BlackRock's Tony Kim on AI's Next Winners? Chips, Memory, Robotics & Quantum

Summary

BlackRock’s Tony Kim frames AI as a compute-centric regime that is reshaping market structure, data centers, and the semiconductor supply chain at once. He argues the old cloud/SaaS infrastructure was built for a different era, while AI demands far larger servers, denser power delivery, shorter data paths, and a shift from copper to light. Memory is becoming a major bottleneck as models grow more agentic and memory-intensive, but fabs take years to build, creating a supply-demand mismatch. On the investing side, he says most capital belongs in the next three-year AI cycle, while a smaller allocation should be reserved for frontier technologies that could converge around 2030, including quantum, SMRs, fusion, and orbital data centers. He also sees robotics as a major coming category, with China leading company formation and IPOs, and thinks social robots could be a huge market for companionship and aging populations.

Notes

• AI has shifted the economy from software-centric to compute-centric value creation.
• He uses market-cap buckets to argue chips and hardware are now a massive standalone pool of value.
• He is publicly focused on next-gen accelerators, quantum, optics, and AI chip co-design.
• He describes AI infrastructure as a full rebuild, not an incremental upgrade.
• The cloud era was optimized for hosted software, not AI-native compute loads.
• AI data centers are being redesigned around much shorter communication distances.
• The physics bottlenecks are power density, energy, and grid capacity.
• He sees a move from copper interconnects toward optics inside the data center.
• Model-silicon co-design is becoming the new norm for leading foundation labs.
• Compute architecture is being tailored more tightly to model parameters and inference needs.
• He says AI is pushing the industry toward memory-rich systems, not compute-only systems.
• The memory stack he cites includes SRAM, DRAM, stacked DRAM, HBM, and high-bandwidth flash.
• He attributes the RAM shortage to AI systems becoming more brain-like and memory intensive.
• Memory supply is constrained by a multi-year fab build cycle.
• He says today's demand is colliding with 3-4 year memory capacity lead times.
• His allocation framework puts the majority of capital into the current three-year AI window.
• He wants the rest of the portfolio in longer-duration optionality that can still matter in five years or more.
• 2030 is his recurring milestone for frontier technologies.
• He specifically ties orbital data centers to the 2030 timeline.
• He argues chips are not commodities because the industry is highly concentrated and highly profitable.
• He says chip margins exceed software, pharma, industrials, and telecom.
• Robotics is, in his view, a brain-plus-body system rather than a single product category.
• He separates the robot stack into language/memory and motor-control/world-model components.
• He says China has a large robotics company pipeline and may produce 30 to 40 IPOs this year.
• He contrasts China's robotics funding path with the U.S., saying China uses public markets earlier.
• He is especially interested in social robots for loneliness, education, and elder care.
• He thinks social robots may be more important than industrial robots over time.
• He links robotics demand to aging populations and falling birth rates.
• He says enterprises are being redesigned around tokens, data, and a context layer.
• He explicitly maps enterprise AI as ontology/context on top of data with tokens flowing through agents.
• He defines token flow as the pathway from compute to foundation labs to apps and services.
• Inference clouds, edge clouds, and edge AI occupy the last-mile token role.
• He thinks the best app businesses may wrap whole workflows rather than sell point solutions.
• He thinks private-equity roll-ups are targeting customer bases and removing inefficient cost layers.
• Traditional service industries may be refactored end-to-end if AI adoption remains low.
• He expects the compute wall, memory wall, and data-center redesign to keep advancing over the next year.
• He wants continued proof points on data-center reimagination.
• He thinks big foundation labs may go public in the next 12 months.
• Orbital data centers are one of his preferred next frontier bets.
• He argues moving compute into space could reshape how terrestrial data centers are built.
• He says he was given latitude and trust by people who helped him enter BlackRock and investment banking.
• His mentors include historical figures he studies, rather than living operators.