Investors Put Pressure on Anthropic and Jensen Huang Has a Problem

Following Nasdaq Slump, Investors Want to Cut Anthropic’s Valuation in Half

Anthropic’s investors are now discussing a valuation of just one trillion dollars for the planned IPO, according to reports citing unnamed sources. In October, the company reportedly still targeted two trillion dollars. This puts the target valuation today roughly one trillion dollars below its value from a month ago. Previously, investors had signaled a fair value closer to 1.5 trillion dollars to banks. Three major backers reportedly consider a 500-billion-dollar discount necessary.

The trigger was Thursday’s sell-off. The Nasdaq lost between 5.5 and 6.4 percent that day, depending on the report, sliding into correction territory. At points, it trailed its October 6 peak by nearly 12 percent. The semiconductor index SOX plunged by 12.3 percent. This was preceded by historic gains for Democrats in the midterm elections. Since then, investors have been weighing the possibility that data center expansion could slow down.

On Friday, prices stabilized slightly. The Nasdaq gained 1.1 percent, with chip stocks like Nvidia, CoreWeave, Intel, and Sandisk recovering some of their losses. Sentiment was supported by strong labor market data: 165,000 new jobs were created in October, and the unemployment rate fell to 4.1 percent. Over the week, however, a 3.4 percent loss remained, and the index continued to trade more than 10 percent below its October high.

Anthropic had planned to go public in the coming weeks, aiming to raise up to 100 billion dollars. The company could postpone the date. However, Michael Field, strategist at Morningstar, considers a postponement more damaging to long-term prospects than a lower valuation. HSBC analyst Max Kettner had predicted the sell-off. He attributes greater weight to the outcome of Anthropic’s IPO for market sentiment than to the elections. He describes the best-case scenario as follows: Anthropic goes public, but at one trillion instead of 1.5 trillion dollars.

The valuation is relevant for the entire industry because Anthropic stands at the center of AI investments. The company has made commitments for compute and infrastructure capacity totaling nearly 900 billion dollars, including with Amazon, Google, Microsoft, and Broadcom. Its financial condition thus affects valuations throughout the supply chain, from chipmakers and networking equipment providers to data center operators and power utilities. → Axios, Capital Brief, Telegraph, The Asia Business Daily

Synthszr Take: Anthropic is the price tag against which every private AI round is benchmarked. If the benchmark valuation is cut in half from two trillion to one trillion dollars, the comparable multiples of all smaller labs slide with it. Their latest rounds were negotiated against this exact anchor. Consequently, the next financing round turns into a down round or falls through entirely. Labs with large compute commitments will be hit first, as their bills keep coming while valuations drop. Compute discipline thus becomes a survival trait.

US Intelligence Agencies Classify Jensen Huang Confidante Alice Huang as Chinese Agent

According to insiders, US intelligence agencies have concluded that Alice Huang, an executive at the company Megaspeed, was recruited by China’s Ministry of State Security (MSS). Her assignment was reportedly to gather intelligence on Nvidia CEO Jensen Huang. US officials reportedly determined that the Nvidia CEO maintained a “close relationship” with the businesswoman. She is said to have passed information to Beijing.

According to the assessments, the operation also involved exerting influence. Alice Huang reportedly attempted to persuade Jensen Huang to lobby for an easing of US export controls that restrict the sale of Nvidia chips to China. Megaspeed is based in Singapore and was led by Alice Huang. Critics describe the company as a Chinese front company, pointing out that it continues to have access to Nvidia’s most powerful chips. The relevant US export authority, the BIS, has not barred Megaspeed from shipments so far. Furthermore, claims on social media suggest that Megaspeed is primarily funded by Alibaba. This claim has not been independently verified.

The case has now drawn the attention of the US Congress. The chairs of three House committees have subpoenaed Jensen Huang for a closed-door briefing scheduled to take place later this month. The subject is his ties to Alice Huang.

In parallel, Nvidia is tightening controls over the whereabouts of its hardware in Southeast Asia. According to insiders, regional customers seeking Nvidia Cloud Partner status will in future have to participate in a program that tracks the locations of their AI chips. Nevertheless, political pressure is likely to mount: Democrats, having won majorities in both chambers, are preparing stricter oversight of Nvidia. The focus is on chip exports and zirkuläre Finanzierung in the AI sector. Elizabeth Warren is considered a potential chair of the Senate Banking Committee. → The Information AM, list-manage

Synthszr Take: China’s access to hardware was apparently never the problem: Megaspeed continues to receive Nvidia’s most powerful chips to this day while the BIS looks on. What was scarce was an influential voice in Washington advocating for relaxed export controls. That voice belonged to Jensen Huang, and the MSS just undermined it itself. From now on, any Nvidia argument in favor of the Chinese market sounds like a briefing from Beijing. Tracking in Southeast Asia comes too late to save lobbying efforts before Congress.

Google Cuts Video Price for Gemini Omni 1.2 Flash by 30 Percent and Enables 15-Second Clips

Google is making its video model Gemini Omni 1.2 Flash generally available via the Gemini API and Gemini for business, cutting the price by 30 percent compared to previous version 1.1. For 720p video, Google now charges $0.07 per second. According to Google, clips can now be generated natively up to 15 seconds long, up from the previous limit of 10 seconds. Length can be chosen in one-second increments between 3 and 15 seconds. New is conversational multi-turn editing: background, product, lighting, or aspect ratio of an existing clip can be altered without regenerating it completely. → Google

Synthszr Take: At $0.07 per second, Google’s wholesale price becomes the benchmark for every creative tool built on the Gemini API. Their markup could previously be justified by workflow: drafting, correction, upscaling. Omni 1.2 Flash now delivers these steps itself, down to modifying just the lighting. Providers are left with little more than the user interface. The margin per second of video ends up with Google Cloud, while the tools compete against their own supplier.

OpenAI Classifies GPT-6 Cyber as 'Critical' and Restricts Access to Vetted Defenders

OpenAI has introduced GPT-6 Cyber, a model designed to find, exploit, and validate software vulnerabilities. Under its Preparedness Framework, the company classifies it as 'Critical' in the cybersecurity category; its predecessor GPT-5.6 Cyber was rated 'High'. According to OpenAI, the model achieves 48.5 percent on the ExploitGym benchmark, compared to 32.1 percent for its predecessor. In an alignment evaluation, the vendor states it did not target services outside the authorized scope in 99.93 percent of tests, and in 100 percent of cases when using the designated deployment prompt. According to OpenAI, researchers used the model to develop 948 exploits for Privilege Escalation on Linux and discovered 81 previously unknown, severe vulnerabilities in Firefox. Mozilla has already released fixes for many of them. → OpenAI

Synthszr Take: 81 severe Firefox vulnerabilities land at Mozilla, an organization with a paid security team and well-established release cycles. Most open-source projects rely on a single maintainer working after hours. Daybreak is intended to reach precisely those people, delivering findings rather than extra hands. With every model leap, the backlog between reported and patched vulnerabilities widens. Security budgets will therefore shift from discovery to remediation—meaning toward providers that write, test, and deploy patches.

Thinking Machines Trains Inkling 2 Small in 68 Hours, Open Weights to Follow

Thinking Machines has released Inkling 2 Small, a multimodal model with 594 billion parameters, of which 17 billion are active per token. It is available immediately via the company’s proprietary platform Tinker for inference and fine-tuning. According to the team, weights are scheduled to follow next week under Apache 2.0 once external safety evaluations are concluded. Artificial Analysis still lists the model as proprietary at launch. According to Soumith Chintala, pre-training ran for 68 hours on GB300 chips, with the main phase covering 8.4 trillion tokens in the NVFP4 number format. The model scores 31 points on Artificial Analysis’s Intelligence Index, 18 more than its predecessor. With open weights, it would be the highest-rated open model outside China, according to Artificial Analysis. → Latent Space

Synthszr Take: The most expensive line on the price sheet is training: $1.50 per million tokens, more than for output. Thinking Machines sells customization; the model is the hook. The launch sequence makes this clear: Tinker first, weights a week later. Under Apache 2.0, any customer can subsequently self-host their fine-tuned model. Tinker must therefore retain through tooling and speed what the license does not lock in.

As Incoming Supervisory Board Chairman, René Obermann Demands Real AI Value from SAP

René Obermann will become supervisory board chairman of SAP in early 2027, manager magazin reports in an in-depth profile. The 63-year-old was previously a startup founder, CEO of Deutsche Telekom, and investor at Warburg Pincus, most recently chairing the supervisory board of Airbus. For the profile, editor Caspar Schlenk spoke with associates including Telekom CEO Tim Höttges and 1&1 CEO Ralph Dommermuth. They describe Obermann as a self-made achiever from modest beginnings who wants to understand every detail. He himself cites his origin as 'Krefeld, behind the train station.' → Tech Update – manager magazin

Synthszr Take: Musk bet nearly half a billion dollars on one political camp and lost. SAP is bringing in the exact opposite: a chief overseer whose capital consists of two corporate careers at Telekom and Airbus, not proximity to political power. The loser in this personnel move is Christian Klein. Obermann is unfamiliar with enterprise software and will therefore want to see every AI announcement translated into revenue. A supervisor who wants to understand every detail reads roadmaps like balance sheets—and there, tangible real-world value is all that counts.

PSI Uses Anthropic AI to Settle Open Conjecture on Quantum Advantage Standing Since 2008

Startup Physical Superintelligence (PSI) claims to have proven the Aaronson-Ambainis conjecture. The conjecture has been open since 2008 and addresses the question of when quantum computers can achieve exponential advantages over classical computers. According to PSI, an internally developed research system found the complete proof on the third day of a run, backed by instances of an unreleased Anthropic model. According to a paper by David Turturean published on GitHub, the main theorem was formalized in the Lean proof assistant, verified by the kernel, and peer-reviewed by researchers. The result pertains to yes-or-no questions: a classical algorithm can approximate on most random inputs the probability that a quantum algorithm answers 'yes', without requiring exponentially more data queries. The theorem holds in the Query-Modell, which counts only read accesses to the input rather than total computation time. → The Quantum Insider

Synthszr Take: The Lean kernel checks every logical step, but not the translation. According to the paper, what was formalized was their own formulation of the conjecture, and that is precisely where it is decided whether PSI solved the 2008 problem or an adjacent one. That verification remains manual work for human experts. If it holds up, the bottleneck visibly shifts: the machine delivers the proof on day three, but selecting and validating the right theorem remains an expensive exercise in human judgment.

Meta Trains Free Models with Wonder, Keeping Top Model for Paying Customers

Starting Friday, all users of Meta’s agent app Muse and of Meta AI are receiving new models trained by the company using its flagship model Muse Wonder. Wonder itself remains reserved for paying subscribers for more demanding tasks. Meta has not named the new models, referring to them as 'specialized' versions of existing models. They adopt two features first introduced with Wonder: Active Memory and Active Focus. Wonder and the more capable variant Wonder Full have been available since October 12. Wonder is accessible via the Meta Model API and the coding agent Muse Code, while Wonder Full is available only in the two most expensive tiers of Muse Code.

Muse launched in September based on the smaller model Muse Spark. Unlike comparable personal agents, Muse offers a free tier. In contrast, OpenAI’s Dots and Anthropic’s Duet are included only in the respective providers' highest pricing tiers. According to Meta, it did not build a downsized general-purpose version of Wonder. The models were specifically trained on behaviors that matter in a consumer app. This includes deciding when the agent should work autonomously and when to clarify with the user. It also includes accepting mid-task corrections without losing context. Using a large model to train smaller models is a common industry practice known as Distillation. However, Meta has not disclosed its specific method.

Active Memory is designed to carry over details from past conversations into new ones. According to Meta, the models distinguish between temporary and persistent information and recognize which details are outdated. In Meta’s internal evaluation, the new models correctly use critical context in 90 percent of tasks, compared to 76 percent for the predecessor model. Active Focus is intended to keep tasks moving even if something goes wrong along the way, such as when a website crashes midway through a booking. Meta reports a 67 percent improvement in completing routine tasks involving unexpected issues, though it provides no baseline figure. Both performance claims originate from the provider itself.

Via the API, Wonder costs $4 per million input tokens and $20 per million output tokens. Meta AI runs inside WhatsApp, Instagram, Facebook, and Messenger. Industry observers estimate that lighter models trained on Wonder will be significantly cheaper than serving Wonder itself across this massive user base. → The New Stack

Synthszr Take: Until Friday, Active Memory and Active Focus were tied to the paid Wonder model; now they run for free inside WhatsApp, Instagram, and Messenger. For third-party assistants, this is the most dangerous competitor imaginable: one that costs nothing and is already installed. Their selling proposition was that their agent remembers and follows through on an interrupted booking. Meta has distilled both capabilities and given them away for free. What remains are niches with proprietary data or access to enterprise systems that are off-limits to Meta AI.