimate-parent look-through, with BIS clarifying on 31 May 2026 that even Singapore or Cayman entities need licenses if the ultimate parent is in mainland China or Macau; (3) end-user and end-use scrutiny; (4) the foreign direct product rule, which reaches TSMC and Samsung output [00:57:10-00:59:06]. "而且这个趋势还会进一步收紧。" (And the trend will tighten further.) [00:59:21] Context: about 600,000 H200s are reportedly in Hong Kong, and Chinese firms will pay "两到三倍的价格去租一个 3 到 5 兆瓦的机房" (two to three times the price to rent a 3-5 MW data center) [00:55:55, 00:56:15].

Compute Market Cycle Whiplash: From Glut in 2025 to Scramble in 2026

"去年整个算力还是一个供大于求的状况，就完全是买方市场...最离谱的一个事情是 25 年有一些模型公司还在把算力退掉。" (Last year compute was in oversupply...the most absurd thing is that some model companies were returning compute in 2025.) [01:00:01, 01:00:06] Now server prices have jumped "从年初的 400 万涨到现在的 2000 万" (from 4 million to 20 million RMB) driven by tighter import channels plus coding and agent workloads [01:00:06]. Financial leasing firms went from fearing defaults to "如沐春风，要大干一场" (basking, ready to go big) [01:00:58]. Intra-company allocation priorities reveal strategy: ByteDance's order is Seedance, Seed, other R&D, then external cloud; Tencent's is Hunyuan, WorkBuddy, Xiaowei, then cloud [01:02:32].

Politics Is the Real Driver: Democratic Midterm Wins and Escalating Investigations

Todd reads the scandal as targeting Trump, not Huang: "这个事根本就不是冲黄来的，实际上最后背后就是川普。" [00:48:50] His four-step playbook: damage reputation, force reassignment, build charges, release publicly [00:44:25]. Manchi supplies the investigative facts: Commerce first, then transfer to the Department of Justice (considering criminal liability), FBI on possible leaks, and Congress, which on 5 November demanded Huang appear 18 November for a closed-door hearing, though all three committee chairs are Republican [00:47:24-00:48:12]. Todd's framing of Huang's task: prove he either didn't know or didn't approve the diversion [00:49:38].

The "Slowdown" Debate and the Chain-Reaction Analogy

Manchi suggests Huang's exit could unintentionally deliver what Musk's six-month pause petition wanted [00:38:48]. Todd counters with Oppenheimer: "一旦开启这个军备竞赛，没有人能停得下来...直到大家竞争到一个均衡点" (once an arms race starts, nobody can stop it...until everyone reaches an equilibrium). [00:39:07] Manchi also notes two-sided harm: US AI slows from Democratic scrutiny while China is hit by tighter embargo [00:41:13].

2. Contrarian Perspectives

Infrastructure Is Done in the Public Markets; the Value Will Shift to Applications That Don't Yet Exist

Against the consensus that AI capex has years left to run, Todd says: "我觉得基建应该是真的走完了，就二级市场上行来讲，基建各方面来看真的走完了，就从股票的角度讲差不多了。" [01:13:14] Supporting it: the hyperscalers are spending roughly what they earn, so the funding engine is spent [00:15:42]. Yet he stays bullish long-term on applications: "离用户越近啊，一般来讲市值越大、营收越大" in a mature industry, but AI is inverted [01:10:27]. "只要这两个条件...模型能力...提高...模型的成本每天都在下降...会出现很多你根本原来意想不到的很多应用。" [01:11:02]

Anthropic at $1.4T Is a Pass — and a Short Candidate at $2T

Todd rejects the consensus darling: "不买，太贵了，不买。2 万亿肯定不买，1.4 万亿也不买。" (No, too expensive. Definitely not at $2T, not at $1.4T.) [01:12:02] His reasoning: "Anthropic 你很难找到加速的第二曲线...真正找到就是写代码这一个应用。" Drug discovery faces physical limits like clinical phases and patient counts, and robotics isn't ready [01:12:14, 01:12:39].

Blocking China's Chips Backfires: "堵不如疏"

Contrary to the restrictionist view, Todd argues "你想真正击溃它，你就比如说大规模地倾销，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。那可能你 5 年、10 年以后再看，你就会发现你其实完全不如就是大规模倾销。" (If you really wanted to crush it, you'd dump chips at scale. Blocking forces it to build its own, and in 5 to 10 years you'll find dumping would have been far better.) [00:21:52] Manchi confirms this was Huang's own long-running argument [00:22:20].

CUDA's Moat Is Overrated Because It's "Just Software"

Most treat CUDA as Nvidia's deepest moat. Todd says AI makes software the cheapest layer: translating between CUDA and Huawei's CANN is now an AI task, not a painful manual one [00:10:23]. If true, switching costs collapse faster than the market assumes.

A Cook-Run Nvidia Would Be Worse for Taiwan's Suppliers Than Huang

Todd relays market banter with real signal: "如果库克来了就大家就完了，就是最严厉的父亲就来了...他把利润全部或者几乎全部集中在苹果自己手里边，然后跟着他做的兄弟其实都是就是干一个苦活，然后毛利都很低。" [00:52:46] Huang's regime shared profits with suppliers; a supply-chain operator CEO could claw margin back, which reshapes the economics of the whole Nvidia ecosystem.

3. Companies Identified

Nvidia — Dominant AI chip maker (70-75% gross margin), now under investigation. Mentioned for exceptional earnings amid a de-rating. "你说英伟达它没有增长吗？这非常夸张，70 左右的毛利，26 年 4000 亿的营收，27 年要 7000 亿的营收。" (Todd) [00:16:23]

AMD — GPU/accelerator maker; the top beneficiary of Nvidia's weakness. "AMD 在 26 年占台积电的先进封装大概 9 个点，然后到 27 年加早 20%。" (Todd) [00:17:57]

Broadcom — ASIC/custom accelerator supplier (including Google TPU). "多让出来的 CoWoS...大概率主要会被 AMD 吃掉，然后次要可能会被博通，也就是 Google 的 TPU 吃掉。" (Todd) [00:17:57]

Google (TPU / GCP) — Hyperscaler with its own accelerator; its TPU pricing threatens Nvidia margins. "英伟达的卡毛利是七成，那 Google 的 TPU 就说我们五成的毛利就可以，甚至我不怎么挣钱我都好商量。" (Todd) [00:30:30]

TSMC — Foundry and advanced-packaging (CoWoS) supplier; the choke point for allocation. Referenced in Todd's CoWoS allocation analysis [00:17:57] and by Yilun as covered by the foreign direct product rule [00:59:06].

CoreWeave — GPU cloud heavily backed by Nvidia credit; flagged as the most exposed. "CoreWeave 这个公司非常激进地扩张...央行一旦开始流动性收缩或者转变...CoreWeave 可能是最先倒霉或者遭殃的。" (Todd) [00:20:35]

Nebius — Neocloud grouped with CoreWeave as a "small bank" dependent on Nvidia liquidity. (Todd) [00:20:20]

OpenAI — Nvidia's major financing counterparty; "2500 亿美金的 8GW" buildout at risk if guarantees are cut. (Todd) [00:06:50]. Yilun notes its compute is allegedly 100x a major Chinese lab's [00:33:26].

Anthropic — Reported at ~$1.4T; silent on the scandal, and tense relationship with Nvidia. Todd would not buy it [01:12:02]. Manchi: "Anthropic 也挺正常，Anthropic 本来和英伟达的这个关系就会比较紧张。" [00:46:53]

Huawei (Ascend / CANN / Huawei Cloud) — China's leading domestic AI chip platform; its biggest customers buy directly. "华为昇腾的卡其实像它最大的客户可能就是字节和 DeepSeek...两家模型公司自己去持有这些卡。" (Yilun) [00:31:19] Annual shipments "大几十万颗、七八十万颗" [00:26:01].

SMIC (中芯国际) — China's foundry; the true bottleneck for domestic GPU supply. "这其实中芯国际、北方华创他们得去解决的问题。" (Manchi) [00:23:22]

Naura (北方华创) — Chinese semiconductor equipment maker, named as a key upstream constraint [00:23:22].

Cambricon (寒武纪) and the "GPU Four Dragons" — Domestic GPU makers that may gain capacity if SMIC frees up. "匀出来的产能或许就可以给 GPU 四小龙或者寒武纪这些公司去使用。" (Yilun) [00:26:01]

Moore Threads (摩尔线程) — Domestic GPU firm named among competitors to Alibaba's T-Head [00:30:30].

Alibaba / T-Head (平头哥) / Alibaba Cloud — Has in-house chips and is pivotal in China compute; also renting compute to Kimi and Zhipu. "阿里云就不得不把更多的算力租给外面的独立基模厂商，包括 Kimi 和智谱。" (Yilun) [01:03:01]

Baidu / Kunlunxin (昆仑芯) — Baidu's chip arm; Tencent buys Kunlunxin cards for external cloud. "腾讯现在对外提供云服务只能用昆仑芯还有海光的卡。" (Yilun) [01:02:59]

Hygon (海光) — Domestic chip supplier used by Tencent's cloud [01:02:59].

Tencent — Notably compliance-focused; bought heavily into H20. "腾讯会说折合成 H20，因为在 H20 合规的时候，腾讯真的买了很多 H20。" (Yilun) [01:02:32]

ByteDance — Early aggressive card hoarder and Ascend's largest customer [01:02:32, 00:31:19].

DeepSeek — Chinese model lab; major Ascend buyer and ecosystem pioneer (TileLang adoption). "梁文锋...提到了就说 25 年的时候他们就开始用 TaiLong" (transcribed; refers to TileLang). (Manchi) [00:24:20]

Zhipu (智谱) — Chinese model lab; pursuing token-revenue-share, 30% reported with AWS. "目前这种 token 分成的探索一般都是 3:7 分成，提供算力的那一方分 7 成，然后模型厂商分 3 成。" (Yilun) [01:05:25]

Moonshot AI / Kimi — Independent Chinese model lab renting compute from Alibaba Cloud [01:03:01].

Fireworks AI — US inference startup whose growth in privately deploying Chinese open models could be hurt by bans. "他们之前有一个增长比较快的业务就是...去给美国的大企业客户做私有化的部署...已经是很知名的一级市场的独角兽了。" (Manchi) [00:34:33]

Megaspeed — Southeast Asian chip reseller, the trigger for the investigation; flagged in the 2025 reporting [00:21:06, 00:42:57].

Palantir — Publicly defended Huang alongside Musk (Manchi) [00:46:31].

Meta — Hyperscaler spender; core staff silent on the scandal (Manchi) [00:46:31].

Mellanox — Networking company Nvidia acquired in 2019, central to its data-center pivot. (Manchi) [00:03:42]

Bitmain (比特大陆) — Bitcoin-mining ASIC maker, used as the precedent for GPU-to-ASIC displacement. (Manchi) [00:09:10]

Apple — Referenced via Tim Cook's supply-chain margin capture [00:52:46].

Delta Electronics (台达) and Taiwanese suppliers — Part of Nvidia's profit-sharing supply chain. (Todd) [00:05:37]

Concurrent Technology (协创数据) — Compute provider in Zhipu's token-sharing model. (Yilun) [01:04:56]

Parallel Technology (并行科技) — Independent cloud cooperating with Zhipu [01:04:06].

Polymarket — Prediction market suggested as a tracker for the next Nvidia CEO. (Todd) [00:53:01]

4. People Identified

Jensen Huang (黄仁勋) — Nvidia founder, now resigned as CEO. Cited as the founder behind at least four to five company-defining pivots. "黄实际上在过去 33 年的过程中里，实际上几次，大概至少 4 到 5 次推动了公司有非常大的转型和变化，其实回头看都是正确的。" (Todd) [00:03:00]

Chris Malachowsky — Nvidia co-founder, no longer in core operations (Manchi) [00:04:10].

Curtis Priem — Nvidia co-founder who retired in 2003 (Manchi) [00:04:10].

Nvidia's CFO (interim CEO) — Conservative successor; "其实 CFO 呢这个人呢就其实一直就比较保守和谨慎...今天早上的新闻嘛，就马上他就开始又提要反对。" (Todd) [00:06:50]

Tim Cook — Floated as a possible Nvidia CEO; feared by Taiwan's suppliers. "库克很很厉害的地方就是，他把利润全部或者几乎全部集中在苹果自己手里边。" (Todd) [00:52:46]

Elon Musk — Publicly welcomed Huang to his company; earlier proposed a six-month AI pause. "马斯克说我很欢迎就是黄仁勋来我们公司工作。" (Todd) [00:45:53]

Lisa Su (苏姿丰) and Lip-Bu Tan (陈立武) — Chinese-American CEOs of AMD and Intel, used to illustrate the diaspora leadership pattern. (Manchi) [00:36:02]

Sam Altman — Early OpenAI figure photographed with Huang and Musk [00:10:50].

Robert Oppenheimer — Analogy for AI's unstoppable chain reaction. (Todd) [00:39:36]

Liang Wenfeng (梁文锋) — DeepSeek founder, whose leaked investor transcript mentioned adopting TileLang in 2025 (Manchi) [00:24:20].

Zhang/Huang Le (黄乐) — Woman linked to Huang and Megaspeed at the center of the scandal (Manchi) [00:42:57].

J.D. Vance (万斯) — Cited by Todd as distancing from Trump over the midterm loss: "共和党里面的老哥万斯就直接说啊这个我们失败啊，主要是因为战争的问题。" [01:08:11]

Donald Trump (川普) — Alleged ultimate target; reportedly suggested Cook as a candidate (Todd) [00:52:46].

Eric Schmidt (referenced as former Google CEO) — Cited on founders versus managers: "Google 的前 CEO...他作为一个职业经理人...更多的工作都是维持这个战略稳定地运行。" (Todd) [00:03:00]

5. Operating Insights

Disclose Your Allocation Priority Stack to Expose Your Real Strategy

Yilun's inside look at how ByteDance and Tencent rank internal compute demand is a template for operators: your resource allocation order is your strategy. ByteDance: Seedance, Seed, other R&D, external cloud; Tencent: Hunyuan, WorkBuddy, Xiaowei, then cloud. "字节他们内部算力调用的优先级可能是 Seedance、Seed，然后内部其他的研发场景，最后是外部的云平台。" (Yilun) [01:02:32]

Fix the Internal Transfer Pricing Problem Before It Kills Cross-Sell

Alibaba's dilemma shows a classic incentive failure: "千问使用阿里云的算力是没有内部结算机制的，所以如果作为销售，...对销售团队来讲是没有任何动力去把更多的卡给千问用的。" (Yilun) [01:03:44] Without internal settlement, sales teams will favor external customers and starve your own strategic product.

Convert Open-Source Distribution Into Revenue With Token Revenue-Share

Zhipu's structure lets third parties host compute while the lab retains pricing control and avoids capex: "你通过官方渠道可以卖的价格是更高的...智谱它没有去持有这一部分算力，它就不承担这个找算力的消耗和成本。" with a 3:7 split favoring the compute provider (Yilun) [01:04:56, 01:05:25]. This is asset-light distribution that also defends against open-source leakage.

Pre-Position for a Cyclical Whipsaw: Buy Capacity Before You "Need" It

Zhipu and DeepSeek's post-mortem: they under-bought compute in 2025 when supply was loose and later wished they'd been more aggressive. Yilun cites constrained funding and IPO timing as reasons [01:06:15, 01:06:46]. Lesson: scarce inputs are cheapest when the narrative says they're abundant.

6. Overlooked Insights

The Leasing Companies Are the Real Transmission Mechanism of China's Compute Market

Mentioned almost in passing, the financial-leasing firms are a systemic channel. Yilun said last year they were nervous that returned compute would make their loans uncollectible; this year they're expanding: "去年我们聊的时候他们很担忧...因为害怕大厂把算力退租掉了，他们放出去的贷款就收不回来了嘛。但今年我看到他们就状态完全不一样，整个人如沐春风，要大干一场。" [01:00:58] These lenders are now pivoting to financing domestic-chip deployments and CPUs for agent scheduling, which signals where the next wave of capital flows. This credit layer is the Chinese analog of Nvidia's own role as lender-of-last-resort, and its risk appetite is a leading indicator for domestic chip adoption and for a potential credit crunch if the cycle flips.

The CPU Demand Tail from Agents Is Quietly Emerging

In both Manchi's closing and Yilun's leasing remarks, CPUs for agent orchestration appear as a new demand vector: "在 agent 变得更主流之后，大家意识到 CPU 的需求在上升" (Manchi) [01:14:58], and Yilun suggests leasing firms "去研究 CPU 的需求。你可以去讲一下 agent 调度的故事" [01:00:58]. While the market obsesses over GPU allocation, the agent era shifts workload toward general-purpose compute, an under-covered theme that could benefit a different set of suppliers than the GPU consensus names.