Open-source AI without Jensen

Nvidia CEO Jensen Huang has resigned, a move the Interconnects newsletter characterizes as a major setback for US development of an open model ecosystem. The publication frames the change as a decapitation of the movement's most visible corporate champion and argues it is time for open model supporters to rebuild their coalition. The piece ties the news to the aftermath of a midterm political swing, arguing the disclosures landed at a moment when they could do maximum damage, and notes that reports of the Nvidia departures were circulating before they became a public storm. It offers no further detail on the circumstances of the resignation itself.
Huang's role in open models, as the newsletter describes it, went well beyond cheerleading. He was willing to spend his influence on the ecosystem's behalf and to intervene when early trouble emerged, and he had been flooding the zone with funding. Interconnects expects that funding to very likely dry up now. The publication casts open models as one of Huang's more ambitious, founder-style bets.
Nvidia's financial interest in open models survives him, per the piece. The company benefits from avoiding concentration of power among a handful of customers large enough to design their own chips and thereby erode Nvidia's profits. But supporting an open ecosystem is, in the newsletter's framing, a tricky puzzle more than an engineering problem.
Nvidia's operations are described as famously dependent on Huang as an operator. The consensus view, according to Interconnects, is that challenges lie ahead, possibly including further high-profile departures through chip-smuggling investigations or competitors poaching talent. The company's top priority is now to stabilize, deliver to customers over the next 24 months, and prevent an implosion. This suggests to me that Nvidia's open model advocacy could become a lower priority than supply commitments and internal consolidation, since a transitional leadership team tends to protect the core business first.
The newsletter argues the open model economy is still in its nascency, making the coming years crucial for establishing use-cases, revenue flow and reasons to keep training next-generation models. It distills its argument into one upside, one reason for optimism and one downside. The positive case: the ecosystem has correctly latched onto ownership of intelligence and enterprise post-training as a crucial lane, though how big the downstream market is remains to be determined. The downside: the long-term promise of open models as truly commoditized intelligence is far from realization while compute is so scarce.
As the piece lays it out, inference startups are themselves struggling to secure chips because Anthropic and OpenAI are buying so much capacity, which leaves little margin competition on open-weight models. The newsletter points to DeepSeek's DeepInfer release as evidence that open-source models can still be served at giant inference margins, meaning they are far from a commodity. In that environment, closed labs with better economies of scale and tighter integration from model to datacenter to chips win. This is a useful point for product teams: open weights reduce licensing lock-in, but they do not automatically reduce serving cost when supply of compute, not software, is the binding constraint.
The piece raises a political risk. It describes a major swell of Democratic funding toward candidates who want to regulate the AI industry under a safety banner, and contends many of these figures hold general views that open models are unsafe without being educated on their benefits. The author's biggest near-term worry is sloppy regulation at the state level, followed by national regulation. The newsletter also argues the ecosystem has been one strong Chinese open-weight model away from a crisis since this summer, and that while no major hack driven by a popular Chinese open-weight model has occurred, it is a matter of when, not if. It calls for someone to encourage caution and an ecosystem perspective when open-weight hawks come hunting.
The newsletter argues that knocking over the open-source AI domino today is easy, while unwinding the resulting concentration of power, reduction in safety research and stifled long-term innovation is far harder. It notes that public data makes the trade-off look like balancing long-term innovation against short-term risk, but says that impression stems from how hard the open model economy is to measure, since its most valuable aspects run on private infrastructure.
The piece draws on a ChinaTalk and SemiAnalysis emergency podcast. SemiAnalysis CEO Dylan Patel argued the gap is not money but talent: most of the strongest researchers sit at Anthropic, OpenAI, DeepMind and Meta, and the rest are split across many startups. Patel said Thinking Machines has a strong team but questioned whether it matches Moonshot, which he described as roughly on par with US labs in skill while carrying a large deficit in compute and deployment. He noted DeepMind, Grok and Meta have regularly trailed the frontier of Chinese labs despite spending more, and said that without a scale advantage you do not win.
The newsletter credits Nvidia's Nemotron effort with the right ambition and strong people but says it has not reached mission accomplished. It sees a positive sign in Thinking Machines' recent Inkling 2 Small, described as the best Western open-weight model in a while, while noting it is far behind Kimi K3.1 Pro, with expectations high for Zhipu's next GLM update. It contrasts this with closed labs such as Anthropic, which it calls some of the greatest businesses ever assembled, and mentions a next funding round at a $40B valuation said to be done but insufficient to keep pace with the closed frontier for long. The newsletter names hyperscaler leaders including Zuckerberg, Andy Jassy and Satya Nadella as the most likely allies.
Why it matters: developers and product teams building on open-weight models depend on a small set of patrons and funders, and the newsletter's argument implies that pipeline may thin just as teams are committing to post-training and ownership of intelligence as a strategy (an inference from its stated funding expectations). Teams should weigh licensing flexibility against the possibility that serving costs stay high while compute remains scarce, and should watch state and national rulemaking as a nearer-term risk than model quality (also an inference). The piece's core claim is that no single successor can fully replace Huang's operational role, so the ecosystem needs new institutional backing.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
A decapitation. It’s time to rebuild.