20VC with Harry StebbingsInference Will Be $10TRN Market: What No One Sees | Software Investors are F***ed | Naveen Rao
20VC with Harry Stebbings
Energy becomes the bottleneck 0:00
Energy supply is hard to build, and the speaker says people underestimate that. He points to about $50 billion to add one gigawatt, which is more costly and complex than a semiconductor fab. The near-term source of data center power has been excess capacity on the grid, but that surplus will run out soon. He says the United States is behind and will need much more buildout.
Compute shifts up stack 9:00
The speaker says the biggest gains in compute now come less from physical chip shrinkage and more from architecture, smaller data types, and model design. He says model-level locality and reduced data movement matter more than they used to. The work should be energy first. He traces the old goal of computers back to ENIAC, where each new machine had to be faster than the last one.
Hardware and intelligence 15:01
There is still room to improve model efficiency because software gains on a standard GPU are limited by the physical design underneath. The speaker says the brain shows a different path: intelligence and hardware co-evolved, so sparse connections can be exploited far more when the machine is built for them. That is why specialized chips for specific workloads, from Cerebras to others, look like the future. He ties this back to Turing’s goal of making an intelligent machine.
Verification and judgment 23:01
AI has widened the surface area a team can cover. A question that once took weeks of research and code can now be asked of a model and answered in an hour, though the output still needs checking. He calls this an AI slop problem, since models can produce many artifacts, code included, that are hard to verify. He also says humans still beat agents at judgment, like a neurosurgeon after years of training.
Trust and training 30:00
[30:00] Agents may become more trusted than people, but not yet. Value comes from the whole machine, not a separate model. Like a neurosurgeon or chef, a machine matters through how it was taught. The important thing is reaching a real, verified result, and the data gathered over years is part of that value.
RL and commoditization 37:31
[37:31] Reinforcement learning matters because you cannot pre-train for every task. It gives verifiable rewards, like a tennis serve that is fast and in bounds. He expects a post-transformer world, with more efficient reasoning methods, and says founder skill still matters, though vision matters more. Data stays valuable, but some intelligence and inference will commoditize, lowering margins while the market grows.
Software margins fade 44:30
Pure software investors are in trouble, because SaaS no longer looks like the place where margins get won. The easy cloud-era play, where you could spin up a product fast and test a small market, is giving way to applications built around an intelligence layer. The speaker thinks the classic software market will shrink over time, even if the labels change.
AI services and robotics 46:30
The new model is software plus services, with AI and humans working together to deliver an outcome. Harvey is the example given: it uses AI, checks the output with humans, and does legal work faster and cheaper than the old method. Big firms may build their own systems, but most companies will still buy out-of-the-box tools. Robotics is seen as much larger than coding, with working prototypes in a few years.
Models and robots 59:31
He thinks model makers have not yet settled what they will become. He does not expect the value to sit in the model alone. He expects companies to build end-to-end devices, like robots that use a model to do one thing very well. In that view, the hardware and the model will stay joined.
Money and conviction 1:00:01
He does not worry much about price when the total market is this large. He prefers to keep some of his wealth in the space because he sees major growth ahead. He also says his views on investing have changed. For deep technical work, he now cares less about market validation and more about visible progress on the technical problem itself.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
