Skip to content
a16za16z

The Chip That Could Unlock AGI.

Naveen Rao is cofounder and CEO of Unconventional AI, an AI chip startup building analog computing systems designed specifically for intelligence. Previously, Naveen led AI at Databricks and founded two successful companies: Mosaic (cloud computing) and Nervana (AI accelerators, acquired by Intel). In this episode, a16z’s Matt Bornstein sits down with Naveen at NeurIPS to discuss why 80 years of digital computing may be the wrong substrate for AI, how the brain runs on 20 watts while data centers consume 4% of the US energy grid, the physics of causality and what it might mean for AGI, and why now is the moment to take this unconventional bet. Timecodes: 00:00 - Trailer 00:56 - Exploring hardware for running AI workloads 02:02 - Why Naveen built lots of software in a "hardware company" 03:22 - Why start a new chip company? 05:13 - How computing systems went digital 09:26 - Why intelligence is a good fit for analog computer systems 12:30 - What tradeoffs Naveen faced in pursuing his own path 15:23 - The Data modalities Unconventional chips will be best for 16:54 - Does this get us closer to AGI? 21:00 - Where Naveen gets his excitement and motivation 22:37 - What makes Naveen confident that Unconventional will work 24:43 - Unconventional's hiring priorities 26:27 - Career advice for young people 28:19 - What Naveen has done best in his companies Resources: Follow Naveen on X: https://twitter.com/NaveenGRao Follow Matt on X: https://twitter.com/BornsteinMatt Stay Updated: Follow a16z on X: https://twitter.com/a16z Follow a16z on LinkedIn: https://www.linkedin.com/company/a16z Follow the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Follow the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see http://a16z.com/disclosures.

Matt Bornsteinhost
Dec 8, 202530mWatch on YouTube ↗

CHAPTERS

  1. Trailer: AI as humanity’s next evolution and the “new chip” premise

    The episode opens with a bold framing: AI as a transformative step for humanity, and Naveen Rao as a builder who’s repeatedly been early to major waves. The trailer tees up the central bet—rethinking compute hardware to make AI more efficient and more ubiquitous.

    • AI framed as a civilizational shift, not just a product category
    • Naveen Rao positioned as an early, contrarian builder across AI and systems
    • Unconventional AI introduced as a new approach to compute for intelligence
    • Core tension: current compute paradigm may not scale (especially on energy)
  2. Why hardware matters for AI workloads: Rao’s “shrink it, speed it up” mindset

    Rao explains how his early career trained him to translate algorithms into efficient, real-time systems by building specialized hardware. That background shapes his view that the right hardware substrate can unlock entirely new applications and form factors for AI.

    • Specialized hardware historically enabled real-time workloads (e.g., compression, wireless)
    • Efficiency and form factor constraints often force hardware innovation
    • Rao’s cross-training: industry hardware + PhD in neuroscience
    • Perspective: if something is too slow/inefficient in software, reconsider the substrate
  3. Full-stack before it was trendy: why he built software inside “hardware companies”

    Rao argues the boundary between software and hardware is artificial—just a line drawn around what’s configurable. He describes an older definition of “full stack” spanning silicon, architecture, low-level software, and applications, and why that breadth enables better system design decisions.

    • “Full stack” as silicon-to-application understanding
    • Hardware/software boundary is a design choice, not a natural law
    • Right-sizing: decide what must be fixed vs. programmable for the problem
    • Breadth across layers helps identify the best leverage points
  4. Why start Unconventional now: first principles, brains, and the energy wall

    Rao says Unconventional isn’t “a chip company” first—it’s an exploration of how learning works in physical systems. He ties the motivation to biology’s efficiency and the growing reality that global compute is becoming fundamentally energy-constrained.

    • Company focus begins with theory and first-principles learning in physics
    • Biology as inspiration: efficiency, adaptivity, dynamic energy usage
    • We’ve largely built the same digital-computing paradigm for ~80 years
    • AI demand is pushing computing into an energy-limited regime
  5. How computing went digital: scalability beat analog’s efficiency

    The conversation contrasts digital and analog: digital represents numbers with bits and trades precision for generality and scalable manufacturing. Analog computing was early and efficient, but variability and scaling challenges pushed the industry toward robust digital abstractions.

    • Digital computing: numeric representation with finite-bit precision and error
    • General-purpose simulation drove digital dominance
    • Analog computers were efficient but hard to scale due to variability
    • Historical parallel: ENIAC-scale tube counts resemble today’s “scale-out” training clusters
  6. Why intelligence may fit analog/mixed-signal: stochastic brains vs deterministic substrates

    Rao argues neural networks are inherently stochastic and distributed, yet we run them on highly deterministic, precision-oriented hardware built for arithmetic. He suggests there may be an “isomorphism” in circuits that better matches intelligence, using physics directly rather than layered abstractions.

    • Neural networks as stochastic machines; brains implement dynamics physically
    • Digital stacks add lossy abstractions (OS/APIs) compared to “intelligence as physics”
    • Analog approaches can compute via the medium’s native dynamics
    • Goal: find circuit-level mappings that naturally support learning and inference
  7. The real constraint: AI at grid scale and the infrastructure bottleneck

    Rao and Bornstein connect AI’s growth to data-center energy use, grid strain, and the difficulty of building generation and transmission fast enough. This creates urgency for compute paradigms that dramatically improve efficiency rather than assuming unlimited power.

    • Data centers already consume a meaningful share of grid energy; AI is accelerating demand
    • Rising risk of brownouts and transmission constraints
    • Estimates imply massive new generation capacity is needed over the next decade
    • Efficiency gains in compute are framed as essential, not optional
  8. Tradeoffs and where analog wins: dynamical systems, time, and “intelligence substrates”

    Rao rejects a binary digital-vs-analog framing and instead argues for workload fit. He highlights dynamical systems—problems with time evolution—as especially amenable to physical computation, while conventional numeric computing remains important elsewhere.

    • Hybrid view: digital remains vital; analog helps where dynamics/time matter
    • Computers “simulate time” numerically; physical systems embody time directly
    • Brains integrate many messy variables yet achieve high precision (athletes as examples)
    • Positioning: build an “intelligence substrate,” not just faster arithmetic
  9. Best-fit modalities and models: transformers today, dynamics-forward models tomorrow

    Unconventional plans to start from what works—transformers and diffusion—while exploring models that explicitly encode dynamics, like diffusion/flow and energy-based approaches. Rao argues transformers exploit GPU-friendly structure but may be parameter-inefficient relative to alternative formulations.

    • Near-term compatibility with transformers/diffusion rather than discarding SOTA
    • Interest in diffusion/flow/energy-based models due to ODE/dynamics structure
    • Hypothesis: map model dynamics onto physical-system dynamics for efficiency
    • Expectation of open-source releases to let others experiment with mappings
  10. Does this move toward AGI: causality, time evolution, and richer primitives

    Rao cautiously argues that systems grounded in dynamics, time, and causality could be a better basis for more general intelligence than static function approximators. He notes today’s models are powerful yet still make “stupid errors,” suggesting gaps in causal understanding.

    • AGI discussion acknowledged as inherently uncertain and “hand-wavy”
    • Claim: dynamics/time-evolution may support better causal reasoning
    • Physical irreversibility and temporal structure as important priors
    • Current AI: useful tools, but not yet comparable to working with a person
  11. Ecosystem strategy: partners, competitors, and manufacturing-scale reality

    Rao outlines a phased ambition: prove a new intelligence paradigm within five years, then ensure it scales in manufacturing to matter at global energy scale. He describes TSMC as a likely partner, and leaves open whether NVIDIA/Google become collaborators or competitors.

    • Two-phase plan: paradigm proof → scalable manufacturable systems
    • Manufacturing scalability is required to impact energy constraints globally
    • TSMC positioned as a key partner for prototyping and production readiness
    • NVIDIA/Google/Microsoft seen as application-frontier players; competition vs collaboration depends on how the paradigm lands
  12. Motivation, confidence, and the “practical research lab” culture

    Rao describes the unique thrill of hardware bring-up and his optimism about AI’s societal upside, positioning himself as strongly non-doomer. He grounds confidence in biology as existence proof plus decades of academic work, and emphasizes an open-ended early culture focused on existence proofs before optimization.

    • Hardware bring-up as a powerful motivator; “turning it on” moment matters
    • AI viewed as enabling deeper collaboration and understanding, despite tradeoffs
    • Confidence sources: brains + 40+ years of research + emerging theory
    • Culture: prioritize existence proofs; avoid premature manufacturability constraints while keeping product focus
  13. Hiring, career advice, and leadership style: agency, breadth, and big hard problems

    Rao details the interdisciplinary team needed—model theorists, systems architects, mixed-signal circuit designers—and the risks of building one of the largest analog chips attempted. He advises young engineers to start in startups for breadth, and explains his leadership approach: attract people to hard problems and maximize agency across the org.

    • Hiring across theory, architecture, and analog/digital mixed-signal implementation
    • Acknowledgment: first prototypes are unpredictable; pushing scale is inherently hard
    • Career advice: prioritize breadth early; startups accelerate cross-stack learning
    • Leadership: recruit for hard challenges, then increase agency and let teams own outcomes

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.