Four days after Jensen Huang resigned, the people who want him back have a war room and a petition. And the board has an Arm CEO on the phone. This is Lantern AI Daily Briefing for Monday, November 16th. Today: the campaign to return Huang to Nvidia, a contract inside Anthropic's IPO filing that either side can now cancel, the arithmetic behind AMD's thirty percent claim against Vera Rubin, and llama.cpp going one point oh. One tap on follow, and we'll be back in your ears before you know it.
From the AI Daily Insights newsletter, summarizing a Reuters report that relayed the Financial Times:
Reuters, relaying the Financial Times, reported on November 15 that Huang had discussed returning as Nvidia CEO with confidants. An ally said supporters were assembling a “war room” of Silicon Valley executives to coordinate lobbying of Nvidia’s board and seek White House backing. Huang supports the effort. Nvidia has not announced a decision to reinstate him.
That was Sunday. Friday's episode covered the resignation itself. Here's what moved since then. The public backers named so far include Elon Musk and Epic Games CEO Tim Sweeney. Business Insider counted more than a hundred and eight thousand names on a Stand with Jensen petition by Saturday afternoon, and the Guardian put it past a hundred fifty-eight thousand on Monday. Business Insider also found obviously suspicious names on the list, which organizers said they were removing. A petition with fake signatures doesn't pick a CEO. The board does. Forbes says Huang holds about three point six percent of Nvidia, with roughly the same voting power. No super-voting shares. He can't overrule anyone. And the pressure runs both ways. The newsletter says the board has approached Arm CEO Rene Haas as a possible successor, though nobody has been appointed. The Next Web reports Treasury Secretary Scott Bessent wants an outside CEO. The independent inquiry, run by Wachtell Lipton, is still going. Two big shareholders, Brad Gerstner of Altimeter and Gavin Baker of Atreides, wrote to directors asking for a slow, deliberate process that keeps a return on the table. Per SBS, they did not demand he come back immediately. The builder angle is the checkbook. Crunchbase counts a hundred twenty-one startup financings with Nvidia or NVentures in 2026, up from ninety-one last year. And Reuters reported that Kress, now interim CEO, had previously challenged Huang's big AI investment and financing deals. If you're counting on a follow-on check from Nvidia, assume it now needs one more signature. Nathan Lambert, in his Interconnects essay Saturday, called the exit a setback for US open models and worries about funding drying up. That's his forecast; no project cuts have been announced. The first hard data point comes tomorrow, Tuesday, when Nvidia reports earnings with the leadership question still open.
Muhammad Zulhusni, writing in CloudComputing News:
Anthropic has increased its cloud computing commitments with Microsoft by up to $4.3 billion, bringing the total value of its agreement to between $47.2 billion and $49.2 billion.
That comes from the amended filing Anthropic submitted on Tuesday the 10th, the same day it launched the roadshow we covered last Wednesday. The minimum Microsoft commitment rises from forty-four point nine billion to forty-seven point two, running through June 2033, and it can't be cancelled outside narrow exceptions. The other hyperscaler minimums: at least a hundred twenty-one point three billion with Google, and a hundred ten billion with Amazon. The Microsoft number is the headline, but the Lambda disclosure is the interesting part. Anthropic signed three deals with Lambda this year, about forty-one point three billion dollars in total. The August one, thirty-four point nine billion, let either side walk if Lambda didn't secure a facility lease by a deadline. Lambda missed it. So that contract is now terminable. And Anthropic's answer, in the filing, is: we do not currently intend to terminate the agreement. Note the word currently. Unlike the hyperscaler deals, the Lambda payments are tied to capacity actually delivered and accepted, so Anthropic isn't paying for a building that doesn't exist. Here's why that option is worth having. Compute and infrastructure expense went from nine hundred forty-four million in the first quarter of 2025 to nine point nine billion in the third quarter of 2026. Of the eighteen point eight billion spent in the first nine months of this year, about six and a half billion went to short-term capacity, which Anthropic says costs substantially more than multi-year deals. They're paying a premium for being short, and they're keeping a lever on a deal that's late. Investors don't seem bothered. Bloomberg reported Sunday that orders for the IPO were about seven times the shares offered, with Anthropic and some existing holders targeting up to ninety point four billion across the offering and the supplier placement. Reuters had the book at six times by Thursday. Pricing is expected Wednesday the 18th. Orders are requests, not cash, until it prices.
Phuong Nguyen, writing in tek54:
What doesn't check out is the word "beats." AMD's figure, by its own fine print, was produced by AMD Performance Labs in October 2026. It is not an MLPerf submission.
AMD added the MI450X to its Instinct line last week, alongside the flagship MI455X, and claimed the MI455X runs DeepSeek-R1 up to thirty percent faster than Nvidia's Vera Rubin. Nguyen did the arithmetic. AMD's number is twenty-one thousand three hundred thirty-three tokens per second per GPU. Nvidia's published MLPerf v6.1 result for the seventy-two-GPU Vera Rubin rack works out to about sixteen thousand four hundred per GPU. Divide one by the other and you get one point three. So the math is fine. The comparison isn't. One number went through MLPerf's adversarial review, where rivals check your submission. The other came out of AMD's own lab. And it's the Offline scenario, which is batch everything, nobody's waiting. No Server result under a latency cap, no power per token, no accuracy check at FP4. And the stock went up about five percent on a slide. To be fair, Nguyen says this is the whole industry's house style, Nvidia included. AMD's own footnote cautions that results shift with the model, system setup and software version. What I'd actually want is tokens per dollar at my latency target. Offline throughput is the number you get when nobody's in a hurry. My users are always in a hurry. If AMD submits MI455X to MLPerf under Server rules, that's a story. Until then, this is a bid for a meeting with procurement. The MI450X itself is the more practical news: same compute, bandwidth and connectivity as the MI455X, with two hundred eighty-eight gigabytes of HBM4. If your model and context fit in that, you get a cheaper slot in the same Helios rack.
The AlexTech.ai editorial team, summarizing the ggml-org post on the Hugging Face blog:
It is the first major version of both projects after three and a half years of development, and it marks a move to a base considered stable for anyone building applications on top of the engine.
This is the least flashy story today and maybe the most useful one. Both projects means llama.cpp and ggml, the tensor library under it, and a huge slice of local inference runs on that pair. The team's post commits to semantic versioning: if an update breaks the API or the ABI, the major version changes. Anyone who has pinned a commit hash and prayed knows what that's worth. The post also says the team is putting more resources into CI, using Hugging Face's Buckets and Jobs. And there's a line on AI-written code: they're integrating language models into their development cycle, but staying in control of the architecture and understanding the implementation details remain top priorities. That's the right order. Use the tools, keep a human who actually understands the code reviewing the merges. The people telling you agents will replace maintainers should go read a llama.cpp pull request queue sometime. Practical takeaway: if you ship llama.cpp or ggml inside a product, now you can pin a major version and expect it to mean something.
Today's Anthropic segment was really a story about long-term capacity contracts, and that's the whole beat of our sibling show, The Data Center Daily. It's a daily briefing on AI compute, covering hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Look for it wherever you found us.
Homework: find the oldest unpinned dependency in your inference stack and fix it before it fixes you. Tomorrow we'll have Nvidia's earnings and whatever the board says, or doesn't, about who runs the company. You've been listening to Lantern AI Daily Briefing, part of the Lantern Podcasts family. Back with you Tuesday.