Overview of 20VC with Naveen Rao
This episode is a wide-ranging conversation between Harry Stebbings and Naveen Rao, founder of Unconventional AI and former Mosaic ML leader, about where AI is heading next. Naveen argues that the real bottlenecks are shifting from model quality to energy, hardware, inference efficiency, data quality, and verification. His core thesis: the next decade of AI will be shaped less by “better software” and more by the ability to convert energy into intelligence efficiently, with robotics, specialized hardware, and post-transformer architectures creating massive new markets.
Key Themes and Big Ideas
Energy is the real constraint
- Naveen believes energy supply, not just chips or models, will be the limiting factor for AI growth in the next few years.
- Building power infrastructure is extremely expensive and slow; he cites gigawatt-scale builds as comparable in complexity to semiconductor fabs.
- He thinks the U.S. has relied too long on excess grid capacity and is now behind China in total energy infrastructure buildout.
Intelligence demand is effectively infinite
- Naveen frames intelligence as an asset with “infinite demand” in an adversarial world.
- As AI becomes more useful, demand for compute, inference, and energy will expand rather than shrink.
- He sees the inference market as potentially reaching $10 trillion within 10 years.
Hardware and infrastructure matter more than pure software
- His view is blunt: if you’re a pure software investor, you may be in trouble.
- He believes the highest-value layer over the next decade will be the hardware stack: chips, systems, cloud, and networking.
- He does not think the AI value chain will commoditize evenly; instead, value will concentrate where energy is converted into intelligence most efficiently.
We are moving into a post-transformer world
- Naveen agrees with the idea that transformers are not the final answer.
- He argues current reasoning approaches are inefficient and that new architectures will emerge.
- He thinks the next breakthroughs will come from co-designing the model, architecture, and physical hardware together.
AI Model, Data, and Compute Insights
Physical efficiency gains are slowing
- Traditional Moore’s Law-style gains are largely over.
- Most recent efficiency gains have come from:
- smaller numerical precision, like 32-bit to 16-bit to 8-bit
- better architecture
- better use of locality and reduced data movement
Data and curriculum are becoming the secret sauce
- Naveen says model quality is increasingly driven by:
- data selection
- training curriculum
- post-training methods like RL and supervised fine-tuning
- He believes specialized datasets and higher-quality data are becoming highly defensible.
Verification is the new bottleneck
- AI can now generate code, models, and artifacts very quickly.
- The problem has shifted from producing output to verifying correctness.
- He sees humans increasingly acting as high-skill reviewers and judges of AI output.
RL matters because real-world tasks need verifiable rewards
- Naveen explains why reinforcement learning is so important:
- many tasks cannot be solved by pre-training alone
- RL works when there is a clear, measurable outcome
- He sees RL as a practical path to making models better at high-value tasks.
Robotics: Naveen’s Most Aggressive Bet
Robotics will be much bigger than coding agents
- Naveen is highly bullish on robotics and thinks it will be 100x bigger than agentic coding.
- His reasoning:
- the physical world is still largely untouched by intelligence
- robots can unlock massive economic value in labor-heavy, physical tasks
- the cost and friction of real-world operations is enormous today
Timeline for robotics
- He expects working prototypes for some tasks within 2–3 years
- He expects the broader market to emerge in 5–7 years
- He is more confident in non-humanoid robotics in the near term than humanoids
What he would invest in
- He prefers the technology stack, not the consumer application layer yet.
- He believes the companies building the models may eventually build the full device or system.
- He sees robotics as tightly coupled with model development and physical product design.
Market Structure and Investing Implications
SaaS and pure software are less attractive
- Naveen thinks the classic SaaS era, where software alone captured huge margins, is fading.
- He believes the new market structure is more about:
- software + services
- AI + humans
- hardware + intelligence
- He thinks software investors need to adapt fast or risk being left behind.
Open models will dominate tokens, frontier models will dominate margins
- His rough view:
- 80% of token traffic will flow through open models in a few years
- frontier models will still capture a disproportionate share of revenue and margin
- He compares this to the smartphone market, where a smaller share of devices captures the majority of profit.
Commoditization is real, but it does not eliminate value
- Naveen thinks:
- some intelligence will commoditize
- inference prices will likely fall
- but the market will still grow dramatically
- In his view, falling unit prices expand the total market through something like Jevons’ paradox.
Specialized companies will win
- He thinks the future belongs to companies that specialize deeply in:
- data
- workflows
- specific physical tasks
- high-value verticals
- Generic tools are easier to copy; specialized systems with strong feedback loops are more defensible.
Founder and Startup Advice
Vision matters more than execution alone
- Naveen thinks the AI era raises the bar for founders.
- In the past, strong execution alone could build a good company.
- Now founders need a real view on where the world is going, because the market is moving too fast.
Market validation is not always the right metric
- He says investors should stop over-indexing on “market validation” for deep-tech companies.
- For hard technical problems, the key question is whether the underlying problem is being solved.
- If the problem is real and the solution works, the market will follow.
Big labs move fast; startups should pursue deeper innovation
- He believes large AI companies are mostly executing quickly on existing paradigms.
- Startups have the freedom to pursue more radical breakthroughs.
- He compares this to drug development:
- big companies do the incremental work
- startups take the biggest risks
Notable Quotes and Ideas
- “We’re going to be out of energy pretty soon in the next few years.”
- “Intelligence, to me, will have infinite demand.”
- “I do think we will be in a post-Transformer world.”
- “If you’re a pure software investor, you’re kind of fucked.”
- “The inference market is going to be a $10 trillion market within 10 years.”
- “The problem has moved from producing to verification.”
- “I actually think robotics is going to be 100 times larger than agentic coding.”
Main Takeaways
- AI’s next bottleneck is likely energy and infrastructure, not model hype.
- The biggest value may accrue to hardware, systems, and inference efficiency.
- Verification will become a central layer in AI workflows.
- The future is likely post-transformer, specialized, and hardware-aware.
- Robotics may become one of the largest new AI markets, far bigger than many software-only categories.
- Founders and investors should focus on deep technical progress, not just surface-level product adoption.
Bottom Line
Naveen Rao’s worldview is extremely bullish on AI, but not in the typical “more software, more apps” sense. He believes the real opportunity lies in rebuilding the compute stack around intelligence, power, and specialization. In his view, the winners will be the companies that can efficiently turn energy into useful intelligence across software, hardware, and eventually the physical world.