PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: NVLink

  • How GPT-5, Claude, and Gemini Are Actually Trained and Served: The Real Math Behind Frontier AI Infrastructure

    Reiner Pope, CEO of MatX and former TPU architect at Google, sat down with Dwarkesh Patel for a different kind of episode: a chalk-and-blackboard lecture on how frontier LLMs like GPT-5, Claude, and Gemini are actually trained and served. With nothing but a handful of equations and public API prices, Reiner reverse engineers an astonishing amount of what the labs are doing. If you have ever wondered why Fast Mode costs more, why context length stalls around 200k tokens, why models seem 100x over-trained, or why hyperscalers are pouring half a trillion dollars into memory, this is the most lucid explanation on the internet.

    TLDW

    Frontier LLM economics come down to two simple budgets: compute time and memory time. Once you write the rooflines on a blackboard, almost everything else falls out of them. Optimal batch size is roughly 300 times your sparsity ratio (around 2,000 to 3,000 tokens for a DeepSeek-style model). A new batch “train” departs every 20 milliseconds because that is how long it takes to read HBM end to end. Mixture of experts strongly favors staying inside a single rack, which is why scale-up domains went from 8 GPUs (Hopper) to 72 (Blackwell) to 500-plus (Rubin). Pipeline parallelism solves weight capacity but does nothing for KV cache, and adds painful per-hop latency, which is why Ilya famously said pipelining is not wise. Because of reinforcement learning and inference economics, frontier models are roughly 100x over-trained versus Chinchilla optimal, and a well-tuned model should output roughly as many tokens during deployment as went into its pre-training corpus. API prices leak the rest: Gemini’s 50% premium above 200k tokens reveals where KV memory time crosses weight memory time, prefill being 5x cheaper than decode confirms decode is memory bandwidth bound, and cache hit pricing tiers map directly to HBM, DDR, flash, and (yes) spinning disk. The lecture closes on a beautiful detour about the convergent evolution of neural nets and cryptographic ciphers.

    Key Takeaways

    • Two equations explain almost everything. A roofline analysis comparing compute time to memory fetch time predicts cost, latency, and architectural choices with shocking accuracy.
    • Optimal batch size is about 300 times sparsity. For a DeepSeek model that activates 32 of 256 experts, that lands around 2,000 to 3,000 tokens per batch. Real deployments go a bit higher to leave headroom.
    • The 20 millisecond train. A new batch departs every 20ms because that is how long it takes to read all of HBM once. Worst-case queue latency is roughly 40ms.
    • Fast Mode is just smaller batches. Pay 6x more, get 2.5x faster decode by amortizing weights over fewer users. There is a hard latency floor at the HBM read time.
    • Slow Mode would not save much. Once you are past the optimal batch size, the cost-per-token plateau is dominated by compute, not weight fetches. You cannot meaningfully amortize KV cache because it is unique per sequence.
    • One rack is the natural MoE unit. Expert parallelism wants all-to-all communication, which strongly favors the scale-up network (NVLink) over the scale-out network (roughly 8x slower).
    • Bigger scale-up domains drove model scaling. The jump from 8 (Hopper) to 72 (Blackwell) to 500-plus (Rubin) GPUs per rack increased aggregate memory bandwidth by 8x, which is why trillion-plus parameter models only became viable recently.
    • Pipeline parallelism is overrated for inference. It saves on weight memory capacity but does nothing for KV cache memory. It also adds milliseconds of latency per hop in decode.
    • Why Ilya said pipelining is not wise. Architectural constraints (cross-layer residuals like in Kimi) and the inability to amortize weight loads across micro-batches make pipelining a hassle in training too.
    • The memory wall is real and paradoxical. Hyperscalers reportedly spend 50% of CapEx on memory, yet racks have far more HBM than a trillion-parameter model needs. The capacity is there for KV cache and batch size, not for weights.
    • Frontier models are roughly 100x over-trained vs Chinchilla. When you minimize total cost across pre-training plus RL plus inference, smaller models trained on more data win.
    • Each model should output roughly all human knowledge. If you equalize pre-training and inference compute, the total tokens served by a model during its lifetime should approximate its training corpus. Roughly 150 trillion in, 150 trillion out.
    • API pricing reveals architecture. Gemini’s 50% premium above 200k context, the 5x decode-vs-prefill ratio, and cache duration tiers all leak detailed information about KV size, memory bottlenecks, and storage hierarchy.
    • KV cache is roughly 2KB per token. Solving Gemini’s pricing equation gives a plausible 1.6 to 2 kilobytes per token at 100B active parameters and 200k context.
    • Decode is memory bandwidth bound, prefill is compute bound. The 5x price gap is direct evidence.
    • Cache pricing maps to memory tiers. The 5-minute and 1-hour cache durations probably correspond to flash and spinning disk drain times respectively. LLM serving uses spinning disk.
    • Context length is stuck near 200k. Memory bandwidth, not compute, is the binding constraint. Sparse attention gives a square-root improvement but is not infinite.
    • Cryptography and neural nets are mathematical cousins. Both rely on jumbling information across inputs. Feistel ciphers led directly to RevNets (reversible neural networks). Adversarial attacks mirror the cipher avalanche property.

    Detailed Summary

    The Roofline: Compute Time vs Memory Time

    Reiner starts with the simplest possible model of LLM inference. The time to do a forward pass is bounded below by the maximum of compute time and memory fetch time. Compute time is the batch size times active parameters divided by FLOPs. Memory time is total parameters divided by memory bandwidth, plus a KV cache term that scales with batch size and context length. From these two equations, almost every economic and architectural fact about modern LLMs can be derived.

    Plotting cost per token against batch size gives a clean picture: at low batch you pay enormous overhead because you cannot amortize the weight fetches, and at high batch you hit a compute floor. There is a sweet spot where memory bandwidth time equals compute time. That sweet spot is what Fast Mode and Slow Mode are tuning around.

    Why Fast Mode Costs More: The Batch Trade-Off

    When Claude Code or Codex offers Fast Mode at 6x the price for 2.5x the speed, what is really happening is that they are running you at a smaller batch size. Smaller batch means weight loads are amortized over fewer users, so cost per token goes up. But latency goes down because each forward pass touches less data. There is a hard floor on latency because you have to read every byte of HBM at least once per token, and that takes about 20 milliseconds on Blackwell-class hardware. There is also a soft ceiling on Slow Mode savings because the unamortizable parts (KV cache fetches, compute) eventually dominate.

    The 20 Millisecond Train

    HBM capacity divided by HBM bandwidth lands consistently around 20 milliseconds across generations of Nvidia hardware. That is the natural cadence at which a frontier model can run a forward pass over all its weights. Reiner uses a memorable analogy: a train departs every 20 milliseconds. Any users whose requests are ready board the train. If the train is full, they wait. If it is empty, it leaves anyway. This is why you do not need millions of concurrent users to saturate a model’s batch. You only need enough to fill a 2,000-token train every 20ms.

    Why Optimal Batch Size Is About 300 Times Sparsity

    Setting compute time equal to weight fetch time and rearranging gives a beautiful result: batch size needs to be greater than (FLOPs / memory bandwidth) times (total params / active params). The hardware ratio is a dimensionless 300 on most GPUs and has stayed remarkably stable from A100 through Hopper, Blackwell, and Rubin. The model term is just the sparsity ratio. For DeepSeek with 32 of 256 experts active, that is 8. So optimal batch is around 2,400 tokens. Real deployments push this to 3x to leave headroom for non-ideal efficiency. At 64 trains per second, that is roughly 128,000 tokens per second per replica, or about 1/1000 of Gemini’s reported global throughput.

    Mixture of Experts Wants to Live Inside a Rack

    MoE all-to-all routing means every token can be sent to any expert on any GPU. The communication pattern strongly prefers the fast scale-up network (NVLink) inside a rack to the slower scale-out network between racks. Scale-out is roughly 8x slower in bandwidth. This is why one rack ends up being the natural unit for an expert layer, and why Nvidia’s progression from 8 GPUs per rack (Hopper) to 72 (Blackwell) to 500-plus (Rubin) has been such a big deal for model size scaling.

    Reiner walks through the physical constraints: cable density, bend radius, weight, power, cooling. Modern racks are pushing every dimension to the limit. Stuffing more GPUs into the scale-up domain is genuinely a hardware engineering problem.

    Pipeline Parallelism: Why Ilya Said It Is Not Wise

    Pipelining splits model layers across racks. It is the natural way to scale beyond the scale-up domain for very large models. But it has problems. In inference, pipelining does not save runtime, it only saves memory capacity per rack, which already is not the binding constraint because trillion-parameter models only need a terabyte and racks have 10x that. In training, pipelining creates the famous bubble (idle GPU time at the start and end of each pipeline pass) and forces micro-batching, which kills your ability to amortize weight loads across the global batch.

    There is also an architectural cost. Models like Kimi use cross-layer residual connections where attention attends to layers a few back, and pipelining makes those patterns very hard to implement cleanly. Ilya’s quip “as we now know, pipelining is not wise” captures all of this.

    The Memory Wall Paradox

    Industry analysts report that hyperscalers are spending 50% of CapEx on memory this year, while smartphones and laptops are seeing 30% volume drops because there is not enough HBM and DDR to go around. Yet a Blackwell rack already has tens of terabytes of HBM, far more than a trillion-parameter model needs. The reason is that all that extra capacity goes to KV cache, batch size, and longer context. The bandwidth, not the capacity, is what matters most for weight loading. This also implies that hardware could be designed with less HBM per GPU if you commit to pipelining the weights, which is a real architectural option for a chip startup like MatX.

    Reinforcement Learning and the 100x Over-Training of Frontier Models

    Chinchilla scaling laws say a model with N active parameters should be trained on roughly 20N tokens for compute-optimal training. But frontier labs do not just minimize training cost. They minimize training plus inference cost across the model’s deployment lifetime. With reinforcement learning added to the mix, the cost equation has three terms: pre-training (6 times active params times tokens), RL (somewhere between 2x and 6x times active params times RL tokens, with a 30% efficiency penalty for decode-heavy rollouts), and inference (2 times active params times inference tokens).

    If you assume those three roughly equalize at the optimum (a heuristic that holds for many cost curves), you get a clean conclusion: the data going into pre-training should be roughly equal to the data going into RL, which should be roughly equal to the tokens served at inference. With 100 billion active parameters and roughly 150 trillion training tokens, that is about 75x past Chinchilla optimal. Reiner rounds it to 100x. This is the most concrete first-principles argument for why frontier models are so deeply over-trained, and it implies that as inference traffic grows, models should keep getting smaller and longer-trained.

    Each Model Should Output All of Human Knowledge

    The most jaw-dropping consequence: if you equalize pre-training and inference compute, then the total tokens generated by a model across its deployment lifetime should approximate the size of its training corpus. GPT-5, served to hundreds of millions of users for two months, will collectively output something on the order of 150 trillion tokens. That is roughly the sum of human knowledge in textual form. Each frontier model is, in this sense, a one-shot universal author of a corpus the size of its source material.

    API Prices Leak Architecture

    This is where the lecture gets really fun. Gemini 3.1 charges 50% more for context above 200k tokens. Setting memory time equal to compute time at exactly 200k context and solving for KV cache size gives roughly 1.6 to 2 kilobytes per token, which is plausible for a model with 8 KV heads, dense attention, and head dimension of 128.

    The 5x premium for output (decode) tokens versus input (prefill) tokens is direct evidence that decode is severely memory bandwidth bound and prefill is compute bound. Prefill processes many tokens per weight load, so it amortizes memory cost over the whole sequence. Decode processes one token per weight load, so it pays full memory cost every time.

    Cache hits priced at one tenth of cache misses tell you that storing the KV cache in HBM (or DDR or flash) is much cheaper than recomputing it from scratch. The two cache duration tiers (5 minutes and 1 hour) probably correspond to memory tiers whose drain times match those durations: flash for the 5-minute tier, spinning disk for the 1-hour tier. Yes, spinning disk is in the modern LLM serving stack, despite being decades-old technology.

    Why Context Length Has Plateaued at 200k

    Context lengths shot up from 8k to roughly 200k during the GPT-3 to GPT-4 era and have stayed roughly flat for the past two years. Reiner argues this is the natural balance point where memory bandwidth cost crosses compute cost. Going to a million tokens is expensive. Going to 100 million tokens (which Dario has hinted is needed for true continual learning via in-context learning) is essentially impossible without either a memory technology breakthrough or a much more aggressive sparse attention scheme. Sparse attention helps with a square-root improvement, but it is not unlimited. Going too sparse trades off too much quality.

    Cryptography Meets Neural Nets

    The episode ends with a lovely intellectual detour. Cryptographic protocols and transformer architectures both rely on jumbling information across all inputs. They are doing inverse versions of the same operation: ciphers take structured input and produce randomness, while neural nets take noisy input and extract structure. Both fields use differentiation as their primary attack vector (differential cryptanalysis on ciphers, gradient descent on neural nets). Adversarial attacks on image classifiers exploit exactly the avalanche property that good ciphers are designed for.

    The most concrete crossover: Feistel ciphers, which let you build invertible functions out of non-invertible ones, were ported into deep learning as RevNets (reversible networks) in 2017. RevNets let you run the entire network backwards during the backward pass, eliminating the need to store activations and dramatically reducing training memory footprint. It is the opposite trade-off of KV caching: spending compute to save memory rather than spending memory to save compute.

    Thoughts

    The most striking thing about this episode is how much can be deduced from a few equations and the public API price sheets of the major labs. The labs treat their architectures as trade secrets, but the moment they price tokens to be close to cost (which competition forces them to do), the prices themselves leak the underlying ratios. Anyone with a pen and paper can reverse engineer the KV cache size, the memory tier hierarchy, and the compute-vs-memory bottleneck profile of a frontier model. There is a lesson here for builders: in competitive markets, the prices tell you almost everything.

    The 100x over-training result has interesting implications for what comes next. If the optimal balance shifts further toward inference (as adoption keeps growing), models should get smaller and longer-trained. That is good news for serving costs and bad news for training-compute-as-moat. The biggest determinant of model quality might increasingly be data quality and RL environment design, not raw pre-training compute. This squares with what is visible publicly: the leading labs are investing heavily in RL infrastructure, evaluations, and synthetic data pipelines.

    The memory wall is the most underrated infrastructure story in AI. Most people think of compute as the bottleneck, but Reiner makes it clear that memory bandwidth is what actually limits context length, which limits how agentic a model can be in practice. If you cannot get to 100 million token contexts, you probably cannot have an AI agent that has been working with you for a month and remembers everything. Either some sparse attention scheme has to give us cheap effective context length, or we need a memory hardware breakthrough, or we have to invent some form of continual learning that does not rely on context windows. None of those paths are obviously easy, and the fact that context length has been flat for two years despite enormous investment suggests we are stuck against a real wall.

    The cryptography parallel is the kind of cross-disciplinary insight that does not show up enough in AI discourse. Treating neural networks as a kind of differentiable cipher reframes a lot of the architecture choices (residual connections, layer normalization, attention) as deliberate efforts to make the function smooth and invertible enough to learn, in contrast to ciphers, which are deliberately designed to resist exactly that. Adversarial robustness research probably has a lot more to learn from cryptanalysis than it currently does.

    Finally, the format itself is a win. Most AI podcasts are conversational, which is great for personality but bad for technical depth. A blackboard lecture with an interlocutor who asks naive questions at the right moments is a much higher bandwidth medium. More of this, please.

  • Jensen Huang on Lex Fridman: NVIDIA’s CEO Reveals His Vision for the AI Revolution, Scaling Laws, and Why Intelligence Is Now a Commodity

    A deep breakdown of Lex Fridman Podcast #494 featuring Jensen Huang, CEO of NVIDIA, covering extreme co-design, the four AI scaling laws, CUDA’s origin story, the future of programming, AGI timelines, and what it takes to lead the world’s most valuable company.

    TLDW (Too Long, Didn’t Watch)

    Jensen Huang sat down with Lex Fridman for a sprawling two-and-a-half-hour conversation covering the full arc of NVIDIA’s evolution from a GPU gaming company to the engine of the AI revolution. Jensen explains how NVIDIA now thinks in terms of rack-scale and pod-scale computing rather than individual chips, breaks down his four AI scaling laws (pre-training, post-training, test time, and agentic), and reveals the near-existential bet the company made putting CUDA on GeForce. He shares his views on China’s tech ecosystem, his deep respect for TSMC, why he turned down the chance to become TSMC’s CEO, how Elon Musk’s systems engineering approach built Colossus in record time, and why he believes AGI already exists. He also discusses why the future of programming is really about “specification,” why intelligence is being commoditized while humanity is the true superpower, and how he manages the enormous pressure of leading a company that nations and economies depend on. His core message: do not let the democratization of intelligence cause you anxiety. Instead, let it inspire you.

    Key Takeaways

    1. NVIDIA No Longer Thinks in Chips. It Thinks in AI Factories.

    Jensen’s mental model of what NVIDIA builds has fundamentally changed. He no longer picks up a chip to represent a new product generation. Instead, his mental model is a gigawatt-scale AI factory with power generation, cooling systems, and thousands of engineers bringing it online. The unit of computing at NVIDIA has evolved from GPU to computer to cluster to AI factory. His next mental “click” is planetary-scale computing.

    2. Extreme Co-Design Is NVIDIA’s Secret Weapon

    The reason NVIDIA dominates is not just better GPUs. It is the extreme co-design of the entire stack: GPU, CPU, memory, networking, switching, power, cooling, storage, software, algorithms, and applications. Jensen explains that when you distribute workloads across tens of thousands of computers and want them to go a million times faster (not just 10,000 times), every single component becomes a bottleneck. This is a restatement of Amdahl’s Law at scale. NVIDIA’s organizational structure directly reflects this co-design philosophy. Jensen has 60+ direct reports, holds no one-on-ones, and runs every meeting as a collective problem-solving session where specialists across all domains are present and contribute.

    3. The Four AI Scaling Laws Are a Flywheel

    Jensen outlined four distinct scaling laws that form a continuous loop:

    Pre-training scaling: Larger models plus more data equals smarter AI. The industry panicked when people said data was running out, but synthetic data generation has removed that ceiling. Data is now limited by compute, not by human generation.

    Post-training scaling: Fine-tuning, reinforcement learning from human feedback, and curated data continue to scale AI capabilities beyond what pre-training alone achieves.

    Test-time scaling: Inference is not “easy” as many predicted. It is thinking, reasoning, planning, and search. It is far more compute-intensive than memorization and pattern matching. This is why inference chips cannot be commoditized the way many predicted.

    Agentic scaling: A single AI agent can spawn sub-agents, creating teams. This is like scaling a company by hiring more employees rather than trying to make one person faster. The experiences generated by agents feed back into pre-training, creating a flywheel.

    4. The CUDA Bet Nearly Killed NVIDIA

    Putting CUDA on GeForce was one of the most consequential technology decisions in modern history. It increased GPU costs by roughly 50%, which crushed the company’s gross margins at a time when NVIDIA was a 35% gross margin business. The company’s market cap dropped from around $7-8 billion to approximately $1.5 billion. But Jensen understood that install base defines a computing architecture, not elegance. He pointed to x86 as proof: a less-than-elegant architecture that defeated beautifully designed RISC alternatives because of its massive install base. CUDA on GeForce put a supercomputer in the hands of every researcher, every scientist, every student. It took a decade to recover, but that install base became the foundation of the deep learning revolution.

    5. NVIDIA’s Moat Is Trust, Velocity, and Install Base

    Jensen was direct about NVIDIA’s competitive advantage. The CUDA install base is the number one asset. Developers target CUDA first because it reaches hundreds of millions of computers, is in every cloud, every OEM, every country, every industry. NVIDIA ships a new architecture roughly every year. No company in history has built systems of this complexity at this cadence. And the trust that NVIDIA will maintain, improve, and optimize CUDA indefinitely is something developers can count on. If someone created “GUDA” or “TUDA” tomorrow, it would not matter. The install base, velocity of execution, ecosystem breadth, and earned trust create a compounding advantage that is nearly impossible to replicate.

    6. Jensen Believes AGI Is Already Here

    When asked about AGI timelines, Jensen said he believes AGI has been achieved. His reasoning is practical: an agentic system today could plausibly create a web service, achieve virality, and generate a billion dollars in revenue, even if temporarily. This is not meaningfully different from many internet-era companies that did the same thing with technology no more sophisticated than what current AI agents can produce. He does not believe 100,000 agents could build another NVIDIA, but he believes a single agent-driven viral product is within reach right now.

    7. The Future of Programming Is Specification, Not Syntax

    Jensen believes the number of programmers in the world will increase dramatically, not decrease. His reasoning: the definition of coding is expanding to include specification and architectural description in natural language. This expands the population of “coders” from roughly 30 million professional developers to potentially a billion people. Every carpenter, plumber, accountant, and farmer who can describe what they want a computer to build is now a coder. The artistry of the future is knowing where on the spectrum of specification to operate, from highly prescriptive to exploratory and open-ended.

    8. China Is the Fastest Innovating Country in the World

    Jensen gave a nuanced and detailed explanation of why China’s tech ecosystem is so formidable. About 50% of the world’s AI researchers are Chinese. China’s tech industry emerged during the mobile cloud era, so it was built on modern software from the start. The country’s provincial competition creates an insane internal competitive environment. And the cultural norm of knowledge-sharing through school and family networks means China effectively operates as an open-source ecosystem at all times. This is why Chinese companies contribute disproportionately to open source. Their engineers’ brothers, friends, and schoolmates work at competing companies, and sharing knowledge is the cultural default.

    9. The Power Grid Has Enormous Waste That AI Can Exploit

    Jensen proposed a pragmatic solution to the energy problem for AI data centers. Power grids are designed for worst-case conditions with margin, but 99% of the time they run at around 60% of peak capacity. That idle capacity is simply wasted. Jensen wants data centers to negotiate flexible contracts where they absorb excess power most of the time and gracefully degrade during rare peak demand periods. This requires three things: customers accepting that “six nines” uptime may not always be necessary, data centers that can dynamically shift workloads, and utilities that offer tiered power delivery contracts instead of all-or-nothing commitments.

    10. Jensen Turned Down the CEO Role at TSMC

    In 2013, TSMC founder Morris Chang offered Jensen the chance to become CEO of TSMC. Jensen confirmed the story is true and said he was deeply honored. But he had already envisioned what NVIDIA could become and felt it was his sole responsibility to make that vision happen. He sees the relationship with TSMC as one built on three decades of trust, hundreds of billions of dollars in business, and zero formal contracts.

    11. Elon Musk’s Systems Engineering Approach Is Instructive

    Jensen praised Elon Musk’s approach to building the Colossus supercomputer in Memphis in just four months. He highlighted several principles: Elon questions everything relentlessly, strips every process down to the minimum necessary, is physically present at the point of action, and his personal urgency creates urgency in every supplier. Jensen drew a parallel to NVIDIA’s own “speed of light” methodology, where every process is benchmarked against the physical limits of what is possible, not against historical baselines.

    12. Intelligence Is a Commodity. Humanity Is Not.

    Perhaps the most philosophical takeaway from the conversation: Jensen argued that intelligence is a functional, measurable thing that is being commoditized. He surrounded himself with 60 direct reports who are all “superhuman” in their respective domains, more educated and deeper in their specialties than he is. Yet he sits in the middle orchestrating all of them. This proves that intelligence alone does not determine success. Character, compassion, grit, determination, tolerance for embarrassment, and the ability to endure suffering are the real differentiators. Jensen wants the audience to understand that the word we should elevate is not intelligence but humanity.

    Detailed Summary

    From GPU Maker to AI Infrastructure Company

    The conversation opened with Jensen explaining NVIDIA’s evolution from chip-scale to rack-scale to pod-scale design. The Vera Rubin pod, announced at GTC, contains seven chip types, five purpose-built rack types, 40 racks, 1.2 quadrillion transistors, nearly 20,000 NVIDIA dies, over 1,100 Rubin GPUs, 60 exaflops of compute, and 10 petabytes per second of scale bandwidth. And that is just one pod. NVIDIA plans to produce roughly 200 of these pods per week.

    Jensen explained that extreme co-design is necessary because the problems AI must solve no longer fit inside a single computer. When you distribute a workload across 10,000 computers but want a million-fold speedup, everything becomes a bottleneck: computation, networking, switching, memory, power, cooling. This is fundamentally an Amdahl’s Law problem at planetary scale. If computation represents only 50% of the workload, speeding it up infinitely only doubles total throughput. Every layer must be co-optimized simultaneously.

    NVIDIA’s organizational structure is a direct reflection of this co-design philosophy. Jensen has more than 60 direct reports, almost all with deep engineering expertise. He does not do one-on-ones. Every meeting is a collective problem-solving session where the memory expert, the networking expert, the cooling expert, and the power delivery expert are all in the room together, attacking the same problem.

    The Strategic History of CUDA

    Jensen walked through the step-by-step journey from graphics accelerator to computing platform. The company invented a programmable pixel shader, then added IEEE-compatible FP32 to its shaders, then put C on top of that (called Cg), and eventually arrived at CUDA. The critical strategic decision was putting CUDA on GeForce, a consumer product.

    This was nearly an existential move. It increased GPU costs by roughly 50% and consumed all of the company’s gross profit at a time when NVIDIA was a 35% gross margin business. The market cap cratered from around $7-8 billion to approximately $1.5 billion. But Jensen understood a principle that many technologists overlook: install base defines a computing architecture. x86 survived not because it was elegant but because it was everywhere. CUDA on GeForce put a supercomputing capability in the hands of every gamer, every student, every researcher who built their own PC. When the deep learning revolution arrived, CUDA was already the foundation.

    How Jensen Leads and Makes Decisions

    Jensen described a leadership philosophy built on continuous reasoning in public. He does not make announcements in the traditional sense. Instead, he shapes the belief systems of his employees, board, partners, and the broader industry over months and years by reasoning through decisions step by step, using every new piece of external information as a brick in the foundation. By the time he formally announces a strategic direction, the reaction is not surprise but rather, “What took you so long?”

    He applies this same approach to his supply chain. He personally visits CEOs of DRAM companies, packaging companies, and infrastructure providers. He explains the dynamics of the industry, shares his vision of future demand, and helps them reason through why they should make multi-billion-dollar capital investments. Three years ago, he convinced DRAM CEOs that HBM memory would become mainstream for data centers, which sounded ridiculous at the time. Those companies had record years as a result.

    Jensen’s “speed of light” methodology is his framework for decision-making. Every process, every design, every cost is benchmarked against the physical limits of what is theoretically possible. He prefers this to continuous improvement, which he views as incrementalism. He would rather strip a 74-day process back to zero and ask, “If we built this from scratch today, how long would it take?” Often the answer is six days, and the remaining 68 days are filled with accumulated compromises that can be challenged individually.

    AI Scaling Laws and the Future of Compute

    Jensen broke down the four scaling laws in detail. The pre-training scaling law, which depends on model size and data volume, was thought to be hitting a wall when the industry worried about running out of high-quality human-generated data. Jensen argued this concern is misplaced. Synthetic data generation has effectively removed the ceiling, and the constraint is now compute, not data.

    Post-training continues to scale through fine-tuning and reinforcement learning. Test-time scaling was the most counterintuitive for the industry. Many predicted that inference would be “easy” and that inference chips would be small, cheap, and commoditized. Jensen saw this as fundamentally wrong. Inference is thinking: reasoning, planning, search, decomposing novel problems into solvable pieces. Thinking is much harder than reading, and test-time compute is intensely resource-hungry.

    Agentic scaling is the newest frontier. A single AI agent can spawn sub-agents, effectively multiplying intelligence the way a company scales by hiring. The experiences and data generated by agentic systems feed back into pre-training, creating a continuous improvement loop. Jensen described this as the reason NVIDIA designed the Vera Rubin rack architecture differently from the Grace Blackwell architecture. Grace Blackwell was optimized for running large language models. Vera Rubin is designed for agents, which need to access files, use tools, do research, and spin off sub-agents. NVIDIA anticipated this architectural shift two and a half years before tools like OpenClaw arrived.

    China, TSMC, and the Global Supply Chain

    Jensen provided a thoughtful analysis of China’s tech ecosystem. He identified several structural advantages: 50% of the world’s AI researchers are Chinese, the tech industry was born during the mobile cloud era (making it natively modern), provincial competition creates internal Darwinian pressure, and the culture of knowledge-sharing through school and family networks makes China effectively open-source by default.

    On TSMC, Jensen emphasized that the deepest misunderstanding about the company is that its technology is its only advantage. Their manufacturing orchestration system, which dynamically manages the shifting demands of hundreds of companies, is “completely miraculous.” Their culture uniquely balances bleeding-edge technology excellence with world-class customer service. And the trust that Jensen places in TSMC is extraordinary: three decades of partnership, hundreds of billions of dollars in business, and no formal contract.

    Jensen also discussed the AI supply chain more broadly. NVIDIA has roughly 200 suppliers contributing technology to each rack. Jensen personally manages these relationships, flying to supplier sites, explaining industry dynamics, and helping CEOs reason through multi-billion-dollar investment decisions. When asked if supply chain bottlenecks keep him up at night, he said no, because he has already communicated what NVIDIA needs, his partners have told him what they will deliver, and he believes them.

    The Energy Challenge and Space Computing

    On the energy front, Jensen proposed a practical approach to the power problem. Rather than waiting for new power generation, he wants to capture the enormous waste already present in the grid. Power infrastructure is designed for worst-case peak demand, but 99% of the time it runs far below capacity. AI data centers could absorb this excess capacity with flexible contracts that allow graceful degradation during rare peak periods.

    On space computing, NVIDIA already has GPUs in orbit for satellite imaging. Jensen acknowledged the cooling challenge (no conduction or convection in space, only radiation) but sees it as a future frontier worth cultivating. In the meantime, he is focused on the lower-hanging fruit of eliminating waste in the terrestrial power grid.

    On AGI, Jobs, and the Human Future

    Jensen stated directly that he believes AGI has been achieved, at least by the practical definition of an AI system capable of creating a billion-dollar company. He sees it as plausible that an agent could build a viral web service that briefly generates enormous revenue, just as many internet-era companies did with technology no more sophisticated than what current AI agents produce.

    On jobs, Jensen was both compassionate and clear-eyed. He told the story of radiology: computer vision became superhuman around 2019-2020, and the prediction was that radiologists would disappear. Instead, the number of radiologists grew because AI allowed them to study more scans, diagnose better, and serve more patients. The purpose of the job (diagnosing disease) did not change, even though the tools changed completely.

    He applied this principle broadly: the number of software engineers at NVIDIA will grow, not decline, because their purpose is solving problems, not writing lines of code. The number of programmers globally will grow because the definition of coding is expanding to include natural language specification, opening it up to potentially a billion people.

    His advice to anyone worried about their job is straightforward: go use AI now. Become expert in it. Every profession, from carpenter to pharmacist to lawyer, will be elevated by AI tools. The people who learn to use AI will be the ones who get hired, promoted, and empowered.

    Mortality, Succession, and Legacy

    The conversation closed with deeply personal reflections. Jensen said he really does not want to die. He sees the current moment as a “once in a humanity experience.” He does not believe in traditional succession planning. Instead, he believes the best succession strategy is to pass on knowledge continuously, every single day, in every meeting, as fast as possible. His hope is to die on the job, instantaneously, with no long period of suffering.

    He described a vision for a kind of digital continuity: sending a humanoid robot into space, continuously improving it in flight, and eventually uploading the consciousness derived from a lifetime of communications, decisions, and reasoning to catch up with it at the speed of light.

    On the emotional experience of leading NVIDIA, Jensen was candid about hitting psychological low points regularly. His coping mechanism is decomposition: break the problem into pieces, reason about what you can control, tell someone who can help, share the burden, and then deliberately forget what is behind you. He compared this to the mental discipline of great athletes who focus only on the next point.

    His final message was about the relationship between intelligence and humanity. Intelligence, he argued, is functional. It is being commoditized. Humanity, character, compassion, grit, tolerance for embarrassment, and the capacity for suffering are the true superpowers. The word society should elevate is not intelligence but humanity.

    Thoughts

    This is one of the most substantive CEO interviews of 2026. What makes it remarkable is not just the breadth of topics but the depth of reasoning Jensen demonstrates in real time. You can actually watch him think through problems on the spot, which is rare for someone at his level.

    A few things stand out. First, the CUDA origin story is one of the great strategic narratives in tech history. The decision to absorb a 50% cost increase on a consumer product, watching your market cap collapse by 80%, and holding the course for a decade because you understood the power of install base is the kind of conviction that separates generational companies from everyone else.

    Second, Jensen’s framing of the four scaling laws as a flywheel is the clearest articulation anyone has given of why AI compute demand will continue to accelerate. Most people understand pre-training. Fewer understand test-time scaling. Almost nobody is thinking about agentic scaling as a compute multiplier. Jensen has been thinking about it for years and already designed hardware for it before the software ecosystem caught up.

    Third, the discussion on jobs deserves attention. The radiology example is powerful because it is a completed experiment, not a prediction. The profession that was supposed to be eliminated first by AI instead grew. The mechanism is straightforward: when you automate the task, you expand the capacity of the purpose, and demand for the purpose increases. This does not mean there will be no pain or dislocation. Jensen acknowledged that explicitly. But the historical pattern is clear.

    Finally, the philosophical distinction between intelligence and humanity is the kind of framing that could genuinely help people navigate the anxiety of this moment. If you define your value by your intelligence alone, AI commoditization is terrifying. If you define your value by your character, your compassion, your tolerance for suffering, and your willingness to keep going when everything goes wrong, then AI is just the most powerful set of tools you have ever been given.

    Jensen Huang is 62 years old, has been running NVIDIA for 34 years, and shows no signs of slowing down. If anything, his conviction about the future is accelerating alongside his company’s growth.

    Watch the full episode: Lex Fridman Podcast #494 with Jensen Huang

  • How NVIDIA is Revolutionizing Computing with AI: Jensen Huang on AI Infrastructure, Digital Employees, and the Future of Data Centers

    NVIDIA CEO Jensen Huang discusses the company’s role in revolutionizing computing through AI, emphasizing decade-long investments in scalable, interconnected AI infrastructure, breakthroughs in efficiency, and the future of digital and embodied AI as transformative for industries globally.

    NVIDIA is transforming the landscape of computing, driving innovation at every level from data centers to digital employees. In a recent conversation with Jensen Huang, NVIDIA’s CEO, he offered a rare look at the strategic direction and long-term vision that has positioned NVIDIA as a leader in the AI revolution. Through decade-long infrastructure investments, NVIDIA is not just building hardware but creating “AI factories” that promise to impact industries globally.

    Decade-Long Investments in AI Infrastructure

    For NVIDIA, success has come from looking far into the future. Jensen Huang emphasized the company’s commitment to ten-year investments in scalable, efficient AI infrastructure. With an eye on exponential growth, NVIDIA has focused on creating solutions that can continue to meet demand as AI expands in complexity and scope. One of the cornerstones of this approach is NVLink technology, which enables GPUs to function as a unified supercomputer, allowing unprecedented scale for AI applications.

    This vision aligns with Huang’s goal of optimizing data centers for high-performance AI, making NVIDIA’s infrastructure not only capable of tackling today’s AI challenges but prepared for tomorrow’s even larger-scale demands.

    Outpacing Moore’s Law with Full-Stack Integration

    Huang highlighted how NVIDIA aims to surpass the limits of traditional computing, especially Moore’s Law, by focusing on a full-stack integration strategy. This strategy involves designing hardware and software as a cohesive unit, enabling a 240x reduction in AI computation costs while increasing efficiency. With this approach, NVIDIA has managed to achieve performance improvements that far exceed conventional expectations, driving both cost and energy usage down across its AI operations.

    The full-stack approach has enabled NVIDIA to continually upgrade its infrastructure and enhance performance, ensuring that each component of its architecture is optimized and aligned.

    The Evolution of Data Centers: From Storage to AI Factories

    One of NVIDIA’s groundbreaking shifts is the redefinition of data centers from traditional storage units to “AI factories” generating intelligence. Unlike conventional data centers focused on multi-tenant storage, NVIDIA’s new data centers produce “tokens” for AI models at an industrial scale. These tokens are used in applications across industries, from robotics to biotechnology. Huang believes that every industry will benefit from AI-generated intelligence, making this shift in data centers vital to global AI adoption.

    This AI-centric infrastructure is already making waves, as seen with NVIDIA’s 100,000-GPU supercluster built for X.AI. NVIDIA demonstrated its logistical prowess by setting up this supercluster rapidly, paving the way for similar large-scale projects in the future.

    The Role of AI in Science, Engineering, and Digital Employees

    NVIDIA’s infrastructure investments and technological advancements have far-reaching impacts, particularly in science and engineering. Huang shared that AI-driven methods are now integral to NVIDIA’s chip design process, allowing them to explore new design options and optimize faster than human engineers alone could. This innovation is just the beginning, as Huang envisions AI reshaping fields like biotechnology, materials science, and theoretical physics, creating opportunities for breakthroughs at a previously impossible scale.

    Beyond science, Huang foresees AI-driven digital employees as a major component of future workforces. AI employees could assist in roles like marketing, supply chain management, and chip design, allowing human workers to focus on higher-level tasks. This shift to digital labor marks a major milestone for AI and has the potential to redefine productivity and efficiency across industries.

    Embodied AI and Real-World Applications

    Huang believes that embodied AI—AI in physical form—will transform industries such as robotics and autonomous vehicles. Self-driving cars and robots equipped with AI will become more common, thanks to NVIDIA’s advancements in AI infrastructure. By training these AI models on NVIDIA’s systems, industries can integrate intelligent robots and vehicles without needing substantial changes to existing environments.

    This embodied AI will serve as a bridge between digital intelligence and the physical world, enabling a new generation of applications that go beyond the screen to interact directly with people and environments.

    Sustaining Innovation Through Compatibility and Software Longevity

    Huang stressed that compatibility and sustainability are central to NVIDIA’s long-term vision. NVIDIA’s CUDA platform has enabled the company to build a lasting ecosystem, allowing software created on earlier NVIDIA systems to operate seamlessly on newer ones. This commitment to software longevity means companies can rely on NVIDIA’s systems for years, making it a trusted partner for businesses that prioritize innovation without disruption.

    NVIDIA as the “AI Factory” of the Future

    As Huang puts it, NVIDIA has evolved beyond a hardware company and is now an “AI factory”—a company that produces intelligence as a commodity. Huang sees AI as a resource as valuable as energy or raw materials, with applications across nearly every industry. From providing AI-driven insights to enabling new forms of intelligence, NVIDIA’s technology is poised to transform global markets and create value on an industrial scale.

    Jensen Huang’s vision for NVIDIA is not just about staying ahead in the computing industry; it’s about redefining what computing means. NVIDIA’s investments in scalable infrastructure, software longevity, digital employees, and embodied AI represent a shift in how industries will function in the future. As Huang envisions, the company is no longer just producing chips or hardware but enabling an entire ecosystem of AI-driven innovation that will touch every aspect of modern life.