Anthropic opens its IPO roadshow as Nvidia's board reviews Jensen Huang

Anthropic sets a price range for its Nasdaq listing

Anthropic formally started marketing its initial public offering on 10 November (US time). It plans to sell 825 million Class A shares at between US$70 and US$78 each, and it has applied to trade on the Nasdaq Global Select Market as "ANT". The company itself is issuing 801,351,354 of those shares. Existing holders are selling the remaining 23,648,646, and Anthropic won't receive any money from that portion.

Underwriters may get a 30-day option to buy up to 123.75 million more shares. Goldman Sachs and Morgan Stanley are the contacts for the preliminary prospectus. The registration statement went to the SEC on 26 October.

What it means for NZ organisations: Once Anthropic is listed, it'll have to publish financial results on a regular schedule. Procurement teams that rely on Claude will see much more of their supplier's finances.

SoftBank posts a ¥1.1 trillion quarterly loss

SoftBank released its July to September results on 10 November. It lost ¥1.1 trillion, about US$7 billion, compared with a ¥2.5 trillion profit in the same quarter a year before. After plans changed, it booked a US$3.7 billion loss on advance payments for US data centre equipment, which included a site in Pike County, Ohio. Financing costs more than doubled to ¥462 billion. The shares rose 14% before the figures came out, but they're still down by almost half since early June.

What it means for NZ organisations: Even the biggest backers of AI infrastructure are changing their build plans partway through. Keep capacity agreements short and make sure you can get out of them.

Nvidia's independent directors review reports about Huang

On 8 November, Nvidia's board said a committee of independent directors would investigate media reports that founder and CEO Jensen Huang had been the target of a foreign intelligence effort. Outside counsel is advising the committee. The directors said they had no reason to think national security or company secrets were compromised, and that the inquiry would still look at all the information available. Huang supports the review and has agreed to cooperate. The board said it has confidence in him and his leadership team.

What it means for NZ organisations: Most cloud AI capacity runs on Nvidia hardware. If you hold large GPU reservations, follow the board's updates for any sign that the leadership questions affect supply.

ChatGPT Enterprise lands on AWS Marketplace

Since 9 November, organisations can buy OpenAI's ChatGPT Enterprise through AWS Marketplace and pay for it on their AWS bill. The listing bundles Chat, ChatGPT Work, Codex and dots. Eligible buyers can also get ChatGPT for Healthcare and ChatGPT for Financial Services there. Prices come only through private offers. Qualifying spend can count toward an AWS EDP or PPA commitment, depending on each customer's agreement with AWS.

What it means for NZ organisations: If you've already committed spend with AWS, ask your account team whether ChatGPT seats can count against it before you sign a separate licence.

Fireworks AI discloses another breach

Inference provider Fireworks AI has disclosed a second incident. In the first, in early October, attackers got into environment variables that held API keys and tokens for some accounts. In the second, a third-party cloud provider that hosts Fireworks' R&D machines was compromised on 23 October. On 24 October the intruder used that access to get onto Fireworks development machines and run a workload there. Fireworks detected it the same day.

The exposed data includes resource identifiers, names and descriptions, billing and shipping addresses, invoices, payment history, balances and credit usage. For some customers it also includes incomplete model weights and datasets. No card numbers were taken because Fireworks doesn't store them. Palo Alto Networks' Unit 42 is helping with the investigation.

What it means for NZ organisations: Rotate any credentials stored on third-party AI platforms and review their usage logs. Ask your hosting providers which subcontractors can reach your training data.

Meta prices Muse Ember for agent workloads

Meta launched Muse Ember, a low-cost model for agents, on 9 November. It costs US$0.10 per million input tokens, US$0.50 per million output tokens and US$0.004 per million cached input tokens. It handles text, images, audio and video, has a 1 million token context window with up to 512,000 output tokens, and is available as muse-ember-1.0 through Muse Code and the Meta Model API. Meta says it matches Claude Haiku 5.5's best result for about a third of the cost. That's Meta's own figure.

What it means for NZ organisations: Routing, extraction and tool-calling steps could get cheaper. Test it on your own data before you switch.

Google extends its TPU 8 family

Google announced two more eighth-generation chips on 9 November. The TPU 8tx works with AMD EPYC processors and keeps data and model weights isolated in hardware while they're being processed. Google says its latency is half that of the TPU 8t. The TPU 8i Max is built for very large context windows and has double the SRAM and HBM bandwidth of the TPU 8i. Google hasn't given availability dates.

What it means for NZ organisations: Health, finance and public sector teams should ask Google Cloud when this confidential inference hardware will reach a region they can use.

Anthropic ships Fable 5.2 and Mythos 5.2, then Claude Meetings

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. It calls them its most capable models for demanding technical work and says they improve on coding, finance and science.

The next day, Claude Meetings went into beta for Team and Enterprise customers, starting with Google Meet. Claude shows up in the call as a visible participant and announces itself in the chat. While people talk, it can draft plans, sketch timelines on a whiteboard, answer questions in the chat or pull information from connected apps. Notes and action items appear in claude.ai after the call. It won't join calls with external guests unless an admin has approved it, and it leaves if an external guest joins.

What it means for NZ organisations: Test the new models on your own workloads before moving production traffic to them. Check any meeting assistant pilot against your consent practices and the Privacy Act 2020.