By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Sourcery with Molly O'Shea

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

Tony Kim is Managing Director and Head of the Global Technology Team within Fundamental Equities at BlackRock.

Molly O'SheahostTony KimguestMolly O'Sheahost

Jul 24, 20261h 8mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Tony Kim’s “BC to AD” framing: AI flips tech into a compute-first world
Tony Kim introduces his core thesis: the AI era triggered a discontinuity that radically increased the value and importance of compute hardware. He argues the tech market’s center of gravity has shifted away from a software-centric regime toward chips, hardware, and the physical compute stack.
- •Tony describes 2023 as a breaking point where “everything changed” for compute demand
- •He claims the base layer of compute increased dramatically, changing server economics
- •He lays out a market-cap framing that emphasizes how hardware/compute now dominates value creation
- •He says the pre-AI era had the inverse dynamic, with software capturing more of the value
- •He flags robotics—especially in China—as a major coming wave tied to embodied AI
On-location at RAISE Summit: why he’s in Paris and what panels he’s doing
Molly and Tony set the scene in Paris at the RAISE Summit and discuss how Tony became involved. Tony outlines the four panels he’s participating in and how they map to his focus on the physical layers of AI.
- •Tony explains he filled in as a speaker last year and decided to support the event’s growth
- •He describes RAISE as rapidly scaling and becoming a major AI conference in Europe
- •He lists panels spanning accelerators, quantum computing, optics for data centers, and chip/XPU co-design
- •Molly highlights the event’s scale and notable names appearing on stage
- •Tony emphasizes his agenda is heavily tilted toward hardware/compute rather than software
Macro through-lines: compute primacy, data-center redesign, and co-design of models + silicon
Tony lays out the major themes he sees driving AI’s next phase: compute becomes the dominant lever, data centers must be rebuilt for new physics constraints, and model development is increasingly intertwined with hardware architecture. He also points to power and energy architecture as a parallel transformation.
- •He argues compute and models are symbiotic, making compute the primary driver of capex and valuation
- •He says AI forces a staged rebuild of data centers, with knock-on effects across the stack
- •He describes power-density constraints (bandwidth, power, heat) as design limiters
- •He frames a shift from copper interconnects to optics (“light”) to address scaling limits
- •He highlights co-design: silicon designed to match model requirements and vice versa
“RAMpocalypse”: why memory becomes the next bottleneck
Tony explains why memory intensity is rising sharply as AI systems evolve, especially with more persistent memory and agentic behavior. He also highlights a timing mismatch: it takes years to add memory-fab capacity, while demand is immediate.
- •Tony says AI development is beginning to mirror aspects of the human brain, with increasing emphasis on memory
- •He expects growing demand for many memory types as part of compute architectures
- •He describes a supply/demand “duration mismatch,” since fabs take years to build
- •He predicts memory’s importance will rise relative to the current compute-only narrative
- •Molly adds that with few major memory players, underwriting supply catch-up is difficult
The “complete rebuild”: cloud’s old economics vs AI’s token-and-capex future
Tony contrasts the 2000–2020 cloud era—built on CPUs and hard drives and viewed as cheap commodity compute—with the AI era’s expensive, power-hungry infrastructure. He argues AI shifts margin structure, pushing the industry from asset-light to asset-heavy models.
- •He characterizes classic cloud as effectively reselling CPUs with hard drives, enabling high-margin software layers
- •He says AI data centers jump from megawatts to gigawatts and from smaller sites to massive server farms
- •He claims AI turns compute servers from thousands of dollars into million-dollar systems
- •He introduces the idea that AI data centers are “token factories,” changing how value is monetized
- •He argues this transition compresses margins at the top of the stack and forces a wholesale rebuild
How Tony allocates capital: “now” vs “future,” and why so many bets converge on 2030
Tony describes his portfolio decision-making as balancing near-term AI opportunities against longer-dated frontier bets. He notes multiple technology roadmaps appear to converge around 2030, while emphasizing he seeks nonlinear, asymmetric outcomes in the frontier bucket.
- •He frames most effort within a near-term window, but insists on evaluating durability beyond it
- •He stresses that beliefs about the future drive valuation multiples, not just short-term forecasts
- •He describes allocating some time and capital to frontier technologies with long gestation periods
- •He observes many roadmaps converging on 2030, spanning multiple categories he tracks
- •He introduces a five-year “still cool?” test as a way to avoid fading narratives and multiple compression
Chips weren’t commodities: consolidation, pricing power, and a hardware renaissance
Tony pushes back on the idea that semiconductors are commodities, arguing profitability and industry structure prove otherwise. He says AI has reignited excitement in physical-world disciplines—from rack design to materials science—and exposed talent shortages in specialized hardware domains.
- •He argues “chips are a commodity” was a false premise and that chip firms have unusually high profitability
- •He describes decades of consolidation leading to only a few dominant players in many categories
- •He says lack of venture funding historically reduced new entrants, increasing incumbent power
- •He points to renewed enthusiasm for physical infrastructure: servers, fiber, power, racks, packaging, materials
- •He mentions a shortage of specialists in areas like custom memory design and chip architecture
Robotics stack: building a robot ‘mind’ and separating brain vs body
Tony explains robotics as an extension of AI: labs develop ‘brains’ while manufacturing ecosystems build bodies. He sketches a two-part cognitive architecture—language interaction plus motor/world modeling—and positions embodiment as the next major wave.
- •He frames robotics as intelligence embodied: a brain plus physical form factors
- •He describes two “brains”: language/interaction and a motor/world model for motion and perception
- •He separates labs building cognition from manufacturing that scales bodies, limbs, and hands
- •He argues robotics parallels AI’s stack evolution, just applied to the physical world
- •He implies cross-combinations are possible when brain and body are modular
Why China is early in robotics: company count, IPO pipeline, and manufacturing leverage
Tony claims China has a large number of robotics companies and expects many potential IPOs there relative to the U.S. He attributes part of the difference to market structure and emphasizes Asia’s manufacturing strengths as a key advantage in producing robot bodies at scale.
- •He claims there are roughly 130–140 robotics companies in China and many potential IPOs this year
- •He contrasts that with very few potential U.S. robotics IPOs in the same timeframe
- •He notes China’s public markets may be used earlier as a funding mechanism than in the U.S.
- •He argues the West may lead in model development, while Asia leads in manufacturing scale
- •He describes “mix and match” dynamics: pairing a Western brain with a Chinese-manufactured robot body
The biggest robotics market might be companionship, not factories
Tony shares a thesis that social and companionship robots could become a major market, driven by loneliness and aging-population dynamics. He suggests robots don’t need perfect humanoid form factors to deliver value if they provide intelligence and empathy.
- •He expresses particular interest in social robots for companionship, education, and elderly care
- •He connects the idea to loneliness and to demographic pressures he sees globally
- •He argues appealing designs and smaller form factors may be more practical than perfect humanoids
- •He imagines robots that can converse deeply, remember stories, and preserve life histories
- •Molly adds examples of small, telepresence-like companion devices seen in consumer contexts
Speech and voice infrastructure: rapid growth as “brain” interfaces expand
Molly highlights that speech model companies are growing extremely fast, and Tony agrees. They place voice APIs and speech systems as critical interfaces for how humans will interact with AI systems and agents.
- •Molly says speech model companies are growing faster than many peers in Silicon Valley
- •Tony agrees and reinforces the trend’s strength
- •Molly names AssemblyAI as an example of a fast-growing company in this category
- •They link voice progress to broader themes of AI ‘brains’ and human interaction layers
- •The conversation pivots from model interfaces toward downstream data impacts
“Token flow” enterprise: context/ontology layers, data foundations, and agent orchestration
Tony proposes a framework for the future enterprise built around the movement of tokens through compute, models, and application harnesses. He describes enterprises building proprietary context layers (also referred to as ontology) on top of their data foundations to orchestrate agents.
- •Tony predicts enterprises will be redesigned around tokens plus a data foundation
- •He describes a context/ontology layer embodying proprietary knowledge and institutional memory
- •He says agents will operate “wild,” interacting through the context layer into enterprise data
- •He defines “token flow” as following who creates tokens, serves tokens, and repackages tokens with context
- •He argues companies need a position in token flow to sustain margin and relevance
PE roll-ups as AI disruption: buying customers, refactoring workflows, and efficiency bets
Molly asks about private-equity-style roll-ups as an AI disruption strategy. Tony describes the logic: acquire customer bases and legacy operations, then use AI to restructure service delivery, remove inefficiencies, and sell an integrated solution rather than point products.
- •Molly suggests roll-ups can function like building a new product by buying distribution/customers
- •Tony says the play is to restructure service delivery and strip out inefficiencies and cost
- •He describes a shift toward selling full-stack outcomes rather than identifying as a single layer (app/service/compute)
- •He notes uncertainty about how well this strategy will work and says he’s watching closely
- •He frames it as a potential “next shoe to drop” for industries with low AI adoption
Next 12 months: concerns, optimism, labs going public, and orbital data-center milestones
Tony says he expects the same core forces—compute demand, memory constraints, and data center redesign—to continue powering through periodic macro scares. He’s excited by continued proof points in data-center reimagination, possible public listings of major labs, and progress toward orbital data centers.
- •He notes recurring market “scares” but expects the compute and memory buildout to persist
- •He wants to see continued validation of the data center redesign trajectory
- •He says he’d like major foundation labs to go public within the next 12 months
- •He highlights interest in progress toward orbital data centers and their implications
- •Molly frames his outlook as part prediction, part “manifestation,” which he reframes as contemplation
Who shaped Tony’s worldview: people who took chances—and historical figures as mentors
In closing, Tony reflects on mentorship as less formal coaching and more about people who believed in him and gave him latitude. He also says many of his ‘mentors’ came from studying history—leaders and creators he read about growing up.
- •Tony says he didn’t have classic mentors, but benefited from people who took a chance on him
- •He mentions someone who brought him into BlackRock and gave him broad freedom to operate
- •He recalls early-career supporters who encouraged major life moves, including going west before dot-com
- •He says he tries to pay it forward by encouraging young people and giving them breaks
- •He identifies historical figures he studied in libraries as core influences on his worldview

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.