2026.11.17 (美东) 盘后

热点

热点追踪

  • 英伟达宣布黄仁勋离职,克雷斯任临时 CEO。马斯克加入 17 万人联署呼吁回归。黄保留 1,320 亿美元股份,减持透明度受关注。
    最新 11/16 15:40·起始于 11/12 15:23
  • 台湾加权指数收盘大跌,创下史上第二大单日点数跌幅。供应链担忧引发外资抛售,但有分析师指本地制造优势仍稳固。
    最新 11/16 16:16·起始于 11/16 16:16
  • 黄仁勋考虑复职获特朗普支持,马斯克呼吁回归,微软市值超越英伟达。财长要求大幅重组,财报前纳指震荡,市场情绪谨慎。
    最新 11/16 17:03·起始于 11/16 16:13
  • 英伟达三季报营收预计 1090 亿美元,分析师担忧前景;段永平称黄仁勋回归"不是零概率"。路透指需证明实力超越黄仁勋个人,市场关注临时 CEO 科莱特。
    最新 11/16 17:11·起始于 11/16 09:32

海豚研究

Dolphin Research
Dolphin Research

Good morning. It is 2026-11-16 SGT, and Dolphin Research walks you through global markets at a glance.

Trading focus: Nvidia leadership saga escalated over the weekend: FT says Jensen Huang is discussing a return, while Treasury Sec. Besant urges hiring an external CEO and demands a response within 10 days on China sales. NVDA fell 2.0% Fri, down ~13% for the week.

PHLX Semis -3.4% Fri. AMD -8.8%, Broadcom -4.8%, Cisco -6.2%, with the prior four-day winners broadly giving back gains. Nebius -9.9%, while Cerebras defied the tape, +11.2%.

Macro tone: Oct U.S. PPI came in at 6.5% YoY, a touch below estimates, yet several Fed officials warned of broader inflation pressures. The 10Y UST rose ~28bps on the week to ~5.14%, while the 2Y was up ~33bps.

Majors moves: S&P 500 -0.40%, Nasdaq -0.91% Fri, but still up 2.51% and 2.27% for the week. Meta led mega-cap tech Fri, +2.9%, while Dell fell 7.2%.

Near-term watch:

Pre-mkt today: BIDU, XPeng, BEKE earnings; China Oct macro data.

After Tue close: NVDA earnings (first since Jensen stepped down); U.S. Oct retail sales.

Wed–Thu: Tencent earnings, Anthropic pricing. Alibaba, Lenovo earnings; Anthropic lists.

$英伟达(NVDA.US) , $AMD(AMD.US) , $博通(AVGO.US) , $思科(CSCO.US) , $Arm(ARM.US) , $美光科技(MU.US) , $闪迪(SNDK.US) , $Meta(META.US) , $微软(MSFT.US) , $谷歌-A(GOOGL.US) , $应用材料(AMAT.US) , $台积电(TSM.US) , $Nebius(NBIS.US) , $Cerebras(CBRS.US) , $腾讯控股(00700.HK) , $标普 500 指数(.SPX.US) , $纳斯达克 100 指数(.NDX.US) , $费城交易所 半导体 ETF - iShares(SOXX.US) , $恒生科技指数(STECH.HK)

Performance windows: U.S. stocks and global assets reflect 11/13 (Fri) close vs. prior day. Asia reflects today open vs. prior close.

1) Ahead of NVDA earnings, 'who runs it' is now harder to price than 'how much it sells'

Arm -4.6% / Nvidia -2.0%

The weekend leadership plot pulled both ways: FT reports Jensen Huang has discussed a return with close allies, who formed a war room to lobby the board and seek White House support. A petition for his reinstatement surpassed 130k signatures by noon on the 15th.

On the other side, per Bloomberg, Sec. Besant asked Nvidia to hire an external CEO and respond within 10 days whether seemingly unlawful China sales boosted revenue. Media say the board has sounded out Arm CEO Haas, while Nvidia says it has not formally approached any candidates.

Financing also loosening: reports say interim CEO Kress cut the data center guarantee to OpenAI from $250bn to about $105bn.

Bloomberg notes Nvidia's $500bn AI financing platform has not closed a single deal since its Aug launch, with a 25% backstop cap far below Broadcom's ~72.5% in transactions tied to Anthropic. Semafor also reports the U.S. DOJ is probing Nvidia on antitrust grounds, separate from inquiries into Megaspeed and Super Micro.

Institutions argue that if Jensen truly departs, it would be a negative for the semi ecosystem. He has been pivotal in mobilizing financing and buildout, striking supplier deals, and advancing leading-edge nodes together with partners.

Flows: per offshore trading desks, NVDA fell 13% on the week and is down 33% since early Oct, wiping about $500bn in market cap. AMD and Broadcom together added over $500bn. Long-only money kept selling NVDA, rotating into hyperscalers and other mega-cap tech.

Dolphin take: Tue's print will be the first statement after the company 'lost its soul'. The quarter is done and orders were maxed, so the numbers likely stand; what matters is whether the FY28 roughly 70% growth guide gets trimmed, and whether 2027 demand is borrowing from 2028.

If the board follows the Sec.'s push for an external operator, the CEO-to-CEO value of bespoke HBM and CoWoS capacity lock-ins is hard to replicate. Even if near-term results hold under existing contracts, NVDA's mid-to-long-term multiple should compress. (Original by Dolphin)

2) AI complex 'catch-down': prior four-day winners broadly reversed Fri, with Cisco orders adding pressure

Nebius -9.9% / AMD -8.8% / Cisco -6.2% / Credo -5.9% / Celestica -5.4% / Broadcom -4.8% / Marvell -4.7% / Arista -4.5%

Per offshore desks, Fri looked like a broader semi and AI theme catching down to NVDA: the SOX had rallied ~840bps over the prior four days, which in hindsight stretched the rubber band. Macro also tightened: VIX back to 20, 10Y UST at 5.15%, USD broke higher.

Positioning was not heavy, yet de-risking continued: hedge funds net sold U.S. equities for a second week, and the U.S. tech long-short ratio fell to 1.5 (vs. 2.14 peak in Jun), the lowest since records began in 2016. 12M momentum rebalancing is set to net sell semis and buy software.

Cisco was the direct spark for the networking complex: revenue and guide beat, but product orders slowed from 35% to 21%, below just-above-30% expectations. Shares spiked ~10% after-hours before fading.

Institutions estimate hyperscaler orders this quarter at $2.2bn–$2.4bn, down 40%+ vs. ~$4bn last quarter. FY27 hyperscaler AI revenue outlook was raised from ~$7.5bn to ~$9.0bn, though management admitted part reflects customer requests to bring forward backlog deliveries.

Dolphin take: Fri's semi selloff looks more like a catch-down than new negatives. AMD and Broadcom together added ~$500bn over the prior four days, with the leadership-change tailwind not yet realized before rate hikes and de-leveraging took it back. More concerning are Cisco's slowing orders and pulling forward demand, shifting the network chain from price hikes and allocation to order durability. NVDA's guide Tue will decide whether this drawdown is a shakeout or a turn.

3) Memory: DRAM collusion suit proceeds, NAND spot slides further; memory stocks keep lagging

SanDisk -5.6% / Western Digital -5.0% / Seagate -3.1% / Micron -2.5%

A judge rejected Samsung, SK Hynix, and Micron's motion to dismiss a class action alleging DRAM collusion, accusing the trio of cutting legacy DRAM supply and pivoting to HBM.

Supply: Korean media say SK Hynix plans another HBM4 capacity expansion, adding at least 20k wpm, citing strong demand from AMD and Broadcom. SK Hynix fell 3.0% Fri in Seoul.

Pricing divergence: BofA says the DRAM upcycle thesis stands, while NAND spot has clearly softened and could be the first to reprice this cycle. 512Gb TLC wafer spot is down 33% from the 8/17 peak.

China controls limited impact: Nomura estimates that even if China AI chip demand disappears entirely, global HBM demand in 2028 would be down just 1.7%–2.5%, and DRAM down 0.36%–0.54%.

Dolphin take: The Jensen saga is tapping the brakes on the AI chain, hastening memory's feast into its second half. LT contracts lock volume and floors but cap spot upside, shifting focus from how high to how long. With prices stabilizing, multiples have limited room to rerate up or down, implying range trading. With NAND spot loosening first and litigation adding noise, near-term beta gives way to LT contracts and buybacks. (Original by Dolphin)

4) New cloud takes another hit, Cerebras rallies: cards are available, data centers are not

Cerebras +11.2% / IREN -6.7% / CoreWeave -6.1%

New cloud is viewed as a consensus loser from NVDA's leadership change: Deutsche Bank says a new management team may reduce financing for new cloud and AI labs, raising GPU inventory risk from 2028.

Reports say Applied Digital is asking tenants like Oracle and Meta for multi-billion-dollar guarantees to secure power, with financing guarantees spreading from NVDA to more blue-chip clients.

Cerebras posted a far-above-consensus Q4 guide ($310mn–$330mn). On the call, it said the delayed OpenAI commitment was under 5% (~37MW), with nothing canceled, and expects to catch up by 2H27.

Management's words: everyone's data centers are delayed, including OpenAI and hyperscalers. The CEO also said Anthropic ported models to AMD in about three days, undercutting incumbents' software moats.

Institutions add: Microsoft's compute deployments slipped from Q3 to Q1 next year due to third-party delays. Asked if Anthropic can self-build, the CoreWeave CEO said running data centers is far harder than it looks.

Dolphin take: Cerebras' chips are not the issue; the bottleneck is data centers. Both this quarter and next quarter guides look solid, with core profitability turning positive, and '3Q was the GPM trough' already validated with a Q4 guide of 47%–49%. But cloud revenue is dragged by site delays with flat QoQ next quarter. For new cloud, the pricing uplift is in the price; the race shifts to volume cadence and ROI, and only those that energize on time deserve the multiple. (Original by Dolphin)(Original by Dolphin)

5) Anthropic IPO this week: investors question not the IPO valuation, but where future growth comes from

Per Jianshi Breakfast, CEO Dario Amodei kept a low profile on the roadshow, emphasizing under-promise and over-deliver. He said timing and market conditions were not a key concern, and acknowledged compute capex as the most fundamental business challenge.

Institutional feedback suggests few dispute the valuation so far, with questions focused on revenue uplift, open-source and pricing competition, and the feasibility of self-build compute. Estimates point to quarterly QoQ revenue add rising from ~$2.7bn in Q4-26 to ~$8.6bn in Q4-27.

The SpaceX compute contract could drive compute GPM negative initially, but margins are expected to stabilize after Q2 next year. Street expects solid aftermarket performance, especially if pricing lands near $80 or below.

Sell-side Q4 revenue estimates cluster around ~$19.5bn, and most investors see IPO valuation as undemanding with room to run. Excluding one-offs like distillation, price cuts, and seasonality, Q4 QoQ revenue growth could be in the mid-30% area and may mark the quarterly trough, with acceleration through Q1–Q3.

Price cuts did not hurt business: Opus 5.5 reduced customer task costs by 40%, which management said was margin-neutral, with run-rates recovering within two weeks. Only 3% of listed companies are customers today.

Shift to Direct: ~45% of revenue is distributed via hyperscalers, while owned compute yields roughly $8bn–$11bn per GW annually, vs. ~$15bn–$16bn for leased. Direct compute starts in Q2-27, with only ~1.5GW of next year's ~12GW added compute coming from Direct.

The cash flow toll: Street models capex rising from an annualized sub-$2bn in 2026 to over $60bn by 2030, with FCF still negative by 2030. SpaceX compute costs about $50bn per GW, roughly 3x that of new cloud partners.

Sell-side sees OPM falling from about +15% in Q3 to -30% in Q4 and -45% in Q1, returning to breakeven by Q3-27.

The biggest question is what drives growth after Coding: heavy Claude Code users spend about $1,500 per month on average, +50% QoQ in Q3, but Dario says coding is becoming more mature.

Implied by market expectations, net new ARR would need to accelerate to ~$10bn per month in 2027 and $13bn–$15bn in 2028. Management has provided little on the growth path.

Dolphin take: Anthropic's pricing will stress-test the AI financing chain this week. With NVDA pulling back guarantees and new cloud seeking blue-chip backstops, can the market award a high multiple to a lab with hypergrowth but compute-heavy economics. Beyond day-one pop, watch for a clearer revenue driver breakdown for LT growth and a delivery cadence for self-build compute to optimize costs.

Dolphin will publish a deep-dive Part II on Anthropic (valuation) soon; Part I covers core fundamentals (Original by Dolphin).

6) Dolphin Hot Topic | Jensen out: why are China assets paying the price if the soul lost is Nvidia's

Zhipu -10.0% / MiniMax -8.2% / Tencent -3.1% (HK, 11/13 close)

Dolphin believes the campaign against Jensen may be just the starter pistol, with the real target being the Remote Access Security Act (RASA) shelved for nearly a year. It would classify Chinese clients' remote rentals of overseas compute as exports under Commerce oversight.

Reports that the Commerce Dept. is moving to plug this loophole over the weekend align with that direction. If remote compute is cut, the impact goes beyond Megaspeed's sub-$3bn annualized buys to as much as $40bn, or roughly 10% of NVDA revenue.

The NYT also says Lutnick previously pushed to narrow related controls, involving a Malaysia data center that may serve Tencent.

Dolphin take: Jensen stepping down plus RASA progress implies near-term outcomes not of China only using domestic cards, but of China having no cards to use. Stock runs continue; incremental demand is choked; grey channels cost more. Chinese cloud and IDC capacity won't come online, and overseas workarounds close. Domestic cards lack near-term capacity even for inference; NVDA's losses are longer-term, as open-source pathways to counter big customers' in-house chips get cut off, benefiting Ascend, AMD, and hyperscalers with in-house silicon, with open-source possibly splitting into U.S. and China hardware stacks. (Original by Dolphin)

1) PPI slightly below, but hikes are still priced

PPI: U.S. Oct PPI rose 6.5% YoY, the highest since Nov 2022 and a touch below the 6.6% consensus. MoM at 0.3% vs. 0.4% expected; core PPI held at 5.0% YoY.

Fed hawks: Chicago Fed's Goolsbee called for rebuilding anti-inflation credibility, and Boston Fed's Collins said further tightening could be needed soon. Markets still price a 25bps hike in Dec as base case, with about a 15% chance of 50bps.

Financial Stability Report: the Fed said valuation pressures remain elevated, with equity multiples partly reflecting expectations for superintelligence. AI asset prices and investment booms could amplify volatility.

USTs and USD: the 10Y is around 5.14%, up ~28bps on the week. The 2Y is ~4.84%, up ~33bps, the largest weekly move in two years; DXY set 19-month highs for three straight days, with JPY briefly through 160.

This week: Tue night U.S. Oct retail sales, Wed FOMC minutes. If retail is firm and minutes hawkish, 10Y above 5% will keep pressuring high-multiple tech.

2) High-vol assets (11/13)

PHLX Semis -3.4%: AI complex catch-down; see Trading focus 2.

Bitcoin -4.3%, Ether -4.4%: BTC broke below $68k intraday, a three-month low.

Brent -2.5% WoW: down nearly 19% over two weeks; IEA says the global oil market flipped to surplus in Q4.

Silver -1.9%, Gold -4.55% WoW: COMEX gold saw its biggest weekly drop in six months.

3) Compute and memory prices (11/16)

GPU rental: Silicon Data new-cloud index on 11/15 shows B200 at $6.87 per GPU-hour, +12% vs. 10/17 and +0.9% WoW. B300 at $7.67, H100 at $3.10, H200 at $3.50, both largely flat.

Token costs: Silicon Data LLM token spend index on 11/14 was $0.79 per mn tokens, -4.8% over seven days and -17% vs. 10/16. Closed-source at $2.02 (+1.8%), open-source at $0.49.

DRAM: DDR4 8Gb spot at $44.54, -5.3% vs. the 10/27 peak $47.04, -4.9% WoW, yet still +8% since late Jul.

NAND: 512Gb TLC wafer spot at $14.27, -33% vs. the 8/17 peak $21.21 and -24% over four weeks. BofA calls NAND the first category in this cycle that needs repricing.

4) Asia open and today's setup (11/16)

Asia open: Nikkei 225 +0.3% at the open; KOSPI +0.3% (KOSDAQ +0.9%). SK Hynix flat; Samsung Elec +0.2%; Tokyo Electron +0.3%; SoftBank Group +1.5%.

Semi catch-down spillover: with U.S. semis -3.4% and memory/networking leading declines, SK Hynix, Samsung, Japan/Korea equipment, and Taiwan networking/server chains likely open under pressure. China A-share optical module names dropped over 8% Fri; watch for stabilization today.

Controls and China assets: Treasury asked NVDA to respond within 10 days on China sales, and Commerce is plugging the overseas rental loophole. HK cloud, IDC, and LLM names likely stay pressured, while domestic compute beneficiaries have a floor; China Oct data (retail and TSF) are a swing factor.

Rates: with the 10Y UST at ~5.14% and USD stronger, high-multiple growth valuation remains under pressure. Flows likely favor cash-generating mega platforms.

Meta META|$1.64tn|+2.9%

Offshore desks say long-only funds rotated from NVDA into hyperscalers. Soros Fund Management added nearly 130k Meta shares in Q3, lifting its stake almost 6x. Microsoft MSFT|$3.89tn|+0.2%

Outlook, Teams, and Microsoft 365 suffered a widespread outage Fri, peaking at over 130k reports. Microsoft cited network expansion upgrades, not a cyberattack; lawmakers have requested details.

Ignite conference in San Francisco this week.

Alphabet GOOGL|$3.64tn|-1.0%

Rumors from some trackers point to search growth at ~9%, below the ~10% expected.

Institutions roughly estimate Anthropic at ~10% of Google Cloud revenue, contributing low-teens percentage points to growth (ex-TPU). Anthropic's shift to self-build compute would become a growth headwind for GCP.

Amazon AMZN|$2.88tn|-0.3%

ULA's Vulcan rocket failed 10 minutes after liftoff, costing Amazon 40 Leo satellites, the first mission failure for the Boeing–Lockheed JV.

Institutions roughly estimate Anthropic at ~10% of AWS revenue, contributing nearly 25% of its Q3 growth. A pivot to self-build compute would similarly weigh on AWS growth.

Applied Materials AMAT|$294.0bn|-2.7%

For calendar 2026, WFE growth guide moved to 40%+, with a first-ever 2027 WFE growth guide at 30%+. Institutions translate that to roughly $200bn–$208bn, below buy-side at $210bn–$230bn.

Management says 2027 cloud capex will grow over 50% to well above $1tn. Its largest customer continues to provide an 8-quarter rolling forecast, and none have asked to delay deliveries.

Next quarter has 14 weeks; on a 13-week basis, Dolphin estimates revenue roughly -1% QoQ, with semi systems up only ~1%. GPM at 50.4%, up just 10bps QoQ. (First Take)

Dolphin take: A near-term pause does not derail the upcycle; management sounds more confident. Two-year growth guides suggest downstream pull remains strong, but ex the extra week, systems are flat QoQ, constrained by cleanroom and supply chain. Logic is nearly 60% of systems; if logic cools, memory must step up, which hinges on memory capex. The 54% 2030 GPM goal looks tough at the current pace of QoQ expansion. (First Take)(Earnings review)

TSMC TSM|$2.07tn|-1.7%

Deutsche Bank says Nvidia accounts for roughly 25% of TSMC revenue, and freed-up CoWoS/3nm/2nm capacity can shift to AMD, Broadcom, and SpaceX.

DB also notes actual data center build capacity of ~30GW per year is below chip demand expectations north of 40GW, limiting downside risks.

The company plans a $500mn fund to invest in chip startups and upstream players, saying it wants to support the next Nvidia.

SMIC 0981.HK|

Co-CEO said full-year revenue should exceed $12bn, with AI-adjacent chips (BCD, specialty memory, controllers, etc.) driving faster-than-expected demand. Q3 ASP per wafer rose 1.8% QoQ.

On the call: some tool deliveries slip into next year, and price hikes are taking hold. Management expects AI-related supply-demand to remain very tight into 2027, with non-handset pricing rising next year, not falling. (Trans)

Intel INTC|$317.1bn|-2.5%

Soros Fund Management added 2.73mn shares in Q3, lifting the stake by over 27x. It cut NVDA by 19% in the same period.

Dell DELL|$219.1bn|-7.2%

Evercore says Dell's FY27 AI server target (~$74bn) is heavily reliant on NVDA and makes it one of the most exposed to leadership risk. Bernstein, however, put Dell on its preferred list as directly oversold.

Dell reports on 11/24.

Tesla TSLA|$1.54tn|-1.8% / SpaceX SPCX|$1.89tn|-0.6%

Bloomberg says merger talks are accelerating, with a possible Dec announcement. Tesla opened Cybercab orders to selected individuals at a $29,990 starting price, with first deliveries in Dec; NHTSA has launched a probe.

Starship's 15th test used the launch tower chopsticks to catch the ship for the first time. But the V3 booster had an anomaly on its maiden flight and exploded on splashdown; Bernstein says full reuse is key to the SpaceX valuation.

Baidu BIDU|+0.7%

Q3 results due today, see earnings preview at the end. Baidu recently launched a 1,024-card supernode, strengthening the domestic GPU substitution narrative.

Nvidia NVDA|After 11/17

Watch: first earnings since Jensen stepped down, hosted by interim CEO Kress. Consensus revenue around $108.7bn (+91% YoY); market focus on guide, hyperscaler tone, and CEO search progress.

Dolphin watchpoints: whether the FY28 ~70% growth guide is revised, the size of cloud backlog (over 2 trillion last quarter), the GPM drag from memory price increases, and the accounting for China and remote compute revenue.

Baidu BIDU|11/16

Watch: UBS says internal DAUs for agents rose from 5k to 37k, with token usage rising but not yet monetized. Kunlun Chip spin-off progress.

Dolphin watchpoints: AI Cloud (RMB 7.3bn in Q2, +50% YoY), pace of legacy ad declines (-19% in Q2), capex and buybacks.

Tencent 0700.HK|11/18

Watch: with the overseas rental loophole closing and NYT naming a possible Malaysia DC, the market will watch how compute access affects capex and AI timelines.

Dolphin watchpoints: capex (RMB 52.8bn in Q2, 26% of revenue) and FCF (RMB -13.8bn in Q2), plus gaming and ads growth (+22% in Q2).

Alibaba BABA|11/19

Watch: Dolphin sees remote compute cuts as a major headwind for Alibaba's cloud growth outlook. Once fully domestic, Huawei chip supply remains uncertain; last week Alibaba confirmed T-Head's spin-off filing, with institutions estimating 700k unit shipments in 2027. (Original by Dolphin)

Dolphin watchpoints: external cloud revenue growth (45% last quarter), AI product revenue (RMB 12.376bn last quarter), capex and FCF (RMB -44.67bn last quarter).

<End here>

Recent Daily Recap

2026-11-13: Dolphin Daily | Jensen out: AMD and Broadcom feast, new cloud suffers

2026-11-12: Dolphin Daily | Retail lifts new cloud, fresh scandals weigh on Nvidia

Risk disclosure and disclaimer: Dolphin Research Disclaimer and General Disclosure

Dolphin Research
Dolphin Research

11 月 12 日,黄仁勋辞去英伟达总裁、CEO 和董事职务,CFO Colette Kress 兼任临时 CEO。$英伟达(NVDA.US)

当天英伟达低开高走,最终收跌 4.3%,但阿里跌 7.1%、万国数据跌 11.4%、Nebius 跌 16.8%,比正主跌得还狠。

除了换帅的风格差异(激进的老黄 vs 保守的 CFO),可能会导致 AI 投资扩张放缓。

更严重的预期演绎是,围剿老黄只是发令枪,可能的真正目的是在为一部在参议院躺了快一年的法案《远程访问安全法案》(Remote Access Security Act,下称 RASA)重新推进而扫清障碍。

这里面看空资金算的不仅仅是英伟达会丢失多少中国收入,而是整个开源生态的蛋糕缩小。

一、RASA 是什么?

先来了解下 RASA,具体限制的是哪里。

现行出口管制管的是 “芯片能不能运进中国”,是实体设备的采购限制。

但漏洞在于,芯片可以留在海外机房,而中国客户通过云服务、裸机服务租赁 “远程使用”。因此 RASA 要做的,是把这种远程访问也算作出口,交给商务部去管,视作同等限制。

截至目前 RASA 法案仍处于被搁置状态,参、众两院的版本存在差异,参议院版本的限制更狠:

虽然 RASA 不是因为这次老黄的丑闻才有的,但明显这次给了它新的政治推力。因此,很难不怀疑项庄舞剑意在沛公:先有 RASA 的推动意愿,再有老黄丑闻曝光。

而在老黄丑闻曝光的同时,更多的政治力量也在试探:

1)不止于 RASA,限制细节已经推进。据 The Information,白宫 10 月下旬要求在《AI Overwatch Act》里加一条,责成商务部堵住远程访问漏洞;这部法案已写进参议院版国防授权法,预计明年 1 月前过关。

海豚君捋一下,也就是说:

a.《RASA》赋予了商务部监管海外远程云算力的法定权限(授权性立法);

b.《AI Overwatch Act》则更进一步要求国会直接责成商务部制定具体的限制规则(规定监管必要性和监管细节)并执行起来。

2)英伟达让步已有苗头。 11 月初,英伟达被媒体爆出,要求云服务商客户中,非中国客户至少占 51%。虽然,这是英伟达自己制定的供应商条款,并非来源于直接的法律约束,但说明公司已经在为最坏情形做准备。

在同一时期,Oracle 也被媒体报道,取消了来自腾讯的 70 亿 AI 芯片租赁订单。

3)老黄作为最大的反对者离场。商务部官员曾称,英伟达在 RASA 上的劝诫游说上 “尤其用力”,而国会在限制对华芯片上的共识正在增强。因此老黄的卸任,是推进远程算力监管路径上重要障碍的消除。

二、中国 IDC、中国云大厂、英伟达都会很受伤?

1. 短期业绩的小账:对英伟达收入影响不小

最新财报显示,中国(含香港)占英伟达收入的比重已从高点 28% 降到 8%(只剩合法收入)。但这次发难得关键是,中国客户使用的东南亚,甚至台湾、日本等地的 “远程” AI 算力:

(1)英伟达出票给新加坡的收入曾从 23 年 10% 的份额,在 25 年一度跳到 20%–22%,而同期中国(含香港地区)收入份额下降到 5%。

FY3Q26(2025 年二季度)口径调整后(从 “按客户开票所在地” 改为 “按终端客户所在地和发货所在地”),新加坡小到并入其他地区。

当时英伟达解释是,客户只是用新加坡集中开票,产品几乎都运往别处(99% 来自美国客户的订单)。

(2)最新的 FY2Q27(2026 年二季度)财报披露,中国台湾地区的收入 269.85 亿美元,占比 28%,而上一季只有 14.7%,又跳升了 14pct。

公司批注为更多的整机交易是通过台湾 ODM(鸿海、广大、纬创等)结算,而不是云厂直接向英伟达采购,86% 的台湾收入来自欧美地区的终端客户。

海豚君来粗算一下:在 2Q26FY 改口径之前,新加坡和中国(含香港)两地区合计份额基本在 20-30% 波动。

而中国限售之后,虽然丢了一部分收入,但中国地区从 20% 下降到 6%,这里面 14% 份额的大部分,大约 10% 从新加坡渠道收回来了。

除此之外,就是改口径之后,台湾地区份额的 “间歇性” 跳升了,3Q26FY 跳高了 6pp,随后恢复,然后 2Q27FY 又跳高了 13pp。

这里纽约时报认为,Megaspeed 正是通过英伟达服务器合作伙伴渠道,才获得采购近 50 亿美元芯片(从 FY25 Q2 至 FY26 Q4 一年半时间采购了 48 亿美元),早年比如直接在新加坡开票,后来在台湾的 ODM。

据 New Street 预测,如果远程云算力需求被切断,对英伟达的收入影响就不只是 Megaspeed 获取的年化不到 30 亿美元(这只占 0.1% 的收入),而是高至 400 亿、占比 10% 的收入需求。

市场预期 2027 财年 10-30% 的渠道收入损失,以中位数 20% 计,假设新需求能够立即补上其中的一半,这其实和海豚君上面粗算的 10 个点中国份额预计得差不多。

这比 RASA 通过之前,追查并切断东南亚租卡等商业行为的假设下,预计 60 多亿收入减少,影响要明显严重得多(New Street 预估芯片硬件转运 + 离岸租用占比 4%,即 160 亿收入中,60% 被云厂接走,剩下 40% 缺口,即 1.6%)。

海豚君认为,最差的第三档假设下:在目前 CoWoS 仍然紧张、算力需求爆棚的时候,缺口似乎能被其他需求填补,但这势必会影响未来需求的趋势,也就是英伟达的增速指引:

英伟达 Q3 财报发布在即,明年 70% 的增速是否会调整?就算 27 年能满足,但是否是 28 年之后的需求前置带来?

目前 SemiAnalysis 追踪的 GPU 服务器租赁价格上,B200 自 10 月底开始转跌,不过 B300 的租金依旧坚挺,尤其是在受限的中国地区,半年时间涨了 2.5 倍。

三、开源生态的大账:“盗版养用户” 的逻辑破灭

用海外远程云平台来提供高端算力的服务,是英伟达所讲述的开源生态逻辑关键核心。

这一年,中国大模型所领导下的开源生态繁荣,老黄多次力挺站台、并且身体力行的为开源生态添砖加瓦,尤其是在 2026 年的相关动作相当紧凑:

a. 7 月发 x 帖公开支持开源;

b. 8 月被爆正在开发万亿参数的开源模型 Nemotron 4;

c. 9 月宣布收购 Hugging Face;

d. 最近的 10 月在新加坡某论坛上,赞扬 DeepSeek、Kimi、GLM 等中国开源模型表现出色。

推开源而非闭源,是长期有利于英伟达的。开源 “把稀缺品做成白菜价”,模型越便宜、客户越多,GPU 需求越大。

而且,相比于闭源龙头主导生态,开源模型能让中小企业和主权项目自己训练、自己部署模型,把英伟达客户结构打散,这样英伟达能保持相对话语权。

因此,开源生态会大幅缓解英伟达的大客户集中度情况,从而减少客户自研芯片带来的长期风险。

其实老黄的战略思路,本质上有点类似微软当年的 “盗版养用户” 逻辑——中国用户反正要用盗版,不如让他们用 Windows,习惯养成了再慢慢收钱。因此护城河不是那个正版的 Windows 光盘,而是习惯和生态。

而英伟达的对应物是 GPU 和 CUDA:主流和明面上,高端 GPU 是被禁了,但通过 “远程云服务” 又给中国的客户们开了一道 “不可明说” 的口子,让前沿训练还是留在英伟达的芯片体系内。而且通过这道口子,也能看清楚中国开源模型的研发进度。

老黄这几年反复讲的一句话就是:与其把中国开发者逼进华为生态,不如让他们继续用美国技术栈。海外机房的远程算力,客观上就是在 “养开发者习惯”。

不过进一步对比,其实还是有差异,这也导致了中国在 AI 芯片上推进国产化的意愿会更强、动作也会更快:

(1)微软的盗版 Windows,大部分中国客户没有付费,因此依赖性更强。但英伟达的远程云算力,中国客户仍然要付一笔昂贵的费用,有成本动机去做国产化的开发。

(2)微软当时采取该策略,没有合规风险,只有短期商业收款能力和长期赚钱潜力的业务选择;而英伟达直接面对的是美国国安压力,承担政治和法律风险;

(3)微软的 Windows 在文件格式和软件兼容上锁定较强,但 AI 芯片端,开发者写的是 PyTorch,用 PyTorch 来调用 Cuda,因此换硬件的门槛在下降,中国客户无论是在监管还是行业内部,都有意在持续推动国产替换。

因此 “养用户” 这套打法,在 AI 时代的半衰期远比 Windows 时代短。而美国政府推动 RASA,对这套打法更是精准一击。

四、短期利好闭源,但开源可能裂成两套硬件栈

如果连 “远程使用” 都不行了,那么头部中国开源模型只能更快地优先适配国产芯片。虽然这已经在发生,但不得不承认,中国 AI 芯片无论是在性能还是产能上都不足以满足前沿模型的训推:

(1)智谱 GLM-5 从第一天起适配 7 款国产加速器;

(2)DeepSeek 为 V4 Pro 开源了昇腾上的推理内核;

(3)美团称 LongCat 2.0 完全在昇腾 910B 上预训练。

分开看,性能就算开始追赶、缩小差距,但最紧急的 “量” 上,国产卡短期完全接不住:

东南亚 79 万张 H100 按 2.8 倍的 B300/H100 折算后,等同于 28 万张的 B300 卡。如果全部断供,那么同等算力下,回流到华为昇腾 910C 的需求约 50 万–62 万张,接近昇腾 910C 今年约 60 万颗的规划产量;

但更严重的瓶颈是 HBM 上,中国存储厂的 HBM 今年大约只够做 25 万–30 万张成品卡。也就是说这里面的 30 万张成品卡所需 HBM 缺口,大概率靠之前锁定或其他渠道获得的海外 HBM 来供应。

小结:中国 CSP+ 模型短期可能 “很受伤”;英伟达远期 “地位走弱”

老黄卸任 + RASA 推进,短期效果不是 “中国只能用国产卡”,而是 “中国用不了卡”——存量照跑、增量被掐、灰色渠道更贵。

海外通道被政策锁住,需求被挤回国内,等同于国产算力被政策圈定了确定性需求。但对于中国:

a. 对中国的云厂商(如阿里而言),以及上游的 IDC 中心来说——产能建不出来,云业务收入,海外路径没了,客户训练服务压力走大;国内的卡甚至连推理需求都满足不了。

这对阿里影响尤其大,全部转向国内采购的时候,不说产能,华为到底有多大的意愿给阿里提供多少芯片都不好说。

b. 模型厂来说,叙事也跟着转向——中国前沿模型厂不能再把 CUDA 算子优化当卖点,甚至要进一步否认自己有英伟达卡,而是改口讲国产栈原生,其中还得承担性能差带来的额外成本。

最终结局就是开源世界被分成中美两套硬件栈:一套围绕 CUDA,一套围绕昇腾等中国算力平台。

而对英伟达而言:

a. RASA 真正的损失在长期: 英伟达靠开源对冲大客户自研,而开源里最活跃的一支力量在中国。网线一拔,这条反馈回路断掉,受益的是昇腾,也是 AMD 以及那些自研芯片的大客户(Google、Amazon、Meta 等)。

b.换帅进一步放大了风险。 开源是典型的创始人式长期押注,不赚钱、难量化,只有老黄这种拥有绝对话语权和长期战略眼光的统帅才能多年推进,正如当初的 Cuda。

但 CFO 往往保守,以及未来从外部引进的职业经理人 CEO,又有多大动力去推进开源生态?首当其冲就是自家的开源模型 Nemotron,目前不仅技术落后而且还不赚钱,可能被第一刀砍下。

失去了黄仁勋的英伟达,好比失去了最重要的灵魂,风险和价值损失根本无法准确评估:不仅仅是海豚君在这篇文章中所写的开源生态下的长期增长愿景,还包括短中期视角下,产业链端 CEO to CEO 的特殊合作(比如核心关键的 HBM、CoWos 产能锁定等)。

<正文完>

本文的风险披露与声明:海豚君免责声明及一般披露