Summary

Kim marks 2023 as the BCE-to-AD break for AI — “It’s like BCE, Anno Domini… bam, ’23 happens, everything changed” — and says market value has shifted toward hardware. His rough tally: ~$10T in software/services/internet, $22–23T in the Mag Seven, and $30T+ in chips and hardware — “before AI, in BCE era, it was probably reversed.” The base compute layer rose roughly 10,000×: “a $10,000 server is a million-dollar server.”
The data center is being rebuilt around raw physics: a trillion dollars of CapEx this year and ten trillion over five years “to move data centimeters and millimeters. That’s AI.” Transmission distances collapse from kilometers to millimeters, while bandwidth, power, and heat rise with the smaller distances — forcing a regime change “from copper to light,” 800-volt power architectures, and eventually solid-state transformers.
Memory is the next primacy: RAMpocalypse reflects chips and models increasingly mirroring the human brain, which Tony says has “a lot more memory storage than maybe compute.” He says agentic harnesses and institutional memory are driving memory intensity higher. Molly notes there are only 3 main memory players and says SK Hynix is about to go public, while acknowledging it might be public by release. Fabs take 3–4 years to build against a shortage today — a “mismatch of demand, supply, duration” causing market angst. “Today we’re all talking about compute, compute, compute. I think the primacy of memory will become even more important.”
The margin stack has inverted: classic cloud was “really reselling CPUs with hard drives” under fat SaaS margins; AI compute factories sell tokens and squeeze the app layer. The models “like it or not, have consumed the market cap out of software and services… it’s like the Borg” — hence SaaSpocalypse. And if scaling holds at an order of magnitude a year — “10 times 10 times 10, that’s 1,000 times in three years” — defensibility must be rethought: “Moats are always breached, aren’t they? So it’s more about offense.”
Portfolio construction: roughly 90% of Kim’s investment allocation goes into the three-year “vortex of AI,” but he also needs to believe in life beyond it, and many frontier themes converge on 2030. Million-qubit logically error-corrected quantum, SMRs with regulatory approval, orbital data centers, 800V, and various AGI estimates cluster around that period. His filter for exit multiples: “You always wanna be betting on not what’s cool today. Will you still be cool in five years?”
On robotics, Kim sees a major Asian manufacturing angle: “The Chinese are coming” — 130–140 robotics companies in China, 30–40 potential IPOs there this year versus “zero, one, two” in the US. A possible endgame is to “mix and match a Chinese physical robot with a Western brain,” and he says he knows that is happening. His hot take: the biggest unlock may be companionship robots — approachable, half-sized C-3PO-like machines with “the intelligence of Shakespeare and Einstein” — for loneliness, education, and the elderly, given Asian birth rates below 1.0 against a 2.1–2.2 replacement rate.
Twelve-month outlook: “every six months there’s a scare” (SaaSpocalypse, RAMpocalypse, financing, CapEx), but the buildout “will just plow through,” with foundation-lab IPOs — “huge market appetite” — and orbital data center proof points the things to watch. His current thesis is explicitly provisional: “If we talk again in six months, it might be completely different.”

Deep dive

1. 2023 was the BCE-to-AD break — and market value shifted toward hardware

Kim’s periodization is the episode’s spine: “AI happens. It’s like BCE, Anno Domini… bam, ’23 happens, everything changed.” The base layer of compute went up roughly 10,000× — “a $10,000 server is a million-dollar server” — and Silicon Valley moved “from a software-centric world to a compute-centric world,” with models and compute “symbiotic and synonymous with each other.”
His rough map of about 1,500 companies over $1B market cap: “ten, twenty, thirty” — $10T+ in non-Mag-Seven software/services/internet, $22–23T in the Mag Seven, and $30T+ in chips and hardware. “I don’t think people realize that we are that compute hardware-centric now… Before AI, in BCE era, it was probably reversed.”
The mechanism of the transfer: “The models themselves, like it or not, have consumed the market cap out of software and services. It’s like the Borg” — and for those models to exist, they must live on the compute stack.

2. Cloud economics inverted: from reselling CPUs to selling tokens

Kim’s demystification of the 2000–2020 cloud era: “everyone says it’s software, but it was really reselling CPUs with hard drives.” Compute was an afterthought — a thin layer under massive SaaS margins — which “fueled a twenty-year run in cloud and SaaS.”
The new data center “is an alien data center to this data center”: revenue is moving from asset-light, high-margin businesses toward asset-heavy, lower-margin, big-dollar token factories, which “takes a lot of margin out of that top layer of the stack” and triggered SaaSpocalypse. “There’s a lot of apocalypses, you know.”
If scaling laws hold at an order of magnitude a year — “10 times 10 times 10, that’s 1,000 times in three years” — capability keeps compounding, so the question is where margin and defensibility live. His answer rejects the standard frame: “Moats are always breached, aren’t they? So it’s more about offense. Can you move faster?”

3. The physics: a trillion dollars to move data millimeters

The data center’s distance collapse — kilometers to meters to centimeters to millimeters — and with each order of magnitude, “the bandwidth goes up, the power goes up, the heat goes up.” The irony as he tells it: “the $1 trillion of CapEx this year and the $10 trillion over the next five years are coming to move data centimeters and millimeters. That’s AI.”
Consequences he’s tracking: “we’re going from a regime of copper to a regime of light,” a power revolution around behind-the-meter generation and the grid, “the rise of 800 volts,” and ultimately solid-state transformers — since “every time you have these step-downs in voltage, you lose efficiency and energy.”
The other structural theme is co-design — silicon tightly matched to model specs and vice versa, “the new path that many of the leading foundation labs are pursuing” — the subject of his Broadcom panel on the new “Jalapeño” chip. His four RAISE panels were D-Matrix on accelerators, PsiQuantum on quantum computing, Lumentum on optics, and Broadcom on XPU and AI co-design.

4. RAMpocalypse: chips and models are converging on the brain

Kim says chip and model development “is starting to mirror the human brain.” Early models were all parallel compute with little memory; now personal AIs, agentic harnesses, institutional context, and stored memory are increasing memory intensity. The human brain, he says, “has a lot more memory storage than maybe compute,” while acknowledging that depends on how one looks at synapses and neurons. Chip architecture is now “all about arbitrating memory in some form with your compute” — SRAM, DRAM, stacked DRAM, HBM, and high-bandwidth flash.
The tradeable tension is duration: “it takes three or four years to build a chip fab or a memory fab, but yet there’s a shortage today” — a mismatch of demand, supply, and duration “causing a lot of angst in the market.” His conclusion: “I think the primacy of memory will become even more important.”
Molly’s setup — that there are only three main players and that SK Hynix was about to go public, while noting it might already be public by publication — raises the underwriting question: how long does a demand premium built on limited supply last?

5. Capital allocation: 90% in the vortex, but many themes converge on 2030

His three-tier framework: roughly 90% of his investment allocation goes into the three-year “vortex of AI” — who’s winning, losing, ascending, stagnating — but even there, “you must be betting on the future as well,” because belief in years four through six “has a huge impact on your multiple.” Cheap-but-atrophying assets may not be the best allocation of capital.
The frontier tier draws on his own pre-genAI record: he made AI investments in 2019, 2020, and 2021 “where you didn’t know that this LLM thing was gonna happen,” guided by the intuition that some form of AI compute would be needed. Six or seven years later, “the AI accelerator wars have begun.”
The pattern he keeps hitting: “All roads converge to 2030” — utility-scale, logically error-corrected million-qubit quantum; SMRs with regulatory approval; orbital data centers beginning to scale; 800V; solid-state transformers; and AGI estimates ranging from 2028 to 2030. His filter for everything: “Will you still be cool in five years? … The multiple that people will pay will go down and your growth rates are decelerating, and now you’re in a bind.”
He calls this his current thesis and says it might be completely different if they talk again in six months.

6. Chips were never a commodity — “the ring of power. It was lost, and then it was found”

His revisionism on semis: “It’s called Silicon Valley for a reason… people forgot.” Chips carry “the highest profitability of any sector — higher margins than software, pharmaceuticals, industrials, telecom, anything,” and venture capital, until recently, had not funded these companies — so hundreds of companies collapsed into a few survivors with, in his view, duopolistic power and huge pricing power, “the complete opposite of commodity.”
AI has spawned a hardware renaissance: “servers are cool, fiber is cool, power is cool, rack design is cool… materials science is cool.” The bottleneck is people — “people in the lost art of analog computing” are “like blacksmiths,” and someone at one of the biggest memory companies told him: “We cannot get people to design custom memory… we gotta repurpose some of these software programmers into memory co-design architects.”

7. Robotics: 140 Chinese companies, Western brains, and a bet on loneliness

His anatomy of the robot: two brains in one — a world model for motion and perception plus an LLM as translator — built by labs and then embodied. Hands are “probably the hardest thing,” but the body itself “is a manufacturing hardware business.” China is coming into the field with 130–140 robotics companies and 30–40 potential IPOs this year versus “zero, one, two in the United States” — partly because shallow private markets push Chinese firms public earlier, “and they’re coming in waves.”
The West is probably ahead on model development, in Kim’s framing, while the Asian manufacturing complex — Japan, Korea, and China — could use EV-like physical-scale manufacturing to lower robot production costs. Molly contrasted hydraulic Atlas, which can pick up a fridge, with Figure’s focus on package sorting and commercial work; Kim corrected her that Atlas is Korean, owned by Hyundai through Boston Dynamics. The possible permutation he confirms: “you can mix and match Chinese physical robot with a Western brain, and I know that’s happening.”
His focus is less on industrial manufacturing robots and more on social embodiment — loneliness, education, and the elderly. Demographics (Asian birth rates well below 1.0 versus 2.1–2.2 replacement; nursing-home companies among the world’s best-performing companies) support the opportunity for an approachable, half-sized robot — R2-D2 and C-3PO rather than the Terminator — “that has the intelligence of Shakespeare and Einstein and speaks every language,” able to converse empathetically with his mother and record life histories.

8. Token flow is the positioning test — and PE roll-ups are the next shoe to drop

His model of the future enterprise: tokens in, a data layer embodying proprietary and third-party data, a context layer of the company’s cumulative knowledge (Molly suggested Palantir calls this an ontology), and then “agents go wild.” The investable rule: “Either you create tokens — compute. You serve the tokens — foundation labs. Or you put a harness, package, context around the token — app services… If you’re not in that flow, it’s a problem.” App companies are struggling to find their place; some inference and edge-cloud companies have inserted themselves as last-mile token providers.
Molly traced downstream effects through speech APIs such as AssemblyAI, the data layer at Databricks and Snowflake, and databases such as MongoDB; Kim said he was invested in many of those companies and agreed they were in token flow.
The alternative is full abstraction: “I will take on all of your insurance. Pay me X… you used to pay 100, I’ll charge you 20.” On the PE roll-ups Molly cited — General Catalyst’s Creation Fund and Long Lake’s reported purchase of American Express Global Business Travel — Kim’s honest hedge was that buying companies with customers and stripping out inefficiency is “maybe… a business,” but “the jury’s out. I’m watching it.” The broader rubric: 500-person companies generating the revenue of 10,000, forcing incumbents to adjust or sell.
His close: through the six-monthly scares, the buildout “will just plow through, and our fears will subside”; he’s excited for foundation labs going public (“huge market appetite”) and orbital data center progress that “could have huge implications” for terrestrial buildouts. On mentors, the people who shaped him were historical figures he studied in libraries — Caesar, Alexander, Napoleon, Beethoven, and Churchill — plus those who took a shot on him. “Dead people and kind people. How’s that?”

Why AI Agents Could Finally Reinvent the Credit Card Redefining Chip Architecture with Arm CEO Rene Haas