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

Summary

• Kim dates AI like a calendar reset — “it’s like BCE, Anno Domini… bam, ‘23 happens, everything changed” — with the base layer of compute going up 10,000X as “a $10,000 server is a million-dollar server.” His market-cap map: roughly $10T in software/services/internet, $22–23T in the Magnificent Seven, and $30T+ in chips and hardware — “ten, twenty, thirty” — and “I don’t think people would realize that we are that compute hardware-centric now.” Pre-AI, he says, the ordering was probably reversed.
• The trillion dollars of CapEx this year and “ten trillion over the next five years” is, in Kim’s framing, being spent “to move data centimeters and millimeters — that’s AI.” Data-center distances are shifting from kilometers to millimeters; as distance shrinks by an order of magnitude, bandwidth, power, and heat rise, creating logarithmic effects on design — and forcing a shift “from a regime of copper to a regime of light,” 800-volt power architectures, and eventually solid-state transformers.
• “RAMpocalypse” reflects chip and model design starting to mirror the human brain, which is increasingly memory-intensive — “the primacy of memory will become even more important.” Chip architectures arbitrate memory (SRAM, DRAM, HBM, high-bandwidth flash) with compute, while a three-to-four-year fab build time collides with today’s shortage — a “mismatch of demand, supply, duration” that is “causing a lot of angst in the market.” Molly notes that there are only 3 main memory players.
• His allocation framework: roughly 90% of his investment goes into the three-year “vortex of AI… the now,” but the multiple depends on believing in year five — “Will you still be cool in five years?” He also reserves resources and time for “nonlinear asymmetric potential” frontier bets, and “all roads converge to 2030”: million-qubit error-corrected quantum, SMRs, AGI, 800V power, and orbital data centers all cite the same date.
• Semiconductors were never a commodity — chips carry “the highest profitability of any sector,” consolidated into duopolies as venture capital stopped funding them. “It’s not the revenge of the nerds, it’s the lost ring of power that was found” — and the physical sciences are cool again. A major memory-company contact told him, “We cannot get people to design custom memory”; Kim says the industry must “repurpose some of these software programmers into memory co-design architects.”
• On robotics, “the Chinese are coming”: 130–140 robotics companies in China, 30–40 potential IPOs there this year versus “zero, one, two” in the US. Kim says the West is probably ahead on robot brains while Asia’s manufacturing complex may have an advantage in the body — so “you can mix and match Chinese physical robot with a Western brain, and I know that’s happening.” His contrarian soft spot: half-size C-3PO-style social robots for loneliness and elderly care, not just industrial humanoids.
• The enterprise stack compresses to tokens in, a data layer, a context/ontology layer, and agents — “follow the flow of tokens… If you’re not in that flow, it’s a problem.” Voice APIs and inference-cloud companies are in the flow, while app companies are struggling to find their place; Molly separately says speech-model companies such as AssemblyAI are growing extremely quickly. The token-flow lens also informs the AI-era PE roll-up model — “Today, you pay 100; I’ll charge you 20” — though on roll-ups “the jury’s out.” Next 12 months: AIpocalypse, war, interest rates, and CapEx scares may recur, but compute “will just plow through,” and he wants to see big labs go public and orbital data centers take a forward step.

Deep dive

1. 2023 was the calendar break: compute rose 10,000X and ate software’s market cap

• Kim’s periodization is the episode’s spine: pre-AI cloud (2000–2020) was “really reselling CPUs with hard drives” — compute was so thin and cheap that margins went to SaaS on top. Then “AI happens. It’s like BCE, Anno Domini… bam, ‘23 happens, everything changed” — the base layer of compute went up 10,000X, “a $10,000 server is a million-dollar server,” and DRAM went from a smartphone commodity to something he now needs to pack, including expensive HBM, onto AI systems.
• His rough market-cap arithmetic: he estimates about 1,500 companies globally with market caps above $1B, perhaps 2,000 if China is added, then puts roughly $10T in software/services/internet, $22–23T in the Magnificent Seven, and $30T+ in chips and hardware — “ten, twenty, thirty… I don’t think people would realize that we are that compute hardware-centric now.” Before AI it was probably reversed.
• The mechanism behind SaaSpocalypse: the new compute factory sells tokens, and that “takes a lot of margin out of that top layer of the stack.” The models themselves, “like it or not, have consumed the market cap out of software and services… it’s like the Borg.”
• On defensibility, with scaling at roughly an order of magnitude a year (“ten times ten times ten, that’s a thousand times in three years”): “Moats are always breached, aren’t they? So it’s more about offense. Can you move faster?”

2. The data-center rebuild: ten trillion dollars to move data millimeters

• Kim’s physics framing: data-center connections are shifting from kilometers to building-to-building, rack-to-rack, and within the chip — down to meters, centimeters, and millimeters. As distance shrinks by an order of magnitude, bandwidth, power, and heat rise, creating logarithmic effects on design. “The irony of it all — the trillion dollars of CapEx this year and the ten trillion over the next five years that are coming — is to move data centimeters and millimeters. That’s AI.”
• The consequences cascade through every layer: “we’re going from a regime of copper to a regime of light.” Data centers now range from small facilities in cities to giant server farms in Texas, while a power revolution follows — the grid, behind-the-meter generation, new sources, “the rise of 800-volt,” and ultimately solid-state transformers, since every voltage step-down loses efficiency.
• Another major theme is co-design: tightly integrating silicon with model specifications and letting model specifications inform compute design — “the new path that many of the leading foundation labs are pursuing.” Kim plans to discuss it with Charlie at Broadcom, citing the recently launched Jalapeño chip.

3. RAMpocalypse: the machine is growing a brain, and brains are mostly memory

• Kim’s core analogy: chip and model development “is starting to mirror the human brain.” In the early days, models had abundant parallel compute but little memory; now personal AIs, agents, and enterprise context are adding memory. “The human brain is very memory-intensive… Today, we’re all talking about compute, compute, compute. I think the primacy of memory will become even more important.”
• Chip architectures arbitrate memory in concert with compute — SRAM, DRAM, stacked DRAM, HBM, and high-bandwidth flash — with these memory and storage methods tightly packed into the chip and computer architecture.
• The tradeable tension is duration: fabs take three to four years to build against a shortage today — “this mismatch of demand, supply, duration… is causing a lot of angst in the market.” Molly frames the underwriting question around only three main memory players, a large demand premium, and how long it will last; she also says SK Hynix might be public by the time the episode is released.

4. Portfolio construction: roughly 90% in the “vortex of the now,” the rest converging on 2030

• The three-year window — “the vortex of AI… the now” — absorbs 90-plus percent, or roughly 90%, of Kim’s investment: who’s ascending, declining, stagnating. But even there, “the belief in a future has a huge impact on your multiple” — cheap-looking, atrophying assets fail the opportunity-cost test.
• His discipline against momentum: “You always wanna be betting on not what’s cool today. Will you still be cool in five years?” Follow today’s trend with a known half-life and you exit into decelerating growth and a compressing multiple — “now you’re in a bind.”
• The frontier sleeve is informed by history: he made AI investments in 2019–21, before GenAI, on the intuition that some form of AI compute would be needed, and “now the AI accelerator wars have begun.” Today’s equivalent long-dated bets share a date: “All roads converge to 2030” — utility-scale, logically error-corrected million-qubit quantum, SMRs with regulatory approval, AGI for classical computing (“2030, 2029, 2028, whatever”), 800V architectures, solid-state transformers, and orbital data centers beginning to scale. He wants “nonlinear asymmetric potential,” not incremental change.

5. The lost ring of power: chips were never a commodity, and the physical world is cool again

• Kim’s revisionism: “People always said chips are a commodity, but yet they have the highest profitability of any sector” — higher margins than software, pharmaceuticals, industrials, or telecom. Hundreds of chip companies consolidated over 20 years into a few powerful players; because venture capital never funded these companies, few new entrants emerged and survivors became “behemoths with huge pricing power, the complete opposite of commodity.” His preferred metaphor: “It’s not the revenge of the nerds, it’s the lost ring of power that was found.”
• The renaissance now extends to everything physical — “servers are cool, fiber is cool, power is cool, rack design is cool” — with an acute talent shortage. At dinner, someone from one of the biggest memory companies told Kim, “We cannot get people to design custom memory.” Kim’s response is that the industry must “repurpose some of these software programmers into memory co-design architects.” Analog computing, he quips, is like being a blacksmith.

6. Robotics: China floods the body, the West may lead the brain, and loneliness is the market

• His anatomy of a robot: two brains in one — a world model for perception, motion, and the physical world, plus an LLM-like intelligence and language layer that can act as a translator. The motor-function system controls movement and reactions; the two systems are then embodied in a body. The body — arms, legs, limbs, and especially hands — is a manufacturing hardware business built by labs and manufacturers together.
• The numbers: 130–140 robotics companies in China, “30, 40 potential IPOs this year in China alone… and there’s what? Zero, one, two in the United States.” Kim says China’s shallow private markets push companies public earlier and that they are coming in waves. He frames a possible split: the West is probably ahead in model development, while Japan, Korea, and China may have an advantage in manufacturing through their EV and industrial base. The endgame: “mix and match Chinese physical robot with a Western brain, and I know that’s happening.”
• His self-described soft spot, offered against the industrial consensus: social robots for loneliness, elderly care, and education. In Asia in particular, birth rates are “well below 1.0” against the 2.1 or 2.2 needed to stay even, and nursing-home companies are among the best-performing companies in the world. A half-size, approachable C-3PO-like robot “with the intelligence of Shakespeare and Einstein” could converse empathetically with his mother and “record their life histories.” No perfect hand articulation required.

7. Token flow decides who survives — plus a 12-month outlook and “dead people and kind people”

• Kim’s enterprise end-state: tokens in, a data foundation, a context layer — Molly supplies Palantir’s word, “ontology” — and then “agents go wild.” His filter for every business model: “Follow the flow of tokens. Either you create tokens” through compute, “serve the tokens” through foundation models, or put a harness, package, or context around them through application services. “If you’re not in that flow, it’s a problem.” Voice APIs and inference/edge clouds are in the flow; app companies “are struggling to find their place.”
• Molly’s examples include rapidly growing speech-model company AssemblyAI and downstream data/database beneficiaries such as Databricks, Snowflake, and MongoDB. Kim says these businesses need to locate themselves in the token flow.
• The alternative is abstracting the whole stack into outcomes — “I will do all of your claims processing. I will do all of your insurance processing. Pay me X. Today, you pay 100; I’ll charge you 20” — which is informing AI-era PE and venture roll-ups. Molly cites General Catalyst’s Creation Fund and says Long Lake had just bought Amex GBT, “I think.” Kim is intrigued but unconvinced: a new company with 500 people could generate the revenue of 10,000 by rethinking traditional workflows, but “the jury’s out… I’m watching it.”
• His next 12 months: “It seems like every six months there’s a scare” — AIpocalypse, war, interest rates, too much CapEx, not enough financing — but he’s optimistic the compute wall, memory wall, and data-center redesign “will just plow through, and our fears will subside.” He’s excited to see big foundation labs go public, given the “huge market appetite,” and to see the next forward step toward orbital data centers, which could “unlock a rethink” of terrestrial builds.
• On the closing mentorship question, Kim rejects the “legendary career” framing: he is “just trying to survive.” He had people who believed in him and gave him latitude, plus historical figures — Caesar, Alexander, Napoleon, Beethoven, and Churchill — absorbed in libraries as a somewhat ostracized kid. “It’s kind of dead people and kind people. How’s that?”

Verification Notes

• Kim’s company-count setup is internally unclear: he says roughly 1,500 companies globally, perhaps 2,000 if China is added, then says the figure he is using excludes China; the market-cap figures are explicitly rough approximations.