Daily Brief Nov 12, 2026 · How we checked →

CNBC: NVIDIA board weighed ousting Huang, announced nothing

Source: CNBC

Today's front-page cartoon: one single picture that brings together today's stories.

At a glance

1. CNBC reported that NVIDIA's board met on the evening of November 11 to decide whether to keep CEO Jensen Huang, three days after a board statement of confidence in Huang; by early November 12 NVIDIA had made no announcement.

2. Anthropic says in its IPO roadshow video that the software work of moving Claude onto AMD GPUs took three days, down from several months for dozens of engineers. No third party has measured the performance.

3. A reporter at The Information says the US Commerce Department could finalize rules on remote use of advanced chips as early as this month; the rule text hasn't been published.

Also today: #4 — venture investor Tomasz Tunguz estimates how gross profit splits across the layers of AI.

Today's main line

1. NVIDIA's board voiced confidence in Jensen Huang on November 8, then met three days later on whether to keep Huang, and no company document covers the meeting

CNBC reported on the evening of November 11 that NVIDIA's board was meeting to decide whether to replace CEO Jensen Huang (CNBC, 2026-11-11). NVIDIA holders and buyers can judge whether the board will replace its CEO only from media reports; the November 17 earnings report is the first scheduled occasion where the company itself will speak.

In the full story: how the outlets' accounts differ, which documents NVIDIA has issued, and where investors stand

2. Anthropic's IPO roadshow video says Claude was ported to AMD GPUs in three days and gives no performance figures; AMD is also buying Anthropic shares at the IPO

Concept illustration: in a wide frame under a dark banner lettered ANTHROPIC beside a white angular mark, two people sit typing at desktop computers on the left; two tall black server racks each carry a sign lettered AMD, the left one beside the desks with a dark screen, the right one with a speech bubble on its screen above a label reading CLAUDE, and between them a standing person faces right toward a red laptop on a low stand whose screen reads MYTHOS. Below the frame at the bottom left, part of a grey projector is visible. Printed on the image: 3 days Porting Claude to AMD GPUs, down from months for dozens of engineers: Anthropic's roadshow claim. — Anthropic says in an IPO roadshow video that Anthropic's model Mythos got Claude running on AMD GPUs with almost no human intervention. The video says only that Claude runs and gives no performance figures, and no third party has measured performance. According to Anthropic's amended IPO registration, AMD will buy Anthropic shares at the IPO price when Anthropic lists.

Our inference: if the three days holds, big buyers with frontier models could port to other chips for far less; ordinary developers are unaffected.

In the full story: what the video actually says, how Jensen Huang sees CUDA, and AMD's financial ties to Anthropic

3. A reporter at The Information says the US Commerce Department could finalize rules this month restricting Chinese customers' remote rental of advanced chips; the text isn't public

Concept illustration: a hand at lower right holds a clipboard topped by a black banner lettered THE INFORMATION in white, with a red symbol at its left end. On the clipboard, at left, a booth lettered COMMERCE DEPT stands behind a wooden crate lettered CHIPS, and a horizontal black-and-white striped barrier arm extends from a post beside it; in the middle, a man in a dark coat with a badge on a lanyard walks to the right carrying a second striped barrier arm on his shoulder; at right, a woman types on a laptop at a desk, and a red cable runs from the desk along the floor to the bottom of a server rack inside a small building, with chip icons on each of its shelves. — The Information reporter Leo Schwartz wrote on X that the US Commerce Department is consulting other agencies on rules restricting Chinese customers' remote rental of advanced chips. The rules' text isn't public, and the reporters' posts don't say which chips and countries the rules cover or when the rules take effect. Being finalized doesn't guarantee the rules will be issued.

In the full story: how the rules differ from a sales ban, whether the White House or Commerce is stalling, and why the Treasury got involved

4. Venture investor Tomasz Tunguz estimates that 40% of gross profit across the AI value chain lands with applications, and the 28% for compute rental providers is before GPU depreciation

Concept illustration: a desktop monitor stands on a desk with a keyboard in front of it; on the screen, below a black title bar with large white lettering, a small laptop showing a speech bubble, a tall stack of red coins with the label APPS under it, a shorter stack of beige coins with the label COMPUTE RENTAL under it, and a small three-unit server cabinet stand left to right on one horizontal line, with three grey horizontal lines below the labels. A pale pink panel stands behind the monitor on the right. Printed on the image: 40% Applications' share of AI value chain gross profit; compute rental providers 28%, Tunguz estimates. — Venture investor Tomasz Tunguz published a model on November 11 that converts the revenue of five layers of the AI value chain, from power plants to applications, into US dollars per watt per year to estimate each layer's share of gross profit. Every layer's share rests on an application gross margin that Tunguz assumes rather than measures. The share for compute rental providers is before GPU depreciation. CoreWeave CEO Intrator's new contract pricing is relayed only by Tunguz.

In the full story: how five layers are put on one unit, where the depreciation step doesn't add up, and when price rises can pass through

Also happened

1. Not verified by us yet: The Information reporter Stephanie Palazzolo says in a post that inference providers unable to get chip allocations increasingly want to buy chips and run their own data centers; the post doesn't name which ones (Stephanie Palazzolo, 2026-11-11).

2. Not verified by us yet: Reuters reports that a Chinese university on the US export blacklist bought NVIDIA H200s in October; according to analyst Paul Triolo, the quantity was two (Reuters, 2026-11-11; Paul Triolo, 2026-11-11).

3. Not verified by us yet: According to a post by Sravan Kundojjala excerpting earnings calls, semiconductor inspection equipment makers KLA and Onto both say rising DRAM prices are weighing on gross margins (Sravan Kundojjala, 2026-11-11).

4. Not verified by us yet: Elon Musk wrote in a reply on X that he would gladly take compute back but thinks the opposite will happen; the reply doesn't name Anthropic (Elon Musk, 2026-11-11).

Deep dive: NVIDIA's two backstops aren't the same thing: one buys AI cloud compute on a schedule, the other pays only after an OpenAI lease default (November 12, 2026)

The market assumes: Cristian Dina, a reporter at the tech outlet The Next Web, wrote in August that NVIDIA is backing the buildings, not OpenAI's rent, so the only thing to watch is what the campus is worth (The Next Web, 2026-08-17). The evidence says: NVIDIA's August current report says the guaranty it gave on a data-center lease for an OpenAI affiliate could require payment only if the tenant defaults, and ends once OpenAI reaches a satisfactory credit rating; whether NVIDIA pays, and when the guaranty ends, both depend on the tenant (NVIDIA current report, 2026-08-17).

NVIDIA's financial support for customers falls into two kinds by how the contracts are written. The first is commitments to buy compute from cloud providers that rent out AI compute: US$36 billion as of July 26, paid on an annual payment schedule, with NVIDIA receiving the compute it pays for. The commitments shrink only when other customers or NVIDIA's own research use up the capacity (NVIDIA quarterly report, 2026-08-26). The second is guaranties that could require payment only if a tenant defaults. The largest is the OpenAI affiliate lease guaranty above, whose initial commitment the August current report caps at US$105 billion; the quarterly report lists another guaranty of the same kind, US$3.5 billion, on data-center leases for AI cloud partners. Our inference: the two kinds are triggered by different events, so expected losses should be estimated separately; only a broad collapse in AI demand could trigger both at once, and only then do they belong in the same worst case.

What would prove this wrong: if, by December 31, 2026, a new NVIDIA quarterly report rewrites the guaranty so that payment is due without a tenant default, or rewrites the compute purchase commitments so they no longer shrink when other customers use the capacity, this reading is wrong. The deadline is SecondSource's own. The nearest checkpoint is NVIDIA's November 17 earnings report.

What this means for you: In estimating NVIDIA's contingent obligations, the two commitments have separate triggers; only a stress case in which demand weakens across the board adds their caps together.

The full deep dive is published in Chinese and Japanese only; there is no English edition.

Chips & semiconductors

1. October revenue at four Taiwanese server assemblers was above a year earlier at all four, from 46.02% higher at Inventec to 117.56% at Wistron; Wiwynn and Inventec fell from September.

Concept illustration: a wide panel on a black base has a black top band, where blue lowercase lettering 'wistron' (with a green i) on the left is joined by a short line ending in a dot to a white box lettered 'WISTRON' in black capitals; inside the panel, a short stack of beige flat layers stands on a shelf line at left and a taller stack of red layers at right, each with a white band down its middle. Behind the panel on the right is a beige flat-roofed building with a roll-up shutter door. Printed on the image: +117.56% Wistron October revenue vs a year earlier, compiled by Wistron. — Wistron, Quanta, Wiwynn and Inventec, four Taiwanese server assemblers, all reported October revenue above a year earlier, while Wiwynn and Inventec fell from September and the filings don't say why. The company-compiled monthly figures have no product breakdown and so don't show how much comes from AI servers.

2. Quanta's CFO says US states' limits on data-center construction haven't led any customer to change orders so far, and fourth-quarter rack shipments may dip slightly.

Concept illustration: a large white sign with curved sides stands on a dark base, its black top band carrying a blue ring-shaped mark with red stripes beside white lettering QUANTA. On the sign, a label reading THIRD QUARTER sits above a long row of black rack cabinets on a shelf line, and below it a label reading FOURTH QUARTER sits above a shorter row of red rack cabinets that stops before the right end of its shelf line. — Quanta CFO Yang Junlie said no customer had adjusted orders so far because of US states' limits on data-center construction, according to the earnings-call account in the Economic Daily News, a Taiwanese business daily. Quanta also says rack shipments peak in the first quarter of next year.

Named commentary

1. TF International Securities analyst Ming-Chi Kuo says Jensen Huang looks unlikely to come through unscathed, and the pragmatic course is for NVIDIA to gradually reduce its reliance on Huang's personal decisions and prepare a succession.

Concept illustration on a mustard background, with two stacked panels that show the same office: each has a sign lettered NVIDIA beside a green emblem at the top, a signboard lettered DECISIONS on the left, four trays of white folders with lines running to a stack of papers on a desk, and a dark silhouetted figure, seen from behind, seated in a red office chair; two large black opening quotation marks are stacked at the right edge beside the upper panel. In the lower panel, the line from the bottom tray is dashed instead and runs along the bottom past the desk, with a single folder on it, to an empty chair drawn only as a dashed outline; beside the chair is a signboard lettered SUCCESSION, and a hand in a dark sleeve reaches in from the right edge and points a long stick at the dashed chair. Printed on the image: Kuo's personal judgment: NVIDIA should gradually rely less on Huang's decisions, prepare succession. — Ming-Chi Kuo, a TF International Securities analyst who has long tracked Apple and the semiconductor supply chain, says Jensen Huang looks unlikely to come through unscathed and the pragmatic course is for NVIDIA to gradually rely less on Huang's personal decisions and prepare a succession. The view is Kuo's personal judgment with no data attached; Kuo names no candidate or timeline and doesn't disclose holdings or client relationships.

2. Commentator Zvi Mowshowitz, after reading the system card Anthropic released with its new models, argues that Anthropic should determine now that the new model Mythos 5.2 has reached the AI R&D acceleration threshold Anthropic set.

Concept illustration on a peach background: at the top, a sign lettered THRESHOLD on the left is joined by a horizontal red bar to a sign on the right lettered ANTHROPIC next to an orange asterisk-shaped mark, with large black quotation marks drawn above that sign. Below the bar stands a tall stack of white sheets with a sign lettered R&D SPEED in front of it, and the top of the stack sits just under the red bar; to the left of the stack, a table holds a laptop whose screen shows a speech bubble and a card lettered MYTHOS 5.2, and on the right a black silhouette of a person holds an open book in one hand and points up and to the left with the other, toward the top of the stack near the right end of the bar. Printed on the image: 1.88x Anthropic's internal estimate of AI's R&D speedup vs a 2x threshold Anthropic set, per Mowshowitz. — Commentator Zvi Mowshowitz argues that Anthropic should determine now that the new model Mythos 5.2 has reached the AI R&D acceleration threshold Anthropic set, after reading the system card Anthropic released with new models. This is Mowshowitz's personal argument, every figure behind the argument comes from Mowshowitz's account of the system card, and Mowshowitz has long argued for slowing AI development.

Research & models

1. Epoch AI, an independent research group that tracks AI compute and capability, estimates that the AI compute six major chipmakers add each quarter is growing 2.5 times a year, down from about 6 times in 2023 by Epoch's estimate.

Concept illustration: a large panel with a thick dark frame stands on two short feet, and a black band along its bottom carries the white lettering EPOCH AI. Inside the panel, on one horizontal baseline, a low block of thin cream-coloured layers sits on the left, and a block of thin red layers of about the same width rises on the right, far taller and reaching almost to the top of the panel; each block's top has a few short slanted marks at both ends and a small black rectangle in the centre. — Epoch AI, an independent research group that tracks AI compute and capability, estimates that the AI compute six major chipmakers add each quarter is growing 2.5 times a year, down from about 6 times in 2023. The sign of a slowdown isn't statistically significant, and both growth rates are Epoch's estimates, counted when chips are sold.

2. A preprint has a strong model write a skill library for a weaker one, and when the teacher profiles the student model first, task success reaches 86.46%, against 81.04% without profiling.

A preprint has the strong model Claude-Sonnet-5 profile the weaker DeepSeek-V4-Flash before writing DeepSeek-V4-Flash a skill library, a set of ready-written procedures an agent can call while working; according to the paper's abstract, when the two models swapped skill libraries, each model did better with the library written for that model, a claim given without figures. The task-success results come from the authors alone as averages of four main runs, and no one else has repeated the results; the abstract doesn't give the task set or the number of tasks, and the paper tests only three models.

Product moves

1. Databricks has made Genie Ontology, which organizes enterprise context for data agents, generally available, and on a benchmark Databricks ran, Genie scores 91%, against 58% and 45% for two unnamed rivals from frontier AI labs.

Concept illustration: on a white display stand against a pale blue-grey background, a framed board shows three bars: a medium-height grey bar on the left, the tallest bar, in red, in the middle, and the shortest bar, in grey, on the right, each topped by a small T-shaped mark. Below the board, three grey laptops stand on the stand's ledge, each screen showing a few lines of text and a speech bubble (the middle one's bubble is linked to three small boxes), and each laptop has a small white card beside it, the middle card lettered GENIE and the other two plain; the front of the stand carries a red shape made of stacked layers and large dark letters reading DATABRICKS. Printed on the image: 91% Genie score vs 58% and 45% for two unnamed rivals, on a test Databricks and Turing built and scored. — Databricks has made Genie Ontology generally available, a feature that lets data agents learn a company's internal definitions and usage automatically from dashboards, SQL and notebooks. Databricks and partner Turing built and scored the benchmark comparing Databricks' Genie agent with two unnamed agents from frontier AI labs, and the two rivals used Databricks' toolkit without Genie's MCP interface.

2. CoreWeave, a cloud provider that rents out GPUs, has made CoreWeave Database, a managed PostgreSQL service that runs inside customers' clusters, generally available, and the announcement gives no performance or pricing figures.

Concept illustration: a gray boxy cabinet stands on two short feet against a pale green background, with a top panel showing a blue angular mark beside the lettering COREWEAVE and, below its screen, three panels, the outer two with vent slots. The screen holds three cards side by side: gray-headed TRAINING and INFERENCE cards, each with a three-dot speech-bubble icon and gray lines, and a red-outlined card headed COREWEAVE DATABASE with a red tiered-cylinder icon and a table grid, all three joined underneath by a thin gray line. Printed on the image: Database runs inside customers' clusters, alongside training and inference, per CoreWeave. — CoreWeave, a cloud provider that rents out GPUs, has made CoreWeave Database, a managed PostgreSQL service, generally available. CoreWeave's announcement gives no performance or pricing figures.

From the archive

1. [Looking back] (SecondSource's public reading of November 6) On Alibaba's open-source Qwen3 models, penalizing only the final output during training, without touching the chain of thought, doesn't keep the chain of thought monitorable.

Concept illustration in two side-by-side panels showing nearly the same scene: at left a person in black clothing with a badge on a lanyard looks upward, a red block lettered HINT stands between them and a black laptop lettered QWEN3 and ALIBABA beside an orange symbol, and above are a cloud-shaped thought bubble at upper left linked to the laptop by a trail of small circles and a speech bubble at upper right whose tail points to the laptop, each bubble holding three grey lines, with a red bar on top of both bubbles in the left panel but only in the speech bubble in the right panel. In the right panel a hand reaching in from the lower right holds a pen against the bottom edge of the speech bubble, and a black band running under both panels is lettered MATS GEODESIC RESEARCH. Printed on the image: Both teams: models still used hints or cheated; chain of thought omitted it; monitor missed it. — Two research teams, Drori and colleagues and Mitrani Hadida and colleagues, each penalized only the final output while training Alibaba's open-source Qwen3 models, without touching the chain of thought; both papers are self-reported experiments that can't be extrapolated to frontier models. Mitrani Hadida and colleagues explicitly set out to reproduce the Drori team's result but succeeded in only slightly more than half of the model and dataset combinations.

Correction: the November 9 issue needs four changes. Main-line item 1 said H200 sales had "no public figures, so the risk can't be calculated": per-company amounts are indeed not public, but NVIDIA's August quarterly report states that licensed shipments were under 1% of the latest quarter's data-center revenue, that is, under US$890 million. Main-line item 2 said ASML's statement that supply can't meet demand appeared only in relays; in fact ASML's October 14 earnings press release says so. The same item and At a glance said lead times "could stretch" to twice the longest ever, while the Chosun Ilbo original says lead times for key equipment have already stretched to as much as twice as long. Main-line item 1 and At a glance said the letter from three House committee chairs asked about counterintelligence measures and Huang's contacts with Megaspeed executives, but left out the letter's second paragraph, which says China's infiltration and acquisition of advanced US technology is the core of the three committees' investigative work this Congress. The letter doesn't name the H200, chips or export licenses; that is unchanged. The archived pages on the site have been rewritten, and the conclusions are unchanged.

SecondSource isn't a news digest: AI reads primary sources across fields, every judgment comes with its original source, and mistakes are corrected in public. The focus is what has changed and which claims still lack evidence, not what happened today.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.