WaveSession

TensorWave

WaveSession is a conversation series with the people setting the direction of an open ecosystem around AI. In this show we interview the founders, researchers, builders, and capital providers making the calls on what gets built, how it scales, and who are getting sneak peeks into the future. Each episode moves across the stack, from the chip to the grid, the software to the firmware, the language to the physics beneath it. These are the calls being made in real time, by the people making them. Each session is an honest account, where we explore beyond the superficial, not a highlight reel. The guests are testing approaches that haven't been proven yet, because building open AI infrastructure at this scale means solving problems that don't have a playbook: technical, financial, and human, all at once. WaveSession treats this moment as what it is: a technological revolution, built in the open, with societal consequences we get into as they unfold. It invites the audience to challenge, question, and shape where it goes.

Episodes

  1. 1d ago ·  Video

    Data Centers Don't Waste Water. An Ex-Google Engineer Explains.

    Could today's data centers produce twice the AI without building more? Former Google data center lead and DeepMind alum Jim Gao, CEO of Phaidra, talks about power, cooling, water use and the metric that matters most now: tokens per watt. Data centers are at the center of this month's news. Long before anyone called them AI factories, Jim Gao was designing the cooling systems inside Google's data centers. Back then, the whole team was about 80 people, and a 5-megawatt shipping container counted as cutting edge. In this episode of WaveSession, recorded at the Yotta Conference in Las Vegas, Gao joins Jeff and Ilya. They talk about how data centers grew from containers to gigawatts, why power, cooling and compute need to run as one system, and why he believes existing infrastructure could produce twice as many tokens. Gao takes on the public perception problem head-on, including being heckled at New York Climate Week. He argues that data centers have become a physical stand-in for people's anxiety about AI. He explains why modern closed-loop facilities use far less water than most people assume. He also points to towns like Pryor, Oklahoma, and Quincy, Washington, that data centers helped revitalize. He also goes back to his early days at Google, where he became responsible for the company's PUE, the core measure of data center efficiency. In 2012, fresh off Andrew Ng's Coursera machine learning course, he snuck an extra slide into a quarterly review. It pitched using ML to make sense of the tens of thousands of data streams in a single data center. That idea led him to DeepMind and eventually to co-founding Phaidra. From there, the conversation turns to why Jensen Huang's "AI factory" framing is right, and why tokens per watt is now a business imperative. Power, cooling and IT have historically run in near-total isolation, a setup Gao calls "bonkers" at today's scale. Working with NVIDIA, Phaidra's AI agents predict thermal spikes from AI workloads and cool ahead of them, eliminating 80% of those spikes. He explains how raising liquid cooling temperatures by 10°C can cut chiller energy in half, and how dynamic oversubscription could turn stranded cooling capacity into compute. This episode covers: The water misconception and closed-loop coolingWhy tokens per watt is the metric that mattersDynamic oversubscription and doubling token outputEliminating 80% of thermal spikes with NVIDIAGoogle's data centers, from 5MW shipping containers to 30MW warehousesThe 2012 pitch that brought machine learning to data center operationsThree eras of data center controls: binary, variable, and AI-drivenAnd more! Takeaway: Power is now the limit on how much intelligence we can produce. The biggest gains won't just come from building more. They'll come from running power, cooling and compute as one intelligent system. Did the water numbers surprise you? Tell us in the comments. About the GuestJim Gao is the co-founder and CEO of Phaidra, which builds AI agents for AI factories and mission-critical industrial facilities. He spent over a decade at Google as a mechanical engineer responsible for the company's global PUE. He pioneered the use of machine learning in data center operations before joining DeepMind.  Chapters 0:00 The Data Center Water Misconception0:50 How Jim Gao Got Into Infrastructure2:28 Early Days at Google8:16 Applying Machine Learning to Data Centers10:45 Three Eras of Data Center Control14:28 Data Centers as "AI Factories"15:27 Power: The Defining Constraint on Intelligence20:33 Cutting Cooling Energy 50% with NVIDIA & CoreWeave25:00 The Water Myth Debunked29:30 The Real Promise of AI31:30 How Data Centers Revitalize Communities35:44 Data Centers Driving Renewables & Nuclear40:02 Power, Cooling & IT as One System42:17 What Is Phaidra?52:11 Dynamic Power Allocation with NVIDIA57:36 Safety & the Path Forward1:00:24 A New Utility Era

  2. Nov 5 ·  Video

    Why Diffusion, Not Transformers, Might Win the Speed War

    Diffusion models rewired image generation years ago. Now they're coming for text, and Inception Labs co-founder and CTO Aditya Grover says the results are already reshaping voice AI, customer support, and coding agents. In this episode of WaveSession, recorded at the Ai4 Conference in Las Vegas, Jeff and Ilya sit down with Grover to unpack how diffusion-based LLMs like Inception's Mercury models are delivering reasoning-model quality at a fraction of the latency and cost of traditional autoregressive systems. Grover traces Inception's origins back more than a decade to Stanford, where he and co-founders Stefano Ermon and Volodymyr Kuleshov were early researchers in generative modeling, present for the invention of diffusion itself, back when GANs still dominated computer vision. The conversation digs into why diffusion, despite powering nearly every modern image and video generator, took years to break into text, and how removing the causal mask that forces left-to-right token generation unlocks parallel generation at scale. Grover explains this isn't a replacement for transformers but a different way of training them, a distinction that matters enormously for latency-sensitive applications like conversational voice AI, where the industry today is largely stuck using non-reasoning models just to hit acceptable response times. From there, the conversation turns to where diffusion LLMs are winning in production right now: voice and customer support, search and retrieval, and orchestration of coding sub-agents that need to run fast, cheap tasks like context compaction. Grover frames the current market shift succinctly: enterprises are moving from "token maxing" to "value maxing," prioritizing intelligence per dollar over raw scale. He and the hosts also get into the hardware side of the equation, how purpose-built AI chips from companies like Cerebras and Etched compound with diffusion's efficiency gains, and how systems engineering is having a resurgence as the industry rethinks end-to-end architecture instead of just chasing bigger models. The discussion broadens into some of the most contested questions in AI strategy: open-weight versus closed-weight models, and whether the U.S. is doing enough to stay competitive as Chinese labs push the frontier of open models. Grover, notably the co-founder of a closed-weight model company, argues forcefully for a healthy open ecosystem, comparing the debate to earlier fears about releasing ChatGPT at all. He also gets candid about scaling Inception from a research lab into a 40-person company: hiring philosophy, the tension between researcher autonomy and product accountability, and why the only feedback that actually matters is whether a user is happy. This episode covers: How removing the causal mask lets diffusion LLMs generate tokens in parallel instead of one at a timeWhy diffusion delivers reasoning-model quality without reasoning-model latencyThe shift from "token maxing" to "value maxing" in enterprise AI economicsWhere diffusion LLMs are already winning in production: voice, customer support, search, and coding sub-agentsHow purpose-built AI chips compound with diffusion's speed advantageThe open-weight vs. closed-weight debate, and why Grover argues for a healthy open ecosystem despite building a closed-weight companyScaling Inception from a research lab to a 40-person product companyAnd more!Takeaway: Speed isn't just a UX win in AI, it's a scientific compounding advantage. Faster inference accelerates reinforcement learning iteration, and the industry's shift from raw scale to intelligence-per-dollar suggests the next real breakthroughs may come from architecture, not just size. About the Guest Aditya Grover is the co-founder and CTO of Inception Labs, a deep-tech AI startup pioneering diffusion-based large language models (Mercury). A former PhD student in Stefano Ermon's lab at Stanford University, Grover was among the early researchers working on generative models before diffusion became mainstream, with a research background spanning generative AI and reinforcement learning. Under his technical leadership, Inception has scaled diffusion LLMs from sub-billion-parameter academic prototypes to production-grade models now deployed by enterprises in voice, customer support, search, and coding applications, with expanding distribution across AWS, Azure, and emerging cloud platforms. Chapters0:00 Cold Open: Speed Is Everything 1:00 Meet Aditya Grover, Co-Founder & CTO of Inception Labs 1:23 From Stanford to Startup 3:36 Transformers vs. Diffusion Models 5:12 What Parallel Generation Unlocks 7:15 The Voice Latency Problem 11:33 Top Use Cases: Support, Search & Coding Agents 13:12 Hardware & the Rebirth of Systems Engineering 18:15 Open-Weight vs. Closed-Weight Models 22:30 The Gap Between AI Cost and Value 25:52 How Diffusion Avoids Wasted Compute 29:12 The Bitter Lesson of AI 33:16 From Researchers to Business Builders 39:04 What Keeps Him Up at Night 43:30 How Inception Uses AI Internally 50:00 Why Speed Defines Quality 53:32 What's Next for Mercury

  3. Oct 29 ·  Video

    Why "World Models," Not LLMs, May Be AI's Next Frontier

    Before Waymo became a familiar sight and before ChatGPT redefined what AI could do, Oliver Cameron was already betting his career on machines that could predict the physical world. In this episode of WaveSession #5, Odyssey co-founder and CEO Oliver Cameron, who previously built and sold Voyage to Cruise, joins Jeff and Ilya to unpack why "world models" may be the next great frontier in AI: systems trained to understand physics, causality, and the real world instead of just text. Cameron traces his path through AI's most consequential decade, starting with his jump into driverless cars in 2015, when self-driving vehicles, not language models, represented the cutting edge of applied AI. He walks through founding Voyage, which deployed real self-driving technology in retirement communities and military bases, and the lesson that shaped his entire career: in safety-critical AI, you tackle the hardest environment first. That philosophy led to Voyage's sale to Cruise, where Cameron spent two and a half years scaling driverless systems in San Francisco, before a pivotal insight in 2023 sent him in a new direction. At Cruise, Cameron and his future Odyssey co-founder noticed something remarkable: the "prediction models" powering driverless cars, systems forecasting how the world would evolve seconds into the future, were extraordinarily good at their narrow task but useless outside a car. That became the founding thesis of Odyssey: next-state prediction of the physical world should be a general-purpose technology, not a single-application tool. The conversation gets into the technical architecture behind Odyssey's approach, autoregressive diffusion transformers trained on internet-scale video, and how it differs fundamentally from language models trained on biased, human-written text, and from other world model approaches like 3D generators and JEPA-style architectures. They also get into the business and research realities of building a foundation model company from scratch. Cameron explains why Odyssey deliberately avoids narrow specialization, why gaming data unexpectedly improves a model's ability to drive cars, and why physics, not any single application, has to be the foundation everything else is built on. Jeff and Ilya trade insights on data marketplaces, compute scarcity, and open-weight video models as an accelerating research ecosystem. Cameron closes with candid thoughts on scaling company culture without losing research taste, why optimism is a strategic advantage in AI, and a preview of Odyssey's upcoming flagship model plus new research into multi-agent world models simulating millions of participants at once. This episode covers: Why world models learn physics from video instead of language from text, and what that changesThe lesson that shaped Cameron's career: tackle the hardest environment first, not lastHow the "prediction models" behind driverless cars became the founding insight for OdysseyWhy Odyssey is built for breadth, not specialization, and how gaming data improves driving performanceWhy text was the easy starting point for AI, not the most complete view of reality"The harness": how Odyssey constrains a general world model toward specific tasks without over-restricting itWhy betting on faster-than-expected progress is itself a competitive strategyAnd more!Takeaway: The industry built language models first because text was cheap and compressible, not because it was the most complete way to understand the world. World models trained on video may be the next real unlock, teaching AI the physics, causality, and dynamics that text was never built to capture. About the Guest Oliver Cameron is the co-founder and CEO of Odyssey, an AI research company building foundation world models, general-purpose AI systems trained to understand and simulate the physical world. Before Odyssey, Cameron founded Voyage, a self-driving car company focused on deploying autonomous vehicles in retirement communities and other real-world environments, which was acquired by Cruise, where he spent two and a half years scaling driverless technology. Earlier in his career, Cameron was involved in AI education at Udacity. He co-founded Odyssey in late 2023 alongside CTO Jeff (formerly of self-driving company Wayve), and the company recently raised $310 million to advance its research into world models, autonomous systems, and multi-agent simulation. Chapters0:00 Intro & Welcome 2:21 Jumping Into Self-Driving Cars (2015) 10:49 Why You Should Tackle the Hardest Problem First 15:41 Data Lessons: Self-Driving vs Language Models 17:28 From Cruise to Founding Odyssey 18:55 What Are World Models? 22:10 Language Models vs World Models 26:20 How Odyssey Differentiates 34:31 Tailwinds: Data, Compute & Scaling Laws 46:50 Building Odyssey: Team & Research Taste 55:41 What Keeps a Founder Up at Night 1:01:18 $310M Raise, Odyssey 3 & What's Ahead 1:02:48 Multi-Agent World Models & Simulation Theory

  4. Oct 22 ·  Video

    Why Firmware Is Going Open Source, Fast

    What actually happens between the moment a server powers on and the instant your operating system boots? In this episode of WaveSession #3, Chris Walter, co-founder of firmware engineering firm 9Elements and founder of the Open Source Firmware Foundation, pulls back the curtain on BIOS, BMC, and root-of-trust firmware, and explains why the once-closed, NDA-locked firmware industry is rapidly going open source. Chris walks through his unconventional path into firmware, starting with a master's thesis attempting to brute-force a cryptographic algorithm using FPGA clusters, and evolving through side-channel attacks, hardware security auditing, and embedded sensor development before landing in BIOS, BMC, and root-of-trust engineering. That security-first background shapes his central thesis: firmware is the invisible foundation underneath every claim a system makes about being secure. If the boot chain is compromised, everything built on top of it, confidential computing, secure enclaves, container isolation, falls apart with it. The conversation gets into why hyperscalers like Meta and Google, followed by mid-sized on-prem operators, are driving the shift from proprietary firmware toward open standards. Custom silicon, multi-host server designs, faster time-to-OS, and the ability to patch security vulnerabilities on your own timeline are pushing companies to demand transparency and unified firmware stacks instead of juggling incompatible vendor blobs. Chris and Ilya trace the parallels to past industry inflection points, the rise of Linux, software-defined networking, and software-defined radio, arguing that firmware is hitting its own tipping point now, compressed into years instead of decades. Efficiency and AI take over the back half of the conversation. Chris explains how open BMC access already lets operators fine-tune fan curves and power draw at scale, and why firmware's deterministic, well-documented nature makes it surprisingly well-suited to AI-assisted development. 9Elements uses Claude extensively for automated testing and firmware workflows, even as legal NDAs around proprietary silicon limit what can be fed into LLMs in the first place. They close on the "second life" of decommissioned hyperscaler hardware, the acquisition upheaval hitting legacy firmware vendors like Phoenix and AMI, and Chris's prediction for a far more fragmented, competitive firmware ecosystem over the next decade, with AMD's OpenSIL initiative as an early proof point. This episode covers: Why firmware is the invisible root of system security, and what happens when that layer is compromisedWhy 9Elements had to go open source to survive, reverse-engineering silicon behavior just to support early customersHow open firmware gives decommissioned hyperscaler hardware a "second life" instead of becoming an unusable brickWhy firmware's deterministic nature makes it a surprisingly strong fit for AI-assisted development, and where NDAs still limit itParallels between today's firmware shift and past tipping points like Linux and software-defined networkingWhy longer hardware lifecycles, 7 to 12+ years in production, are forcing companies toward open, self-serviceable firmwareWhere the firmware industry is headed by 2030, including AMD's OpenSIL and the fallout hitting legacy vendorsAnd more!Takeaway: Firmware has quietly become the thing every other security claim depends on, and the industry that used to lock it behind NDAs is now racing, at Linux-and-SDN speed, to open it up before longer hardware lifecycles and hyperscale demands make the old model unworkable. About the Guest Chris Walter is the co-founder of 9Elements, a firmware engineering company specializing in open source BIOS, BMC, and root-of-trust firmware for server and embedded systems. He also co-founded the Open Source Firmware Foundation, a nonprofit umbrella organization supporting open firmware projects, and helps organize the annual Open Source Firmware Conference. Chris's background spans hardware security research, side-channel and fault-injection analysis, and embedded firmware development, giving him a security-first lens on how modern computing infrastructure boots, verifies, and protects itself. Chapters 0:00 Why the Firmware Industry Is Broken0:48 Meet Chris Walter of 9Elements4:07 What Is Firmware?7:56 Open Source Firmware Innovation10:04 BIOS, BMC & Root of Trust13:57 Why Nine Elements Went Open Source18:24 Hyperscalers Driving Open Firmware22:01 Power Efficiency & Fan Curves29:11 Binary Blobs & the IP Debate36:18 AI in Firmware Development47:23 Giving Hardware a Second Life50:34 The Future of Firmware57:00 AMD's OpenSIL Initiative58:48 How to Get Involved

  5. Oct 15 ·  Video

    From Contract AI to Agentic AI: Jerry Ting's Enterprise Playbook

    From Grammarly-before-Grammarly to leading Agentic AI at Workday, Jerry Ting's journey reveals what it actually takes to bring AI into the world's largest enterprises. In this episode of WaveSession #2, Jeff and Ilya sit down with the Evisort co-founder to unpack contract AI, enterprise trust, and why "don't trust us, test us" became a winning sales philosophy. Ting's path to founding Evisort started long before the current AI boom, back when he was building rudimentary grammar-correction tools for his father as an undergrad studying computational linguistics. That early obsession with natural language processing collided with a very human problem at Harvard Law School: reading dense, error-prone contracts by hand. Ting and his MIT-based co-founder spent years in 2016 manually labeling legal clauses and training models before "evals" and "accuracy" were even industry buzzwords, a level of foundational rigor Ting says gives founders from that era a real edge in today's faster-moving AI landscape. The conversation gets into what actually drives enterprise AI adoption and expansion. Ting credits Evisort's growth, customers expanding fourfold post-sale, to forward-deployed engineers who combined legal domain expertise, technical architecture skills, and change-management instincts. He's candid about where many AI startups go wrong: staying too narrow, selling a single point solution, and losing customers to platform players once IT leadership gets involved. They also tackle the murky question of AI ROI: Ting argues that framing AI as "hours and dollars saved" gets you in the door for six-figure deals, but real strategic alignment with the CFO and CEO is what unlocks seven-figure contracts. Ting and the hosts explore the human side of this shift too: job displacement anxiety among law students and entry-level workers, the rise of internal "AI transformation" roles sitting inside legal, finance, and HR rather than IT, and why hands-on experimentation, not passive consumption of AI content, is the only real path to fluency. Ting shares a vivid story of training 500 Workday salespeople to use AI for account research in real time, and reflects candidly on his own transition from founder to Workday executive, including the emotional complexity of an acquisition and why "resting and vesting" was never really an option for him. They close with Ting's outside view on VC and private equity firms building arms to acquire and transform legacy "Middle America" businesses with AI, and why he believes adoption is no longer optional for anyone in a decision-making seat, even amid European regulatory skepticism. This episode covers: Why "don't trust us, test us" became Evisort's core sales strategy in an era of AI skepticismWhy value-based ROI framing has a ceiling, and what actually unlocks seven-figure enterprise dealsHow forward-deployed engineers drove 4x post-sale expansion at EvisortWhy AI transformation roles are emerging inside legal, finance, and HR instead of ITWhy hands-on usage, not watching demos, is the only real path to AI fluencyThe emotional reality of Ting's transition from founder to Workday executive after the acquisitionHow VC and PE firms are converging to acquire and transform legacy businesses with AIAnd more! Takeaway: Enterprise AI adoption isn't won by the most sophisticated model. It's won by trust, hands-on proof, and the domain experts inside legal, finance, and HR who actually understand the workflow being transformed. About the Guest Jerry Ting is the co-founder of Evisort, an AI-powered contract lifecycle management platform he built after identifying inefficiencies in contract review while attending Harvard Law School. Under his leadership, Evisort grew to serve roughly 30% of the Fortune 500 before being acquired by Workday. Ting now serves as General Manager of Agentic AI at Workday, where he leads AI engineering and product strategy. He also teaches at Harvard Law School's innovation lab, advises startups and private equity firms on AI strategy, and is pursuing public company board opportunities focused on enterprise AI transformation. Chapters 0:00 Cold Open: Why Law Tests the Frontier of AI0:37 Meet Jerry Ting: From Harvard Law to AI Founder4:38 The AI Renaissance & Why Contracts Were the Perfect Problem8:45 “Don’t Trust Us, Test Us”: Evisort’s Go-To-Market13:47 Forward Deployed Engineers & Value-Based Selling19:32 Reimagining Work & Upleveling Employees23:34 AI Upskilling: “Just Touch It”29:31 Selling Evisort to Workday34:36 The Refounding of Tech38:31 Enterprise AI Obstacles & Job Displacement45:24 VC vs. PE: Winning the Vertical AI Race50:23 The Rise of AI Transformation Roles

  6. Oct 8 ·  Video

    Sebastian Thrun on Physical AI and the Future of Work

    Long before ChatGPT made AI a household word, Sebastian Thrun was already building it, winning the DARPA Grand Challenge with Stanley, founding Google X and Waymo, and launching Udacity to democratize tech education. In this episode of WaveSession #1, Thrun joins Jeff and Ilya for a wide-ranging conversation on physical AI, the collapse of the hardware-software divide, and why the next decade of innovation will look nothing like the last. Thrun opens by revisiting the 2005 DARPA Grand Challenge, where his team's self-driving car, Stanley, navigated 130 miles of Mojave Desert terrain using little more than a trunk-mounted computer and roughly 100 parameters, a stark contrast to today's trillion-parameter models. He explains how that early work exposed the hard tradeoffs of physical AI: unlike a hallucinating chatbot, a hallucinating self-driving car can kill someone, which is why reliability, not just capability, became his team's obsession. That same thread carries into his account of building Waymo, now 17 times measurably safer than human drivers. From there, the conversation moves into Thrun's years running Google X, where he helped incubate Google Brain, Google Glass, and a string of moonshot projects built on the belief that technology follows an exponential curve and pessimists are almost always wrong. He's candid about failure here: Google Glass didn't fail because the underlying vision was wrong, he argues, but because the form factor and go-to-market were wrong, a distinction he credits to a passage in Mark Pincus's book about how entrepreneurial intuition tends to be right even when specific ideas are wrong. Thrun connects this directly to his current obsession: building an "AI engineer" capable of collapsing the fragmented, over-engineered way hardware and software teams work today into a single, radically more efficient design process. The conversation also digs into what AI means for careers, education, and the workforce. Thrun argues that specialists, lawyers, pilots, accountants, are the most exposed to automation, while generalists who can adapt across domains will thrive, a shift he says is already visible in his own companies, where engineer-to-product-person ratios have flipped. He's equally direct about the societal stakes: this isn't a technology revolution, he says, it's a societal one, requiring redistribution, dialogue, and deliberate effort to keep opportunity from concentrating in too few hands, grounded in his experience scaling Udacity across the Middle East, including a measurable GDP impact in Egypt. Toward the end, the conversation turns personal. Thrun talks about his hiring philosophy (he wants to hear what you screwed up, not just what you're proud of), his belief that blame is corrosive and self-reflection is a competitive advantage, and his own admission that even after decades of building companies, he still feels like an eager undergraduate. He closes with two practical rituals for staying oriented amid constant change: "non-pessimist Tuesday" and "no-blame Thursday." This episode covers: Building Stanley and winning the 2005 DARPA Grand Challenge with roughly 100 parametersWhy physical AI has almost zero tolerance for error compared to language modelsInside Google X: moonshots, Google Brain, and why Google Glass failed as a product but not as a conceptWhy specialists are most exposed to automation while generalists will thriveWhy Thrun calls this a societal revolution, not just a technological oneHis hiring philosophy: asking candidates what they screwed up, not what they're proud of"Non-pessimist Tuesday" and "no-blame Thursday": simple rituals for staying oriented amid constant changeAnd more!Takeaway: Two decades at the frontier of AI taught Thrun the same lesson over and over: intuition is usually right even when the specific idea is wrong, reliability matters more than raw capability, and the real risk of this moment isn't what AI can do, it's who gets left out as it reshapes everything. About the Guest Sebastian Thrun is a serial entrepreneur, AI researcher, and educator who has spent over two decades at the forefront of applied artificial intelligence. He led the Stanford team that won the 2005 DARPA Grand Challenge, founded and led Google X (Alphabet's "moonshot factory"), and co-founded Waymo, now one of the world's leading autonomous vehicle companies. He also founded Udacity, an online education platform focused on making technology training accessible globally, with measurable economic impact in markets including Egypt, Saudi Arabia, and Palestine. Thrun previously served as a professor at Stanford University and Carnegie Mellon University, and continues to work hands-on building AI-driven robotics and engineering tools. Chapters0:00 — Opening: AI Access for Everyone 0:52 — The Stanley Era: Early Autonomous Vehicles 5:25 — From Chauffeur to Waymo: Physical AI Safety 7:54 — Google X & the Moonshot Factory 12:21 — Transforming Organizations: AI 1.0 vs. 2.0 15:33 — The 20x Productivity Shift in Software 17:20 — Democratizing Education & Commoditized Intelligence 22:30 — Google Glass: Intuition, Failure & Project Wingman 29:21 — Specialists vs. Generalists in the AI Era 33:55 — Advice for New Engineers & Graduates 38:21 — Intuitions vs. Ideas: Lessons from Failure 42:39 — Frontier Models Eating Vertical SaaS

  7. Oct 1 ·  Video

    The Real Bottleneck Isn't AI Chips. It's Power.

    We spent two years worrying about GPU supply. That worry is over, and almost nobody has caught up to what replaced it. In the premiere episode of wavesession, Jeff and Ilya cut through the noise around AI's future and land on the one constraint nobody's pricing in: energy. Not silicon. Not compute. Power. This conversation makes the case, in detail, for why the grid, not the chip, is now the hard ceiling on how far AI can scale. The thesis starts with a simple historical pattern: energy intensity has always tracked human prosperity. Jeff and Ilya argue that compute, the ability to deliver intelligence at scale, is becoming its modern successor. They trace how decades of deferred maintenance on the U.S. power grid, much of it running on infrastructure from the 1950s and 60s, is now colliding head-on with a demand curve nobody designed for. A decade ago, a 30-megawatt data center was considered large. Today's hyperscale campuses are pushing past 100 megawatts, built for a load the grid was never built to carry. There's a culture problem underneath the technical one, too. Utilities run on a "five-nines" reliability mindset: slow, cautious, built for stability over decades. Data center operators run on mission-critical, can't-go-down urgency. Two industries, two completely different clocks, now being forced to speak the same language at the exact moment it matters most. This episode covers: Why chip supply solved itself while energy infrastructure never caught upThe historical link between energy abundance and prosperity, and why compute may be its modern equivalentA Duke study showing the U.S. grid could free up roughly 100 gigawatts of hidden capacity if data centers curtailed demand by just 2.5 to 3% at peak, no new power plants requiredHow hyperscalers kept power demand flat for nearly a decade during the cloud buildout through efficiency gains, and whether that pattern can repeat for AIThe moment betting on unproven AI infrastructure stopped looking risky and started looking inevitableWhy efficiency and sustainability have to be built in now, not deferred to some future AGI breakthrough that may never show upAnd more!Takeaway: The AI industry solved its chip problem. It hasn't solved its power problem, and that gap, not GPU supply, is what will actually determine how far and how fast AI can scale from here. Chapters 0:00 — Cold Open: Kevin O'Leary & the AI Hype Problem1:54 — The Computing Stack: GPUs, CUDA & FPGAs10:50 — The FPGA to GPU Pivot & Going All In on AMD16:31 — AI's Impact on Society & Public Fear22:03 — Energy Intensity as a Prosperity Indicator24:37 — The Power Bottleneck: GPUs Abundant, Power Scarce25:22 — Grid Complexity & Data Center Curtailment33:44 — Efficiency Opportunities in AI Data Centers43:01 — Ilya's Vision: Building the Fourth Utility44:11 — Jeff's Vision: Bridging the Adoption Gap50:10 — From "What Can We Compute" to "What Should We Compute"

  8. Sep 16 ·  Video

    The wavesession Podcast: AI Infrastructure, Data Centers, Chips & What's Next

    Introducing wavesession. The people setting the direction of the open AI ecosystem are making calls right now. Founders, researchers, builders, capital providers. They're deciding what gets built, how it scales, and who gets a preview of what's coming next. wavesession is where those conversations happen. Each episode moves across the stack. From the chip to the grid, the software to the firmware, the language to the physics underneath it. These are the people making those decisions in real time. This isn't a highlight reel. There are no polished talking points. It's an honest account of what it actually takes to build open AI infrastructure at this scale, where the technical, financial, and human problems don't come with a playbook. This is a technological revolution happening in the open. It comes with consequences we're all going to live with. wavesession is an invitation to listen in, challenge what you hear, and help shape where it goes. New episodes coming weekly, starting September 30th.

    The wavesession Podcast: AI Infrastructure, Data Centers, Chips & What's Next
  • 1 Trailer
    7 Episodes

Trailer

Ratings & Reviews

5
out of 5
2 Ratings

About

WaveSession is a conversation series with the people setting the direction of an open ecosystem around AI. In this show we interview the founders, researchers, builders, and capital providers making the calls on what gets built, how it scales, and who are getting sneak peeks into the future. Each episode moves across the stack, from the chip to the grid, the software to the firmware, the language to the physics beneath it. These are the calls being made in real time, by the people making them. Each session is an honest account, where we explore beyond the superficial, not a highlight reel. The guests are testing approaches that haven't been proven yet, because building open AI infrastructure at this scale means solving problems that don't have a playbook: technical, financial, and human, all at once. WaveSession treats this moment as what it is: a technological revolution, built in the open, with societal consequences we get into as they unfold. It invites the audience to challenge, question, and shape where it goes.

Information

  • Creator
    TensorWave
  • Years Active
    2026
  • Episodes
    8
  • Rating
    Explicit
  • Copyright
    © TensorWave
  • Show Website
    WaveSession

You Might Also Like