← älter | neuer →
Here It Is: The Largest IPO of All TimeSynthszr
synthszr #325 from Thursday, November 19, 2026

Here It Is: The Largest IPO of All Time

  • • Anthropic sets a new record for IPOs with $81.3 billion
  • • Google introduces Gemini 4 Neon – a new model for AI applications
  • • Mistral focuses on AI solutions for industrial engineering and regulation

Anthropic raises $81.3 billion: largest IPO of all time

Anthropic has set the offer price for its IPO at $83 per share. The range was previously $70 to $78. The company increased the offering from an initially planned 825 million to 980 million shares, raising $81.3 billion. In June, SpaceX had initially raised $75 billion in what was until then the largest IPO, totaling $85.7 billion after the exercise of the Greenshoe-Option. Anthropic is granting the underwriting syndicate, including Goldman Sachs and Morgan Stanley, a 30-day option for up to 147 million additional shares. Trading on the Nasdaq begins on November 19 under the ticker symbol “ANT”, with CEO Dario Amodei and the co-founders ringing the opening bell.

At the offer price, Anthropic is valued at around $1.5 trillion on a fully diluted basis. In its last private funding round in May, which brought in $65 billion, the valuation was still at $965 billion. A few months ago, the company was reportedly targeting $2 trillion. Of the 980 million shares, around 956.4 million come from Anthropic itself and around 23.6 million from existing shareholders. According to the prospectus, the seven co-founders are each selling just under 3.4 million shares, amounting to roughly $282 million per person. Their remaining stakes were each worth more than $18 billion even at the high end of the initial price range.

The largest shareholders are strategic partners Amazon with nearly 15 percent and Google with nearly 14 percent, with both stakes each worth more than $100 billion. Concurrently with the IPO, a group comprising AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC is buying non-voting shares for $26 billion at the offer price. Anthropic is also a customer of expensive compute systems from Nvidia, Google, Amazon, and AMD. No venture capital firm holds more than 5 percent. Sequoia, former Google CEO Eric Schmidt, and sovereign wealth fund GIC are each said to own around 1 percent. As far as is known, Spark Ventures is the only investor with a board seat. According to insiders, Sequoia, Iconiq, Lightspeed, Accel, and Altimeter intended to purchase billions worth of additional shares in the IPO.

The prospectus reveals a steep rise in revenue: $41 million in 2023, $386 million in 2024, and $4.585 billion in 2025. In the first half of 2026, an additional $16.286 billion was generated. Over the same period, Anthropic recorded a net loss of $80 billion. By the end of June, the company had raised a total of $141.1 billion. The single largest cost block is compute power: purchase and lease commitments amount to more than $70 billion over the next twelve months and more than $450 billion through 2030. Another $450 billion will follow through 2045. → CNBC, Anthropic, TechCrunch, The Guardian

Synthszr Take: Nvidia, Google, AMD, and TSMC are buying $26 billion in shares of a customer that has committed more than $450 billion to them for compute power through 2030. The money goes in a circle and ends up back with the supplier. The $83 offer price is therefore also an insider price among business partners. On Thursday, money that was not earned in any data center will enter the picture for the first time. Only the price of ANT will show whether 1.5 trillion is a valuation or a supply chain with a ticker symbol.

Google launches Gemini 4 Neon at Opus 5.5 level

Google released Gemini 4 Neon on November 18, a mid-tier model positioned between the compact Gemini 4 Hydrogen and the flagship Gemini 4 Carbon. Neon is available via the paid Gemini API in Google AI Studio and Android Studio, as well as through Gemini for Business. Subscribers to Google AI Pro and Ultra receive access in the Gemini app and in the Google Antigravity developer platform. Pricing is $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15, a 95 percent discount. The context window spans one million tokens. Carbon was released about a month ago, and Hydrogen the previous week. According to Artificial Analysis, Neon is Google’s fourth model within seven weeks.

On the Artificial Analysis Intelligence Index, Neon reaches 61 points at the highest reasoning tier, “xhigh”. That puts it in third place behind OpenAI’s GPT-6.2 Astra at 62 and Anthropic’s Claude Fable 5.2 at 65 points. Artificial Analysis independently confirmed the ranking. On the Vals Index, Google reports 70.9 percent and second place, though the public Vals standing still dates back to before the release. On two engineering benchmarks, Google claims Neon is roughly one percentage point ahead of Claude Opus 5.5. Google also reports 65.2 percent on Terminal-Bench 4.1, 46.2 percent on FrontierCode 2.0, and first place on the Coding Agent Index from Artificial Analysis with 58.6 points. According to the company, the model also leads on Legal Research Bench and Tax Agent Bench and demonstrated its lowest rate of hallucinated content to date in internal Antigravity tests.

In terms of list price, Neon is 50 percent more expensive than Claude Sonnet 5.5 and GPT-6.2 Sol, both of which charge $2 for input and $10 for output per million tokens. Compared to Opus 5.5 at $4 and $20, Neon is 25 percent cheaper. The comparison shifts with the calendar: Carbon costs $2 and $10 at an introductory price through November 30, making it cheaper per token than Neon. Starting December 1, Carbon rises to $4 and $20. At OpenAI, a long-context tier of $4 for input and $15 for output applies to the entire request once input exceeds 272,000 tokens.

Google itself emphasizes cost per task. At the default “High” tier, Neon reportedly costs 32 percent less per task than Carbon. Artificial Analysis estimates the advantage at around 17 percent with slightly higher intelligence. Thai Tran, AI Product Lead at Glean, stated in Google’s announcement that Neon matched or outperformed Carbon at three times the speed and 60 percent of the cost. Google did not disclose absolute processing times in seconds or minutes. → Google, VentureBeat, implicator, The New Stack, SiliconANGLE, The Deep View

Synthszr Take: Google’s own price sheet is the strongest argument against Carbon. According to Artificial Analysis, Neon performs slightly above Google’s own flagship model, and starting December 1, Carbon will cost $4 instead of $3 for input. Tiers like small, medium, and large therefore reveal very little about performance anymore. This hits Anthropic harder: Opus 5.5 demands the same price premium as Carbon, yet in two engineering benchmarks maintains only about a one-percentage-point lead over Neon.

Mistral builds AI models for industrial engineering

French AI provider Mistral announced on Tuesday that it is developing new models specifically for industrial engineering and other regulated industries, reports The Deep View. According to the company, these sectors will gain access to sovereign frontier models that they can fully control, own, and customize like open models. Mistral describes industrial engineering as the second wave of AI after software development. CEO Arthur Mensch explained the move by stating that engineering, like finance and the energy sector, must run on proprietary data, specialized models, and systems “that no one else can shut off.” The technical foundation is Mistral Large 4, nicknamed “Le Chonk,” which was trained in Mistral’s own European data centers and, according to the vendor, was capable of competing with the best closed models at launch. → The Deep View

Synthszr Take: “Systems that no one else can shut off” is aimed directly at OpenAI’s business model. A semiconductor manufacturer or aerospace supplier will not build its engineering pipeline on an API whose model version the vendor can retire at will. In certified processes, every model change necessitates a new qualification. Consequently, the first supplier stays for years. Mistral only needs to secure a place in simulation pipelines early enough. However, without a set timeline, OpenAI still has time to follow up with an on-premise offering.

Agent Sprawl Turns AI Adoption into an IT Problem

According to an analysis by Ksenia Se and Will Schenk in the Turing Post, employees in enterprises are now building agents faster than IT can track them. The catalyst is Microsoft’s Ignite conference, where Microsoft identified so-called Agent Sprawl as a growing source of shadow IT. As evidence, the authors cite FICO CIO Mike Trkay, who told the Wall Street Journal that employees at FICO create dozens of agents every day. His concern is that different agents provide conflicting answers to the same business question. GitLab CIO Manu Narayan told the same newspaper that he accepts short-term sprawl because of the insights that might emerge from it. → 🔳 Turing Post

Synthszr Take: Two teams, two definitions of “renewed”: this problem existed long before the first agent. Agents merely make it visible because they deliver contradictory figures in seconds rather than in a quarterly review. Microsoft framed this at Ignite as shadow IT—a control problem for IT. But the bottleneck lies in business terminology. Whether pilots ever turn into production deployments will be decided right there. No registry of agents can replace a clarified definition of customer.

Nadella Wants as Many AI Models as Companies, Licensing Microsoft’s MAI Weights to Do It

Microsoft CEO Satya Nadella wants enterprises to own their own AI models. According to Every, enterprise customers will soon be able to license the weights of Microsoft’s in-house MAI models and run them in their own Azure environments. Nadella said this during a panel discussion following his keynote at Microsoft Ignite. Customers will be able to customize the models using their own data via Fine-Tuning and decide for themselves when to upgrade to new versions. Nadella describes the goal as transforming the model business into a “database business.” → Every

Synthszr Take: One-fiftieth of the cost is the figure that will now be on the table in every procurement conversation with OpenAI and Anthropic. Nadella’s database comparison reveals the calculus: the customer pays for the license and runs the model for years on their own, on Azure. Every fine-tuned MAI model siphons volume away from standard API subscriptions and turns it into compute load in Microsoft’s cloud. Model providers lose both revenue and data in the process, since fine-tuning will take place on the customer’s end.

Nvidia Interim CEO Kress Pulls Forward Pay Raises to Stem Attrition

At her first all-hands meeting, Nvidia’s interim CEO Colette Kress announced salary increases and additional equity grants for employees in December. According to Wired, which obtained remarks from the meeting, this is several months earlier than the usual schedule. The announcement came in the opening minutes of the roughly 50-minute meeting, the first since Jensen Huang’s resignation the previous week. According to Kress, the most frequent of the 700 questions submitted in advance concerned compensation. In second place was the desire for greater board transparency. Board chair Harvey Jones reiterated that Huang’s resignation was accepted due to his personal conduct and lack of candor. → Wired

Synthszr Take: An expedited pay round is an expense Nvidia’s margins will barely feel. Things get expensive if the architects of the next chip generation defect to AMD, Broadcom, or the hyperscalers' custom silicon teams. So Kress is buying time until a new CEO is appointed. The catch lies in the equity grants: their value hinges on the share price, and the share price hinges on a succession process the board explicitly refuses to rush. Money only retains talent as long as the equity remains credible.

SparkToro Measures Purchase Intent in 6.9 Percent of ChatGPT Conversations

In 6.9 percent of ChatGPT conversations within a US panel, users show purchase intent. That is the finding of an analysis by SparkToro, reported by Search Engine Journal. The study is based on 41,961 conversations across 12,321 individuals between January 1 and October 29, 2026, provided by Verso, a service that analyzes connected ChatGPT chat histories. In 4.9 percent of conversations, users were planning a purchase and comparing options; in 2.1 percent, the topic was a specific brand, product, or service. In the demographically weighted sub-sample, 87.9 percent of ChatGPT responses mentioned not a single brand or company. → STACKED MARKETER

Synthszr Take: Purchase intent in 6.9 percent of conversations is plenty of raw material for an advertising business. What’s missing is the tradeable unit. Google sold the keyword—a cleanly delineated signal backed by an auction. A chat delivers a conversation thread, and 87.9 percent of the responses don’t even mention a brand. Before OpenAI can sell ads, it must define which moment in the dialogue gets auctioned off. Until then, that purchase intent sits idle, lacking the infrastructure to turn it into revenue.

FBI Warns: Open AI Agents Like Artex Breach Corporate Networks in Under Three Hours

According to Axios, freely available AI agent frameworks such as Artex and Cairn are giving attackers access to corporate networks within hours. The FBI reports intrusions using Artex that took less than three hours from initial reconnaissance of a target to privileged access. Amazon reports that attackers used Artex to steal credentials for AI models and harnessed those compute resources for further attacks. CrowdStrike says it observed a threat group using Artex to attack more than 200 organizations worldwide within eleven days. Palo Alto Networks attributes the second open-source framework, Cairn, to Chinese-speaking threat actors, who reportedly conducted 25 operations against at least 19 organizations between August and September. According to Axios, the attacker only needs to specify the target and objective. → Axios AI+

Synthszr Take: Artex funds its attacks with stolen model credentials, meaning the victim foots the compute bill. As a result, the marginal cost of a breach drops toward zero, and more than 200 organizations in eleven days is by no means a ceiling. Attackers now scale like software, while defenders still scale like headcount. Whether this is already having an impact will become apparent in the next cyber insurance claims reports. If the number of successful breaches rises measurably there, premiums will be recalculated before a CISO has even defended their budget.

Google DeepMind Opens Genie 4 for Real-Time World-Building at Nearly $200 a Month

Google DeepMind is rolling out Genie 4 to eligible Google AI Ultra subscribers. The new World Model has been running since November 18, 2026, via Project Genie, a browser-based experimental environment for interactive worlds. Genie 4 was originally unveiled in October, but access had previously been restricted to select testers. The system does not render a finished video from a prompt, but instead generates the scene continuously as the user explores it.

The core innovation is real-time editing. Users can type new prompts mid-session to change the weather, introduce a character, or redirect the action without restarting. Genie 4 is also the first version with native real-time audio. Ambient sounds like wind and footsteps are generated alongside the video, and character dialogue responds to events in the scene. According to Google, characters move more realistically, and objects respond better to physical forces. These claims are based on Google’s own demonstrations; independent tests are not yet available.

Access is more restricted than the announcement suggests. According to the Google One plan overview, Project Genie is limited to the Ultra 20x tier, which costs $199.99 a month in the US. Subscribers to the cheaper Ultra 5x tier are left out. Google AI Pro members can only join a waitlist.

Genie 3 remains available as an option and is also receiving updates. Sessions now last up to three minutes, up from the previous 60 seconds. Multiple worlds can be chained together for longer explorations. A 360-degree panorama can serve as a starting point to keep the world consistent. In-world events can be triggered using natural language.

Also new is a Project Genie agent built on Google’s multimodal models. Users describe their ideas in conversation, and the agent generates visuals and custom background music. Inside the world, Gemini monitors on-screen action, suggests next objectives, or triggers events. A community gallery features curated worlds and creations from other users that can be explored, remixed, and published. These include “Living Memory,” which turns personal photos into walkable spaces, the gothic mansion puzzle “Villa Giallo,” and the ant colony simulation “The Life of Ants,” created with filmmaker Anand Gandhi. → Decrypt, blockchain, MakeUseOf

Synthszr Take: A community gallery with remix capabilities is a platform promise, and platforms thrive on scale. Google is gating it behind the Ultra 20x tier for $199.99 a month. This keeps the circle of world-builders minuscule, and a gallery of ant colonies and mansion puzzles will grow slower than Roblox releases new games. The price tag protects scarce compute, but squanders the very network effect Google once used to make YouTube huge.

Washington Attributes Outage Wave at FAA, Chase, and NSA to Patches for AI-Discovered Vulnerabilities

Bank executives and the NSA told members of the US Congress that last week’s wave of outages stemmed from attempts to patch security vulnerabilities discovered by AI models. This was reported by three anonymous sources familiar with the confidential discussions. The patches reportedly overloaded the affected systems temporarily. One informed congressional aide said: “They are trying to push through a huge number of patches all at once, more than the system can handle.” This account has not yet been officially confirmed.

Between November 12 and 17, systems failed at six major organizations. According to Downdetector data, it was the week with the most outages in twelve months. The aviation authority FAA suffered the most severe damage: communications and radar failed at 17 of its 22 en route air traffic control centers, and the nationwide Ground Stop led to approximately 20,000 canceled flights. It was the longest nationwide ground stop since September 11, 2001. The FAA attributes the disruption to excessive connection requests during a router upgrade. It was able to replicate the issue in Arkansas, but does not yet know the root cause.

At Microsoft, more than 80 percent of network gateway servers became unresponsive during the global outage of Outlook and Teams. The company cited a software bug from a July update that surfaced under increased traffic and reduced capacity. At JPMorgan Chase, more than 4,700 incident reports were logged by 2 p.m. Eastern Time on Monday. Payments via Zelle were also delayed at other banks, and customers of Bank of America and Wells Fargo were similarly unable to access their accounts. Chase cited a software bug in a locally operated hardware system. An outage at United Airlines delayed around 150 flights.

On Tuesday, NSANet went down, the NSA’s internal network for transmitting classified cryptographic material. Non-essential staff were told to report to Fort Meade in staggered shifts. The NSA itself cited scheduled maintenance as the reason. In contrast, Intelligence Coordinator Jay Clayton stated that the network went down following a patch for PHP vulnerabilities discovered the previous month using AI. On the same day, Apple reported outages affecting the App Store, Apple Music, iCloud Mail, and AppleCare, pointing to an internal update.

Clayton ruled out a coordinated cyberattack or state-sponsored activity. Secretary of State Marco Rubio said there is “zero evidence” that the incidents are connected or stem from hostile actors. Democratic Senator Mark Warner spoke of tremendous pressure to patch AI-discovered vulnerabilities immediately and often simultaneously, on systems not built for that speed. Homeland Security Secretary Markwayne Mullin said new technology is being layered onto very old systems, overloading them. → Fox News, The Washington Sun, Fox Business, Intelligence Report, CNBC, Townhall, implicator

Synthszr Take: A single faulty CrowdStrike patch was once enough to paralyze airports worldwide, and the industry learned its lesson: roll out in stages, test beforehand. This week shows how quickly that rule falls apart under pressure. AI discovers vulnerabilities faster, and attackers have comparable models. So banks and the NSA patch simultaneously on systems never built for it. The bottleneck has shifted from finding flaws to safely deploying fixes. Staging environments and rollback capabilities therefore belong in the security budget.

Mentioned in this article

Subscribe free. Unsubscribe the second it sucks.

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