For the first time in thirty-three years, someone other than its founder is running Nvidia. And it took about four days to get there. You're listening to Lantern AI Daily Briefing, and it's Friday, November 13th. On the rundown: Jensen Huang's exit from Nvidia, Google's new Gemini 4 Hydrogen workhorse model, SpaceX committing more than seventy-four billion dollars to two supercomputing campuses, and a billion-dollar Series A for Mirendil, a lab started by former Anthropic researchers. If today's show was useful, follow us wherever you're listening — the next one will be waiting.
Alex Volkov, writing in his ThursdAI newsletter:
This morning Nvidia announced that Jensen Huang has resigned as President, CEO and member of the board, effective immediately. He founded the company 33 years ago and turned it into the most important company in AI. CFO Colette Kress is now the interim CEO and keeps the CFO job, Harvey Jones (former CEO of Synopsys) becomes the first-ever board chair, and the board is searching for a permanent CEO among internal and external candidates.
That morning was Thursday. Volkov lays out the timeline: on Sunday, Huang emailed staff admitting an affair, and the board opened an independent investigation into whether he was the target of a foreign intelligence effort. The Information reported that the network of the woman at the center of it, Alice Huang, no relation, helped move about seven and a half billion dollars of Nvidia chips to China. Then the New York Times reported a second affair, and the Wall Street Journal reported the board was preparing to oust him. Two cautions. Nvidia's release says resigned and gives no reason. And as Volkov's own panel pointed out, none of the allegations has been proven. For builders, the practical question is allocation. Kurzia's AI-for-VCs newsletter, citing Tom's Hardware, says Nvidia has no succession plan behind Kress, and predicts tighter export enforcement and slower allocations. That's a forecast, not a fact. But if your roadmap depends on a GPU delivery date, I'd want that date reconfirmed in writing this month. The policy side moved the same day. MIT Technology Review reported Thursday that Representative Bill Foster says Nvidia representatives told him privately that the chip-location tracking in his Chip Security Act is technically possible. Meanwhile the Semiconductor Industry Association, which counts Nvidia as a member, called those tracking mandates potentially infeasible back in March. Nvidia didn't respond to the magazine. Foster's bill is one of three chip-security measures under consideration for the Senate's National Defense Authorization Act. Technically possible, Foster says, but never put in writing. Volkov's panel worried most about open source, since Huang was one of the loudest voices for open models. My take: the GPUs keep shipping regardless. What changes is who inside Nvidia is willing to pick fights in Washington, and on whose behalf. One more data point from Kurzia: Nvidia joined Agile Robots' five-hundred-million-dollar Series F the same day. The deal machine is still running. Whether it keeps that pace without its chief dealmaker is the question for anyone holding an Nvidia strategic check.
Ars Technica, on Thursday's launch:
Google claims that Hydrogen represents a major improvement versus the previous Flash model. It runs in one of four thinking levels (low, medium, high, and the new xhigh), and even in high mode, it's 2.5 times faster than Flash 3.8. Plus, it reduces "fabricated claims" by more than 5x, according to Google's analysis.
Start with price. A dollar fifty per million input tokens and seven fifty output, flat across the whole million-token context, per OfficeChai. That's a quarter below Sonnet 5.5's list price. And unlike Argon and Carbon, there's no staged rollout through cyber researchers. Hydrogen replaced 3.8 Flash across Google's products on day one. The benchmarks are all Google's. OfficeChai counts Hydrogen ahead of Sonnet 5.5 on seven of the ten evaluations Google chose to publish. Sonnet keeps the terminal and computer-use tests: sixty-three point nine to fifty-seven on Terminal-Bench 4.1, and eighty-three point nine to seventy-seven point six on the OSWorld offline subset. On the Artificial Analysis index, Google says fifty-one at the default high setting and fifty-four at xhigh, against Argon's fifty. Note the opponent is Anthropic's mid-tier model, not Fable. The part I'd actually try is Orchestrator in Antigravity. Carbon plans, Hydrogen runs the workers and verifiers. Google says that on Terminal-Bench 4.0 it got close to Carbon's success rate on twenty-four percent less quota, nineteen percent faster. And there's a clock on this. As we covered Wednesday, Carbon's introductory price doubles after November 30th. Kurzia makes the point that anything still pinned to Google's top tier for subagent work sees that bill double unless it moves down. Hydrogen is Google telling you where to move it. One soft spot to test: OfficeChai notes terminal-driven tasks have been a recurring weakness for Google's models. If your agents live in a shell, run your own evals before you switch anything.
PTI, in a report carried by The Hindu BusinessLine:
In a statement issued on Thursday, SpaceX outlined an investment of over $24 billion for a supercomputing facility in Texas's Rio Grande Valley and roughly $50 billion for a separate campus in Cochise County, Arizona.
Two campuses, two power strategies. The Starbase site in South Texas will generate its own electricity without drawing from the public grid, and a public hearing notice lists a gas-fired power plant alongside the supercomputing facility. Arizona will work with local electric cooperatives and could scale to six gigawatts at full buildout. Six gigawatts at one site. For scale, earlier this month we cited a New York Times report that Anthropic expects more than three and a half gigawatts across its entire footprint by the end of 2026. That same report put about a hundred twenty billion dollars of Anthropic's commitments with SpaceX. Nothing in Thursday's statement says who these campuses are for, so I won't draw that line for them. Read the conditions, too. The twenty-four billion covers only the initial phases in Texas. Arizona's fifty billion is still subject to county approvals. The Starbase City Commission takes up a property tax abatement on December 9th, a Wednesday. And Arizona is planning dry cooling with no industrial groundwater use, which reads like a campus designed around the fights we've been watching. Fair. Redbud Consulting counted thirty-eight of forty-four restrictive data-center ballot measures passing on Election Day. Off-grid gas in Texas and dry cooling in Arizona answer two of the loudest objections: the grid bill and the water.
Mark Rutherford, writing in Tech Times:
Mirendil, an AI lab founded by former Anthropic researchers, announced a $1 billion Series A led by Kleiner Perkins and Clearlake Capital on Thursday, Nov. 12, but the release names no valuation. The training bill the company cites, less than $10 million in compute, covers only its smallest model.
A billion-dollar Series A, a hundred forty-one days after a two-hundred-million-dollar seed at a one-billion valuation. Kurzia lists Google among the participants. Bloomberg reported talks in September at a five-billion valuation, but Rutherford says no source confirms the final terms match. When a round this size skips the valuation, I start asking what else it skipped. The headline technical claim is that AI agents wrote more than ninety-nine percent of the code behind its pretraining run, with a single researcher directing. Mirendil itself says the model didn't build itself and that it hasn't reached fully autonomous recursive self-improvement. And its founding team at launch in June was twenty people, not one. Then the evidence. Bits-per-byte on seven code repositories published after training. Mirendil-Base scores point two two two, Kimi-K3 scores point one nine three. Lower is better, so it's second, on a test Mirendil picked and ran. No coding task scores, no independent reproduction. That's a prediction-loss number, not a will-it-merge number. And the sub-ten-million bill covers Mirendil-Base, which is a hundred to five hundred billion parameters. Larger models, up to ten trillion, were still training as of October 21st. What's genuinely interesting is the process, a small team using agents to run a pretraining pipeline. The billion dollars is a bet that the process scales to the big models. Nobody outside the company has checked that yet.
If the SpaceX segment left you wanting the power-plant details, our sibling show The Data Center Daily lives on exactly that. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Find it in the same app you're using right now.
Weekend assignment: point Hydrogen at whatever's running your subagents today, and count the retries, not just the invoice. And when we pick back up, we'll be tracking Nvidia's CEO search and Anthropic's pricing, which CNBC has reported is planned for Wednesday the 18th. That wraps this week of Lantern AI Daily Briefing, a Lantern Podcast. We'll see you Monday, November 16th.