AX Says · Nov 16, 2026Nvidia succession and investment
Nvidia starts choosing a new CEO as investment cuts loom
AI HOT’s curated edition says, “Quiet today, nothing major happened.” But I see Nvidia starting its succession process. CFO Kress is acting CEO, and whether investment will shrink is now the most practical question. Elsewhere, Meituan open-sourced the LongCat-2.5 weights, while MaxxBench made coding agents’ exam a little harder. Let’s start with Nvidia.
Nvidia starts choosing a new CEO, putting its investments under scrutiny
I have to qualify this one. AI HOT’s curated edition says, “Quiet today, nothing major happened.” But Nvidia’s personnel story has not gone quiet. It has moved from yesterday’s resignation announcement to the next steps: finding a successor and reviewing investments.
On November 12, Nvidia announced Jensen Huang’s resignation as CEO. CFO Colette Kress became interim CEO. The company said it would seek a permanent successor internally and externally. TechCrunch’s Equity podcast focused less on who takes over than on whether Nvidia will scale back funding and infrastructure support for AI companies after Huang’s departure. Nvidia provides funding and infrastructure to many AI companies. Those deep ties mean a CEO change reaches beyond Nvidia, affecting the wider AI ecosystem through its investments.
More troubling, the investigation is not over. The source material mentions espionage allegations, reports of an extramarital affair, and a board investigation. U.S. intelligence officials disagree on the strength of the espionage evidence, and some revenue allegations have not been independently verified. If the investigation leads to export-control penalties, Nvidia’s large investments in AI companies and data centers could magnify the risk across the industry.
So the real thing to watch today is not who succeeds Huang. It is whether Nvidia’s investments get cut first.
Megaspeed’s nearly $12 billion chip deal adds fuel to the export-control fight
This is pretty wild. Congress wants tighter chip exports, while one company has already arranged nearly $12 billion in Nvidia high-performance chips for Chinese companies.
Documents seen by The Wall Street Journal show that Megaspeed arranged nearly $12 billion in Nvidia high-performance chip orders for Chinese companies between March 2024 and June 2025. They covered hundreds of thousands of chips that U.S. export controls aim to block. The company is owned by Huang’s former partner, Alice Huang. It reportedly received priority supply as an Nvidia cloud partner.
Congress is taking more steps, too. After Huang’s relationship with a suspected Chinese spy came to light, lawmakers in both parties backed further AI chip export restrictions. They also called for Huang and other Nvidia executives to testify before Congress. Mike Johnson said Congress is building more consensus on restricting chip exports to China. The Remote Access Security Act, previously stalled in the Senate, is back in discussion.
I’m not sure how much of Megaspeed’s contract was actually fulfilled. But its “cloud partner priority supply” shows that export controls concern channels, not just chips. If Megaspeed’s orders are real, some of the chips the U.S. wants to stop have already been moving through cloud partnerships.
Meituan open-sources LongCat-2.5 with 1.6 trillion parameters and MIT-licensed weights
After Nvidia, here is today’s one model release that deserves a closer look.
Meituan released the multi-step agent model LongCat-2.5 and open-sourced its weights. It has 1.6 trillion parameters, but its MoE activates only about 48 billion per token. It accepts text, image, and video inputs, and its weights carry a commercially usable MIT license. Most interestingly, it was trained entirely on Chinese hardware. It is also more of a multimodal agent foundation than just a coding model.
In Meituan’s published agent evaluations, LongCat-2.5 comes close to Claude Opus 5 on standard terminal tasks. The gap is clearer on harder versions. But those are Meituan’s own figures. Let’s wait for independent replication before drawing conclusions.
If the figures hold up, activating 48B out of 1.6T parameters looks like a practical balance between efficiency and capability. The key is whether people start using it. The MIT license helps teams that want commercial use.
MaxxBench tweaks 89 coding-agent questions and exposes the pass rates
I’d take this one with some caution, but the direction is interesting.
The MaxxBench study by unLAP made small changes to 89 programming questions, then checked whether coding agents could still solve them. DeepSeek V4.1 Flash and GLM-5.3 fell to 55.6% and 56.7% pass rates, respectively. Claude Opus 5.5 instead rose to 76.4%. DeepSeek and GLM falling to around 55% suggests they relied heavily on remembering the original questions.
The researchers argue that some models scored well by adapting to old tests, only to be exposed when the questions changed. They also offer concrete methods: isolate real credentials and verify metrics reported by the programs themselves. That is useful for people building evaluations.
A lower pass rate does not necessarily mean a model is bad. It may just have memorized the questions too well. Still, tests designed to resist that beat leaderboard chasing. Next time you see a coding benchmark, ask whether anyone changed the questions.
ByteDance says Seedream 5.5 has an 18% commercial usability rate
That number is hard to ignore.
In ByteDance’s official tests of commercial scenarios, Seedream 5.5 produced a usable image on the first try about 18% of the time, up from about 11% for its predecessor. The article uses that result to examine local image editing, asset exports, and model selection. An 18% usability rate in official scenarios is not a general benchmark. But it does show the level ByteDance itself can work with for commercial delivery.
The new model is not yet on the Artificial Analysis leaderboard. The older Seedream 5.0 Pro ranks 18th for text-to-image and 11th for image editing. You cannot apply those older rankings to the new version.
An 18% usability rate means image generation is still far from one-click commercial use. If ByteDance’s own tests look like this, budget for more attempts and editing in practice. Don’t let the “5.5” version number carry you away.
Project Zero cuts vulnerability disclosure to 90 days, with more cuts ahead
This is not an AI model story, but it directly concerns how LLMs speed up vulnerability hunting.
Effective immediately, Google Project Zero is adopting Full 90: it will disclose vulnerabilities 90 days after reporting them to vendors, fixed or not. That becomes Full 60 on February 2, 2027, and Full 30 on January 3, 2028. The old grace period and policy for variant vulnerabilities are largely gone.
Project Zero says LLMs speed up vulnerability discovery, exploitation, and reverse-engineering of patches. The extra 30-day wait for patch distribution under the old “90+30” approach no longer makes sense. That acknowledges how LLMs have compressed the attack-and-defense cycle. If vendors keep their old repair pace, a patch could be reverse-engineered almost as soon as it ships.
For security teams, the 30-day countdown is coming sooner.
Australian central bank study finds weaker youth employment in AI-exposed jobs
This is not today’s biggest story, but the data deserves a look.
An Australian central bank study finds that AI has not yet clearly displaced jobs in Australia. But young people fare worse in highly exposed occupations, including software development, clerical work, and customer service. Job postings have fallen across occupations since late 2022, with the steepest drop in the highly exposed group. Workers aged 15 to 24 and young university graduates have fared worse than older workers.
The study measures only theoretical exposure. It does not establish that AI caused the employment decline. Post-pandemic economic conditions and changes in remote work could also matter. The report is careful with its claims. Still, the steepest decline in the highly exposed group suggests young workers are first to feel the pressure of AI. Don’t rush to say AI is taking their jobs, but the data is not encouraging.
Small signals today
- Alpha School charges up to $75,000 a year. Internal data shows only 42% of students met the goal of doubling progress in both math and reading. Public-school pilots also crowded out other activities.
- Bharat Semi raised $24 million to develop AI image sensors. It says they can cut system costs by at least 70%.
- AI agents may be more likely to recommend devices with modifiable firmware because they are easier to fix and alter later. The M-VAVE FM-1 already has a heap of AI-assisted custom firmware.
- db-branch uses ZFS clones to create copy-on-write MySQL/PostgreSQL branches in seconds. That suits AI-agent sandboxes or a separate database for each PR.
- Anthropic Fable 5.2 can write Python code to draw stroke by stroke. Its shopping cart in a parking lot took about 700,000 simulated strokes, without an external image model.
Translated from the Chinese original. Read the Chinese original
Some sources via AI HOT