Skip to content

AX Says · Nov 19, 2026IPO and Leadership

Anthropic's $83 IPO price values it at $1.5 trillion.

Today's biggest number isn't a model benchmark score. It's Anthropic's IPO price: $83 a share and a fully diluted valuation of $1.5 trillion. That's above its May funding valuation, but below the previously rumored $2 trillion target. Nvidia has plenty going on too. Jensen Huang has just left, H200 sales to China have stopped, and profits still hit a record. Google, meanwhile, has released Gemini 4 Neon to tackle complex workflows. Let's start with the IPO.

Anthropic prices its IPO at $83, short of its $2 trillion ambition

I checked the math twice: Anthropic's IPO price isn't the previously discussed $70 to $78. It's $83, putting its fully diluted valuation at $1.5 trillion. Its May funding round valued it at just $965 billion, so the IPO price is far higher. But the rumored target a few months ago was $2 trillion. This number is above the last valuation, below the story.

According to the offering documents, the IPO includes 980 million Class A common shares. The company is selling 956,351,354 shares; existing shareholders are cashing out on the remaining roughly 23,648,646 shares. Anthropic gets none of that money. Its ticker is ANT, and trading starts today on Nasdaq.

The high price suggests solid demand. Falling short of $2 trillion also suggests public-market investors are discounting Anthropic's "safety and research" story. How much it loses each year, and when Claude can turn a profit, remain the real questions.

OpenAI's GPT-6.2 is cheaper, but it isn't number one

OpenAI released GPT-6.2 Astra and Sol today. Astra scored 62 on the Artificial Analysis Intelligence Index, 3 points behind first-place Fable 5.2. Early testers are split on it too.

The thrifty one is Sol. At about 15% of the per-task cost, it matches the capabilities of September's Astra. Both models offer 1M context and 128K output. The value gap between flagship and workhorse is no longer just a marketing claim about lower prices. The benchmarks show it.

Still, I'm discounting this one for now. The gap between 62 and 65 isn't small enough to ignore. OpenAI hasn't released the strongest model this time. It has put something it can reliably deliver on the shelf. Cheaper Sol looks built for enterprises running tasks at scale, not for making individual users scream with excitement.

Gemini 4 Neon targets complex workflows, not just benchmark scores

Google released Gemini 4 Neon today. The name sounds like a new phone model. In fact, it's aimed at complex software engineering, knowledge work, and scientific workflows. Developers, enterprises, and paid subscribers can all use it.

It scored 61 on Artificial Analysis's Intelligence Index, placing third. It ranks second on the Vals Index. Its 1336 points on the Agent Arena Index put it second worldwide. More practical is the cost: at the default high reasoning level, it costs 32% less per task than Carbon.

That same day, Google opened up Project Genie and offered Genie 4 for the first time. Users can change worlds with prompts while exploring and generate sounds that change with the scene. Genie 3's single-exploration limit also grew from 60 seconds to 3 minutes. It now supports connected worlds, 360-degree panoramic input, and real-time events.

My read is that Google doesn't just want to chase OpenAI on benchmark tables. It's pairing Neon and Genie for longer tasks where users can change the environment as they go. Whether that works is a question for after developers have spent a quarter with them.

Nvidia's first earnings after the leadership change break records, but Huang isn't coming back

Nvidia reported fiscal 2027 third-quarter revenue of $115.8 billion, up 103% year over year and above its previous $108 billion guidance. Interim CEO Colette Kress expects fourth-quarter revenue of $130 billion, plus or minus 3%. She also raised expected fiscal 2028 revenue growth from 70% to about 75%.

The Verge's timeline makes the scale clearer. Quarterly profit peaked at just $1.24 billion during the crypto boom and reached only $3 billion during the GPU shortage. The latest quarter brought $65.3 billion. The investment figures are even wilder. The value of its private-company holdings rose from $22 billion in January to $65.5 billion. It spent $16 billion buying stakes in public and private companies last quarter, and $58 billion in the first nine months of this fiscal year.

The personnel story looks entirely different. More than 200,000 people signed a petition supporting Huang's return, and Anduril's Palmer Luckey called for it publicly. But executives said on the earnings call that the company has no plan to reappoint Huang, even if he asks to return. OpenAI's Sam Altman said he couldn't imagine the AI industry without Huang. Greg Brockman said he looked forward to continuing their work together. Microsoft's Nadella expressed confidence in Kress's team.

The signal I see: Nvidia's results remain strong, but the handover is already affecting its approach to future investments, chip exports, and ecosystem ties. Semafor reports that some directors are still pushing for Huang's return. The candidate list has narrowed to Arm's Rene Haas and Marvell's Matt Murphy. This isn't settled. It's still playing out.

Nvidia stops H200 sales to China and tightens terms for cloud providers

The same day, Nvidia confirmed it had stopped delivering licensed Hopper data-center products to China, including H200 AI chips. The White House had approved licenses last December. Interim CEO Kress announced the decision on the earnings call. She gave no reason for the halt and didn't say whether Nvidia would continue doing business in China. About 600,000 H200 chips are stuck in Hong Kong because Beijing has not approved large-scale imports.

Kress also clarified when cloud providers qualify for compute purchase guarantees. New cloud service providers must first secure funding, deploy the compute capacity, begin operations, and pass acceptance checks. Only then does Nvidia's purchase commitment take effect. Construction delays void the guarantee. She said these terms aren't new; they've always been in the contracts.

Together, these moves make Nvidia's current style clear. Under Huang, it spent widely and worked to keep exports moving. Now it wants to see delivery for every dollar and reassess every market. China has gone from "restricted, but still selling" to "licensed, but not selling." The next step may be drawing a firm boundary, not negotiating trade terms.

Microsoft releases seven models, with a discount behind the 40% price

Microsoft released seven in-house models at Ignite, covering coding, security, images, speech, transcription, and decision routing. It has also started replacing outside models with its own in its products.

The price-fighter is MAI Code 2. Microsoft says its average performance across four benchmarks reaches 98% of GPT-6.1 Sol's, at just 40% of the cost. But there's fine print: that price includes Microsoft's own 50% promotional discount, which lasts until May 17, 2027.

That "98% / 40%" claim needs a discount of its own. Whether the price holds after the promotion ends will determine if Microsoft's models can really take share from OpenAI. For now, it looks more like a low-price push to bring developers and enterprises into the MAI ecosystem. Once they're used to it, Microsoft can sit down to discuss renewals.

Muse reportedly reads private messages unexpectedly, and NGINX has remote-code flaws

Meta's Muse recently stumbled in internal employee testing. Internal chats obtained by Reuters show Muse unexpectedly reading private messages, apparently mixing account context, and even outputting sensitive financial information. Meta says the account mix-ups were model hallucinations and the issue has been addressed. Reuters couldn't confirm whether the bank balance information was real user data or made up by the model.

The issue has drawn in Apple and Google too. After Reuters retested it, both companies began considering changes to permission controls. Soon, Meta won't be the only one deciding whether an agent can see your calendar, private messages, and bank information.

Separately, Wiz Research and other teams independently found two unauthenticated remote-code-execution vulnerabilities in NGINX: CVE-2026-123633 and CVE-2026-123007. Both scored 9.2 under CVSS 4.0. The risk comes from the server's own HTTP/2 handling, not a badly written application.

OpenAI also published an outline of its frontier-model risk management that day. It sets out four layers from training through public deployment. Research-stage models have no internet access; internal and external deployments progressively add alignment measures and monitoring. If evidence is insufficient, OpenAI will pause training or delay release. Put these together, and today's theme is clear: AI can do the work, but permission boundaries and code flaws are lining up to be exposed.

Small signals today

  • Anthropic reviewed the chaos testing before Claude Docs launched. Claude Code continuously stress-tested staging; across 2 billion requests, the service saw just 60 server errors and no rollback.
  • Besides the Microsoft and Greg Brockman news, there's another chip item: in under five months, OpenAI models helped improve PPA for AMD's flagship GPU IP by up to 14%. That's equivalent to about an extra year of progress. Some results have been validated in silicon; OpenAI is also working with Synopsys.
  • OpenAI detailed the Codex sandbox. It uses operating-system mechanisms to restrict file-system and network access. The built-in workspace allows writing within the workspace and broad read access, but blocks internet access by default. It may still read sensitive files or change local dependencies. #[17]
  • Horizon Labs agents based on Opus 5.5 autonomously created 38 models in seven weeks. They received 54k downloads and were used by 143 GitHub repositories.
  • Citi, JPMorgan Chase, Bank of America, PNC, Citizens, and other banks changed their terms. Customers bear account losses caused by third-party AI agents' actions or leaked credentials.

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks