PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: AI inference cost

  • Inkling: Thinking Machines Lab Releases Its First Open-Weights Model, a 975B Multimodal Mixture-of-Experts With Controllable Thinking Effort That Can Fine-Tune Itself on Tinker

    Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first open-weights model trained from scratch. Inkling is a 975 billion parameter Mixture-of-Experts transformer (41B active) with a context window of up to 1 million tokens, native multimodal reasoning over text, images, and audio, and a dial for controllable thinking effort. The lab is explicit that Inkling is not the strongest model in the world. It is pitched as something arguably more useful: a broad, balanced, customizable foundation you can fine-tune on Tinker, with the full weights on Hugging Face. The announcement even includes a demo where Inkling fine-tunes itself and swaps in its own new weights.

    TLDR

    Thinking Machines Lab released Inkling, a 975B-total, 41B-active Mixture-of-Experts model pretrained on 45 trillion tokens of text, images, audio, and video, alongside a preview of Inkling-Small (276B total, 12B active). The release covers the model’s generalist benchmark profile across reasoning, agentic coding, tool use, vision, and audio; a controllable thinking effort setting that lets developers trade performance against tokens (matching Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens); an encoder-free multimodal architecture using dMel spectrograms and hMLP image patches; a training recipe combining Muon and Adam with weight decay coupled to the learning rate; RL scaled past 30 million rollouts with log-linearly improving reasoning and an emergent compression of the chain of thought; an epistemics push covering calibration, forecasting (where it beats several frontier models), abstention, and censorship resistance; the strongest FORTRESS adversarial safety score among compared open-weights models; a headline-grabbing demo of the model fine-tuning itself into a lipogram assistant via Tinker; and day-one availability on Tinker (at a 50% discount), Hugging Face, and inference partners including Together, Fireworks, Modal, Databricks, Baseten, vLLM, SGLang, and llama.cpp.

    Thoughts

    The most striking thing about this launch is its honesty. Nearly every frontier release leads with a claim to be the best at something, and the fine print walks it back. Thinking Machines Lab says plainly that Inkling is not the strongest model available, open or closed, and then makes the case that “strongest” is the wrong axis for most real buyers. If you are going to run a model millions of times inside a product, what you care about is the cost curve, the adaptability, and whether you can shape it to your workflow. That framing conveniently matches their business (Tinker sells fine-tuning), but it also matches how production AI actually gets deployed, where cost and latency are binding constraints and a benchmark crown is trivia.

    The self-fine-tuning demo deserves more attention than it will probably get. Asked to become a lipogram assistant that never uses the letter “e” (a behavior prompting alone cannot reliably produce), Inkling wrote its own training objective and scoring function, generated its own synthetic data, launched the run on Tinker, evaluated the result against its base self, and then staged a weight swap so the improved checkpoint took over the session. That is a closed loop of specify, train, evaluate, and self-update, packaged as a cute product demo. The loop is the primitive behind every serious conversation about recursive self-improvement, and here it is running as a marketing asset with a 27 minute wall clock. The gap between “toy objective” and “economically meaningful objective” is now a question of reward design, not plumbing.

    Controllable thinking effort is the feature I expect developers to care about most. Instead of publishing a single score, TML publishes a curve: sweep the effort setting from 0.2 to 0.99 and watch performance trade against generated tokens. Inkling reportedly matches Nemotron 3 Ultra on Terminal Bench 2.1 while spending about a third of the tokens. Benchmarks reported as single points hide exactly this, and a model that reaches a target score cheaply beats a model that scores two points higher at triple the cost in any high-volume workload. Expect effort curves to become standard marketing for open models, the way context length became standard a couple of years ago.

    The epistemics section is quietly the most differentiated part of the release. TML trained calibration directly, running RL against proper scoring rules on resolved real-world questions, and pairing a rubric grader with a claims grader that does agentic web search to verify each factual assertion. The result is a model that beats GPT-5.5 and Claude Opus 4.8 on ForecastBench without search and holds its own on Prophet Arena. A model that knows when to say “I don’t know” is more useful across messy real-world domains than one that confabulates confidently, and it is notable that a lab whose stated mission is extending human will and judgment treats calibrated uncertainty as a first-class training target rather than a safety afterthought. The censorship-resistance training, validated on Cognition’s Propaganda and Censorship Eval, extends the same idea: trustworthiness as a capability you train, not a policy you bolt on.

    Finally, the open-weights safety tension is handled with unusual candor. Inkling posts the strongest adversarial FORTRESS score among the open models compared while keeping benign over-refusal low, and it was tested externally for CBRN, cyber, and loss-of-control capabilities. But everyone in this space knows fine-tuning can strip safety behavior from open weights, and TML ships a fine-tuning platform for this exact model. Their acknowledgment that they are actively studying how safety behavior survives fine-tuning on Tinker is the right thing to say, and it is also the open question that will define whether “safe open weights” is a coherent category at all.

    Key Takeaways

    • Inkling is Thinking Machines Lab’s first from-scratch, open-weights model: a Mixture-of-Experts transformer with 975B total parameters, 41B active, and a context window up to 1M tokens.
    • It was pretrained on 45 trillion tokens spanning text, images, audio, and video, and reasons natively over text, images, and audio without separate encoders.
    • A preview of Inkling-Small ships alongside it: a 276B-parameter MoE with just 12B active parameters that matches or beats its larger sibling on several benchmarks thanks to an improved pretraining recipe.
    • TML explicitly positions Inkling as a base for customization rather than the strongest overall model, leaning on multimodality, efficient thinking, and Tinker fine-tuning as the differentiators.
    • The launch demo shows Inkling fine-tuning itself: it wrote its own training objective and data, ran the job through the Tinker API, evaluated the result, and hot-swapped to its own new weights inside the OpenCode harness.
    • The self-fine-tuning target was a lipogram assistant that never uses the letter “e,” a behavior chosen precisely because prompting alone cannot reliably achieve it; the full loop completed in about 27 minutes.
    • Controllable thinking effort is a core feature: a setting swept from 0.2 to 0.99 traces a full performance-versus-tokens curve instead of a single benchmark point.
    • On Terminal Bench 2.1, Inkling matches Nemotron 3 Ultra’s score at roughly one third of the generated tokens, the release’s flagship efficiency claim.
    • Inkling was trained to run inside a variety of coding and agent harnesses, with tool sets and schemas randomized during training to reduce sensitivity to any particular harness.
    • On Design Arena’s blinded human-evaluated Agentic Web Dev leaderboard, Inkling scores 1257, among the strongest open-weights models and tied with Claude Opus 4.6.
    • Headline benchmark scores at effort 0.99 include SWEBench Verified 77.6%, SWEBench Pro Public 54.3%, Terminal Bench 2.1 63.8%, GPQA Diamond 87.2%, AIME 2026 97.1%, and HLE 29.7% text-only (46.0% with tools).
    • Agentic and general scores include MCP Atlas 74.1%, Tau 3 Banking 23.7%, and BrowseComp 77.1% with context management.
    • Vision results are strong for an open model: MMMU Pro 73.5%, CharXiv RQ 78.1%, rising to 82.0% when the model uses a Python tool for zooming and cropping during visual reasoning.
    • Audio results place it among the strongest open-weights audio models: VoiceBench 91.4%, MMAU 77.2%, and Audio MC 56.6%, well ahead of Qwen3-Omni and Nemotron Nano-Omni on the last.
    • The multimodal stack is encoder-free: audio enters as discrete dMel spectrograms and images as 40×40 pixel patches through a four-layer hMLP, both passed through a lightweight embedding layer and processed jointly with text tokens.
    • The MoE design largely follows DeepSeek-V3: 256 routed experts plus 2 shared experts per layer, 6 routed experts active per token, with a sigmoid router and auxiliary-loss-free load balancing.
    • Attention interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads, and uses a learned relative positional embedding instead of RoPE, which TML found extrapolates better to long sequences.
    • Short convolutions are applied after the key and value projections and on the attention and MLP residual branch outputs, an unusual architectural touch aimed at efficiency and long-context performance.
    • Training used a hybrid optimizer strategy, Muon for large matrix weights and Adam for everything else, with weight decay coupled to the square of the learning rate to keep weight magnitudes stable.
    • Post-training was bootstrapped with a small SFT phase on synthetic data generated by open-weights models including Kimi K2.5, with the large majority of compute spent on large-scale RL.
    • RL was scaled past 30 million rollouts across two long continuous runs, with reasoning performance on a held-out aggregate (AIME, HLE, GPQA, and others) improving log-linearly throughout.
    • Effort control was trained by varying the system message and per-token cost across rollouts, teaching the model to modulate its own thinking budget.
    • An emergent effect appeared during RL: the chain of thought compressed over training, dropping articles and connectives into a telegraphic style, driven purely by efficiency pressure rather than any targeted reward.
    • Inkling was TML’s first major training effort and ran on NVIDIA GB300 NVL72 systems; the lab says future models will push compute scale further across pretraining and RL.
    • Calibration was trained directly with RL against proper scoring rules on a large corpus of resolved real-world questions, treating well-placed confidence as a capability rather than a byproduct.
    • On ForecastBench without search, Inkling’s Brier Index of 61.1 beats GPT-5.5 (59.1) and Claude Opus 4.8 (54.6), and it stays competitive with search enabled and on Prophet Arena.
    • Instruction following was trained with two automated graders working together: a rubric grader scoring against a checklist and a claims grader that verifies each factual claim via agentic web search, improving helpfulness and reducing hallucination simultaneously.
    • Abstention-aware rewards on short-form factual QA taught the model to answer when confident and hedge or decline when not, with some prompts explicitly forcing or forbidding hedging so the user’s preference wins.
    • Inkling was trained to answer directly on topics subject to censorship, and Cognition’s Propaganda and Censorship Eval found strong censorship non-compliance.
    • On FORTRESS, Inkling posts the strongest adversarial refusal score (78.0%) of any compared open-weights model while keeping benign compliance high (95.9%), and scores 98.6% on StrongREJECT.
    • Safety testing covered CBRN, cyber, and loss-of-control capabilities plus human-AI threat vectors like sycophancy, vulnerable users, and manipulation, verified by commissioned external testers.
    • Inkling is available for fine-tuning on Tinker today with 64K and 256K context options at a 50% limited-time discount, plus a free Inkling Playground chat interface in the Tinker console.
    • Full weights are on Hugging Face, including an NVFP4 checkpoint for efficient inference on NVIDIA Blackwell, with API availability via Together, Fireworks, Modal, Databricks, and Baseten and inference support in SGLang, vLLM, TokenSpeed, and llama.cpp.
    • TML frames Inkling as the first in a family and as the intended background reasoning model for its previously announced real-time interaction models system.

    Detailed Summary

    What Inkling Is and Why It Exists

    Thinking Machines Lab frames its mission as building AI that extends human will and judgment, and Inkling as the logical next step after shipping the Tinker customization platform, previewing an interaction-focused AI system, and publishing research. Inkling is a Mixture-of-Experts transformer with 975B total and 41B active parameters, a context window up to 1M tokens, and pretraining on 45 trillion tokens of mixed text, image, audio, and video data. The lab is upfront that it is not the strongest model available. The pitch is breadth plus adaptability: a generalist trained across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks rather than tuned to dominate one leaderboard, offered with full weights so people can make it their own. It launches with a preview sibling, Inkling-Small, at 276B total and 12B active parameters.

    The Self-Fine-Tuning Demo

    To demonstrate what customization means, TML asked Inkling to fine-tune itself. Running inside the OpenCode harness with access to Tinker, the model was told to become a lipogram assistant that never uses the letter “e.” Inkling drafted the plan, wrote an objective file with a scoring function (any response containing “e” scores zero), generated synthetic training data, launched a supervised fine-tuning run through the Tinker API, evaluated the checkpoint against its base self, and then staged a self-update so the supervisor relaunched the session on the new weights. The pipeline passed in about 27 minutes, and the updated model answered a test question about launching an LLM without a single “e.” It is a whimsical objective wrapped around a serious primitive: a model autonomously specifying, running, and adopting its own weight updates.

    Agentic Coding and Tool Use

    TML trained Inkling to operate inside many coding and agent harnesses, randomizing tool sets and schemas during training so the model does not overfit to one environment. The release showcases three demos: a one-shot job-application web app that then hosts an embedded browser-use agent operating its own interface; a nine-page, cohesively designed PDF food and travel journal produced from a single editorial prompt with web-verified details; and a server-authoritative multiplayer snake game refined over 40 iterations of feedback from GPT Codex acting as a reviewer. On benchmarks, Inkling posts 77.6% on SWEBench Verified, 54.3% on SWEBench Pro Public, and 63.8% on Terminal Bench 2.1, competitive within the open-weights field, and 1257 on Design Arena’s human-judged web dev leaderboard, in the same band as Claude Opus 4.6.

    Controllable Thinking Effort

    Rather than reporting a single operating point, TML sweeps Inkling’s effort setting from 0.2 to 0.99 and plots score against mean generated tokens on Terminal Bench 2.1, HLE, and IFBench, with competitors shown at their default settings. The headline result is efficiency: Inkling reaches Nemotron 3 Ultra’s Terminal Bench score at roughly a third of the tokens. The argument is that cost and latency are binding constraints in production, especially for interactive collaboration, so the full cost curve, not the peak score, is what developers should evaluate. Effort can be set from within the agent harness, and the ability was trained by varying system messages and per-token costs across RL rollouts.

    Native Multimodality Without Encoders

    Inkling is designed to serve as the background reasoning model for TML’s interaction models system, which requires real-time voice and vision collaboration. The multimodal components are trained from scratch with an encoder-free architecture: audio arrives as discrete dMel spectrograms and images as 40×40 pixel patches through a four-layer hMLP, both mapped through a lightweight embedding layer and processed jointly with text. The model transcribes speech, follows spoken instructions, reasons over long recordings, and answers questions about charts and diagrams, optionally using a Python tool to zoom and crop images mid-reasoning. Scores like 91.4% on VoiceBench and 82.0% on CharXiv RQ with Python place it among the strongest open-weights multimodal models, though still behind Gemini 3.1 Pro.

    Epistemics: Calibration, Forecasting, and Censorship Resistance

    TML groups calibration, instruction following, and censorship resistance under the banner of epistemics. Calibration was trained with RL against proper scoring rules on resolved real-world questions, and it shows: Inkling’s ForecastBench Brier Index of 61.1 without search beats GPT-5.5 and Claude Opus 4.8, and its Prophet Arena score sits close to the frontier. Instruction following used two complementary automated graders, a rubric checklist and a claims grader that verifies factual assertions through agentic web search, so recall-spraying to hack rubrics gets penalized by the factuality check. Targeted abstention-aware QA datasets taught the model to say “I don’t know” or give hedged best guesses when appropriate, while still complying when a user demands a forced guess. Finally, the model was trained to answer directly on censorship-prone topics, with Cognition’s Propaganda and Censorship Eval finding strong non-compliance with censorship patterns.

    Safety for an Open-Weights Release

    Inkling was trained to an internal behavioral spec across all modalities and then checked by commissioned external safety testers. Evaluations covered dangerous capabilities (CBRN, cyber, loss of control) and human-AI threat vectors including sycophancy, vulnerable users, and harmful manipulation. On FORTRESS, which pairs adversarial harmful requests with benign look-alikes, Inkling posts the strongest adversarial score among the compared open models (78.0%) without collapsing on the benign side (95.9%), and it scores 98.6% on StrongREJECT. TML acknowledges the open question hanging over every open-weights release: how safety behavior holds up under fine-tuning, which it says it is actively studying on Tinker.

    Architecture and Training Recipe

    The MoE layout follows DeepSeek-V3: 256 routed experts and 2 shared experts per layer with 6 routed experts active per token, a sigmoid-based router, and auxiliary-loss-free load balancing. Attention interleaves sliding-window and global layers 5:1 with 8 KV heads, and positions are encoded with a learned relative positional embedding that TML found outperforms and out-extrapolates RoPE. Short convolutions appear after the key and value projections and on residual branch outputs. Optimization was hybrid, Muon for large matrices and Adam elsewhere, with hyperparameter schedules drawn from the lab’s modular manifolds research and weight decay coupled to the square of the learning rate to keep weight norms stable. Post-training bootstrapped from a small SFT phase on synthetic data from open models including Kimi K2.5, then spent the bulk of compute on large-scale RL. Everything ran on NVIDIA GB300 NVL72 systems.

    RL at Scale and the Emergent Compression of Thought

    TML scaled asynchronous RL past 30 million rollouts across two long continuous runs, with performance on a held-out aggregate of reasoning evals improving log-linearly the whole way. Along the way an unplanned behavior emerged: the chain of thought became progressively more concise, shedding grammatical overhead into a telegraphic style (“We need to understand” becomes “We need determine”) while remaining comprehensible and leaving final answers unaffected. No reward targeted this; token efficiency pressure alone drove the compression, echoing an observation Cognition made while training SWE-1.7. It is a vivid example of optimization discovering its own shorthand.

    Inkling-Small

    The preview of Inkling-Small is arguably the sleeper story: with 12B active parameters against Inkling’s 41B, it matches or exceeds the larger model on a surprising number of benchmarks, including GPQA Diamond (88.3% vs 87.2%), IFBench (83.4% vs 79.8%), and CharXiv RQ with Python (83.4% vs 82.0%). TML attributes this to pretraining data and recipe improvements made after the big model trained, with both models sharing the same post-training stack. The clearest gaps favoring big Inkling are factuality (SimpleQA 43.9% vs 20.9%), Terminal Bench, and Tau 3 Banking. Full weights for Inkling-Small will be released once testing finishes, and its cost and latency profile targets high-volume workloads like coding, LLM grading, and synthetic data generation.

    Availability and the Ecosystem Play

    Inkling is on Tinker today with 64K and 256K context options at a limited-time 50% discount, plus a free Inkling Playground chat interface with integrated web search in the Tinker console so developers can get a feel for the model before committing to a run. The cookbook gained native Inkling support and three new audio recipes, and a new tml-renderer handles chat templates, tool calls, reasoning content, and multimodal inputs. Deployment partnerships span Together, Fireworks, Modal, Databricks, and Baseten for APIs; RadixArk for SGLang and Miles; Inferact for vLLM; Lightseek for TokenSpeed; Unsloth for llama.cpp; and Hugging Face for transformers integration. Full weights are on Hugging Face in both the original checkpoint and an NVFP4 checkpoint for NVIDIA Blackwell inference.

    Notable Quotes

    “Our mission is to build AI that extends human will and judgment.”

    Thinking Machines Lab, opening the Inkling announcement

    The company’s north star, and the lens through which the whole release (customization, calibration, open weights) is framed.

    “Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning.”

    Thinking Machines Lab, positioning the release

    A rare piece of launch-day honesty from a frontier lab, and the strategic thesis of the whole release.

    “Picking the right base model to fine-tune is a qualitative judgment that combines measurable benchmarks with the unique feel of a model that comes from playing with it.”

    Thinking Machines Lab, on why the Inkling Playground exists

    An argument that vibes are data, from the lab that built a playground into a fine-tuning console.

    “Cost and latency are often binding constraints in real-world applications, and low latency in particular is crucial for enabling collaboration and improvement through iteration.”

    Thinking Machines Lab, on controllable thinking effort

    The case for evaluating models on their full effort-versus-performance curve instead of a single benchmark point.

    “A model that’s confident in every answer it gives, including when it’s missing info and confabulates, forces the user to double-check everything.”

    Thinking Machines Lab, on why calibration was a training target

    The clearest one-line justification for treating calibrated uncertainty as a capability rather than a nicety.

    “Together, the two graders improve helpfulness and reduce hallucination at the same time, rather than trading one for the other.”

    Thinking Machines Lab, on pairing a rubric grader with a web-searching claims grader

    A neat solution to rubric hacking: verify every claim with agentic search so spraying plausible facts stops paying.

    “Safety is crucial for open-weights models. We’re continuing to study safety behavior and capability uplift in customizable models, including how safety behavior is impacted by fine-tuning on Tinker.”

    Thinking Machines Lab, on the open question of fine-tunable safety

    The acknowledgment that safety trained into open weights must survive the very customization the product sells.

    “Inkling is just the start: our first release in a model family we will continue to build on.”

    Thinking Machines Lab, on the roadmap

    Together with the GB300 compute note, a clear signal that larger and stronger family members are coming.

    Read the full announcement, including the interactive demos, effort curves, and complete benchmark tables, on the Thinking Machines Lab blog.

    Related Reading

  • Jensen Huang on Nvidia’s Future: Physical AI, the Inference Explosion, Agentic Computing, and Why AI Doomers Are Wrong

    Jensen Huang sat down with the All-In Podcast crew at GTC 2026 for one of the most wide-ranging and candid conversations he’s had in years. From the Groq acquisition to $50 trillion physical AI markets, from defending Nvidia’s pricing to gently calling out Anthropic’s communications missteps, Huang covered everything. Here’s a complete breakdown of everything said — and what it means.

    ⚡ TL;DW

    • Nvidia has evolved from a GPU company into a full-stack AI factory company, and its TAM has expanded by 33–50% just from new rack configurations.
    • Inference demand is exploding — Huang says compute will scale 1 million times, and analysts who model 7–20% growth “don’t understand the scale and breadth of AI.”
    • The Groq acquisition positions Nvidia to run the right workload on the right chip — GPU, LPU, CPU, switch, all orchestrated under Dynamo, the AI factory OS.
    • Physical AI (robotics, autonomous vehicles, industrial automation) is Nvidia’s play at a $50 trillion market — and it’s already a ~$10 billion/year business growing exponentially.
    • OpenClaw (Claude’s open-source agentic framework) is, in Jensen’s view, the new operating system for modern computing.
    • Jensen pushed back hard on AI doomerism — and diplomatically but clearly called out Anthropic’s communications as too extreme.
    • Robots are 3–5 years away from being “all over the place.” Jensen hopes for more than one robot per human on Earth.
    • Dario Amodei’s $1 trillion AI revenue forecast by 2030? Jensen says he’s being too conservative.
    • His advice to young people: become deeply expert at using AI. English majors may end up winning.

    🔑 Key Takeaways

    1. Nvidia Is No Longer a Chip Company

    Jensen Huang made clear that Nvidia’s identity has fundamentally shifted. The company is now an AI factory company — building not just GPUs but the entire computing stack: GPUs, CPUs, networking switches, storage processors (BlueField), and now LPUs via the Groq acquisition. The operating system tying it all together is called Dynamo, named after the Siemens machine that powered the last industrial revolution by turning water into electricity. Huang’s point: Dynamo is doing the same thing for AI — turning raw compute into intelligence at industrial scale.

    2. The Inference Explosion Is Real and Massive

    A year ago, Huang predicted inference would scale enormously. He’s now doubling down: from generative AI to reasoning models, compute requirements grew roughly 100x. From reasoning to agentic AI, another 100x. That’s 10,000x in two years — and Huang says we haven’t even started scaling yet. He believes the ultimate trajectory is 1 million times more compute than where we started. Analysts who project 20–30% revenue growth for Nvidia fundamentally don’t understand what’s coming.

    3. Disaggregated Inference Is the New Architecture

    The technical centerpiece of GTC 2026 was disaggregated inference — the idea that the AI processing pipeline is so complex (prefill, decode, working memory, long-term memory, tool use, multi-agent coordination) that it should run across heterogeneous chips, not just a single GPU rack. Nvidia’s Vera Rubin system is built for this: multiple rack types handling different workloads. Jensen says Nvidia’s TAM grew by 33–50% just from adding those four new rack types to what was previously a one-rack company.

    4. The $50 Billion Factory Produces the Cheapest Tokens

    Critics argue that Nvidia’s inference factories cost $40–50B versus competitors at $25–30B. Huang’s rebuttal is clean: don’t equate the price of the factory with the cost of the tokens. A $50B Nvidia factory producing 10x the throughput of a $30B alternative means Nvidia’s tokens are actually cheaper. When land, power, shell, storage, networking, and cooling are already fixed costs, the delta between GPU options is a small fraction of total spend — but the performance difference is enormous.

    5. OpenClaw Is the New OS for Modern Computing

    Jensen spent serious time on Claude’s open-source agentic framework (referred to throughout as “OpenClaw”). His view: it’s not just a product announcement — it’s a computing paradigm shift. OpenClaw has a memory system (short-term scratch, long-term file system), skills/tools, resource management, scheduling, cron jobs, multi-agent spawning, and external I/O. These are the four foundational elements of an operating system. His conclusion: for the first time, we have a personal AI computer — and it’s open source, running everywhere.

    6. Agents Mean Every Engineer Gets 100 Helpers

    Jensen’s internal benchmark at Nvidia: if a $500K/year engineer isn’t spending at least $250K worth of tokens annually, something is wrong. He compared it to a chip designer refusing to use CAD tools and working only in pencil. His vision: every engineer will have 100 agents working alongside them. The nature of programming shifts from writing code to writing ideas, architectures, specifications, and evaluation criteria — and then guiding agents toward outcomes.

    7. Physical AI Is a $50 Trillion Opportunity

    This is the biggest framing in the talk. Physical AI — robotics, autonomous vehicles, industrial automation, agriculture, healthcare instruments — represents the technology industry’s first real shot at a $50 trillion market that has been “largely void of technology until now.” Nvidia started this journey 10 years ago, it’s now inflecting, and it’s already approaching $10 billion/year as a standalone business. Huang expects this to grow exponentially.

    8. Robots Are 3–5 Years Away from Ubiquity

    Huang was asked about the “lost decade” of robotics — Google buying and selling Boston Dynamics, years of underwhelming progress. His take: America got into robotics too soon, got exhausted, and quit about five years before the enabling technology (AI “brains”) appeared. Now the brain is here. From a “high-functioning existence proof” (what we have now) to “reasonable products,” technology historically takes 2–3 cycles — meaning 3 to 5 years. He also flagged China’s formidable position in robotics hardware: motors, rare earth elements, magnets, micro-electronics. The world’s robotics industry will depend heavily on China’s supply chain.

    9. Jensen Thinks Dario Amodei Is Too Conservative

    Dario Amodei publicly predicted that AI model and agent companies will generate hundreds of billions in revenue by 2027–28 and reach $1 trillion by 2030. Jensen’s response: “I think he’s being very conservative. Way better than that.” His reasoning? Dario hasn’t fully accounted for the fact that every enterprise software company will become a reseller of AI tokens — a logarithmic expansion of go-to-market that will dwarf what any AI lab can sell directly.

    10. The AI Moat Is Deep Specialization

    When asked what the real competitive moat is at the application layer, Jensen said: deep specialization. General models will handle general intelligence. But every industry has domain expertise that needs to be captured in specialized sub-agents, trained on proprietary data. The entrepreneur who knows their vertical better than anyone else, connects their agent to customers first, and builds that flywheel — that’s the moat. He framed it as an inversion of traditional software: instead of building horizontal platforms and customizing at the edges, AI enables you to go vertical-first from day one.

    11. Jensen’s Gentle but Clear Critique of Anthropic’s Communications

    Asked what advice he’d give Anthropic following the Department of Defense controversy that created a PR crisis, Jensen praised Anthropic’s technology and their focus on safety — then offered a measured but pointed critique: warning people is good, scaring people is less good. He argued that AI leaders need to be more circumspect, more humble, more moderate. Making extreme, catastrophic predictions without evidence can damage public trust in a technology that is “too important.” His implicit warning: look what happened to nuclear energy. A 17% public approval rating for AI is the beginning of that same problem.

    12. China Policy: Back to Market, With Conditions

    Nvidia had a 95% market share in China — and lost it entirely due to export controls, falling to 0%. Jensen confirmed that Nvidia has received approved licenses from Secretary Lutnik to sell back into China, has received purchase orders from Chinese companies, and is actively ramping up its supply chain to ship. His broader point: the risk isn’t selling chips to China — the real risk is America becoming so afraid of AI that its own industries don’t adopt it while the rest of the world surges ahead.

    13. Taiwan, Supply Chain, and Geopolitical Risk

    Jensen laid out a three-part strategy for de-risking around Taiwan: (1) Re-industrialize the US as fast as possible — he said Arizona, Texas, and California manufacturing is accelerating with Taiwan’s help as a strategic partner. (2) Diversify the supply chain to South Korea, Japan, and Europe. (3) Demonstrate restraint — don’t press unnecessarily while building resilience. He also noted that Taiwan’s partnership has been genuine and deserves recognition and generosity in return.

    14. Data Centers in Space

    Not science fiction — Nvidia already has CUDA running in satellites doing AI imaging processing in orbit. The near-term thesis: it’s more efficient to process satellite imagery in space than beam raw data back to Earth. The longer-term architecture for space-based data centers is being explored, with radiation hardening already solved. The main challenge is cooling — in the vacuum of space, you can only use radiation cooling, which requires very large surface areas.

    15. Healthcare: Near the ChatGPT Moment for Digital Biology

    Jensen believes digital biology is approaching its own ChatGPT inflection point — the moment where representing genes, proteins, cells, and chemicals becomes as natural as language modeling. He flagged companies like Open Evidence and Hippocratic AI as examples of where agentic healthcare is already working. His vision: every hospital instrument — CT scanners, ultrasound devices, surgical robots — will become agentic, with “OpenClaw in a safe version” running inside each one.

    16. Open Source and Closed Source Will Both Win

    Jensen pushed back on the idea that open source vs. proprietary is an either/or question. It’s both, necessarily. Proprietary models (OpenAI, Anthropic, Gemini) will continue to serve the general horizontal layer — and consumers love having options with distinct personalities. But industries need open models they can specialize, fine-tune, and control. The open model ecosystem, including Chinese models, is “near the frontier” and growing fast. His framework: connect to the best available model today via a router, and use that time to cost-reduce and fine-tune your specialized version.

    17. Advice for Young People: Master AI, Go Deep on Science

    Jensen’s advice for students deciding what to study: deep science, deep math, and strong language skills — because language is the programming language of AI. He made a striking claim: the English major might end up being the most successful professional in the AI era. His one non-negotiable: whatever you study, become deeply expert at using AI tools. And he used radiologists as proof that AI doesn’t destroy jobs — when AI did 100% of the computer vision work in radiology, demand for radiologists went up, not down, because the total number of scans possible exploded.

    📋 Detailed Summary

    The Groq Acquisition and Disaggregated Inference

    The conversation opened with the Groq acquisition — a deal Chamath jokingly said made him “insufferable” during the six-week close. Jensen explained the strategic logic: as Nvidia evolved from running large language models to running full agentic systems, the compute problem became radically more complex. Agentic workloads involve working memory, long-term memory, tool use, inter-agent communication, and diverse model types (autoregressive, diffusion, large, small). No single chip type handles all of this optimally.

    The solution is disaggregated inference — routing different parts of the processing pipeline to the most efficient hardware. Groq’s LPU chips are particularly suited to certain inference tasks. Nvidia’s Vera Rubin system now encompasses five rack types where it used to be one: GPU compute, networking processors, storage processors (BlueField), CPUs, and now LPUs. Jensen’s TAM math: the addition of those four rack types grew Nvidia’s addressable market in any given data center by 33–50% overnight.

    The operating system managing all of this is Dynamo, which Jensen introduced 2.5 years ago — a deliberate reference to the Siemens dynamo machine that powered the first industrial revolution. Dynamo orchestrates workloads across this heterogeneous compute landscape, optimizing for cost, speed, and efficiency.

    Decision-Making at the World’s Most Valuable Company

    Asked how he allocates attention and makes strategic calls at a $350B+ revenue company, Jensen gave a surprisingly simple framework: pursue things that are insanely hard, that have never been done before, and that tap into Nvidia’s specific superpowers. If something is easy, competitors will flood in. If it’s hard and unique, the pain and suffering of building it becomes a moat in itself. He explicitly said he enjoys the pain — and that there’s no great invention that came easily on the first try.

    Physical AI and the Three Computers

    Jensen framed Nvidia’s physical AI strategy around three distinct computers:

    1. The Training Computer — for developing and creating AI models.
    2. The Simulation Computer (Omniverse) — for evaluating AI systems inside physics-accurate virtual environments (required for robotics and autonomous vehicles that can’t be tested purely in the real world).
    3. The Edge Computer — deployed in cars, robots, factory floors, teddy bears, and telecom base stations. Jensen flagged that the $2 trillion global telecom industry is being transformed into an extension of AI infrastructure — turning radio base stations into AI edge devices.

    Physical AI is, by Jensen’s estimate, the technology industry’s first real crack at the $50 trillion industrial economy. He started the investment 10 years ago. It’s now approaching $10 billion annually and growing exponentially.

    OpenClaw as the New Operating System

    Jensen’s analysis of OpenClaw (Anthropic’s open-source agentic framework, referred to as “Claude Code” / “Open Claude” throughout) was one of the most intellectually interesting sections of the interview. He traced three cultural inflection points:

    1. ChatGPT — put generative AI into the popular consciousness by wrapping the technology in a usable interface.
    2. Reasoning models (o1, o3) — shifted AI from answering questions to answering them with grounded, verifiable reasoning, driving economic model inflection at OpenAI.
    3. OpenClaw — introduced the concept of agentic computing to the general population. But more importantly, it defined a new computing architecture: memory (short and long-term), skills, resource scheduling, IO, external communication, and agent spawning. These are the four elements of an operating system. OpenClaw is, in Jensen’s view, the blueprint for what a personal AI computer looks like — open source, running everywhere.

    He also flagged that Nvidia contributed security governance work to OpenClaw alongside Peter Steinberger — ensuring agents with access to sensitive information, code execution, and external communication can be properly governed with appropriate policy constraints.

    The Agentic Future and Token Economics

    Jensen’s internal benchmark for token spending at Nvidia was striking: a $500K/year engineer who isn’t spending $250K/year in tokens is underperforming. He framed this as no different from a chip designer refusing to use CAD software. The implication for enterprise economics is profound: the cost basis of AI in a company isn’t an IT line item — it’s a multiplier on every knowledge worker’s output.

    He also addressed Andrej Karpathy’s “autoresearch” concept — the idea of AI systems that autonomously run research experiments. A guest described completing, in 30 minutes on a desktop, a genomics analysis that would normally constitute a seven-year PhD thesis. Jensen’s response: this isn’t a fluke. It’s the beginning of a fundamental shift in what “doing science” means.

    His forecast on compute scaling: generative to reasoning = 100x. Reasoning to agentic = 100x. Total in two years = 10,000x. And the end state isn’t even close yet — he believes the long-run trajectory is 1 million times current compute levels.

    AI’s PR Crisis and Anthropic’s Comms Mistakes

    This segment was diplomatically delivered but substantively sharp. Jensen opened by genuinely praising Anthropic — their technology, their safety focus, their culture of excellence. Then he drew a distinction: warning people about AI capabilities is good and important. Scaring people with extreme, catastrophic predictions for which there’s no evidence is less good, and potentially very damaging.

    He pointed to the nuclear analogy: public fear of nuclear energy, driven partly by technology leaders’ own alarming statements, effectively killed the US nuclear industry. America now has zero new fission reactors while China builds a hundred. AI’s 17% public approval rating in the US is the beginning of the same dynamic. Jensen said the greatest national security risk from AI isn’t what other countries do with it — it’s the US being so afraid of it that American industries fail to adopt it while the rest of the world surges ahead.

    His prescription for AI leaders: be more circumspect, more humble, more moderate. Acknowledge that we can’t completely predict the future. Avoid statements that are extreme and unsupported by evidence. Our words matter in a way they didn’t used to — technology leaders are now central to the national security and economic policy conversation.

    China Policy: Return to Market

    One of the more concrete news items in the interview: Nvidia is returning to the Chinese market. Jensen confirmed they had a 95% market share in China — and fell to 0% due to export controls. They’ve now received approved licenses from Secretary Lutnik, Chinese companies have issued purchase orders, and Nvidia is ramping its supply chain to ship.

    His framework for the right AI export policy outcome: the American tech stack — from chips to computing systems to platforms — should be used by 90% of the world as the foundation on which other countries build their own AI. The alternative — an AI industry that ends up like solar panels, rare earth minerals, motors, and telecom infrastructure (all dominated by China) — is a national security catastrophe.

    Self-Driving and Competitive Positioning

    Jensen laid out Nvidia’s strategy in autonomous vehicles: they don’t want to build self-driving cars — they want to enable every car company to build them. Nvidia supplies all three computers: training, simulation, and the in-car edge computer. Their autonomous driving AI system, called “Al Pomayo,” introduced reasoning capabilities into autonomous vehicles — decomposing complex scenarios into simpler ones the system knows how to navigate.

    On competition from customers (Google TPU, Amazon Inferentia, etc.): Jensen isn’t worried. His argument is that 40% of Nvidia’s business comes from customers who don’t just want chips — they need the full AI factory stack. CUDA isn’t just a chip instruction set; it’s a system. Companies that have tried to build their own silicon have found that chips without the full stack don’t solve the problem. Meanwhile, Nvidia is gaining market share, including pulling in Anthropic and Meta as Nvidia customers, and AWS just announced a million-chip order.

    Robotics: 3–5 Years to Everywhere

    Jensen’s robotics take was both bullish and grounded. America invented modern robotics, got too early, got exhausted, and quit just before the AI brain appeared that would make it work. That brain is here now. From the current “existence proof” stage to “reasonable products,” he sees 3–5 years. His aspiration: more than one robot per human on Earth. The use cases he described range from factory floor automation to virtual presence (using your home robot as an avatar while traveling), to lunar and Martian factories run entirely by robots with materials beamed back to Earth at near-zero energy cost.

    China’s position in robotics is formidable and can’t be wished away: they lead in micro-electronics, motors, rare earth elements, and magnets — all foundational to building robot hardware. The world’s robotics industry, including the US, will depend heavily on China’s supply chain for hardware components even if American software and AI lead.

    Revenue Forecasts: Dario Is Too Conservative

    When the hosts described Dario Amodei’s forecast of hundreds of billions in AI model/agent revenue by 2027–28 and $1 trillion by 2030, Jensen said simply: “Way better than that.” His reason: Dario hasn’t fully factored in that every enterprise software company will become a value-added reseller of AI tokens — OpenAI’s, Anthropic’s, whoever’s. The go-to-market expansion that comes from every SAP, Salesforce, and ServiceNow reselling AI is logarithmic, not linear.

    Healthcare: Near the Inflection Point

    Jensen named three layers of Nvidia’s healthcare involvement: (1) AI biology/physics — using AI to represent and predict biological behavior for drug discovery; (2) AI agents — agentic systems for diagnosis assistance, first-visit intake, and clinical decision support (he named Open Evidence and Hippocratic AI as leading examples); (3) Physical AI for healthcare — robotic surgery, AI-enabled instruments, and the vision of every hospital device (CT, ultrasound, surgical tools) becoming agentic. He sees digital biology as approaching its ChatGPT moment — the point where representing genes, proteins, and cells computationally becomes as natural and powerful as language modeling.

    Career Advice: Go Deep, Use AI

    Jensen closed with career guidance. His core advice: study deep science, deep math, and language — because language is now the programming language of AI. He made the counterintuitive claim that English majors may end up being the most successful professionals in the AI era because the ability to specify, guide, and evaluate AI outputs is an artform — and it’s not trivial. The person who knows how to give AI enough guidance without over-prescribing, who can recognize a great AI output from a mediocre one, and who can orchestrate teams of agents toward outcomes — that’s the most valuable skill.

    He used the radiologist story as his closing proof point: when computer vision was integrated into radiology, demand for radiologists went up, not down. The number of scans exploded, hospitals made more money, and more patients got diagnosed faster. AI didn’t replace radiologists — it made them bionic and made the whole system bigger. He expects the same pattern everywhere: every job will be transformed, some tasks will be eliminated, but the total pie grows dramatically.

    💭 Thoughts

    Jensen Huang is doing something rare among tech CEOs: he’s genuinely trying to build the mental model people need to understand what’s happening — not just sell products. The disaggregated inference argument, the three-computer framework, the OS analogy for OpenClaw, the token economics benchmark — these aren’t talking points. They’re conceptual tools for thinking clearly about a landscape most people are still squinting at.

    The most underappreciated part of the interview is the AI PR section. Jensen is essentially sounding an alarm without panicking: if America’s technology leaders keep scaring the public with AI doomerism, we will repeat the nuclear mistake. We’ll regulate ourselves into irrelevance while China builds the infrastructure we refused to build. The 17% approval number he cited should frighten every AI optimist in the room. Fear of a technology, once embedded culturally, is very hard to dislodge.

    The Anthropic critique was surgical. He didn’t name the specific controversy, didn’t pile on, and praised their technology extensively. But the message was clear: extreme safety warnings, even well-intentioned ones, carry real costs in the public square. That’s a genuinely hard tension for safety-focused AI companies, and there’s no clean answer — but Huang’s instinct that humility and circumspection serve better than catastrophism seems directionally correct.

    The physical AI thesis deserves more attention than it gets. Everyone is focused on the software intelligence race — OpenAI vs. Anthropic vs. Gemini. But Jensen is pointing at a $50 trillion industrial economy that AI has barely touched. Robotics, autonomous vehicles, agricultural automation, smart hospital instruments — this is where the real mass of economic value is locked. And Nvidia’s ten-year head start on the enabling infrastructure for physical AI may turn out to be more durable than any software moat.

    Finally: the robot optimism is infectious and probably correct. The world is genuinely short millions of workers. The enabling technology — AI brains good enough to drive perception, reasoning, and action in unstructured physical environments — just arrived. The hardware supply chain is largely intact. And the economic incentive to automate is stronger than it’s ever been. Three to five years feels aggressive. But so did “ChatGPT will change everything” in 2021.