Most weeks in tech are a list of unrelated launches. This one is a pattern. The company that supplies the chips behind the AI boom lost its CEO overnight. One of the biggest AI labs is days from a public listing. And AI agents have started doing real work in two places where mistakes are expensive: spending money and breaking into networks.

I’ve written this for developers, not investors, so each story ends with what it means for people who ship software. Everything here is current as of November 16, 2026, and a few items have deadlines this week, which I’ve flagged.

1. Nvidia’s founder resigns

The biggest shock first. Jensen Huang resigned as Nvidia’s chief executive on Thursday, after revelations that he’d had two extramarital affairs, one of them with a woman accused of spying for China. That woman is reported to be an executive at Megaspeed, a Singapore-based data company and Nvidia customer. CFO Colette Kress is serving as interim CEO while the board searches for a permanent successor, ending a 33-year run for the co-founder. The board has said it will consider both internal and external candidates, and it named Harvey Jones, formerly of Synopsys, as the company’s first-ever board chair.

The market reaction was immediate. Nvidia shares fell about 10% in premarket trading after the announcement, on top of a decline that had already erased roughly $600 billion in value. Shareholders have also filed a class action against Huang and the company, alleging that the relationship wasn’t disclosed to investors. The politics are heating up too: a CNBC headline from Sunday reports that Treasury Secretary Bessent is calling for a dramatic shakeup and an outside CEO, while other headlines on the same page report growing calls in the tech world for Huang to come back.

Date to watch: Nvidia said it would still release its quarterly results as scheduled on November 17. Tomorrow’s earnings call will be the first time investors hear management speak without its founder.

What it means for developers: Nvidia’s roadmap, CUDA ecosystem and supply commitments don’t vanish because the CEO does. But a company this central to AI infrastructure now has a leadership vacuum, an internal investigation and a legal fight. If your product or budget depends on one vendor’s GPUs, this is a good week to ask what your second option is, even if you never use it.

2. Anthropic’s IPO is days away

The other big business story is the listing everyone has been waiting for. Anthropic filed its S-1 on October 26 and applied to list on the Nasdaq Global Select Market under the ticker ANT, with the price range and share count left blank. Since then the terms have arrived. An amended filing on November 10 set the offering at 825 million Class A shares at an estimated $70 to $78 each, alongside a $26 billion private placement at the IPO price. That works out to a fully diluted valuation of roughly $1.26 trillion to $1.41 trillion.

The numbers behind it are huge. Anthropic reported $35.3 billion in revenue for the twelve months ended September 30, 2026. Google, Nvidia and TSMC are among the cornerstone investors.

Dates to watch: Pricing is expected on the evening of November 18, with first-day trading on the Nasdaq on November 19. The timing is awkward. The Wall Street Journal reported that Anthropic planned to press ahead despite a tech selloff. And the Nvidia scandal landed a week before the IPO, with Nvidia reported to have planned a $10 billion investment in the company. OpenAI is reportedly targeting its own listing in 2027.

What it means for developers: A public listing changes how a vendor behaves. Expect more disclosure about revenue, risk and customer concentration, and more pressure on margins, which tends to show up as pricing changes and tighter usage limits. If you build on any AI provider’s API, treat model prices as a variable, abstract your provider behind a thin interface, and keep an eye on the S-1 risk factors, which are surprisingly readable.

(Transparency note for my own readers: I used an AI assistant to help research and draft this post, and one of the companies discussed above makes that assistant. I’ve tried to report the IPO details the same way I’d report any other company’s.)

3. The model race: Google takes the benchmark lead

While the business drama played out, the models kept moving. Google announced Gemini 4 Carbon on October 28. On the Artificial Analysis Intelligence Index it scores 59, a single point ahead of Claude Opus 5.5 and two ahead of OpenAI’s GPT-6 Astra. Google priced it at $2 per million input tokens and $10 per million output tokens, undercutting OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.2 even after prices double at the end of November. The introductory period ends on November 30, after which the price becomes $4 and $20 per million input and output tokens. If you’re budgeting for December, plan on the higher number.

Two caveats matter more than the headline score.

First, it isn’t generally available. As of late October, Google described an initial rollout to selected cyber defenders, followed by trusted partners, with no announcement of general access. That staged approach reflects where this generation is heading: the strongest models are being released as security tools first.

Second, benchmarks disagree. A separate SemiAnalysis ranking puts Claude Opus 5.5 first, with Gemini 4 Carbon second and GPT-6 Astra third. And as one outlet put it carefully, these are Google’s published results, and benchmark scores alone don’t guarantee a model will find every vulnerability or produce a correct fix in real-world systems. Different leaderboards measure different things, so any “number one” headline deserves a second look.

The other labs aren’t standing still. OpenAI announced GPT-6 Cyber on November 5, a model focused on identifying and validating software vulnerabilities and helping to fix them. OpenAI also said its GPT-6.1 Terra meets the Critical threshold in cybersecurity under its Preparedness Framework. The common thread is that every major lab is now shipping cyber-capable models with restricted access.

What it means for developers: Don’t pick a model from a leaderboard. Build a small evaluation set from your own tasks, run it against two or three candidates, and re-run it when prices or versions change. The model that wins on a public index and the one that’s cheapest for your workload are often different.

4. Agents get a wallet

The most practical launch of the week is about money. OpenAI launched ChatGPT Wallet on Friday, which lets AI agents pay for purchases without accessing the user’s card details, with a security check required on every transaction. The details:

  • It’s available to US-based ChatGPT Pro subscribers who hold a US-issued Visa, Mastercard or American Express card.
  • Agents use card tokens from those networks, with biometric approval on each purchase.
  • OpenAI also released a Commerce Partner Playbook to help businesses prepare their sites and payment systems for purchases initiated by agents.
  • OpenAI isn’t taking a cut of transactions at launch and is focused on building merchant relationships.
  • Plans include eventually allowing preauthorized transactions.

That last point is the one to watch. A per-purchase biometric check is a safe starting design, because a human stays in the loop. Preauthorization removes that loop, and the security question changes from “can the agent pay?” to “what stops it paying for the wrong thing?”

Elsewhere in the agent world, a weekly digest citing Google DeepMind reports that its Gemini Robotics 2.5 and a proof of concept with Mercedes-Benz and Apptronik put Apollo 2 humanoid robots on active production lines in Hungary and Germany, picking brackets from bins and packing them into sequencing containers, with a reported 96% success rate. That’s a vendor-reported figure from a proof of concept, so treat it as a promising signal and not a production guarantee.

What it means for developers: If you run an online store, an agent is about to become a new kind of customer. Clean product data, predictable checkout flows and stable identifiers will matter more than clever UI. If you build agents, design the approval step first, not last.

5. Security: AI agents are now attackers

This is the story I’d read twice. The FBI issued a FLASH alert on November 13 warning that criminals are weaponizing ARTEX, an autonomous penetration-testing framework, to accelerate cyberattacks worldwide.

ARTEX was built for authorized security testing. It can turn high-level goals written in plain language into machine commands executed by AI agents, with minimal human involvement beyond the initial objective. The FBI’s concerns are specific:

  • Actors have used the open-source framework since at least August 2026 against financial, government, technology, defence and healthcare organisations, in some cases going from first targeting to privileged access in under three hours.
  • The FBI said the framework lets actors with minimal technical expertise run complex, multi-target operations at the same time.
  • The original developer released it on GitHub in July and pulled it on October 8 after reports of abuse, but the FBI says the withdrawal didn’t stop its spread through forks, archived copies and a commercial edition.
  • Researchers found more than 1,600 publicly exposed ARTEX console instances in early November.

Amazon added a detail that sounds like science fiction. Its threat intelligence team says attackers used ARTEX agents for about a month to steal API keys from exposed AI apps such as LiteLLM and Dify, then verified and pooled the stolen keys to pay for more LLM requests. In other words, the agents funded their own hunting with stolen credentials. Amazon says the targets included LLM gateways, proxies and AI app builders, with the dominant campaign sustaining itself this way.

The usual patching treadmill continues in the background too. Microsoft’s November Patch Tuesday includes an actively exploited Outlook spoofing flaw that can expose NTLMv2 hashes, and Chrome 157 fixes a zero-day that was already under attack. Separately, TSMC confirmed employee data leaked through a cloud provider, and Maersk confirmed a breach of an inherited legacy system.

What it means for developers: If you’ve deployed any self-hosted AI tooling, such as a gateway, a chat UI, a workflow builder or a local model server, check what’s reachable from the internet today. A practical checklist:

  1. Inventory exposed services. Scan your own public IPs and domains for AI dashboards and admin consoles that were deployed for a demo and forgotten.
  2. Put authentication in front of everything. No AI tool should be reachable without a login.
  3. Treat API keys as leakable. Use scoped, per-project keys with spending caps, store them in a secrets manager, and make rotation a one-command job.
  4. Set billing alerts. A sudden spike in token spend is often your first sign of a stolen key.
  5. Patch the boring things. Update Chrome and apply the Outlook fix. Automated attackers go after known holes first.

6. Quick hits

A few smaller stories worth knowing:

  • Open weights keep coming. Runway released Praxis-1 as Apache 2.0 open weights on November 11. A permissive license means you can use it commercially without negotiating terms, which is the part most people overlook.
  • Science news. DeepMind introduced AlphaGenome 2 on November 11, which makes variant-effect prediction conditional on cell type. Another AI-powered study this week found a hidden signal in routine heart scans that could predict sudden cardiac death, outperforming conventional screening. In quantum computing, researchers reported a machine that stores information in microscopic vibrations. These are research results, not products, but they show where the money and talent are going.
  • Chips. Infineon announced functional silicon-carbide power devices processed on 300-millimeter wafers on November 9. That’s a manufacturing milestone, and it matters because power electronics decide how efficient data centers and EVs can get.
  • Regulation. One newsletter reports that New York’s AI safety law has taken effect, bringing new oversight requirements for companies including OpenAI and Anthropic. Elsewhere, Mexico’s Digital Economy Law took effect on November 5. If you serve users in either place, have someone read the actual text.
  • Meetings. Otter AI added support for more than 30 languages. That’s relevant if your team spans regions, as many African dev teams do. What to watch this week Date Event Nov 17 Nvidia reports quarterly earnings, its first without Huang Nov 18 (evening) Anthropic’s IPO is expected to price Nov 19 Anthropic’s expected first day of trading as ANT Nov 30 Gemini 4 Carbon’s introductory pricing ends The bigger picture

If I zoom out, three trends connect all of this:

  1. AI is now a capital-markets story. Trillion-dollar valuations, founder exits and class actions mean AI companies are being judged like any other giant public business, with the volatility that comes with it.
  2. Capable models ship to defenders first. Staged access for cyber models acknowledges that the same capability that finds bugs can exploit them.
  3. Agents moved from demos to consequences. They can now spend your money and attack your servers. The engineering challenge is no longer “can it do the task?” but “can we constrain, observe and undo what it did?”

For those of us who build things, the takeaway is mostly boring and valuable: keep your dependencies replaceable, keep your secrets rotatable, keep a human approval step wherever money or access is involved, and keep your own evaluations instead of trusting someone else’s leaderboard.

What story from this week matters most for your work? Tell me in the comments.