← älter | neuer →
JensenGate: China’s New Models Were Trained on Nvidia’s Blackwell ChipsSynthszr
synthszr #320 from Saturday, November 14, 2026

JensenGate: China’s New Models Were Trained on Nvidia’s Blackwell Chips

  • • Resistance to Anthropic grows as IPO approaches.
  • • Chinese AI models show fingerprints of Nvidia’s Blackwell chips.
  • • Alice Huang’s husband allegedly organized a $7.5 billion smuggling operation.

Industry Pushback Against Amodei’s Anthropic Grows

Anthropic CEO Dario Amodei discussed his company’s IPO plans in a public conversation with JPMorgan CEO Jamie Dimon. According to industry sources, the IPO is scheduled for later this month. It is expected to rank among the largest initial public offerings of all time.

In tandem with the path to the public market, a conflict with major parts of the tech industry has escalated. Anthropic and OpenAI had lobbied at the White House to neither develop nor freely distribute certain AI models due to their potential risks. Microsoft and Nvidia see massive business potential in precisely these models. In late July, Microsoft CEO Satya Nadella and then-Nvidia CEO Jensen Huang exchanged drafts of a letter advocating for openly available models, according to four insiders. Two dozen companies signed before the letter became public. The following week, 200 more joined, including OpenAI. Amodei refused to sign.

Amodei has warned for years of extreme risks posed by artificial intelligence, even as his company builds some of the industry’s most capable systems. According to conversations with more than 20 people across the tech industry, resentment among competitors has now turned into open hostility. Anthropic is widely seen as aggressive in dealing with customers and rivals: The company has cut off individual customers' access to its service and lobbied against free alternatives from other providers. According to these accounts, other tech companies and investors banded together over the past year to take action against Anthropic. Many of those interviewed spoke only on condition of anonymity to avoid jeopardizing their business relationships with Anthropic. → Barron’s Online, New York Times

Synthszr Take: Pushback against Anthropic is above all a dispute among dependents. The letter’s initiators, Microsoft and Nvidia, are themselves invested in Anthropic and sell compute capacity to the company. Many critics voice their concerns only anonymously to protect their business relationships. At its core, the fight is over where margins are captured in the AI stack: Nvidia profits from every model running anywhere, Microsoft from every model it hosts, and Anthropic from keeping the frontier scarce. On both sides, principle aligns with the business model. That doesn’t disprove either side’s argument, but it makes everyone biased. The acid test will arrive with the first two quarterly reports following the IPO: If enterprise revenue stagnates, the safety-first stance becomes a business risk. If it grows, the hostility remains symbolic posturing against a supplier they continue to pay.

Six Chinese AI Models Bear Fingerprints of Nvidia’s Blackwell Chips

A forensic analysis of published model weights concludes that DeepSeek, Moonshot AI, MiniMax, and Ant Group launched their latest models on Nvidia’s Blackwell GPUs. A US license is required for the export of these chips to China; no granted licenses are known. The investigation was conducted by Will DePue of Unemployment Labs and published on November 12, 2026. US government officials had previously accused DeepSeek and Moonshot of using Blackwell without presenting technical evidence. According to the author, the analysis relies solely on publicly accessible weights rather than leaks or insider statements.

Six models are affected. DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300 chips with 148 Streaming Multiprocessor units. DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ling-3.1-flash match GB200 or GB300 chips with 152 of these compute units. Kimi K3.1-Pro starts from the same seed as K3 and could be a post-trained checkpoint of it. Predecessor models from the same labs consistently exhibit Hopper patterns. A Hopper finding does not distinguish between H800, H100, and H200.

The timeline pinpoints the transition quite accurately. DeepSeek-V3 from December 2024 reads as Hopper, while the two V4 models from April 2026 read as Blackwell. At Moonshot, the transition occurs between Kimi K2.5 in January and K3 in July 2026. Kimi K2.7-Code, released six weeks before K3, is based on Hopper weights but contains a small section with B200/B300 patterns. MiniMax switched between April and June 2026, and Ant Group between August and October 2026. The latest evaluable models from Alibaba, Zhipu, Tencent, Xiaomi, and StepFun continue to show Hopper patterns.

The methodology begins at initialization. Before training, a random number generator creates the initial values from a seed, and GPU threads distribute them across the weight positions. For large initializations, 2,048 threads run per compute unit, bringing Hopper to 270,336, B200/B300 to 303,104, and GB200/GB300 to 311,296 threads. Depending on the chip class, the same random numbers end up in different locations. Weights for tokens that never appear during training keep their initial values virtually untouched. The researchers identify the seed and hardware layout on one portion of these weights and verify the combination on a held-out set.

For statistical confidence, the report notes that across roughly 6 trillion tested combinations, a random match would reach around 8 sigma. The strongest findings are reported to fall between 49 and 1,385 sigma. → Unemployment Labs, Wccftech, WinBuzzer, unite

Synthszr Take: Twelve thousand Blackwell chips distributed across four labs, with no known export license, are enough to train an entire generation of cutting-edge Chinese models. A single training run required just 3,000 to 4,500 GPUs. Procuring hardware at this scale via indirect routes is evidently possible. Alibaba, Zhipu, and Tencent, by contrast, continue to train on Hopper. Export controls thus act as a sorting mechanism inside China: determining which labs stay ahead and rewarding those with superior procurement channels.

Alice Huang’s Husband Allegedly Built $7.5 Billion Smuggling Ring for Nvidia Chips

Shanghai fund manager Kang Wei allegedly introduced his wife Alice Huang to former Nvidia CEO Jensen Huang and built a smuggling ring for Nvidia GPUs valued at $7.5 billion. Tom’s Hardware reported the development, citing reporting from The Wall Street Journal. Kang founded investment firm Shanghai Hexi in 2010. Alice Huang joined in 2016 and received a 25 percent stake in the firm from Kang’s holdings, according to the Economic Times. In 2019, the sale of a gaming developer was channeled through a British Virgin Islands holding company registered solely under Huang’s Chinese name. Sources suspect she acted as a nominee signing on behalf of third parties. Following the US export ban on Nvidia chips to China, the buyer, 7Road, spun out its cloud division to Singapore. → Tom’s Hardware

Synthszr Take: Megaspeed took just eight months to go from an outsourced cloud division in Singapore to ownership by a firm named Swiftdata. This chain of corporate restructuring and resale will be suspected behind every GPU buyer by export control officials going forward. Enforcement is shifting from the chip itself to the end customer. Data centers in Southeast Asia will have to disclose their ultimate beneficial owners, even when operating AMD hardware. For neoclouds in the region, compliance becomes a major overhead cost that directly eats into margins.

OpenAI Launches ChatGPT Wallet and Turns Chatbot into a Payment Terminal

According to The Information, OpenAI has launched a wallet feature for ChatGPT. The outlet reported the launch in a short briefing. The available material does not mention feature scope, fees, payment partners, or countries where the wallet is available. A wallet typically refers to a digital account where users store payment methods and process purchases. OpenAI already enables purchases directly in chat with Instant Checkout, a feature developed jointly with Stripe. → The Information

Synthszr Take: A subscription brings in money whether the user buys or not. A wallet only brings in money when they buy. With this, OpenAI adopts the very incentive that slowly unmoored Google Search: the answer that leads to checkout becomes more valuable than the best answer. Merchants integrating via the Agentic Commerce protocol will in future pay the exact entity that decides whose product the agent recommends. Neutrality thus becomes a bargaining chip.

Meta Lets Muse Make Phone Calls, but Only with a $20+ Subscription

Meta is now allowing its AI assistant Muse to make phone calls on behalf of users, though only for paying subscribers in the US. According to Digital Trends, Meta’s Chief AI Officer Alexandr Wang announced the feature on November 12, after Meta had tested it with a small group of users since September. Muse is intended to handle tasks such as booking restaurant tables, sorting out issues with insurance or cable providers, and making calls to government agencies like the DMV. The subscription named Power costs $20 per month and includes 500 million Muse tokens per week. The Maximum tier is $100 per month and increases the allocation to three billion tokens. → Techpresso

Synthszr Take: A call to the DMV mostly consists of being on hold, and Muse has to listen in in real time for every minute of it. That is why Meta charges in tokens here rather than ad impressions. An agent that gets work done does not generate a feed into which ads can be inserted. The tiering from 500 million to three billion tokens per week is a price tag for compute time. The advertising model that built Meta does not fund its agents.

Google Launches Gemini Smart Glasses, Selling Them Through 360 Warby Parker Stores

According to its own announcement, Google is launching smart glasses with integrated Gemini, more than a decade after its first attempt with Google Glass. The frames are made by Warby Parker and Gentle Monster and are designed to look virtually indistinguishable from regular glasses at first glance. Built into the frame are a small camera and speakers near the ears. Via the voice command “Hey Gemini” or a tap on the temple, the AI captures what the wearer sees and responds auditorily. Applications highlighted by Google include navigation via Google Maps, phone calls without reaching for a smartphone, and real-time translation in more than 70 languages. → TAAFT - There’s An AI For That

Synthszr Take: The most important number in this announcement is 360. That is how many Warby Parker stores will offer the glasses for try-on, and that is where the market will be decided. Eyewear requires prescription lenses, fitting advice, and frames people actually want to wear. Meta understood this early on with Ray-Ban and secured optical retail as a distribution channel. Google is now following suit. An AI model can be built into any frame; a retail network of opticians cannot. The mass market will be won by whoever grinds the lenses.

Coding Agents Write Cleaner Code but Reimplement Libraries in 73 Percent of Cases

Coding agents now write significantly less buggy code per line than a year ago, but systematically adopt poor patterns in design decisions. This is reported by researcher Gabriel Orlanski in an evaluation of several benchmarks cited by Techpresso. The basis includes SlopCodeBench, which uses 36 Python tasks and 196 intermediate steps to measure how agents extend their own code as requirements evolve. The strictest solve rate, which requires passing all previous tests, peaks at just 16.3 percent according to Orlanski, rising by about 0.9 points per month. The crux of the criticism comes from the new LibraryUseBench: for every tested agent, more than 73 percent of programs contained a bespoke reimplementation of core library functions, usually with bugs. → Techpresso

Synthszr Take: Cleaner lines make code reviews more dangerous. A linter catches the swallowed exception, but not a custom reinvention of a library function found in over 73 percent of agent programs. Catching such flaws requires holding the entire architecture in your head. As a result, the review burden shifts to the most expensive people on the team, whose capacity many companies are currently planning to reduce. Even the prompt that pushes the rate down to 40 percent is senior work—just done upfront.

Every CEO Dan Shipper Declares Human Attention the Bottleneck of AI

Every CEO Dan Shipper has introduced a formula he calls “the most important formula in AI.” It measures how much a model learns about its user’s preferences per unit of human attention. The approach was developed by Every’s Frontier team, which tests new models and workflows internally and documents its experiments publicly. According to Shipper, attempts to teach an AI one’s personal taste usually fail in one of two ways: a quick thumbs-up provides too little information to the model, while detailed feedback requires more effort than most people are willing to invest. Consequently, an AI follow-up question is only worthwhile if the user answers it and that answer helps the model understand their preferences. → Every

Synthszr Take: Shipper’s formula is a cost calculation for the only raw material AI cannot produce itself. Models generate drafts in infinite quantities, but only a human with limited minutes can evaluate them. This shifts the terms of competition: the leader will be the provider that distills the most taste out of every answered clarification, because this preference knowledge cannot be copied by any competitor. Product teams that still collect feedback as a customer satisfaction metric are burning attention without building an effective model of their user.

FBI Warns: North Korea’s “FaceStealer” Scam Hijacks Identities and Distributes Malware

The FBI is warning of a scam named “FaceStealer,” attributed to North Korean actors. According to PCMag, it allows perpetrators to take over victims' identities and deploy malware onto their devices. Based on available information, the entry point consists of supposed fixes for audio issues. Victims are tricked into running or installing an alleged repair that actually contains malware. Using the compromised devices, the attackers can abuse the victims' accounts and personal data.

Synthszr Take: An alleged audio fix works as bait because operating systems and communication services have delegated troubleshooting to the user for years. Without sound, you look for help yourself and end up executing third-party code with full privileges. The FBI warning pushes responsibility yet another step down the ladder. Platform operators could provide built-in audio diagnostics and consistently block unsigned scripts. As long as they fail to do so, the user effectively foots the bill with their identity.

Biographer and Insiders: Huang Hid Affairs Behind 60 Direct Reports at Nvidia

In a guest essay, Jensen Huang’s biographer Stephen Witt attributes the Nvidia founder’s resignation to hubris and a culture of silence. Witt is the author of the Huang biography “The Thinking Machine.” He explicitly describes the departure as a story of personal hubris, not a story of technology. The trigger was revelations about at least two affairs.

The first relationship involved Alice Huang. She ran a Singapore-based company that owns data centers in Southeast Asia, and U.S. intelligence agencies classify her as a Chinese spy. These data centers are suspected of having helped Chinese AI developers circumvent U.S. export restrictions on Nvidia hardware. According to one report, equipment worth around $7 billion may have been illegally funneled to China this way. Witt considers it hard to imagine that Huang was unaware of these activities.

Huang is also alleged to have had a second affair with Kimberly Powell, an executive in Nvidia’s healthcare division who reported directly to him. According to a former employee, the relationship was an open secret within the company, even though both of Huang’s children hold leadership positions there. Witt speaks of an “omertà” within Nvidia that enabled Huang’s behavior. Only intelligence information and journalistic investigations broke this silence. Witt expects more women to come forward. At the same time, Huang cultivated the image of a humble boss who deflects credit to employees and references his past as a dishwasher. During a shared breakfast, he reportedly tipped a waitress $1,000.

According to three insiders, Huang personally drafted and approved the company’s early denials regarding the relationship with Alice Huang. In doing so, he also misled General Counsel Tim Teter and John Rizzo, the head of policy communications. Rizzo had publicly described the relationship on October 29 as “neither unusual nor unexpected.” Huang resigned on Thursday after the board launched an investigation.

According to insiders, the concealment was also made possible by Nvidia’s management structure. Around 60 executives reported directly to Huang, and on some days he answered up to 1,000 emails. Teams in Santa Clara and around the world were said to be heavily siloed from one another. One insider described this as “Jensen by design”: Huang deliberately structured the company so that no single executive had a complete overview of his actions. According to these accounts, Huang was also personally involved in public communications regarding sensitive topics such as export controls and chip sales to China. → New York Post, New York Times

Synthszr Take: Around 60 direct reporting lines were seen at Nvidia as a speed advantage. That same structure led even the General Counsel to stand behind denials that Huang had drafted himself. Uber with Kalanick provided the blueprint: founder worship becomes dangerous as soon as no one is close enough to push back. The successor inherits an organization without a layer to catch mistakes. It has to be built first, and that costs the speed investors have priced in.

Mentioned in this article

Can you build magical companies with magical technology?

3RD EDITION. 452 PAGES, 20 EUR (German and English).

Can you build magical companies with magical technology?

The core thesis of CODE CRASH is: when code becomes cheap, intent becomes scarce — the clarity about what should be built. Now also becoming scarce is the ability of organisations to absorb what the agents deliver. The third edition of CODE CRASH describes the layers of agentic AI — intent, context, tools, checks and human judgement — in the interplay of teams and organisations.

codecrash.ai →

Subscribe free. Unsubscribe the second it sucks.

High-signal news across AI, business, UX, and tech. Every morning.