ModelsArchitectures & capability
Jensen Huang Resigns as Nvidia CEO; Nathan Lambert Says US Open-Weight AI Has Lost Its Main Backer
Huang's November 12 exit removes the most prominent US supporter of open-weight AI, Nathan Lambert argues. It comes as a Chinese model, Kimi K3.1 Pro, leads open releases and the open model economy still lacks a dependable way to fund frontier training.
From the Interconnects article of November 14, 2026 and cited sources: Jensen Huang resigned as Nvidia CEO on November 12, 2026, after reports of affairs, one of them allegedly with a Chinese spy. [3] [1]
SOURCES[3] Al Jazeera: Jensen Huang out as Nvidia CEO aljazeera.com[1] Interconnects: Open-source AI without Jensen interconnects.ai
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Jensen Huang resigned as Nvidia CEO on November 12, 2026, after reports of affairs, one of them allegedly with a Chinese spy. [3] [1]
- Original source
- Interconnects article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 14, 2026 article published
- Covered
- Nov 14 edition
- Evidence
- 11 sources across 10 sites
- 1Interconnects: Open-source AI without Jenseninterconnects.ai
- 2Linked video: Open-source AI without Jensenyoutube.com
- 3Al Jazeera: Jensen Huang out as Nvidia CEOaljazeera.com
- 4AICoder: Kimi K3.1 Pro open weightsaicoder.com
- 5Interconnects: How big is the open model economy today?interconnects.ai
- 6The Decoder: Inkling 2 Small claims top Western spotthe-decoder.com
RisingHow it’s measured
YouTube video · 11.8K views · 134 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Jensen Huang resigned as Nvidia CEO on November 12, 2026, after reports of affairs, one of them allegedly with a Chinese spy. CFO Colette Kress became interim CEO while FBI and congressional inquiries continue. Lambert argues the larger long-term loss is political: Huang was the leading US champion of open-weight models. Nvidia still has commercial reasons to back them, but its leaders are now focused on stabilizing the company. [3] [1]
The timing is awkward for US open models. Moonshot's Kimi K3.1 Pro, an ungated 2.8T-parameter MoE with 104B active parameters, leads open releases. Its license requires a separate agreement for model-as-a-service firms earning more than $20M. The top Western open model, Thinking Machines' Inkling 2 Small (594B total, 17B active, Apache 2.0), scores 31 on the Artificial Analysis Intelligence Index. Its weights reached Hugging Face on the day Huang resigned. [4] [7] [8] [10]
Lambert's economic case is that open models trail closed ones by a few months and are not reliably cheaper today. Revenue collects with inference hosts rather than with the labs paying for training, which already costs billions per frontier model. On this view, open models win by offering low-cost intelligence that enterprises can adapt and own privately, not by price alone. [2] [5]
The evidence: a leadership change at Nvidia and a Chinese lead in open releases. The implication, which is Lambert's inference, is that US open-weight efforts could lose funding and political cover just as frontier training costs climb.
Executive brief
Nvidia CEO Jensen Huang resigned on November 12, 2026, after reports of affairs, one of them with an alleged Chinese spy. CFO Colette Kress became interim CEO (Al Jazeera). Nathan Lambert argues the bigger long-run loss is political, not commercial. Huang was the main US backer and defender of open-weight AI. Nvidia still has a business reason to back open models, but its leaders now need to steady the company (Interconnects). The timing is awkward. The strongest open model is Chinese (Kimi K3.1 Pro). The best Western one, Thinking Machines' Inkling 2 Small, only just released its weights.
What changed and event timeline
Kimi K3.1 Pro weights go public
Moonshot published an ungated 2.8T-parameter MoE checkpoint. Its custom license requires a separate agreement for model-as-a-service firms with more than $20M in revenue ().
Lambert sizes the open model economy
He tracks token volumes at Baseten, Fireworks and Together AI and finds near-exponential growth. He also flags metrics that are noisy and partly marketing ().
Inkling 2 Small launches with closed weights
Thinking Machines made it available through its Tinker platform and said Apache-2.0 weights would follow after external safety testing ().
Huang resigns; Inkling weights ship
Kress became interim CEO and Harvey Jones board chair, with FBI and congressional probes under way (). The same day, Inkling 2 Small weights reached Hugging Face ().
"Open-source AI without Jensen."
Lambert calls the resignation a "decapitation" of the open-weight movement and urges US coalition-building ().
Capabilities and access
- Kimi K3.1 Pro: 2.8T total / 104B active parameters, MXFP4 weights (about 1.56TB), 1M-token context, vision encoder.
- Inkling 2 Small: 594B total / 17B active, image input, variable reasoning effort, Apache 2.0, no audio (Thinking Machines).
- DeepInfer: DeepSeek's production inference framework, Apache 2.0. It supports DeepSeek-V4.1-Flash on Nvidia SM100 or Huawei Ascend 950 hardware (GitHub).
Read the full section
- Kimi K3.1 Pro: 2.8T total / 104B active parameters, MXFP4 weights (about 1.56TB), 1M-token context, vision encoder. Moonshot publishes serving recipes for vLLM and SGLang (AICoder).
- Inkling 2 Small: 594B total / 17B active, image input, variable reasoning effort, Apache 2.0, no audio (Thinking Machines).
- DeepInfer: DeepSeek's production inference framework, Apache 2.0. It supports DeepSeek-V4.1-Flash on Nvidia SM100 or Huawei Ascend 950 hardware (GitHub).
Technical analysis for researchers and developers
- Inkling: 12.2T pretraining tokens, then one 9-day RL run on 2,048 GPUs, trained in Nvidia's NVFP4 format (Thinking Machines).
- Kimi: 93 layers mixing KDA and gated DSA attention, 896 experts with 16 routed plus 2 shared per token.
- Measurement: Lambert's usage curves rest on sparse fundraising disclosures and OpenRouter data (video, 0:35).
Read the full section
- Inkling: 12.2T pretraining tokens, then one 9-day RL run on 2,048 GPUs, trained in Nvidia's NVFP4 format (Thinking Machines).
- Kimi: 93 layers mixing KDA and gated DSA attention, 896 experts with 16 routed plus 2 shared per token. The MXFP4 weights come from quantization-aware training (AICoder).
- Measurement: Lambert's usage curves rest on sparse fundraising disclosures and OpenRouter data (video, 0:35). His research-adoption figure comes from text searches of arXiv ML papers (3:19). Neither method captures hyperscaler traffic.
Claims and evidence
- Fireworks and Baseten are near Gemini API scale: Lambert's chart compares their token volumes with Google's public disclosures (1:59).
- Open models trail closed ones by a few months: based on Artificial Analysis and Epoch-style capability indices (3:46).
- Kimi matches Opus on terminal coding: Moonshot reports 65.7 on Terminal-Bench 4.0 versus 66.4 for Claude Opus 5.5, but 31.5 versus 65.0 on ProgramBench.
Read the full section
- Fireworks and Baseten are near Gemini API scale: Lambert's chart compares their token volumes with Google's public disclosures (1:59).
- Open models trail closed ones by a few months: based on Artificial Analysis and Epoch-style capability indices (3:46).
- Kimi matches Opus on terminal coding: Moonshot reports 65.7 on Terminal-Bench 4.0 versus 66.4 for Claude Opus 5.5, but 31.5 versus 65.0 on ProgramBench. These are vendor numbers with no third-party check yet (AICoder).
- Inkling is the top Western open model: it scores 31 on the Artificial Analysis Intelligence Index, an independent benchmark (Artificial Analysis).
- Open models carry big serving margins: Lambert attributes this to DeepInfer, but the repository discloses no cost figures (GitHub).
Context and prior work
Lambert's Fireworks talk lays out the economics. Revenue collects at the inference layer, not with the labs that bear training costs (9:47). Lambert also quotes SemiAnalysis's Dylan Patel: the bottleneck for US open models is talent, not money (ChinaTalk/SemiAnalysis).
Read the full section
Lambert's Fireworks talk lays out the economics. Model builders, inference hosts and post-training firms make up a fragmented "trio" (8:14). Revenue collects at the inference layer, not with the labs that bear training costs (9:47). Frontier training already costs billions of dollars per model and could reach tens of billions (10:38). Lambert also quotes SemiAnalysis's Dylan Patel: the bottleneck for US open models is talent, not money (ChinaTalk/SemiAnalysis).
Limitations, safety and contested findings
- Cyber capability: Moonshot says it removed hacking-related training data. This is a reduction, not a removal, and it is self-reported.
- Unproven White House accusation: in July 2026 the White House accused Moonshot of using banned chips and copying an Anthropic model, without presenting evidence (The Next Web).
- No independent support for the policy thesis: no other coverage was found linking Huang's exit to Nvidia's open-model funding or Nemotron.
Read the full section
- Cyber capability: Moonshot says it removed hacking-related training data. On its own exploit test, the released model solved 81% of tasks versus 90.5% for an unpruned version (The Next Web). This is a reduction, not a removal, and it is self-reported.
- Unproven White House accusation: in July 2026 the White House accused Moonshot of using banned chips and copying an Anthropic model, without presenting evidence (The Next Web).
- No independent support for the policy thesis: no other coverage was found linking Huang's exit to Nvidia's open-model funding or Nemotron. That argument, and the regulatory-risk argument, are Lambert's own.
Business and practitioner implications
- Price: open models are not reliably cheaper today. Closed labs cover the cost-performance Pareto frontier (5:21).
- Ownership: the selling point is "private, adaptable" intelligence that enterprises own. No source data is linked (14:49).
- Licensing: Kimi's revenue clause matters for hosting providers. Inkling's Apache 2.0 license carries no such restriction.
Read the full section
- Price: open models are not reliably cheaper today. Closed labs cover the cost-performance Pareto frontier (5:21). Lambert says the open ecosystem must push inference prices toward marginal cost (12:20).
- Ownership: the selling point is "private, adaptable" intelligence that enterprises own. Lambert cites a Harvey legal fine-tune of Kimi K3 that he says beat frontier models on accuracy at lower cost; no source data is linked (14:49).
- Licensing: Kimi's revenue clause matters for hosting providers. Inkling's Apache 2.0 license carries no such restriction.
- Funding watch: reported support for Thinking Machines' rumored $40B round (Interconnects) and Nvidia's open-model spending under Kress.
The source trail.
Sources (11)
Interconnects: Open-source AI without Jensen
interconnects.aiLinked video: Open-source AI without Jensen
linked video auto captions
www.youtube.com