PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: Nemotron

  • SubQ 1.1 Small Explained: How Subquadratic Sparse Attention Hits 98% Retrieval at 12 Million Tokens With 64.5x Less Compute Than Dense Attention

    Subquadratic, a frontier AI research and infrastructure company, has released the model card and technical report for SubQ 1.1 Small, a long-context language model built on a new attention mechanism the company calls Subquadratic Sparse Attention (SSA). The headline claim is unusual in two directions at once: the model retains 98% single-fact retrieval accuracy at 12 million tokens, roughly twelve times the length it was primarily trained on, while cutting attention compute by 64.5x against dense attention at a 1 million token context. The deeper argument in the report is not really about a single model at all. It is about what happens to the entire retrieval-and-orchestration stack once reasoning over a complete artifact stops being prohibitively expensive.

    TLDR

    SubQ 1.1 Small is a small long-context model that replaces the dense attention of an existing open-weight frontier model with Subquadratic Sparse Attention, a learned, content-dependent sparse attention mechanism that scales linearly in compute and memory rather than quadratically. On retrieval it posts 99.12% on NVIDIA’s 13-task RULER suite at 128K tokens and 100% needle-in-a-haystack accuracy at 1M and 2M tokens, holding at 98% out to 6M and 12M tokens while attending to only 0.13% of token pairs. It keeps competitive general ability, scoring 85.4% on GPQA Diamond and 89.7% pass@4 on LiveCodeBench v6, and reaches 13% on the long-horizon AutomationBench Finance agentic benchmark, close to Opus 4.8 and GPT-5.5 and well ahead of mid and small tiers. The efficiency story is a scaling win rather than a constant-factor one: 64.5x fewer attention FLOPs than dense attention at 1M tokens and 56x faster than FlashAttention-2 on a single attention layer. The report frames cheap long-context compute as a research accelerator that let the team run more than one hundred million-token experiments and find a training recipe (long-context continued pretraining is the strongest lever) rather than guess at one, positions SSA against FlashAttention, DeepSeek’s Lightning Indexer line, state space models like Mamba, and hybrids, invokes Sutton’s Bitter Lesson to argue that RAG, chunking, and agentic scaffolding are partly workarounds for context scarcity, and was independently verified by Appen. Deployment is starting with design partners now, with a 2M to 12M token lineup planned by year end.

    Thoughts

    The most interesting move in this report is the framing, not the benchmark. Subquadratic plants its flag on Richard Sutton’s Bitter Lesson and argues that much of the modern AI stack, the retrieval pipelines, the chunkers, the re-rankers, the agentic orchestration, is scaffolding built around a single computational constraint: dense attention costs grow with the square of context length. If that constraint relaxes, a lot of hand-engineered machinery that exists to feed a model the right fragments at the right moment starts to look like the task-specific pipelines that learned representations eventually displaced. That is a genuinely provocative thesis, and it is the right lens for reading the rest of the document. The company is not selling a longer context window as a feature. It is betting that whole-artifact reasoning is a different shape of capability than retrieval over fragments, and that fragmentation destroys the cross-references a contract or a codebase actually depends on before the model ever sees them.

    The part of the paper most teams will undervalue is the claim that the real payoff of efficient attention is not cheaper inference but cheaper experimentation. A dense long-context training campaign is expensive enough that most groups get a handful of attempts and are forced to guess at the recipe. Subquadratic says SSA let them run more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context, which is how they discovered that long-context continued pretraining, not clever post-training, was the dominant lever. If that holds, algorithmic efficiency becomes a first-class scaling variable alongside parameters and data, because capability becomes responsive to iteration velocity rather than raw compute alone. It reframes efficiency from a deployment line item into a research multiplier, and that is a more durable advantage than any single benchmark number.

    The generalization result deserves scrutiny precisely because it is so clean. A model trained overwhelmingly at 1M tokens, with a sliver at 2M and nothing beyond, holds 98% retrieval at 12M. The proposed explanation is that SSA routes attention by content relevance rather than fixed positional pattern, so there may simply be no obvious length boundary once the routing behavior is learned. That is plausible and the report is careful to say the 12M result emerged rather than being designed for. But single-needle NIAH is a deliberately clean probe with one target and a binary answer. The far harder RULER suite is only reported at 128K, the longest standardized length in the original benchmark, so the multi-hop, aggregation, and distractor-heavy capability that whole-artifact reasoning actually requires has public numbers at 128K, not at 12M. The honest read is that precise retrieval generalizes spectacularly and composite reasoning at extreme length is still an open question the report does not over-claim on.

    What lends the report credibility is how much counter-evidence it volunteers. It walks through MiniMax abandoning its hybrid M1 architecture and returning to full attention for M2 after efficient variants showed multi-hop reasoning deficits at scale. It admits that earlier SubQ checkpoints improved retrieval while regressing on knowledge benchmarks, forcing dedicated capability-balancing work. It describes catching a case where the MRCR benchmark moved up while the model felt worse in real workflow spot-checks, and switching its development signal to RULER as a result. That last point is a quietly important methodological argument: benchmark score and deployment behavior diverged enough to change checkpoint selection, which is a warning every team shipping long-context models should internalize. A vendor confident enough to show where its own metrics misled it is more trustworthy than one that only shows the wins.

    A few caveats keep the enthusiasm grounded. AutomationBench Finance at 13% is genuinely strong relative to peers, but it is a low absolute score across the board, including for GPT-5.5 at 18% and Opus 4.8 at 16%, so this is early evidence of agentic transfer rather than proof of a finished agent. The efficiency comparisons isolate a single attention layer rather than full end-to-end model throughput, which is the right way to expose the scaling shape but not the same as a wall-clock serving benchmark. The model is built from an unnamed donor open-weight frontier model, so some of its general-knowledge and coding strength is inherited rather than created here. And the most aggressive claims about the future, a 2M to 12M lineup and much higher sparsity, are roadmap, not released artifacts. None of that undercuts the core result. It just means the right posture is to treat SubQ 1.1 Small as a strong proof of concept for an architecture that, if it scales as advertised, could quietly remove a layer of the AI stack that everyone currently takes for granted.

    Key Takeaways

    • SubQ 1.1 Small is a long-context language model from Subquadratic AI, built on a new attention mechanism called Subquadratic Sparse Attention (SSA), released June 16, 2026 alongside a model card and technical report.
    • SSA is a learned, content-dependent sparse attention mechanism that scales linearly in both compute and memory with sequence length, rather than quadratically like dense attention.
    • The central result is context-length generalization: the model was trained primarily at 1M tokens, with some training at 2M and none beyond, yet retrieval held far past the training window.
    • Needle-in-a-haystack accuracy is 100% at 1M and 2M tokens and 98% at both 6M and 12M tokens, roughly twelve times the primary training length.
    • At 12M tokens the model attends to only 0.13% of token pairs, close to a 1,000x reduction in attention relationships, while still retrieving accurately.
    • On NVIDIA’s 13-task RULER benchmark at 128K tokens, SubQ 1.1 Small scores 99.12%, with the remaining errors concentrated in aggregation-style tasks rather than retrieval.
    • RULER tests beyond single-fact lookup: single-key and multi-key retrieval, common-word and frequent-word extraction, and multi-hop variable tracing across positions.
    • At 1M tokens, SSA requires 64.5x fewer attention FLOPs than dense attention (3.9 PFLOP versus 252 PFLOP per attention layer).
    • On a single attention layer, SSA runs 56x faster than FlashAttention-2 at 1M tokens (966 ms versus 54,164 ms on an H100), reaching parity near 16K tokens and pulling away as context grows.
    • The efficiency gain is a scaling-law win, not a constant-factor speedup: the advantage over dense attention grows as context length increases.
    • On general knowledge, SubQ 1.1 Small scores 85.4% on GPQA Diamond (pass@1), below GPT-5.5 (93.2) and Opus 4.8 (92), near Sonnet 4.6 and GPT-5.4-mini (87.5), and above GPT-5.4-nano (81.7) and Haiku 4.5 (67.2).
    • On coding, it reaches 89.7% pass@4 on LiveCodeBench v6, close to the absolute frontier (GPT-5.5 92, Opus 4.8 92.2) and ahead of the smaller tiers.
    • On AutomationBench Finance, a long-horizon agentic benchmark, it scores 13%, close to Opus 4.8 (16%) and GPT-5.5 (18%) and ahead of Sonnet 4.6 (8%), Haiku 4.5 (3%), and GPT-5.4-mini (0%). Absolute scores are low across all models.
    • The model was not trained from scratch. The team converted an existing open-weight frontier model by replacing dense attention with SSA, then built long-context ability through staged context extension and continued pretraining.
    • Context was extended in stages (262K, 512K, 1M, 2M) using YaRN positional scaling, with long-context continued pretraining performed between extension stages on naturally long data: books, long documents, and repository-scale code.
    • Roughly one trillion tokens of continued pretraining were performed, most of it at the 1M-token stage.
    • Long-context continued pretraining was the most consistent predictor of long-context retrieval gains across the experiments, more so than post-training tweaks.
    • The team ran more than one hundred long-context experiments across six major model generations, which the report argues is only possible because SSA made million-token iteration cheap (under a minute per step).
    • Capability balance was a recurring challenge: gains in long-context retrieval often regressed short-context knowledge and reasoning unless training was explicitly managed for both.
    • Benchmark scores and real deployment behavior diverged. The MRCR benchmark moved up while qualitative workflow spot-checks got worse, so the team switched its primary development signal to RULER.
    • The report frames RAG, chunking, summarization, and agentic orchestration as scaffolding built around context scarcity, drawing an analogy to Sutton’s Bitter Lesson, where hand-engineered mechanisms get displaced by larger-scale learning.
    • SSA is positioned against FlashAttention (a memory optimization that does not change quadratic compute), fixed-pattern sparse attention, DeepSeek’s learned sparse line, state space models, and hybrid architectures.
    • DeepSeek’s Lightning Indexer (used in DSA and CSA) is the closest published comparison. Its quadratic scoring overtakes the sparse attention it feeds around 52,000 tokens, reaching roughly 16x the attention cost at 1M and 190x at 12M.
    • State space models like Mamba achieve linear cost through a compressed fixed-size state, but that compression is lossy and weakens exact retrieval, which is why production efficient models are usually hybrids with some dense attention layers retained.
    • MiniMax is cited as a cautionary case: it moved from a hybrid M1 to a full-attention M2 after hybrids showed multi-hop reasoning deficits at scale and less mature supporting infrastructure.
    • The benchmark results were independently verified by Appen, a third-party evaluation firm.
    • The named use cases are financial analysis and due diligence, legal and contract work, and software engineering (architecture-level reasoning, cross-file refactoring, dependency tracing, planning, review, and long-horizon memory).
    • Sparsity settings were deliberately conservative, tuned for maximum context length rather than maximum sparsity. Limited experiments at 4x the sparsity reported positive early results.
    • The training infrastructure used a memory-scaling ladder: single node, intra-node sequence parallelism, CPU offload, multi-node sequence parallelism, nested offloading, and Ring Attention for the longest contexts.
    • Beyond about 8M tokens, BF16 numerical underflow and stability became practical constraints on evaluation.
    • The technical report is authored by Saul Ramirez, Alex Whedon, Ashmal Vayani, and Phong Vo of Subquadratic AI.
    • Deployment is starting with a first cohort of design partners, with broader rollout through the quarter and a general model lineup ranging from 2M to 12M tokens by the end of the year.
    • The company’s framing line is “Efficiency is intelligence,” and its broader thesis is that the point is not bigger context windows for their own sake but reasoning directly over complete artifacts with less surrounding scaffolding.

    Detailed Summary

    The problem: whole-artifact reasoning and context scarcity

    The report opens by naming a class of tasks it calls whole-artifact reasoning: problems whose structure requires reasoning across a complete artifact rather than over isolated fragments. A legal agreement may define a term on page 2, qualify it on page 12, carve out an exception on page 46, and amend it in a schedule. A function may be defined in one file, called from forty others, and constrained by invariants encoded in the architecture rather than in comments. A financial review may require connecting filings, earnings reports, contracts, and internal records. In each case the difficulty is not locating a passage, it is reasoning over relationships distributed throughout a large artifact. Most production systems do not do this directly. They rely on retrieval pipelines, chunking, summaries, and agentic workflows that partition information and reconstruct fragments at inference time, because dense attention scales quadratically with context length and makes direct reasoning over large artifacts expensive. Subquadratic argues that much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts, and it connects this to Sutton’s Bitter Lesson: sophisticated hand-engineered mechanisms historically get displaced once larger-scale learning becomes practical.

    What SSA is and the three requirements it targets

    Subquadratic Sparse Attention is a content-dependent sparse attention mechanism designed to satisfy three requirements at once, a combination the report argues prior approaches never achieved in a practical long-context system. First, dense-attention-level retrieval and reasoning quality, which requires routing that is content-dependent (determined by the tokens themselves) rather than driven by a fixed positional pattern. Second, subquadratic scaling, where selection, retrieval, and attention are each linear in sequence length so the mechanism is linear end to end, not only within the attention read. Third, full-context training with standard autoregressive generation, so the model can optimize over the entire context during training while keeping efficient token-by-token decoding at inference. The internal mechanism by which SSA achieves this is held back as outside the scope of the report, which focuses instead on the requirements and the experimental program that followed.

    Where SSA sits among prior approaches

    The background section is effectively a taxonomy of long-context modeling. FlashAttention is treated not as a competitor but as the standard dense-attention baseline: it solved the memory problem by never materializing the full attention matrix, but it left the quadratic compute cost untouched, so doubling context still quadruples attention computation. Fixed-pattern sparse attention (sliding-window, strided, as in Longformer, BigBird, and the sliding window in Gemma) scales well but sacrifices content-dependent routing and tends to fail on retrieval benchmarks like RULER. Compression methods like Multi-head Latent Attention reduce KV-cache memory at inference but do not change the quadratic prefill cost. Learned sparse attention, exemplified by DeepSeek’s Native Sparse Attention and its Lightning Indexer, learns where to route but pays a quadratic cost in the indexer itself. State space models and linear attention (Mamba, Mamba-2 and Mamba-3, RetNet, RWKV, gated delta networks) achieve linear cost through a compressed fixed-size state, but that compression is lossy and weak on exact retrieval. Hybrids (Jamba, Kimi Linear, Qwen3 Next, Nemotron) keep a few dense layers to preserve retrieval, which means the quadratic component still dominates at long context. System-level workarounds (RAG, agentic frameworks, recursive language models) move retrieval outside the model entirely. The report’s stated open problem is to combine subquadratic scaling end to end with content-dependent retrieval, arbitrary-position access, and practical ultra-long-context training in one system, which it claims no widely deployed architecture provides and which SSA targets.

    Training: conversion, staged context extension, and continued pretraining

    Rather than training from scratch, the team converted an existing open-weight frontier model that supported a 262K-token context by replacing its dense attention with SSA. They then extended the context window in stages (262K to 512K to 1M to 2M) using YaRN to rescale positional representations, performing long-context continued pretraining between extension stages rather than jumping straight to the final length. The training mixture emphasized naturally long data such as books, long documents, and repository-scale code, packed to the target length with document separators and without masking cross-document attention boundaries. Most continued-pretraining tokens were trained at the 1M-token stage, with roughly one trillion tokens total. Post-training played a separate role: shaping how the long-context capability was expressed while preserving reasoning, coding, and instruction following. The team explored sample-level loss aggregation to keep a few extremely long examples from dominating gradient updates, and staged the post-training corpus across synthetic retrieval tasks, long-context reasoning, coding, educational material, and general instruction following, alternating capability-building phases with recovery phases.

    Results: retrieval, knowledge, coding, and agentic tasks

    On retrieval, SubQ 1.1 Small scores 99.12% on the 13-task RULER average at 128K, with errors concentrated in aggregation-style tasks like common-word and frequent-word extraction. On needle-in-a-haystack, evaluated on 50 held-out UUID samples per length, it scores 100% at 1M and 2M (within the training window) and 98% at 6M and 12M (held out), attending to only 0.13% of token pairs at 12M. On knowledge, GPQA Diamond pass@1 is 85.4%, landing between the small and mid frontier tiers and confirming that long-context optimization need not sacrifice reasoning, a result the report credits to its capability-balancing stages after earlier checkpoints showed retrieval gains coming at the cost of knowledge. On coding, LiveCodeBench v6 pass@4 is 89.7%, and the report notes coding data played a dual role, also improving non-code long-context retrieval because code is dense with the cross-position dependencies that train general routing. On long-horizon agentic work, AutomationBench Finance is 13%, where agents must discover the right endpoints among roughly 500 across 47 applications, make interdependent API calls, follow layered business rules, and ignore seeded distractors, graded on binary end-state correctness with no partial credit.

    Efficiency and the DeepSeek comparison

    Efficiency is measured on one attention layer against a dense baseline on the same backbone. Per-forward-pass attention FLOPs scale from a 2.1x reduction at 32K to 8x at 128K, 31.5x at 512K, and 64.5x at 1M tokens (3.9 PFLOP for SSA versus 252 PFLOP for dense). Measured against FlashAttention-2 in isolation, SSA reaches parity near 16K tokens and pulls away to 56x at 1M, where it runs in 966 ms versus 54,164 ms on an H100. The report devotes a discussion section to DeepSeek’s sparse attention line as the closest published comparison. DeepSeek’s Lightning Indexer is a learned selector, but it is a full-attention distilled transformer, so it scales quadratically: in a V3.2-style configuration the indexer is cheaper than the sparse attention it feeds only below about 52,000 tokens, then overtakes it, reaching roughly 16x the attention cost at 1M tokens and 190x at 12M. SSA targets that same selection role with a selector the report says is dramatically cheaper and linear throughout, and notes SSA could conceptually replace the selector over either uncompressed or compressed representations.

    Efficiency as a research accelerator and the evaluation lessons

    A recurring theme is that the most valuable effect of cheap long-context compute was on the research loop, not just inference. Where a dense campaign would allow a handful of attempts, SSA enabled more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context. That throughput is what surfaced the finding that long-context continued pretraining is the strongest lever, and it leads the authors to argue that algorithmic efficiency should be treated as a first-class scaling variable alongside model and dataset size. The report is unusually candid about evaluation pitfalls. It describes how the MRCR benchmark diverged from deployment behavior, with MRCR-optimized checkpoints often feeling worse on repository-scale code reasoning, multi-document synthesis, and contract analysis, which pushed the team to rely on RULER and a fixed set of qualitative workflow spot-checks as development signals. It also cites MiniMax returning from a hybrid M1 to a full-attention M2 as evidence that reducing asymptotic cost is not sufficient on its own if retrieval quality, reasoning at scale, and system maturity are not preserved at the same time.

    Implications, availability, and what comes next

    The report’s deployment argument is that the most important enterprise implication of long-context models is not larger windows but the ability to reason directly over complete or more-complete artifacts, moving retrieval, re-ranking, and orchestration logic into the model where the task is naturally whole-artifact rather than naturally decomposable. It is careful not to declare retrieval obsolete: for corpora larger than any plausible context window, fast-changing knowledge, and genuinely multi-stage workflows, RAG and orchestration remain the right tools. The narrower claim is that the class of scaffolding that exists only to compensate for context limits gets smaller as efficient long-context models extend the reachable window. The benchmark results were independently verified by Appen. Subquadratic is deploying SubQ 1.1 Small with a first cohort of design partners now, with broader rollout through the quarter and a general lineup spanning 2M to 12M tokens planned by the end of the year, and it flags much higher sparsity as future work.

    Notable Quotes

    “Much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts directly.”

    SubQ-1.1-Small Technical Report, framing retrieval and orchestration as workarounds for an architectural limit

    “The hybrid has moved the line, but not changed its shape.”

    SubQ-1.1-Small Technical Report, on why hybrid models keep their quadratic component at long context

    “A routing mechanism intended to make long context affordable becomes the dominant long-context cost, reintroducing quadratic scaling after providing scalar compute savings.”

    SubQ-1.1-Small Technical Report, on DeepSeek’s Lightning Indexer overtaking the attention it feeds

    “If the cost of long-context experiments is too high, teams are forced to guess at the recipe. If the cost falls far enough, they can search for it.”

    SubQ-1.1-Small Technical Report, on efficient attention as a research accelerator

    “Fragmentation systematically destroys those relationships before the model ever sees them.”

    SubQ-1.1-Small Technical Report, on why chunking hurts whole-artifact reasoning

    “Holding the whole artifact in context changes the shape of the task rather than only the speed of it.”

    SubQ-1.1-Small Technical Report, on the difference between bigger windows and direct reasoning

    “The value of SSA is therefore not only that it makes long-context inference cheaper. It makes long-context experimentation cheaper.”

    SubQ-1.1-Small Technical Report, conclusion

    Read the full SubQ 1.1 Small technical report and model card here.

    Related Reading

    • Subquadratic (subq.ai) the company behind SubQ 1.1 Small and the Subquadratic Sparse Attention architecture, where you can join the waitlist.
    • The Bitter Lesson by Richard Sutton the short essay whose argument the report leans on, that hand-engineered mechanisms lose to general methods that scale with computation.
    • Attention Is All You Need the original Transformer paper that introduced the dense attention whose quadratic cost SSA is built to remove.
    • RULER (arXiv) NVIDIA’s long-context benchmark that the report uses as its primary retrieval signal, and that fixed-pattern sparse methods historically struggle with.
    • Retrieval-augmented generation (Wikipedia) background on the RAG approach that the report frames as scaffolding around context scarcity rather than a permanent fixture.

  • Jensen Huang at Stanford CS153 Frontier Systems on Co-Design, Agentic Computing, Vera Rubin, Open Models, and the Million-X Decade That Reshaped AI Infrastructure

    https://www.youtube.com/watch?v=tsQB0n0YV3k

    NVIDIA CEO Jensen Huang returned to Stanford for the CS153 Frontier Systems class (the room nicknamed itself “AI Coachella”) to lay out, in raw form, how he thinks about the computer being reinvented for the first time in over sixty years. Across roughly seventy minutes of student questions he walks through the codesign philosophy that gave NVIDIA a million-x decade, the architectural through-line from Hopper to Grace Blackwell to Vera Rubin to Feynman, the case for open source foundation models, the realities of tokens per watt and MFU, energy demand running a thousand times higher, the China and export-control debate, and his own biggest strategic mistakes. Watch the full conversation on YouTube.

    TLDW

    Huang argues every layer of computing has changed: the programming model, the system architecture, the deployment pattern, the economics. Co-design across CPUs, GPUs, networking, storage, switches and compilers gave NVIDIA roughly a million-x speed-up over ten years versus the ten-x Moore’s Law era, and that headroom is what let researchers say “just train on the whole internet.” Hopper was built for pre-training, Grace Blackwell NVLink72 for inference and reasoning (50x over Hopper in two years), Vera Rubin is built for agents that load long memory, call tools and need a low-latency single-threaded CPU bolted directly to the GPU, and Feynman extends that to swarms of agents that spawn sub-agents. Open weights matter because safety, sovereignty (230-plus languages no one else will fund) and domain models for biology, autonomy, robotics and climate need a foundation that NVIDIA is willing to seed. Compute is not really the scarce resource (Huang says place the order and the chips ship), the broken thing is institutional budgeting that can’t put a billion dollars into a shared university supercomputer. Energy demand is heading a thousand times higher and this is finally the moment market forces alone will fund sustainable generation. On geopolitics he rejects the GPUs-as-atomic-bombs framing and warns America will end up like its telecom industry if it cedes two thirds of the world. On career he advises seeking suffering on purpose. On strategy he says observe, reason from first principles, build a mental model, work backwards, minimize opportunity cost, maximize optionality.

    Key Takeaways

    • The computing model has been substantially unchanged since the IBM System 360, sixty-plus years ago. Huang’s first computer architecture book was the System 360 manual. AI is the first true reinvention.
    • Old computing was pre-recorded retrieval. New computing is generated, contextually aware and continuous. Cloud was on-demand. Agentic systems run continuously.
    • Codesign is NVIDIA’s central thesis. Inherited from the Hennessy and Patterson RISC era at Stanford, extended across CPUs, GPUs, networking, switches, storage, compilers and frameworks all optimized together.
    • The result of full-stack codesign: roughly 1,000,000x faster compute over ten years, versus a generous 10x to 100x for Moore’s Law in the same period. Dennard scaling effectively ended a decade ago.
    • That million-x speed-up is what unlocked “train on all of the internet” as a realistic AI strategy.
    • After GPT, Huang says it was obvious thinking was next. Reasoning is just generating tokens consumed internally, then using tools is generating tokens consumed externally. Agentic systems followed predictably.
    • Education needs AI baked into the curriculum, not just taught as a subject. Pre-recorded textbooks cannot keep pace with knowledge being generated in real time.
    • Huang says he cannot learn anymore without AI. He has the AI read the paper, then read every related paper, then become a dedicated researcher he can interrogate.
    • Mead and Conway and the first-principles methodology of semiconductor design are still worth learning even though most of the scaling tricks have been exhausted.
    • NVIDIA itself is one of the largest consumers of Anthropic and OpenAI tokens in the world. One hundred percent of NVIDIA engineers are now agentically supported. Huang recommends Claude and similar tools by name and says open-source downloads will not match the integrated product harness.
    • NVIDIA still invests heavily in open foundation models because language and intelligence represent the codification of human knowledge. Five pillars: Nemotron (language), BioNeMo (biology), Alphamayo (autonomous vehicles), Groot (humanoid robotics) and a climate science model (mesoscale multiphysics).
    • Sovereign language models matter. Roughly 230 world languages will never be a top priority for a commercial frontier lab. Nemotron is near-frontier and fully fine-tunable so any country can adapt it.
    • Safety and security require open weights. You cannot defend against or audit a black box. Transparent systems let researchers interrogate models and let defenders deploy swarms.
    • The future of cyber defense is not bigger-model-versus-bigger-model. It is trillions of cheap fast small models like Nemotron Nano surrounding the threat.
    • Domain models fuse language priors with world models. Alphamayo learned to drive safely on a few million miles instead of billions because it can reason like a human about the road.
    • MFU (Model Flops Utilization) is a misleading metric. Huang says he wants low MFU, because that means he over-provisioned every resource and never gets pinned by Amdahl’s law during a spike.
    • The xAI Memphis cluster running at 11 percent MFU is not necessarily a failure mode. In disaggregated prefill plus decode inference you can deliver very high tokens per watt with very low MFU.
    • The right metric is performance, ultimately tokens per watt as a proxy for intelligence per watt, and even that needs adjustment because not all tokens are equal. Coding tokens are worth more than other tokens.
    • Hopper was designed for pre-training. NVIDIA chose to build multi-billion-dollar systems when the largest existing scientific supercomputer cost $350 million, with no proven customer base. It worked.
    • Grace Blackwell NVLink72 was designed for inference, especially the high-memory-bandwidth decode phase. It is the world’s first rack-scale computer and delivered a 50x speed-up over Hopper in two years, against an expected 2x from Moore’s Law.
    • Vera Rubin is designed for agents. Long-term memory wired into storage and into the GPU fabric, working memory, heavy tool use, and Vera, a CPU optimized for low-latency multi-core single-threaded code so a multi-billion-dollar GPU system does not stall waiting on a slow tool call.
    • Feynman is being shaped for swarms of agents with sub-agents and sub-sub-agents, a recursive software topology that demands a new compute pattern.
    • Tokens per watt improved 50x in one generation. Compounding energy efficiency is the lever NVIDIA controls directly.
    • Total compute energy demand is heading roughly a thousand times higher than today, possibly two orders of magnitude beyond that. Huang says he would not be surprised if the estimate is low.
    • For the first time in history, market forces alone are enough to fund solar, nuclear and grid upgrades. Government subsidies are no longer required to make sustainable energy investment rational.
    • Copper interconnect is becoming a bottleneck. Photonics is moving from optional to structural inside racks and across them.
    • Comparing NVIDIA GPUs to atomic bombs, Huang says, is a stupid analogy. A billion people use NVIDIA GPUs. He advocates them to his family. He does not advocate atomic bombs to anyone.
    • If the United States cedes two thirds of the global market to competitors on policy grounds, the American technology industry will end up like American telecommunications, which was policied out of existence.
    • Huang directly rejects AI doom-by-singularity narratives. It is not true that we have no idea how these systems work. It is not true that the technology becomes infinitely powerful in a nanosecond. He calls the rhetoric irresponsible and harmful to the field students are about to enter.
    • On Stanford specifically: if the university president places an order, NVIDIA will deliver the chips. The bottleneck is that no university department has a billion-dollar compute budget because budgeting is fragmented across grants. Stanford’s $40 billion endowment is more than enough to fix that.
    • “It’s Stanford’s fault” is meant as empowerment. If something is your fault, you can solve it.
    • Career advice: do not optimize purely for passion. Most people do not yet know what they love. Pick the job in front of you and do it as well as possible. Even as CEO, Huang says, 90 percent of the work is hard and he suffers through it.
    • Suffering on purpose builds the muscle of resilience. When the company, the team or the family needs you to be tough, that muscle has to already exist.
    • NVIDIA’s first generation of products was technically wrong in nearly every dimension: curved surfaces instead of triangles, no Z-buffer, forward instead of inverse texture mapping, no floating point. The strategic recovery, not the technology, taught Huang the lessons that have lasted decades.
    • The biggest clean strategic mistake Huang names is the move into mobile chips (Tegra). It grew to a billion dollars then went to zero when Qualcomm’s modem dominance shut NVIDIA out of the 3G to 4G transition. The recovery into automotive and robotics (the Thor chip is the great great great grandson of that mobile lineage) was real, but Huang refuses to rationalize the original choice.
    • Forecasting framework: observe, reason from first principles, ask “so what” and “what next” until you have a mental model of the future, place your company inside that model, then work backwards while minimizing opportunity cost and maximizing optionality.
    • Best part of the CEO job: living at the intersection of vision, strategy and execution surrounded by people capable enough to make ambitious visions real. Worst part: the responsibility for everyone who joined the spaceship, especially in the near-death moments NVIDIA had four or five times early on.
    • Underrated insider note: Huang’s first apple pie with cheese, first hot fudge sandwich and first milkshake all happened at Denny’s. The Superbird, the fried chicken and a custom Superbird-style ham and cheese with tomato and mustard are his order.

    Detailed Summary

    Computing reinvented from the ground up

    Huang frames the moment as the first true rewrite of the computer in sixty-plus years. From the IBM System 360 forward, the mental model of writing code, running code, taking a computer to market and reasoning about applications stayed roughly constant. AI changes the programming model itself. Software is no longer a compiled binary running deterministically on a CPU. It is a neural network running on a GPU producing generated, contextual, real-time output. That cascades into how companies are organized, what tools developers use, what the network and storage stack look like, and what an application is even allowed to do. Robo-taxis, he notes, are an application no one would have attempted before deep learning unlocked perception.

    Codesign and the million-x decade

    Codesign is the philosophical center of the talk. Huang traces it to the RISC work of John Hennessy at Stanford, where simpler instruction sets won by being co-designed with the compiler rather than maximally optimized in isolation. NVIDIA extends the principle across every layer simultaneously: GPU architecture, CPU architecture, NVLink and NVSwitch fabrics, photonic interconnects, networking silicon, storage paths, CUDA libraries, frameworks and ultimately the model design. The numbers Huang gives are arresting. Moore’s Law in its prime delivered roughly 100x per decade. By the time Dennard scaling broke, real-world gains had compressed to roughly 10x. NVIDIA’s codesigned stack delivered between 100,000x and 1,000,000x over the same ten-year window. That non-linear speed-up is, in Huang’s telling, the precondition for modern AI: it is what allowed researchers to stop curating training sets and just feed the entire internet to the model.

    Education has to fuse first principles with AI tools

    Asked how curriculum should evolve, Huang argues AI must be integrated into the learning process, not just taught about. He recalls Hennessy writing his textbook by hand a chapter a week while Huang was a student, and says pre-recorded textbooks cannot keep up with the rate at which AI generates new knowledge. He describes his own learning workflow: hand the paper to an AI, then have it read the entire surrounding literature, then treat the AI as a dedicated researcher who can be interrogated. At the same time he defends the classics. Mead and Conway are still the foundation. Most modern semiconductor scaling tricks have been exhausted, but knowing where the field came from sharpens judgment when designing what comes next.

    Open source and the five domain pillars

    Huang gives one of the most detailed public accounts of why NVIDIA invests so heavily in open foundation models even while being a top customer of closed labs. He recommends Claude and OpenAI by name for production coding work, and says 100 percent of NVIDIA engineers are now agentically supported. The open-weights case rests on three legs. First, language is the codification of intelligence, and there are at least 230 languages that no commercial lab will ever prioritize. Nemotron is built near frontier and released so any country or community can fine-tune it. Second, the same representation-learning approach has to be replicated in domains where the data is not internet text, so NVIDIA seeded BioNeMo for biology, Alphamayo for autonomy, Groot for humanoid robotics and a climate model for mesoscale multiphysics. The economics of those fields would never produce a foundation model on their own. Third, safety and security require transparency. A black box cannot be defended or audited, and the future of cyber defense is not bigger-model-versus-bigger-model but swarms of cheap fast small models like Nemotron Nano surrounding the threat.

    MFU is the wrong metric, tokens per watt is closer

    A student raises the leaked memo that the xAI Memphis cluster is running at 11 percent Model Flops Utilization. Huang flips the framing. He says he would rather be at low MFU all the time, because that means he over-provisioned flops, memory bandwidth, memory capacity and network capacity. Bottlenecks shift constantly, so over-provisioning across every dimension is what lets the system absorb a spike without getting pinned by Amdahl’s law. In disaggregated inference, where prefill and decode are physically separated and decode is bandwidth-bound rather than flop-bound, NVLink72 can deliver extremely high tokens per watt while reporting very low MFU. Huang argues the right framing is performance, and ultimately tokens per watt as a rough proxy for intelligence per watt, adjusted for the fact that not all tokens are equal. A coding token is worth more than a generic token.

    Hopper, Grace Blackwell NVLink72, Vera Rubin, Feynman

    Huang gives the clearest public framing of NVIDIA’s roadmap as a sequence of architectural answers to evolving compute patterns. Hopper was built for pre-training, at a moment when NVIDIA chose to build multi-billion-dollar machines while the largest scientific supercomputer in the world cost $350 million and the marketplace for such systems was, on paper, zero. Grace Blackwell NVLink72 was the answer to inference and reasoning: a rack-scale computer that ganged 72 GPUs together because decode needs aggregate memory bandwidth far beyond a single chip. The generation-over-generation speed-up was 50x in two years, twenty-five times what Moore’s Law would have delivered. Vera Rubin is being built explicitly for agents. Agents load long-term memory from storage that has to be wired directly into the GPU fabric, they use working memory, they call tools that run on a CPU, and they wait. So the CPU has to be Vera, optimized for low-latency single-threaded code, because the multi-billion-dollar GPU system cannot afford to idle waiting on a slow tool call. Feynman extends the pattern to swarms of agents with sub-agents and sub-sub-agents, a recursive software topology that will demand its own compute pattern.

    Energy demand and the grid

    Huang’s energy projection is one of the most aggressive numbers in the talk. NVIDIA can compound tokens per watt by 50x per generation through codesign, but the total compute demand is heading roughly a thousand times higher, and Huang says he would not be surprised if the real figure is one or two orders of magnitude beyond that. The reason is structural: future computing is generative and continuous, not pre-recorded and on-demand. The good news, he argues, is that this is the best moment in the history of humanity to invest in sustainable generation. Market forces alone are now sufficient to fund solar, nuclear and grid upgrades. Government subsidies are no longer required to make the math work.

    Adversarial countries, export controls and the telecom warning

    This is the segment where Huang is visibly fired up. He attacks the GPUs-as-atomic-bombs framing on its face. NVIDIA GPUs power medical imaging, video games and soy sauce delivery. A billion people use them. He advocates them to his family. The analogy collapses at the first comparison. He attacks the second framing, that American companies should not compete abroad because they will lose anyway, as a self-fulfilling defeat. Competition makes the company better. The third framing, that depriving the rest of the world of general-purpose computing benefits the United States, also fails on first principles: it benefits one or two American companies at the cost of an entire industry. The cautionary parallel is telecommunications. The United States once had a leading position in telecom fundamental technology and policied itself out of it. Huang’s worry, voiced explicitly to a room of CS students, is that they will graduate into a shell of a computer industry if the same path is repeated.

    AI doom and rational optimism

    In the same arc Huang rejects the science-fiction framing of AI as a singularity that arrives suddenly on a Wednesday at 7pm and ends civilization. He calls those claims irresponsible, says they are not true, and points out that the people advancing them are believed by audiences who then make policy on that basis. It is not true that no one understands how these systems work. It is not true that intelligence becomes infinitely powerful instantaneously. It is not true that there is no defense. His framing, which the host echoes as “rational optimism,” is that the goal is to create a future where people care about computers because the technology students are learning is worth mastering.

    Stanford’s compute problem is Stanford’s fault

    A student presses on the scarcity of compute for independent researchers, startups and universities inside the United States. Huang’s answer is sharp: there is no shortage. Place the order and the chips will arrive. The actual broken thing is institutional. University grants are fragmented across departments. No researcher can raise enough on a single grant to fund a billion-dollar shared cluster, and no one shares. He compares it to showing up at the grocery store demanding a billion dollars of tomatoes today. The solution is planning, aggregation and a campus-scale supercomputer, the way Stanford once built the linear accelerator. The endowment is $40 billion. Pulling a billion off it, contracting cloud capacity and giving every student and researcher AI supercomputer access is, in Huang’s view, obviously doable. When he says “it is Stanford’s fault” the host laughs, but Huang clarifies: if it is your fault you have the power to fix it.

    Career, suffering and resilience

    Asked how a CS student should spend the next few years, Huang pushes back on the standard “follow your passion” advice. Most people do not know what they love yet, because no one knows what they do not know. The bar of demanding joy from every working day is too high. Whatever the job is, do it as well as you can. Even as CEO of NVIDIA he says he genuinely loves about 10 percent of his work. The other 90 percent is hard and he suffers through it. He recommends suffering on purpose, because resilience is a muscle that only builds under load, and when the company, the team or the family needs that muscle, it has to already exist. Earlier in his life that meant cleaning toilets and busing tables at Denny’s. He does it today running a multi-trillion-dollar company.

    The biggest mistakes

    Huang separates technical mistakes from strategic mistakes. NVIDIA’s first generation of products was technically wrong in almost every way: curved surfaces instead of triangles, no Z-buffer, forward instead of inverse texture mapping, no floating point inside. The company wasted two and a half years. But the strategic genius of the recovery, the reading of the market, the conservation of resources and the reapplication of talent, is what taught him strategy. The clean strategic mistake he names is mobile. NVIDIA’s Tegra line grew to a billion dollars of revenue and then collapsed to zero when Qualcomm’s modem dominance locked NVIDIA out of the 3G to 4G transition. Huang explicitly refuses the comforting rationalization that the Tegra effort fed the Thor automotive chip (“Thor is the great great great grandson”). The original decision, he says, was a waste of time. The lesson is to think one or two clicks further about whether a market is structurally winnable before committing the company.

    Forecasting under fog of war

    The final substantive exchange is on forecasting. Huang’s method has four steps. Observe what is actually happening (AlexNet crushing two decades of computer vision research in one shot, GPT producing reasoning by token generation). Reason from first principles about why it works. Ask “so what” and “what next” recursively until a mental model of the future emerges. Place the company inside that future and work backwards. Crucially, expect to be partly wrong. Some outcomes will absolutely happen, some will likely happen, some might happen, and the strategy has to be robust across that distribution. The real cost of any strategic choice is the opportunity cost of the alternatives you did not take, so the discipline is to minimize that cost and maximize optionality while letting the journey itself pay for the journey.

    Thoughts

    The most useful thing in this conversation is the explicit architectural mapping of compute patterns to chip generations. Hopper for pre-training. Grace Blackwell NVLink72 for inference, because decode is bandwidth-bound and a single chip cannot supply it. Vera Rubin for agents, because tool calls stall multi-billion-dollar GPU systems and so the CPU has to be optimized for low-latency single-threaded code. Feynman for swarms. That sequence is not marketing. It is a falsifiable thesis about where the bottleneck moves next, and every other infrastructure company should be measuring themselves against it. If Huang is right that swarms of sub-agents are the next dominant pattern, then the design pressure shifts from raw flops to fabric topology, memory hierarchy and storage-to-GPU latency. That has implications for everyone downstream, including the hyperscalers building competing accelerators.

    The MFU section is the most intellectually generous moment in the talk. The instinct in the AI ops community has been to chase MFU as if it were a virtue. Huang argues, persuasively, that low MFU is consistent with high tokens per watt in a disaggregated inference setup, and that bottlenecks rotate fast enough that over-provisioning every resource is the rational design. That reframing matters because it changes what “scarce” means. Compute is not scarce in the way the discourse treats it. What is scarce is a coherent system designed end-to-end. The xAI 11 percent number, in that frame, is not embarrassing. It is the natural reading of a workload that is mostly decode.

    The Stanford segment is the part most likely to be quoted out of context. “It’s Stanford’s fault” is a deliberately provocative line, but the underlying claim is correct and load-bearing. Compute is not gated by NVIDIA refusing to ship chips. It is gated by the fact that fragmented grant funding cannot aggregate into the billion-dollar order that NVIDIA can fulfill. The implication is that universities and national labs need a structural change in how they pool capital for compute, and that the current model of every researcher buying a handful of cards is genuinely obsolete. Huang’s nudge about pulling a billion off the endowment is concrete enough to be acted on, and other major research universities should read this segment as a direct prompt.

    The geopolitical segment is the highest-stakes one. The telecommunications comparison is correct as a historical pattern, and Huang is one of the very few executives in a position to deliver that warning credibly. The unresolved tension is that the argument applies symmetrically. If American AI dominance is built by selling globally, that includes selling into adversarial states, and the policy question is where the line falls. Huang does not answer that question. He attacks the framing that lets the question be answered badly. That is a meaningful contribution to the discourse even if it does not resolve the underlying tradeoff.

    The career advice section is the part the social-media clips will mishandle. “Seek suffering” reads as macho when extracted. In context it is a specific operational claim about how resilience compounds, and it is paired with the Tegra story where Huang himself paid the price of not thinking one more click ahead. That kind of self-implication is rare in CEO talks, and it is the reason the talk is worth listening to in full rather than only reading the recap.

    Watch the full Stanford CS153 Frontier Systems conversation with Jensen Huang here.