• Favor the AI hardware and compute infrastructure theme over a narrow bet on AI software, with a long-term focus on energy efficiency, inference costs, and specialized chips.
• NVIDIA (NVDA) remains well-positioned in AI hardware, but weigh its strong execution against growing competition from custom chips developed by Microsoft, Google, and Amazon.
• Treat CoreWeave as higher risk: its outlook depends on sustained AI demand, while debt and financing needs are key risks to monitor.
• Watch Bloom Energy (BE) and other data-center power suppliers, but wait for evidence that projects can deliver power reliably and economically before assuming the opportunity is proven.
• Consider robotics a longer-term theme: Rao sees prototypes emerging in two to three years and a broader market in five to seven years, but he provided no public ticker or valuation for Physical Intelligence.

AI Hardware and Compute Infrastructure

• Rao sees hardware as the most attractive part of the AI stack over the next 10 years. He argues that value will come from converting energy into intelligence more efficiently, across chips, systems, cloud infrastructure, and networks.
• He expects specialized chips and systems designed for particular AI workloads, rather than a simple market of interchangeable chips. He also said the physical efficiency gains from shrinking transistors have slowed, leaving room for improvement in model design and computing architecture.
• He believes the market for AI inference could reach $10 trillion within 10 years, and said that a tenfold improvement in inference efficiency could greatly expand the market by making more uses economical.

Takeaways

• The discussion favors exposure to the broader AI hardware and compute stack over a narrow bet on software applications.
• The key measures to watch, based on Rao’s thesis, are energy efficiency, inference costs, and whether specialized hardware can deliver meaningful performance gains.
• These are long-term market forecasts, not guaranteed outcomes; the transcript also emphasizes that energy supply and infrastructure construction could constrain growth.

NVIDIA (NVDA)

• Rao praised NVIDIA’s execution and Jensen Huang’s early recognition of AI as an important computing workload. He said NVIDIA currently captures a large share of the AI stack’s margins.
• He does not expect chips or hardware to become commoditized soon, arguing that efficient conversion of energy into intelligence will remain valuable.
• At the same time, he expects more specialized chips and points to companies building their own processors as part of that trend.

Takeaways

• The discussion is broadly positive on AI hardware demand, but it also points to growing competition from specialized chips and in-house silicon.
• Investors considering NVIDIA should weigh its current position against the possibility that AI workloads become more specialized; Rao did not offer a price target or a specific buy recommendation.

CoreWeave

• The hosts raised concerns about the debt cycle and risks surrounding CoreWeave. Rao said the outlook depends heavily on whether demand for AI services—and therefore GPUs and energy—continues to grow.
• He described current market concerns as a possible “blip,” rather than evidence that the AI boom is ending, and said he expects substantial unmet demand for AI.

Takeaways

• The bullish case discussed depends on continued AI demand and the ability to secure the compute and power needed to serve it.
• The transcript specifically flags debt-cycle concerns, so the company’s financing needs are an important risk to monitor. Rao did not provide a price target or a company-specific recommendation.

Bloom Energy (BE)

• Rao cited a Bloom Energy deal to build power capacity in Arizona as an example of companies pursuing new ways to supply data centers.
• He said that if the approach works, it could provide a needed technological advance, but noted that it had not yet been proven out and that its economics differ from grid power.

Takeaways

• The discussion identifies data-center power supply as a potential investment theme, while treating Bloom Energy’s cited project as unproven.
• The key evidence to watch is whether the project works in practice and whether its economics can support broader deployment.

AI Energy and Power Infrastructure

• Rao said data centers have so far drawn partly on unused capacity in existing power infrastructure. He warned that this excess capacity could become scarce within the next few years.
• He described electricity for data centers as potentially commanding higher prices than power for general use, citing behind-the-meter arrangements as an example.
• He also emphasized that new energy infrastructure is difficult and expensive to build. He cited estimates of about $50 billion per gigawatt and argued that large annual capacity additions may be difficult to achieve.

Takeaways

• Power generation and data-center energy supply are potential beneficiaries of AI growth, but the opportunity depends on projects being built and delivering power economically.
• Energy availability, construction timelines, and high infrastructure costs are explicit constraints in Rao’s argument.

Custom AI Chips and In-House Silicon

• The discussion cited Microsoft, Google, and Amazon as companies developing their own chips, alongside other efforts by AI companies and chip startups.
• Rao expects specialization to increase and said he does not see hardware commoditizing soon. He argued that companies able to tailor the entire computing stack to AI workloads could gain an efficiency advantage.
• He also said he would build more manufacturing capacity of his own if resources were unlimited, to reduce supply constraints and tailor components to specific needs.

Takeaways

• The theme supports examining companies across chip design, AI systems, and hardware manufacturing—not only the largest general-purpose GPU supplier.
• The transcript does not identify a single winner in custom silicon. Competition, execution, and the ability to achieve real efficiency gains are central considerations.

Robotics and Physical Intelligence

• Rao argued that robotics could become 100 times larger than agentic coding, because intelligent machines could address physical-world tasks such as construction, cleaning, and repairs.
• He expects working prototypes for some tasks in two to three years, with a broader market emerging in five to seven years.
• He said he is more interested today in the technology stack—such as models and related infrastructure—than in the application layer. He named Physical Intelligence as the company he would choose if he had to put all his money into one company outside his own.
• He expects robotics companies may ultimately combine a model with an end-to-end device rather than sell a model separately.

Takeaways

• Robotics is a high-upside, longer-horizon theme in Rao’s view; the timelines he gave are his estimates, not established outcomes.
• His comments point toward the enabling technology stack and integrated robot-and-model offerings, rather than betting broadly on today’s robot applications.
• Physical Intelligence received a particularly strong endorsement from Rao, but the transcript gives no valuation, price target, or investment terms.

AI Models, Data, and Specialized Datasets

• Rao said high-quality, specialized data and the training processes used to shape models are important sources of value. He argued that data began as a volume-gathering activity but becomes more valuable when curated for specific tasks.
• He expects some level of AI intelligence to become commoditized: models that are several months behind the frontier can still be useful. He suggested that routine tasks may increasingly use open models, while a smaller number of important tasks rely on frontier models.
• His estimate for three years from now was that open models could handle 80% of token traffic, while frontier models could still account for about half of spending.

Takeaways

• The discussion favors specialized data, model training, and efficient deployment over assuming that the most powerful model will capture all the value.
• Investors can track how companies differentiate through data quality and task-specific performance, while recognizing Rao’s view that some model capabilities and margins may commoditize.

Agentic Coding, Cursor, and Pure-Software Investing

• Rao called agentic coding an overrated market for investors outside the major model providers. He argued that the underlying model may capture more value than the coding interface or “harness.”
• He described tools such as Cursor as not especially sticky, based on what he has observed at his own company, where people move between tools.
• More broadly, he warned that investors focused purely on traditional software or SaaS could be poorly positioned as applications become more dependent on AI capabilities.
• He also described AI-assisted services—such as legal work supported by software and human review—as a more compelling model than simply automating a narrow task.

Takeaways

• Rao’s view is bearish on standalone agentic-coding tools as a category, while remaining positive on AI’s ability to increase productivity and expand the amount of work that gets done.
• A useful distinction from the discussion is between a thin application that can be replaced by a model provider and a business that delivers a verified outcome using AI, human expertise, or a specialized workflow.
• His comments are a point of view, not a prediction that all software businesses will fail.

Microsoft, Google, Amazon, and AI Model Providers

• The transcript mentions Microsoft, Google, and Amazon in the context of building their own chips. It also discusses model providers including OpenAI, Anthropic, Google, and xAI.
• Rao said frontier model companies face pressure to release competitive models quickly, which can make it difficult for them to pursue riskier, fundamental research.
• He expects open models to serve many routine tasks, while frontier models remain valuable for a smaller set of demanding uses.

Takeaways

• The discussion suggests evaluating large technology companies not only as software businesses, but also for their AI models, custom chips, and infrastructure.
• Rao’s outlook implies a possible split: broad use of lower-cost open models, with higher spending concentrated in frontier capabilities. He did not give company-specific price targets or recommendations.

Unconventional AI

• The episode’s introduction says Rao founded Unconventional AI and that its latest funding round valued it at nearly $5 billion. Rao described the company’s goal as building a more energy-efficient computer, designing the model, architecture, and physical hardware together.
• He said the company is pursuing an “energy first” design approach rather than treating existing models or computing architectures as fixed.

Takeaways

• Unconventional AI represents a private-company opportunity tied to the hardware and energy-efficiency themes in the discussion.
• The transcript provides a funding valuation, but no public-market ticker, investment terms, or specific performance results.