← älter | neuer →
Crash or Correction? Anthropic and Nvidia Drag Down the NasdaqSynthszr
synthszr #327 from Saturday, November 21, 2026

Crash or Correction? Anthropic and Nvidia Drag Down the Nasdaq

  • • Anthropic stock loses over 11 percent, Nasdaq suffers sharp setback
  • • Google implements AI sidebar to make search more efficient
  • • Ex-Nvidia CEO Huang driven by fear and humiliation, insiders report

Anthropic Stock Plunges, Nasdaq Erases All 2026 Gains

Anthropic shares dropped more than 11 percent on their second day of trading. Measured against the opening price of $111.50, the stock has shed more than a fifth of its value within two days. The slide came amid a broad sell-off across US stock markets: The tech-heavy Nasdaq Composite fell by more than 3 percent on Friday, wiping out all of its 2026 gains.

Yet Thursday’s debut had looked like a triumph. The previous evening, Anthropic had priced its shares at $83, above the previously targeted range of $70 to $78. The IPO raised $81.3 billion, making it the largest in history according to NBC News and Reuters. Trading under the ticker symbol “ANT”, the stock opened on the Nasdaq at $111.50, 34 percent above the issue price. On a fully diluted basis—including options and shares reserved for employee programs—the company was temporarily valued at well over two trillion dollars, instantly becoming one of the most valuable companies in the US.

The premium did not last long. Over the course of the day, the stock slipped well below its opening price, closing at $97.17—around 17 percent above the offering price, but nearly 13 percent below its opening price. A rule for employees is considered a likely reason: non-founder staff and managers are permitted to sell ten percent of their shares starting on the first day of trading. According to the prospectus, this applies to up to 315 million shares. A Lock-up period applies to the remainder.

On Wall Street, the double-digit jump was initially seen as a signal. Dan Morgan, portfolio manager at asset manager Synovus, described it as a sign of strength following volatile weeks for tech stocks. Yet the hoped-for boost to the broader market failed to materialize. US markets opened significantly higher around the IPO, but then turned negative: the S&P 500 and the Nasdaq Composite both closed around 2 percent lower on Thursday.

Contributing to this was a report from investigative outlet Hunterbrook, whose affiliated fund is shorting chipmaker AMD. Hunterbrook pointed to risks in its China business and warned that AMD is entering the same kind of so-called circular deals as Nvidia, where suppliers help finance their own customers. AMD denies any wrongdoing. The Nvidia rival fell more than nine percent in late trading. Cloud providers Nebius and CoreWeave, which specialize in AI compute capacity, lost eight and five percent, respectively.

On Friday, the sell-off continued. Alongside Anthropic, other major AI stocks suffered significant losses, including CoreWeave, Nvidia, and Oracle. The S&P 500 slipped nearly 2 percent, sitting 10 percent below its recent peak. Following the Dow Jones before it, it entered what is known as a correction. The benchmark index was heading toward its worst week since April 2025, when President Donald Trump announced his “Liberation Day” tariffs. → Reuters, The Information, Fortune, Business Insider, TechCrunch

Synthszr Take: Is this 2001 all over again? Back then, the market crashed when it became clear that capacity had grown faster than demand: carriers had laid fiber years in advance that no one needed yet—the endless streams of social media feeds and Netflix conquered screens only later. Today, there is no indication among users or enterprises that demand for computing power is growing more slowly than the capacity being ordered or built. All major hyperscalers report demand exceeding supply, with Gartner and IDC forecasting shortages until at least 2027. In the third quarter, Anthropic even had to buy $5 billion worth of expensive, short-term compute capacity to meet demand. Without oversupply, a crash of nearly 80 percent like back then is fundamentally unlikely. What we are seeing is a correction, not an implosion.

Google Rolls Out AI Sidebar Globally, Bringing Search Closer to Gemini

Google is introducing a collapsible sidebar on its desktop search homepage worldwide, allowing users to find and resume previous AI searches. Robby Stein, Vice President of Product at Google Search, explained that users are increasingly arriving with questions that cannot be resolved with a single query. According to Stein, tests showed that the sidebar made a significant difference in making progress and continuing to explore the web. He described it as a simpler way to pick up where one left off. The feature had surfaced in tests several weeks earlier, where it was referred to as the “AI Mode Sidebar”.

The sidebar is available exclusively in AI Mode. Clicking on the left area of the AI search box opens it and displays recent search queries, including both AI searches and classic organic searches. Selecting an entry opens a new window in AI Mode. The sidebar can be closed at any time via an icon in the top right corner. Those not using AI Mode will not see the new feature.

With this update, the interface of classic Google Search aligns even further with the appearance of the standalone Gemini app. AI Mode and AI Overviews had already shifted Search toward a chat assistant. In the same week, Google introduced temporary threads to AI Mode. Much like temporary chats in Gemini, such queries are not saved in the search history, with the corresponding button also located at the top right of the page.

The rollout comes at a time when Google’s AI strategy is facing pushback. Segments of users, employees, and publishers already believe that Gemini’s presence in search is too dominant. The company’s market capitalization recently dropped by $500 billion after Google admitted that AI search generates less revenue than organic search. On the recent earnings call, CEO Sundar Pichai described the AI commitment as a “long-term planning process, not a short-term allocation exercise.” Observers expect that the sidebar is likely to boost query volume across Search and AI Mode. → Engadget, Search Engine Roundtable, Thurrott

Synthszr Take: Even a classic organic search opens from the new sidebar as a fresh AI Mode window. Google is effectively turning old queries into AI sessions. The company urgently needs this, as AI search brings in less revenue than organic search, and its market capitalization dropped by $500 billion following that admission. Until monetization holds up, usage must justify the expansion. Publishers are paying the price with clicks that never reach the open web.

New York Times: Former Nvidia CEO Jensen Huang Reportedly Led with Fear and Humiliation

Following the ouster of Jensen Huang as Nvidia CEO, The New York Times reports on a management style that, according to former employees, was built on praise, fear, and public humiliation. Huang reportedly told subordinates that given Nvidia’s wealth, money no longer served as motivation, leaving humiliation as the only remaining lever. He reportedly told one manager that his mother should have aborted him. He reportedly threatened to rip another employee’s arms off and beat him to death with them. The Times says it is relying on documents spanning an extended period: in 2005, former HR chief Michael Winston threatened a lawsuit, calling Huang’s behavior the most abusive he had ever witnessed. Nvidia subsequently paid him more than one million dollars. A spokesperson for Huang called the allegations fabricated. → TechCrunch

Synthszr Take: One million dollars paid to Michael Winston in 2005 shows that Nvidia’s board had known about the pattern for twenty years. It was tolerated as long as the share price overruled every complaint. Uber played out the exact same scenario with Travis Kalanick: the allegations had been around for a long time, but action was only taken once they started costing money. The Times investigation therefore primarily provides the justification for a decision that had already been made. A founder-CEO’s behavior remains a private matter as long as their numbers shine.

Thinking Machines' Inkling 2 Mini Reaches Qwen Level with a Fifth of Active Parameters

According to Artificial Analysis, Thinking Machines has released Inkling 2 Mini, an open-weights model that, as a Looped Transformer, cycles through part of its layers multiple times per token. The Mixture-of-Experts model comprises 26 billion parameters, of which 5.4 billion are active. It is small enough to run on a MacBook or on Nvidia’s DGX Spark. Twelve of its 24 layers are executed three times per token, which, according to Artificial Analysis, increases compute capacity without generating additional output tokens. In the Intelligence Index, the model scores 20 points, putting it on par with Qwen3.8 27B and Nanbeige4.3-33B, which require roughly five and six times as many active parameters, respectively. → Artificial Analysis

Synthszr Take: On the price list, Inkling 2 Mini loses. According to Artificial Analysis, GPT-6 Luna scores 28 index points and costs less per task than Inkling’s $0.59. The unit cost calculation only tilts on owned hardware: there, every request costs merely electricity, and the MacBook is already sitting on the desk anyway. Running layers three times makes this possible because it squeezes more compute out of the same weights instead of demanding more memory. For routine enterprise tasks, the marginal cost thus drops toward zero.

Anthropic: 'Think Thoroughly' Only Makes Claude Slower, Not Better

Claude delivers responses of the same quality without prompt add-ons like 'think hard' or 'take your time'—just faster. That is reported by the newsletter TAAFT, citing Anthropic. According to the provider, Claude thinks through every response anyway, regardless of whether prompted to do so by the user. TAAFT points to a study by the University of Chicago: AI shopping assistants instructed to think longer did not make better decisions; they merely appeared more confident. Anthropic recommends instead naming the objective, describing the desired format of the output, and defining when Claude should seek clarification. As an example, the company cites a retailer with declining sales and four hypotheses regarding the causes. → TAAFT - There’s An AI For That

Synthszr Take: 'Think thoroughly' has outlived its usefulness as a control lever. Anthropic moved the dial from the prompt into a setting. Compute effort thus becomes a conscious choice between Medium and High, visible rather than hidden in the text. For companies with mature prompt libraries, this means taking inventory. The phrases cost wait time and, according to the Chicago study, merely generate overconfidence. What makes the difference is the clearly stated objective. In the case of the retailer, it was enough to disprove all four hypotheses and pinpoint cannibalization by new branch stores.

Meta Reportedly Plans to Charge $300 for Its Muse Charm AI Pendant

According to a report by The Information, Meta plans to charge $300 for Muse Charm. The company teased the Tamagotchi-style AI device at this year’s Connect conference. A Meta spokesperson did not issue a denial to TechCrunch, stating only that they would 'share more soon.' Muse Charm is intended to embody Meta’s agentic AI assistant Muse as a physical figurine that users can carry in their pocket or wear around their neck to converse with and issue commands to. Meta has already launched the Muse software with various avatars designed to give the system a personality. → Techpresso

Synthszr Take: Amazon sells Bee Frontier for $75 and profits from a $99-a-year subscription. For Muse Charm, there is no mention of a subscription, meaning the $300 price tag has to carry the entire load. That will be tight, because every query directed at the pendant incurs compute costs in Meta’s data centers long after the initial sale. The more eagerly customers use the device, the faster the hardware margin evaporates. This can only pencil out through Meta’s advertising business.

Signal Blocks Screenshots to Prevent Siri and Gemini from Reading Private Chats

Signal is restricting screenshots on iOS 27 and select Android devices to prevent built-in AI assistants like Siri and Google Gemini from reading private chats off the screen, Cyber Security News reports. End-to-end encryption only protects messages in transit between devices. Once a chat is opened, the text is visible on the display, allowing an assistant to capture it there without breaking encryption. On iOS 27, the new 'Block Siri AI' setting prevents Siri from scanning Signal content and blocks screenshots within the app as well as in the app switcher. On Android, Signal automatically enables the 'Screen security' feature as soon as an assistant requests access to the screen. → Cyber Security News

Synthszr Take: Apple and Google are both referees and players in this dispute. Siri and Gemini become all the more useful the more they are permitted to read off the screen. A clean opt-out API would allow every app to withdraw its content from this data stream. No platform operator will voluntarily build the tool that starves its own assistant. So the blunt instrument remains: blocking screenshots. Signal is willing to accept this tradeoff; a banking app or collaboration tool hardly will. Any real change is more likely to come from Brussels than Cupertino.

Google Tests Its Chrome Agents in an Artificial Internet with Fake Banks

Google’s Chrome security team has built the “Mirror Dimension,” a sandboxed, synthetic internet in which it systematically attacks its own AI browser agents. According to a post by engineers Aaron Selya and Charles Qian on the Google Security Blog, it consists of hundreds of interactive fake websites, including social networks, online stores, and banking dashboards. Wired frames the project as a response to indirect Prompt Injection—hidden instructions on web pages designed to misdirect an agent. According to Google, each security engineer generates more than 1,000 attack test candidates per day using it, about 30 times as many as manually. The attack payloads are injected via a custom HTTP header named “X-Sandbox-Injection-Key” to keep URLs inconspicuous and prevent the agent from recognizing the test. → AI Weekly Espresso

Synthszr Take: The 1,000 test candidates per day reduce the cost per attack attempt by more than thirtyfold. For companies looking to deploy agents in e-commerce or banking, a different metric matters: the cost per deflected attack. This figure is missing, as Google does not disclose a success rate. Cheap tests shift the bottleneck to evaluation and remediation, where engineers whose workdays don’t get thirty times longer are still doing the work.

Anthropic Mathematician Claims Riemann Proof with Claude, Stanford Researcher Counters Without AI

Levent Alpöge, a mathematician at AI company Anthropic, claims to have proven the Riemann Hypothesis using Claude. Originating in 1859, the conjecture concerns the distribution of prime numbers and the location of the zeros of a particular function. It is considered the most famous open problem in mathematics and is one of the seven Millennium Prize Problems of the Clay Mathematics Institute, which awards one million dollars for each recognized solution. So far, only one of these problems has been officially solved: the Poincaré conjecture.

Alpöge has not published a human-readable version. An unpublished draft reportedly spans more than 250,000 pages, posting which he calls “obviously inappropriate.” Instead, he made the proof available in Lean, a language mathematicians use to verify proofs step by step by machine. If such code compiles, the theorem is considered proven beyond doubt. However, the process has potential vulnerabilities, such as if the hypothesis itself is encoded incorrectly. Kevin Buzzard of Imperial College London reacted with: “Let’s see if it even compiles.”

In his statement, Alpöge writes that the method also proves the Generalized Riemann Hypothesis. With more time, he expects the same for the extended variant. He noted that Claude proposed the crucial step, which involves a chain of constructions of type Theta-Lift of length square root of X. James McKee of Royal Holloway University points out that numerous results previously conditional on this hypothesis would now become theorems. Alpöge announced plans to simplify the argument and write it out properly.

Regarding the timeline, Alpöge writes that he notified Anthropic CEO Dario Amodei “late on Sunday evening” via text message. He apparently waited until after the company’s IPO on Thursday before publishing. Anthropic has not commented on the announcement. In August, the company had stated it did not expect Claude’s techniques to lead to a proof of the hypothesis.

A few hours after Alpöge’s announcement, Alessandro Masullo, a postdoc at Stanford, published his own proof, claiming it was done without AI. He stated he had been working on the problem for twelve years and was about a year away from publication. He was advised “given the current climate” to release the current draft, which might contain errors due to the rush. → Levent Alpöge, Business Insider, New Scientist, The Verge, Scientific American

Synthszr Take: The duel benefits everyone except mathematics. Business Insider gets clicks, Anthropic gets a hero narrative one day after its IPO, and Masullo gets a stage for his “we are still here.” The price is paid in quality: after twelve years, a postdoc publishes a draft that, by his own admission, may contain errors. The bottleneck thus shifts to the reviewers. Whether Alpöge’s Lean code correctly encodes the hypothesis determines the million-dollar prize—and that verification does not scale with reach.

DeepMind Researcher Genewein: AGI Is Within Reach, Superintelligence Could Follow

Tim Genewein, a research scientist at Google DeepMind, believes artificial general intelligence (AGI) is within reach and argues for thinking about the stage beyond right now. In an essay, he describes the potential transition to artificial superintelligence (ASI). He defines AGI as machines that match humans across virtually all cognitive tasks. His thesis: progress could continue at the current pace or even accelerate. He concedes that this is by no means certain.

For ASI, Genewein sets a higher bar. He refers to a system that outperforms large groups of human experts on tasks that would take them years, using less time and fewer resources. As examples, he cites driving scientific progress across an entire research field in a fraction of the time, or chip design breakthroughs equivalent to the annual output of major corporations. ASI would have to deliver this performance across virtually all tasks of human interest. Specialized systems that already triumph at chess or Go, predict protein structures, or prove long-unsolved theorems explicitly do not meet this standard.

Genewein views intelligence as a continuum on which human-level capability is just one point. The theoretical upper limit of general intelligence, he notes, is mathematically well understood. Physical, logical, and complexity-theoretic limits are also known, such as in predicting climate, immune systems, or societies with arbitrary precision. What lies between human level and these limits, however, remains largely uncharted. As an example, he cites chess: solving the game perfectly would require more computing power than exists on Earth, yet a chip the size of a fingernail defeats any human.

Genewein calls the driving force “cognitive industrialization.” By his calculations, the effective compute available for AI grows tenfold annually, meaning a ten-billion-fold increase over ten years. Three factors drive this: compute per dollar has grown by roughly 1.5x annually for six decades; the training compute needed for a fixed performance level decreases by 3x to 6x each year, and for inference by roughly 3x; and spending on AI hardware has grown roughly 2.5x annually over the past decade. Even without advances at the capability frontier, he estimates these resources could make existing systems significantly more effective.

However, by Genewein’s own account, more compute does not guarantee superintelligence. → DeepMind Institute, The Independent

Synthszr Take: Genewein’s ten-billion factor measures the input, while the threshold itself remains unverified. For ASI, he at least provides benchmarks, such as a year’s worth of chip design output. For AGI, “virtually all cognitive tasks” is sufficient for him, and the essay does not reveal which task tips the scales. This allows its arrival to be proclaimed at any time and never disproven. Corporations pacing their roadmaps to this are funding a compute curve and hoping it will eventually turn into a test result.

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.