Leigh Huang

Jensen Huang just painted a striking picture of the future of work at Nvidia: 𝟳𝟱,𝟬𝟬𝟬 𝗲𝗺𝗽𝗹𝗼𝘆𝗲𝗲𝘀 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝗮𝗹𝗼𝗻𝗴𝘀𝗶𝗱𝗲 𝟳.𝟱 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀. Speaking at Nvidia GTC in San Jose, Huang said that within the next decade, Nvidia could nearly double its workforce from 42,000 to 75,000 employees, while those teams collaborate with AI agents at a 𝟭𝟬𝟬-𝘁𝗼-𝟭 𝗿𝗮𝘁𝗶𝗼. The key point was not that AI will replace people. It was that AI agents will take on the repetitive, always-on work humans should not have to do. As Huang put it, these agents will be working around the clock, so people do not have to. That vision says a lot about where work is heading. We are moving beyond AI as just a chatbot, search tool, or writing assistant. The next shift is toward 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀: systems that can reason, plan, and act autonomously to achieve specific goals. And this is no longer theoretical. More companies are betting that AI agents will become part of everyday operations. Leaders across industries are pushing teams to adopt AI faster, not slower. Nvidia is also leaning into this trend with its Agent Toolkit, an open platform designed to help enterprises build and deploy their own AI agents. According to a November 2025 McKinsey survey, 𝟲𝟮% 𝗼𝗳 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀 𝘄𝗲𝗿𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀, although most had not yet scaled them across the business. What stands out most in Huang’s comments is the framing: 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮𝗿𝗲 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗼𝗼𝗹𝘀. 𝗧𝗵𝗲𝘆 𝗮𝗿𝗲 𝗯𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝘁𝗲𝗮𝗺𝗺𝗮𝘁𝗲𝘀. Not human replacements, but force multipliers. If this model takes hold, the most valuable employees in the next decade may not be the ones doing everything themselves. They may be the ones who can best direct, manage, and collaborate with fleets of AI agents. That changes how we think about productivity, leadership, and even organizational design. The future of work may not be human vs. AI. It may be 𝗵𝘂𝗺𝗮𝗻𝘀 𝗹𝗲𝗮𝗱𝗶𝗻𝗴 𝗔𝗜 𝗮𝘁 𝘀𝗰𝗮𝗹𝗲. 𝗔𝗿𝗲 𝘄𝗲 𝗿𝗲𝗮𝗱𝘆 𝘁𝗼 𝗹𝗲𝗮𝗱 𝗔𝗜 𝗮𝘁 𝘀𝗰𝗮𝗹𝗲 — 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘂𝘀𝗲 𝗶𝘁? #AIAgents #FutureOfWork #AI #Leadership #EnterpriseAI #Telia

6

FTC Clears SoftBank Acquisition Expanding AI Chip Capabilities ▶ The FTC granted early termination for SoftBank’s review, clearing its acquisition of Ampere Computing. ▶ Ampere produces server processors used in data centers that support AI workloads. ▶ The acquisition adds to SoftBank’s existing holdings in AI-related chip technologies, including Arm and Graphcore. ▶ The approval positions SoftBank to integrate multiple components of the AI chip design and infrastructure stack. ▶ The transaction follows earlier regulatory scrutiny reported during the summer. ▶ The deal strengthens SoftBank’s role in supplying processors for large-scale AI computing environments. Link to article: https://lnkd.in/eYK2csjY 📚 The AI Policy Newsletter: https://lnkd.in/eS8bHrvG 🦋 Follow me on Bluesky: https://lnkd.in/enpH3UjQ

2

Memory chips emerge as AI infrastructure’s secondary beneficiary: concentrated market, constrained supply, expanding margins, long-term contracts. The picks-and-shovels thesis extends beyond Nvidia to the memory layer. And it shows on Micron’s gross margins.

1

Ryota Sato has been doing some really awesome work here at femtoAI compressing OpenAI's open-source Whisper ASR models using OBS (Optimal Brain Surgery). If you are at the EDGE AI Sand Diego conference next week, you should definitely check out his talk!

18

Red Hat is expanding our work with NVIDIA to deliver Day 0 support for the new Rubin platform! As organizations move AI from experimentation into full production, they need a stable, high-performance foundation. We’re meeting this shift head-on by optimizing our hybrid cloud and AI portfolios (including RHEL, OpenShift, and Red Hat AI) for NVIDIA’s latest architectural breakthroughs.

614

7 Comments

NVIDIA GTC 2026 kicks off today in San Jose – and the AI world is holding its breath. Jensen Huang's keynote promises "a new chip the world has never seen before," and with 30,000+ attendees converging on the event, expectations are sky-high. GTC has become more than a product launch – it's where the AI infrastructure roadmap gets written. This year's focus: agentic AI, robotics, and AI infrastructure. But here's what matters for enterprise teams – GTC announcements typically cascade through procurement, architecture decisions, and vendor roadmaps for the next 6-12 months. For those building AI-powered platforms, this is the week to watch closely. The chips announced today will power the models you deploy tomorrow. #AIInfrastructure #NVIDIA #GTC2026 #AIEngineering #EnterpriseAI

Just wrapped up an inspiring keynote with Dr, Lisa Su , CEO of AMD, where we shared our bold vision for accelerating AI infrastructure in Saudi Arabia. Together, we are building more than just technology—we’re laying the foundation for a sovereign, secure, and scalable AI future. Our joint venture will deliver open, high-performance infrastructure with the lowest total cost of ownership, designed for AI training and inference at scale. This partnership marks a major step toward making Saudi Arabia a global hub for AI innovation—powered by purpose, built for the world. #AIInfrastructure #SovereignAI #AMD #SaudiVision2030 #HUMAIN #PartnershipForTheFuture

922

45 Comments

Thrilled to welcome Gary Hicok as Executive Chair of Utilidata's Board! By embedding AI at the edge, we're creating the visibility and software-defined feedback loops that make power systems truly flexible and responsive. This isn't just about efficiency—it's about fundamentally changing how energy infrastructure operates, unlocking up to 50% more compute capacity in data centers through sub-millisecond control. Gary's 25+ years of pioneering edge computing at NVIDIA make him an incredibly valuable leader to help us scale this vision. He understands that AI factories and modern grids need more than incremental improvements—they need the programmable power infrastructure that Karman enables. Welcome aboard, Gary! Let's make power truly programmable. ⚡ #EdgeAI #DataCenters #DistributedComputing #EnergyInnovation

35

Large language models are “wisdom models” LLMs don’t learn from experience. They’re trained on representations of other people’s experiences: text, speech, and recordings; not direct sensory reality. What they absorb is distilled cultural wisdom, the linguistic residue of human thought and feeling. In the chain Data → Information → Knowledge → Wisdom, they operate near the knowledge–wisdom boundary, but only as encoded in language, never as lived or tested through experience. An LLM has no perception, memory, or continuous self, so it can’t distill experience into knowledge. That step needs feedback, embodiment, and a persistent self-model. Instead, LLMs perform pattern distillation, mapping linguistic regularities from many experiential accounts into a latent space that seems wise because it reflects the aggregated hindsight of others. Though they lack experience, LLMs embody civilization’s collective hindsight, the stored outcomes of human learning, failure, and correction. They’re wise as a library is wise, not through understanding but through containment of what understanding produced. True wisdom requires synthesis, the act of revising knowledge through lived feedback. LLMs can only simulate this recursively in text, not enact it. That’s why embodied or agentic AI seeks to close the loop, giving systems a way to learn through doing, not just speak from what’s been done.

15

5 Comments

Wow. This Hammerspace & Secuvy partnership announcement is a game changer for organizations who want to deploy trusted AI applications now without having to expand their storage hardware investment.

23

1 Comment

10/25: Tenth in the series on 25 Google papers for 25 years of WSC: Google-Wide Profiling The "Always On" Monitor: Most companies turn profiling off in production because they fear the overhead. At Google, we keep it on. This paper (Google-Wide Profiling) explains how to get continuous, fleet-wide visibility with a tiny tax: just 0.01% overhead. Why bother, you might ask? Because "lab" environments can be misleading. By profiling in prod, you find: Hardware differences you didn't expect; "Cold" code that is actually running hot; Library bottlenecks across every binary in the fleet. It’s the difference between guessing where the performance went and _knowing_. And for today's nano-banana poster for the paper, I tried a different prompt: [Make an infographic academic-style poster for the following paper. Use a visual aesthetic similar to the architectural blueprint roll -- Large blue sheets with white technical lines, measurements, and room labels.]

380

12 Comments

Just finished the new Jensen Huang episode of the Dwarkesh Podcast and it’s one of the most revealing conversations I’ve heard about the future of AI, NVIDIA’s moat, and U.S. - China tech competition. A few things that stood out to me: 🎯 Nvidia’s real moat is not just GPUs, but the tightly integrated stack including hardware, networking, software, and ecosystem partners that all co-evolve together. 🎯Jensen makes the case that accelerated computing will keep compounding because every layer of the stack is still being optimized, which means we’re likely still early in the AI infrastructure buildout. 🎯The discussion around whether the U.S. should sell advanced AI chips to China gets very tense; Dwarkesh presses hard on the national security implications and the “compute as enriched uranium” analogy, while Jensen pushes back that China will build AI anyway and that U.S. leadership requires staying engaged. What I appreciated most is how the episode blends: 🔺Strategic questions about export controls and geopolitics. 🔺Concrete details on supply chains, TPUs, and competition. 🔺First-principles thinking about what actually sustains Nvidia’s dominance over the long run. If you’re thinking about AI strategy, infrastructure, or policy, this is well worth the listen and a useful case study in how technical, economic, and geopolitical considerations are colliding in real time. Curious how others are thinking about the chips to China question after hearing Jensen’s side of the argument. While there are no perfect answers here, but I think these are questions every serious AI team should be wrestling with. See the comments for a link to the complete podcast 👇 #ArtificialIntelligence #Semiconductors #AIInfrastructure #TechStrategy #Geopolitics #TechPolicy Dwarkesh Podcast NVIDIA

12

1 Comment

One of the reasons I love #venturecapital is getting to work with exceptional people on truly cutting-edge innovation. Excited to share a milestone from one of our portfolio companies, EeroQ Corporation. Together with Conductor Quantum and NVIDIA, EeroQ recently demonstrated an autonomous quantum computing lab — a system where AI can design, run, and iterate on experiments directly on quantum hardware. Using natural language prompts, an AI agent didn’t just assist researchers — it actively executed experiments, tuned parameters, and analyzed real-time results. That’s a meaningful shift. We’re moving from: → humans operating fragile quantum systems to → AI acting as the control layer for quantum machines This matters because one of the biggest bottlenecks in #quantumcomputing isn’t just hardware — it’s the speed of experimentation. Embedding AI into the loop could dramatically accelerate progress. Early days, but the direction is clear: AI won’t just help build quantum computers — it will help run them. Nick Farina https://lnkd.in/dMgRrvdm

27

2 Comments

I was asked recently: what is one thing that has surprised me about leadership? Now, keep in mind I went to Stanford to be a great engineer, and I started at Intel in the engineering department. Engineering was where I lived. As I moved up within the organization, I had to start spending time with other areas of the business. Departments like marketing, business, legal, communications, etc. Guess which one I spent the most time in? Trick question -- it was the psychology department! A huge part of a leader’s job is seeking to understand and influence people. It's one of the most important skills in this role. I would never have believed this as an 18-year-old engineer! Some of the writings and efforts of Pat Lencioni, the The Table Group and the Five Dysfunctions of a team have been really helpful to me over the years to skill up in this area. 

887

43 Comments

The most impactful and important SDK release Trainiun Neuron team has worked on since inception: Native PyTorch support for Trainium, open source kernel (NKI) compiler and pre-optimized NKI kernel library, Neuron Explorer (IMO, the fastest and most intuitive ML profiler in industry), optimized inference implementation for major MoE and video models (GPT-OSS, Qwen, Pixtral….), complete vLLM v1 support !! All with full support for Trainium3. Using Trainium, whether for research, training , or large scale inference just became more intuitive and simple. https://lnkd.in/gjfhjv4j

626

10 Comments

Same company. Same partner. Second time around. This is my second time building Virbela. The Qualcomm Institute - UC San Diego Division of Calit2 at UC San Diego believed in us early the first time, and that bet paid off for everyone. Virbela grew, got acquired, and delivered real impact. Now we're independent again. Small team, big ambition, rebuilding around AI-powered behavioral simulations. And one of the first organizations back at the table? QI. The early stages of any company are uncertain. The people who show up before the traction, before the proof points, they shape the trajectory. Having a partner who saw the full arc the first time and chose to come back? That means everything and ignites our energy. Grateful to Ramesh Rao, Riley Need, Molly Wofford, Ileana Ovalle, and the teams at the Qualcomm Institute and the UC San Diego Office of Innovation and Commercialization

45

12 Comments

I attended EDGE AI FOUNDATION's Edge AI event which took place in San Diego. I missed this live presentation/keynote by Durga Malladi, EVP at Qualcomm responsible for technology road map for edge and data center solutions. He did a great job of breaking down why edge AI inference is where the practical application of GenAI and agentic AI will happen. It's important to remember that before all the AI supercomputing mania that is driving AI silicon and systems innovation, it was and arguably still is the smartphone. Interestingly, as of late, even AI supercomputing, especially in the inference domain, is becoming increasingly heterogeneous and reflective of what has been happening on the smartphone for over a decade with AI computing architectures. I think Durga does a great job of tying the long legacy of AI on the smartphone to the present and future of edge AI and agentic AI across edges. Check out this talk. It's really good orientation on where edge AI is going. 🎥 WATCH HERE! 👉 https://lnkd.in/ghJjRMYY #edgeAI #edgeaifoundation #agenticai #innovation #inference

31

1 Comment

Data gravity is real. Every enterprise is looking at how to best leverage the data securely and efficiently for AI transformation. Loved Ingo Fuchs’s take on ICMS + NVIDIA BlueField: Bringing compute to where the data lives, offload networking/storage/security to DPUs, and reduce latency and amplify customer value. Leave data where it is, don't move around, Secure and Faster insights, stronger isolation, better CPU/GPU efficiency, much better TCO. At #NetApp we are a proud to always partner of #NVIDIA to deliver on accelerated value for our customers and be part of their #AITransformation journey. #DataGravity #BlueField #DPU #IDI #NetApp

35

2 Comments