Tony Kim

1. Key Themes

The Great Compute Inversion: From Software-Centric to Hardware-Centric World

The single most important macro shift Tony Kim identifies is that the entire value stack of technology has flipped. Compute and hardware now dominate market capitalization in a way that is still underappreciated by most investors.

"There's roughly 10 plus trillion market cap in software services and internet. There's 22, 23 trillion in Mag7 and there's another 30 plus trillion in chips and hardware. I don't think people would realize that we are that compute hardware centric. Before AI, in BCE era, it was probably reversed." 00:00:00

BCE vs. AD: 2023 as the Civilizational Inflection Point

Kim uses a vivid historical framing — Before Computing Era vs. AI era — to argue that 2023 represents a hard discontinuity, not a trend. The base layer of compute did not incrementally improve; it repriced by orders of magnitude.

"AI happens. It's like BCE, Anno Domine. 2023 is going from BC to AD. And bam, 23 happens. Everything changed. So that, what was called the base layer of compute, went up, I don't know, 10,000x. It's a $10,000 server is a million dollar server." 00:19:32

The Data Center as a Physics Problem: Kilometers to Millimeters

Kim reframes the entire data center buildout not as a real estate or power story, but as a fundamental physics problem — the relentless compression of the distance over which data must travel, each order-of-magnitude reduction in distance triggering logarithmic increases in bandwidth, power, and heat.

"The data center is changing from before we would transmit data kilometers. And now building to building. And then it's within the building. And then rack to rack. And then now it's within the rack. And then next it's within the chip. And so you're going from kilometers to meters to centimeters to millimeters." 00:06:16

"The irony of it all, that the trillion dollars of CapEx this year and the 10 trillion over the next five years that are coming is to move data centimeters and millimeters. That's AI." 00:06:46

The Rampocalypse: Memory Is the Next Compute

Kim argues forcefully that compute has been the dominant investment theme but that memory is the next primary bottleneck and value creator. The architecture of AI is evolving to mirror the human brain, which is far more memory-intensive than compute-intensive.

"The human brain is very memory intensive. And today is more compute intensive, but as you see it with Rampocalypse, the memory intensity has just skyrocketed. And going forward, you will see more and more and more memory... I think the primacy of memory will become even more important." 00:12:13

The Copper-to-Light Transition: Optics as Infrastructure

As data center distances compress and power density explodes, the medium of data transmission must change. Kim identifies the shift from copper to photonics/optics as one of the defining infrastructure plays of the next era.

"We're going from a regime of copper to a regime of light. And these things help address this power density and energy and the laws of physics that are pushing data center design to its very limits." 00:07:36

Chip Co-Design with Models: The New Competitive Moat

The most sophisticated foundation labs are no longer just buying GPUs — they are co-designing their silicon to match the specific parameters of their models, and vice versa. Kim sees this as the new defensible competitive frontier.

"This co-design, this notion of the co-design of tightly integrating the design of your silicon to match the parameters and the specs of the model. And the model specs informing the design of the compute. And then this kind of co-design is like the new path that many of the leading foundation labs are pursuing." 00:08:27

2030 as the Convergence Year for Frontier Technologies

Kim has identified a non-coincidental convergence: quantum computing utility scale, SMR regulatory approval, AGI timelines, 800-volt power architecture, solid-state transformers, and orbital data centers all point to approximately 2030. This is his frontier investment horizon.

"All roads converge to 2030. It's like quantum computing, utility scale, logically error corrected, million qubit quantum computer, 2030. SMRs, fusion small nuclear reactors, regulatory approval. You talk to these companies, 2030. When will we hit 800 volt power architectures late 2020s, 2030? Will we have solid state transformers 2030?" 00:32:40

The Token Flow Framework: Where Margins Live and Die

Kim has developed a unifying framework for understanding where value accrues across the entire AI stack. Everything reduces to being in or out of the "token flow" — creating tokens (compute), serving tokens (foundation labs), or packaging tokens with context (applications).

"Either you create tokens, compute. You then serve the tokens, foundation labs. And then you put a harness, package, context around the token. App services, et cetera. So you must be in this token flow to either resell or repackage the tokens with your context in your very specific application. If you're not in that flow, it's a problem." 00:54:42

China's Robotics Wave: 130+ Companies and 30-40 IPOs in 2025 Alone

Kim is watching an explosion in Chinese robotics that is orders of magnitude larger than what is happening in the United States. The competitive threat is real, particularly in the physical manufacturing layer of robots.

"The Chinese are coming and there's 130, 140 robotics companies in China. I see 30, 40 potential IPOs this year in China. Alone this year. And there's what? Zero, one, two in the United States. Maybe this year." 00:00:27

SaaS Margin Compression Is Structural, Not Cyclical

Kim argues that the old SaaS model — building massive application margins on an almost-free compute base — is being structurally eroded by AI. The new stack is asset-heavy with lower margins, and the foundational models themselves are absorbing value that used to flow to software.

"The models themselves, like it or not, have consumed the market cap out of software and services. It's like the Borg... That compute factory is creating tokens and then the model guys are then selling their tokens. And that just takes a lot of margin out of that top layer of the stack." 00:24:45

2. Contrarian Perspectives

Chips Were Never Commodities — That Was Always a False Narrative

The conventional wisdom for two decades was that semiconductors were commodified, low-margin businesses where all value flowed to software. Kim argues this was simply wrong — chip companies have always had the highest margins of any sector.

"People always said chips are a commodity, but yet they have the highest profitability of any company in the world of the sector is chip companies. They're higher margins than software, pharmaceuticals, industrials, telecom, anything. And so this notion that they were a commodity was just a false notion in my opinion." 00:39:56

He further notes that venture capital systematically avoided funding chip companies for decades, which caused the number of competitors to collapse rather than proliferate — the opposite of what you'd expect in a truly commodified market, and exactly what created today's pricing power behemoths.

"Venture capital, i.e. Silicon Valley, until recently, never funded these companies. So there's no money going in... In fact, what you have is a shrinking effect. The number of companies have collapsed. And then those that have survived are behemoths with huge pricing power." 00:40:24

The Most Valuable Robotics Market Will Be Social and Companion Robots, Not Industrial Ones

The dominant narrative around robotics investment focuses on manufacturing automation and industrial use cases. Kim believes the larger and more interesting market is in social embodiment — companionship for the elderly and lonely — a market almost nobody is funding in the West.

"My view is one of the things I'm most interested in in the robotic side is around not so much the manufacturing robot... is around loneliness, around social embodiment, around education, around elderly, around young people. To bring consumer and or commercial like social robots more so than industrial use." 00:47:14

"If you could embody some intelligence, empathy, it can be many different form factors too. It does not have to be the terminator. I think that could unlock a really, really interesting market, a really big market." 00:48:43

Moats Are a Defensive, Outdated Concept — Offense Is What Actually Wins

The entire venture and investing community frames competitive advantage as "moats." Kim rejects this framing as inherently defensive and therefore the wrong mental model for this era.

"Moats are always the breach, aren't they? So it's more about offense in my opinion. Can you move faster?" 00:27:08

Orbital Data Centers Could Fundamentally Restructure How Terrestrial Data Centers Are Built

This is treated as a futuristic curiosity by most investors. Kim believes it is a 2030 reality with profound backward implications on how ground-based data centers are being planned and built today.

"I would love to see in the next 12 months the next forward step toward orbital data centers. Because that also engenders a radical change in data centers. If you keep pushing on that progression, it could unlock a rethink, and moving the burden of terrestrial compute into space. And that can have huge implications of how current data centers are even being built." 01:02:39

3. Companies Identified

NVIDIA (implied)

Major GPU compute provider, referenced implicitly as the dominant supplier of the million-dollar AI servers that replaced $10,000 commodity servers. Core to the entire compute thesis.

Broadcom

Semiconductor and infrastructure software company. Kim is doing a dedicated panel on XPU and AI co-design with Broadcom's leadership, citing their Jericho chip as an example of model-silicon co-design.

"This is what I'll talk with Charlie at Broadcom about. Obviously they did that with the Jericho chip that recently came out." 00:08:53

PsiQuantum

Quantum computing company. Kim selected them specifically for a dedicated quantum computing panel at RAISE, signaling conviction in their approach.

"Another one with PsiQuantum and around quantum computing." 00:01:31

Lumentum

Optical and photonic components company. Kim chose them to anchor his panel on optics and next-generation data center design — directly aligned with his copper-to-light thesis.

"A third with Lumentum and kind of bringing optics to next-gen data center design." 00:01:31

D-Matrix

Next-generation compute architecture company focused on inference acceleration. Kim featured them on his panel on next-gen computer architectures.

"One on accelerators with D-Matrix, kind of the next-gen computer architectures." 00:01:31

SambaNova Systems

AI hardware and software company. Kim has a personal investment relationship with SambaNova; he filled in for their CEO Rodrigo Yang at RAISE the prior year.

"One of my companies I was involved with, SambaNova. Lippu was supposed to be one of the speakers. And he couldn't make it, so I decided to fill in." 00:02:16

SK Hynix

One of three global DRAM/HBM memory manufacturers. Called out as a bellwether for the memory shortage thesis and upcoming public market interest.

"SK Hynix is about to go public. And I don't know if by the time we put this out it might be public, but the big questions around that are, like, how do you underwrite that because the demand premium is massive because there's limited supply?" 00:15:47 (Molly O'Shea)

Figure AI

Humanoid robotics company. Referenced as a leading Western robotics embodiment play, contrasted with Atlas/Boston Dynamics.

"I recently visited Figure AI, the humanoid robotics company." 00:38:41 (Molly O'Shea)

Boston Dynamics / Atlas

Robotics company owned by Hyundai. Kim specifically corrects the common misconception that Atlas is American — it's Korean — underscoring his point about Asian manufacturing advantage in physical robotics.

"You mentioned Atlas robot. That's Korean actually. That's Hyundai that owns Boston Dynamics." 00:46:18

SpaceX

Referenced in context of orbital data centers and its upcoming IPO as a major catalyst in the space infrastructure buildout.

"Orbital data centers, that's also targeting 2030. It's very interesting... especially with SpaceX coming and the whole IPO on that." 00:32:40 / 00:38:41

Assembly AI

Voice AI infrastructure company. Called out by the host as one of the fastest-growing companies in the AI era, benefiting from the primacy of speech models.

"There's a company called Assembly AI that is like growing incredibly fast. And they're great." 00:51:32 (Molly O'Shea)

Databricks

Data and AI platform. Cited as a downstream winner from the agentic data explosion.

"The Databricks, the Snowflakes, they're hitting some of that extra premium in the market." 00:51:58 (Molly O'Shea)

Snowflake

Cloud data platform. Cited alongside Databricks as a downstream beneficiary of agentic data creation.

MongoDB

Database platform. Cited as benefiting from agents creating proliferating app-layer data needs.

"As agents create more apps, now MongoDBs, those databases are reaching." 00:51:58 (Molly O'Shea)

Long Lake / General Catalyst Creation Fund

A new AI-native company created through PE roll-up strategy, acquiring Amex Global Travel Business. Referenced as a live example of the AI-enabled disruption of traditional industries.

"One of them, Long Lake, just bought Amex Global. Their travel business. Which is interesting because you have a small player buying a large player." 00:58:34 (Molly O'Shea)

Cognition (Scott Wu)

AI coding agent company. Referenced in the RAISE speaker lineup, signaling Kim's attention to this space.

Cerebras (Andrew Feldman)

AI compute company using wafer-scale chips for inference. Referenced in the RAISE speaker lineup.

4. People Identified

Head of Global Technology at BlackRock, managing both public and private technology investments. One of the most senior public market technology investors globally. Known for early AI compute bets pre-2020.

"I made some of these AI investments pre-Gen AI, like pre-2020, 19, 20, 21, you know, where you didn't know that this LLM thing was gonna happen." 00:31:14

Charlie (Broadcom)

Referenced by first name only — this is Hock Tan, CEO of Broadcom, with whom Kim is doing a dedicated panel on XPU and AI chip co-design. (Note: "Charlie" likely refers to another Broadcom executive or panel participant given context.)

"This is what I'll talk with Charlie at Broadcom about." 00:08:53

Henri (RAISE Summit founder)

Founder and organizer of the RAISE Summit in Paris. Kim credits him with building Europe's most significant AI conference from a standing start.

"It was the second year of development that Henri had kind of pioneered and built this event." 00:02:40

Eric Schmidt

Former Google CEO. Referenced as an early RAISE speaker that signaled the conference's legitimacy.

"I remember last year seeing Eric Schmidt on stage. And I had not heard of the conference before. And I was just amazed." 00:03:26 (Molly O'Shea)

Jamin (Coatue, CIO of Public Markets)

Referenced as a fellow investor sharing a similar thesis on agent memory proliferation and the "chip flip."

"We had a conversation with Jamin from Coatue. He's the CIO of public markets over there. And they were talking about what the proliferation of agents' memory is only increasing more." 00:15:14 (Molly O'Shea)

Scott Wu

CEO of Cognition. Featured alongside Kim in the RAISE series, signaling prominence in the AI coding agent space.

Andrew Feldman

CEO of Cerebras. Featured in the RAISE series, one of the most prominent alternative AI compute architects.

Rodrigo Yang

CEO of SambaNova Systems. Referenced as someone Kim knows personally and stood in for at RAISE.

CJ Desai

President and COO of MongoDB. Featured in the RAISE series.

5. Operating Insights

Build Your Company's "Ontology Layer" Before Competitors Do

Kim's framework for the future enterprise is precise: tokens in → data foundation → context/ontology layer → agents. The companies that will win are those that successfully encode the cumulative proprietary knowledge of their organization into a structured context layer that agents can operate within. This is the new operational moat, and it requires deliberate architecture decisions now.

"Can you embody all of the knowledge of your company in what they call a context layer, a layer of the secrets and the ways of your company that you embody the cumulative knowledge of your employees into some context layer... You have this ontology layer and this context layer sitting on the data foundation with tokens in. And then everyone builds agents. Agents go wild." 00:53:58

Talent Strategy: Mine Software Engineers for Hardware Roles Nobody Is Training For

The most acute talent shortage is not in AI software — it is in memory co-design, chip architecture, and analog engineering. The supply of trained people is effectively zero. Companies that proactively retrain or recruit software engineers into these physical-layer disciplines will have a structural advantage.

"We cannot get people to design custom memory... We got to repurpose some of these software programmers into memory co-design architects. And so like all of this has happened in the physical, what I call the physical world." 00:42:22

PE Roll-Up + AI Stack Replacement as a Repeatable Playbook

Rather than selling AI tools into incumbents (slow, politically difficult), a more aggressive operator strategy is to acquire incumbent companies for their customer relationships and distribution, then rip and replace the entire operational stack with AI-native architecture, dramatically cutting the cost base and repricing the service.

"Let's take an old industry. Let's buy these companies. And let's bring in this whole new stack, reimagine the stack, and just sell a whole new solution. Today you used to pay $100, I'll charge you $20." 00:57:26

6. Overlooked Insights

The Chinese-Body, Western-Brain Robotics Arbitrage Is Already Happening

Kim makes this point extremely briefly, almost in passing, but it is enormously significant: sophisticated Western robotics players are already quietly combining best-in-class Chinese-manufactured robot bodies (benefiting from Asia's manufacturing scale and EV-era supply chains) with Western AI brains. This is not a future scenario — it is happening now. This has profound implications: the robotics competitive landscape is not US vs. China, it is a potential collaborative arbitrage that sidesteps both the "China can't do AI" and "US can't manufacture" problems simultaneously.

"You might have is ultimately you can mix and match Chinese physical robot with a Western brain. And I know that's happening. I know people — I just want the robotic, those Chinese robots are amazing, right? So why don't we stick a Western brain in there, inside?" 00:46:43

This is an under-discussed M&A and partnership vector that could produce category-defining robotics companies in the next 24-36 months, operating largely outside the current public narrative of US-China tech decoupling.

The Duration Mismatch in Memory Is a Structural Alpha Opportunity, Not Just a Risk

Kim briefly identifies a specific market structure inefficiency: DRAM and memory fabs take 3-4 years to build, but demand is acute today. This creates a predictable, multi-year supply constraint that is not a market failure but a physical inevitability — meaning the shortage and its associated pricing power for incumbents (SK Hynix, Samsung, Micron) is not speculative but mathematically locked in for the foreseeable planning horizon.

"There's a mismatch of what I call duration. Because it takes three or four years to build a chip fab or a memory fab. But yet everyone is a shortage today... There's this like mismatch of demand supply, duration, and this is causing a lot of angst in the market." 00:13:11

The overlooked insight is that this angst itself is the opportunity: investors who understand the duration mismatch are not taking risk, they are collecting a premium from those who misread a structural supply lock-in as cyclical uncertainty. The three-player oligopoly in memory (SK Hynix, Samsung, Micron) combined with 3-4 year fab lead times means the pricing environment for high-bandwidth memory is essentially set through the end of the decade.