In a single stretch of this summer, the three companies that define the AI era are all reaching for the public market at once. SpaceX filed to raise roughly $75 billion at a $1.77 trillion valuation, the largest offering ever attempted. OpenAI filed confidentially in late May at a valuation approaching $1 trillion. Anthropic filed days later at $965 billion, having just passed OpenAI as the most valuable private AI company. Three of the largest equity offerings in history, clustered in one quarter.

Almost every discussion of them asks the same question: are these valuations justified?

That is the wrong question.

The more useful one is what happens if they work.

Bubbles rarely end because investors turn cautious. They end because capital converges. The Nifty Fifty, Japanese equities in the 1980s, internet stocks in the late 1990s, housing in the mid-2000s: different assets, same mechanism. A small set of names absorbs a disproportionate share of global flows, and the narrative becomes self-reinforcing until it isn't.

The convergence is already measurable. The Magnificent Seven now account for 34.8% of the S&P 500, up from 12.5% a decade ago. In 2025 they produced about 42% of the index's entire return. In private markets the tilt is sharper still: artificial intelligence and machine learning now represent 71% of all venture capital deal activity. Is this normal? No.

The reason is not irrationality. Consider the institutions that have to put trillions to work: sovereign funds, pensions, insurers, family offices. Government bonds offer limited upside. Europe is slow. China carries geopolitical risk. The largest public technology companies are already enormous. AI is the one story that reads as unbounded, spanning robotics, autonomous systems, scientific discovery, and national competitiveness. Capital is flooding a remarkably small set of names not because the alternatives are bad, but because they look less compelling.

Most investors are positioned for an AI crash.

The risk I find more interesting is an AI success.

Suppose the three offerings clear, demand exceeds expectations, and the stocks rise. Capital follows performance. More institutions allocate, more funds are raised, more talent migrates, more attention arrives. The winners get larger, and the gravity they exert intensifies.

This is where the careful version of the argument matters.

Capital is not fixed in aggregate; the pools grow, and much of the $660 billion in projected 2026 hyperscaler capex is funded by the giants' own cash flows. The constraint is at the margin. The next fund raised, the next senior engineer hired, the next allocation decision, the next magazine cover: those bend toward the same handful of names. Crowding out happens one marginal choice at a time, and the businesses that lose access to it are not the ones in the headlines.

That is the second-order effect.

The most durable investment opportunities have historically come from neglect rather than enthusiasm. When capital flows in one direction, entire categories become underfunded and unfashionable while their underlying economics stay intact. The value does not disappear. It becomes temporarily unpriced.

So the question worth holding is not "which AI company wins," but "what is being underfunded because AI is winning?" The answer is not energy or data-center infrastructure, which the buildout is pulling capital toward, not away from. It is the ordinary cash-generating businesses whose value does not depend on the buildout at all, and whose cost of capital is quietly rising as attention drains elsewhere.

What this is not.

This is not a claim that the AI valuations are wrong. The earnings are real: the ten largest S&P names generate roughly 31% of the index's earnings and the seven largest produce close to 70% of its economic profit. Concentration backed by cash flow is not 1972. It is also not a market-timing call. The rotation may already be consensus, with Goldman Sachs expecting the Mag 7 to trail the equal-weight index in 2026.

The narrower claim I will defend is this:

If the offerings fail, valuations reset and the lesson is as-usual.

If they succeed, the same force that lifts them pulls capital, talent, and attention away from everything that is not them. The returns from that distortion accrue to whoever is funding the businesses the distortion leaves behind.

Identifying them is the work.

It is this: the organization tried to run agentic AI on top of an operating system designed for humans — and called that a deployment.

It failed the way you would expect. Not with a dramatic crash. With friction. Approvals that took too long. Agents that escalated everything because no one defined what they were allowed to decide. Humans who didn’t know whether to trust the output or second-guess it. Teams that found workarounds. Pilots that never became production.

he World Economic Forum published an analysis last week describing agents as requiring a fundamentally new model of organizational authority — not software tools, not digital employees, but actors with explicitly defined scopes of action, accountability structures, and handoff protocols. MIT Technology Review ran a piece making the same argument from the perspective of management design. Bain published a framework for how enterprises actually move AI agents from pilot to production.

All three arrived at the same uncomfortable conclusion: the organizations failing at agentic AI are not failing because AI is weak.

They are failing because management is unchanged.

I. Two Operating Systems, Running Simultaneously

Here is the clearest way I can frame the problem.

Every organization that is serious about AI is now running two operating systems simultaneously.

OS-1 is the human operating system. It was built over decades — sometimes centuries — and it is optimized for human cognitive limits, human social dynamics, and human accountability structures. Decisions flow through hierarchy because humans need context and buy-in. Approvals exist because trust is established through precedent, not specification. Roles are defined loosely because humans adapt. Communication is often implicit because shared context reduces overhead.

OS-2 is the machine operating system. It is optimized for throughput, consistency, and scale. Agents do not need context — they need scope. They do not need buy-in — they need permissions. They do not adapt to implicit expectations — they execute against explicit specifications. The faster and more autonomous they run, the more precisely their authority boundaries need to be defined.

The problem is that most organizations have deployed OS-2 inside the governance architecture of OS-1.

An agent asks a question. OS-1 says: “this should go to the manager.” The manager is in three other meetings. The agent waits. The latency that made AI deployment feel slow was always a management design problem masquerading as a technology problem.

II. The Korea Problem

I want to make this concrete, because the abstract version is easy to nod at and ignore.

Korea is one of the world’s most instructive test cases for this problem. Not because Korean organizations are uniquely bad at AI — they are not. The country’s investment in AI infrastructure is formidable. The National Growth Fund approved over 4 trillion won in AI chip, data center, and foundational model investments just this past week. FuriosaAI received 800 billion won. XCENA raised $135 million. Asteromorph, Motif Technologies, Exina — capital is arriving, and the founding teams are serious.

But Korean enterprise deployment has a structural constraint that no one is naming directly: the organizational OS that Korean companies built their competitive advantage on is the most human-optimized OS in the world.

Confucian hierarchy, consensus-before-action norms, seniority-weighted decision rights, and implicit communication within trusted networks — these are not weaknesses. They produced Samsung. They produced POSCO. They produced the fastest industrialization in recorded economic history. For the industrial era, this operating system was the advantage.

Agentic AI does not work inside OS-1. It does not work in hierarchical approval chains that were designed to slow down consequential decisions because the cost of a wrong decision was too high to recover from. An AI agent that has to wait three days for manager sign-off on a $200 decision is not an AI deployment — it is an expensive inbox addition.

The Korean companies that crack this first — that figure out how to run OS-2 inside Korean organizational culture without destroying what made OS-1 valuable — will have built something genuinely defensible. Not just a product advantage. An organizational knowledge advantage that competitors will take years to replicate.

That is a startup opportunity. It is also an investment thesis.

III. What Getting It Right Actually Requires

I have been running a small set of AI agents inside my own work for the last year. A Chief of Staff agent. A Content Lead. A Research Lead. They run through a Slack pipeline and they interact with each other.

Building this taught me something that management frameworks rarely capture cleanly: the hardest part of deploying agents is not the technical setup. It is writing down what you are actually trying to accomplish at a level of precision that a machine can act on.

Humans operate on implicit shared context. When I ask a colleague to “handle” something, years of shared experience compress into that one word. An agent has no such compression. It needs the expansion: what does “handle” mean, what are the boundaries, what counts as success, what triggers escalation.

The act of specifying this — precisely enough for an agent to act on it — forces an organizational clarity that most teams have never had to produce. You discover that your processes were not actually as defined as you thought. That the “approval” step was really two decisions running in parallel that no one had separated. That the “review” step meant three different things to three different people.

This is why the organizations that are succeeding with agentic AI are not the ones that deployed the best model. They are the ones that did the organizational design work first.

The World Economic Forum’s ACAP framework — Autonomous Capability Assignment Protocol — formalizes this: for each agent, define authority (what it can decide alone), constraints (what it cannot do without escalation), and accountability (who owns the outcome when it acts). That framework is not an AI framework. It is a management framework. It just required AI agents to force organizations to write it down.

IV. The Investor Lens

From a VC perspective, this creates a sharper filter than most AI investment frameworks I have seen.

The conventional question for AI portfolio companies is: how good is the model? Or, more practically: what is the retention, what is the revenue, how fast is it growing?

Those questions are necessary but not sufficient. The question I am increasingly asking is: what is the organizational theory embedded in this product?

Every B2B AI product has an implicit answer to the question: how should the humans and the agents divide the work? Some products answer it badly — they assume the human wants to stay in the loop on everything, which produces a product that is slower than not using AI at all. Some answer it well — they have clearly defined the high-judgment decisions that stay human, the repetitive execution that runs autonomously, and the handoff protocol between them.

The products with a well-designed answer to that question have lower churn, faster deployment cycles, and higher net revenue retention. Not because the model is better. Because the org design embedded in the product respects how humans and machines actually work differently.

This is the trust architecture question applied at the product layer, not just the infrastructure layer. And it is what I am watching in every portfolio company and every new deal that crosses my desk.

V. The Two Things That Have to Coexist

Here is the version of this that I think is hardest to hold.

OS-1 — the human operating system — is not the enemy. Hierarchy exists for reasons. Consensus-building exists for reasons. Implicit trust networks exist for reasons. These are not inefficiencies to be eliminated. They are the tissue that makes organizations function as human communities rather than just execution machines.

OS-2 — the machine operating system — is not a replacement. It is a parallel track. The organizations that get this right are not the ones that strip out human oversight in pursuit of agent autonomy. They are the ones that are precise about which decisions belong to which system.

High judgment, high stakes, high uncertainty: OS-1. A human who understands context, can tolerate ambiguity, and is accountable for the outcome.

High volume, high consistency, well-specified: OS-2. An agent that executes reliably, escalates precisely, and surfaces its reasoning.

The management design question — the question that CIOs and founders and GPs should be asking right now — is not “how do I deploy AI?” It is: “for each decision in this workflow, which operating system should own it, and what is the handoff protocol between them?”

Marcus Aurelius wrote in the Meditations: “Everything is interlocked, and the bond is sacred, and almost nothing is alien to anything else.” He was writing about the Stoic doctrine of sympathy — the interconnectedness of all things in the cosmos. But the observation holds here in a different register: the human operating system and the machine operating system are interlocked. The failure mode is not choosing the wrong one. The failure mode is pretending you can run both without designing the interface between them.

VI. What I Am Watching

A few things are in motion right now that bear on this.

Anthropic’s IPO filing (confidentially submitted June 1) will force a level of transparency on frontier AI unit economics that private markets have never had to produce. Revenue quality, customer concentration, compute cost structure, gross margins — all of it will be in an S-1. That is going to be the most informative document for AI investors since the Netscape prospectus. Not because the IPO price will be right, but because the disclosure will clarify which AI business models are actually durable.

Korea’s National Growth Fund investments — FuriosaAI, Rebellions, Asteromorph, the CXL fabless companies — are building the compute and memory infrastructure layer. That layer is table stakes. The question I am watching is which Korean AI application companies will build on top of it with a product theory that solves the OS-1/OS-2 interface problem for Korean enterprises. That is where the durable application value will land.

The organizational design research wave (WEF, MIT Technology Review, Bain all publishing frameworks in the same week) is a reliable signal. When management consultancies start publishing frameworks, it means the early adopters have already proven the pattern and the mainstream is about to ask for a roadmap. The companies that got their organizational design right two years ago are about to have a very good 18 months.

Closing

The $4 trillion in Korean AI infrastructure investment, the Anthropic near-trillion-dollar valuation, the wave of agentic AI deployment frameworks — none of it will produce the returns it promises if the organizations deploying AI do not solve the operating system problem.

The model is not the bottleneck. Management is.

The companies — Korean and global — that figure out how to run two operating systems simultaneously, with a clearly designed interface between them, will not just deploy AI successfully. They will have built an organizational capability that compounds. Each iteration of the machine system will be better specified. Each handoff protocol will be tighter. The humans will get better at the judgment work because the machine has taken the execution work off their plate.

That is the advantage that is actually worth paying for. Not the model. The interface.

Ethan Cho is Partner & CIO at TheVentures, a Seoul-based seed-stage VC focused on AI, mobility, and deep tech. He writes about AI economics, Korean venture, and the investor’s craft at 애당초 4개의 시선 on Substack.

Now there’s a new acronym forming: HBF — High Bandwidth Flash.

It’s early. Engineering samples won’t ship until late this year. But the industry is already moving with unusual speed, and the structural dynamics look remarkably similar to HBM’s early days. Here’s what investors and builders need to understand.

What is HBF?

HBF is 3D-stacked NAND flash memory engineered to deliver near-DRAM bandwidth at flash-level capacity. It creates a new tier in the memory hierarchy — sitting between fast-but-expensive HBM and slow-but-cheap SSDs.

The specs tell the story:

• 512 GB per 16-die stack (vs. HBM’s typical 24-64 GB)

• 1.6 TB/s read bandwidth (matching HBM4’s target)

• HBM4-compatible footprint and power profile

• Non-volatile — no refresh cycles, lower power consumption

• 8-16x more capacity than HBM at ~1/8th the cost per GB of bandwidth

In plain terms: HBF gives you an order of magnitude more memory capacity at the same bandwidth, in the same physical space, at a fraction of the cost. And it holds data without power.

Why Now? The Inference Era Demands It

The AI industry is undergoing a fundamental shift from training to inference. Training is compute-bound — you need fast arithmetic, and HBM serves that well. But inference is increasingly memory-bound.

Consider the numbers:

• GPT-4 class models: estimated 1+ trillion parameters, requiring 1-2 TB just to store weights

• Mixture-of-experts architectures: multiple expert modules need to be loaded and swapped rapidly

• Real-time serving at scale: thousands of concurrent inference requests, each needing fast access to massive model data

Today’s HBM tops out around 64 GB per stack. Even with multiple stacks, you’re looking at 128-256 GB per accelerator. That’s not enough. The industry has been working around this with model parallelism and offloading to SSDs, but both add latency and complexity.

HBF solves this by providing terabyte-scale memory at near-HBM speeds, purpose-built for streaming model weights during inference.

Turing Award winner David Patterson has publicly called this the next AI supply bottleneck. The “father of HBM,” KAIST Professor Joungho Kim, predicts HBF will surpass HBM in total memory demand by 2038.

The Architecture: Dual-Tier Memory

The emerging consensus is a dual-tier architecture:

• HBM handles real-time computation — activations, attention matrices, intermediate results that need nanosecond-level access

• HBF stores and streams model weights, embedding tables, and KV caches at terabyte scale

Think of it like the relationship between L1 cache and main memory, but scaled up to the accelerator level. HBM is your hot working set. HBF is your model store, streaming data at speeds that were previously only possible with volatile DRAM.

SK hynix calls their implementation the H3 architecture (HBM + HBF hybrid), and they’re reporting 2.69x performance-per-watt improvement versus HBM-only configurations.

Key Players and Timeline

SanDisk + SK hynix are leading the charge. In February 2026, they jointly launched the HBF Standardization Consortium under the Open Compute Project (OCP). SanDisk is building a pilot production line now, targeting first samples in H2 2026.

Samsung is taking a more cautious approach — prioritizing HBM4E and HBM5 while investing in NAND architectures compatible with HBF. They’re hedging rather than leading.

Timeline:

• Late 2026: Engineering samples

• Early 2027: First inference devices with HBF integration

• 2027-2028: Broader adoption by hyperscalers (Google, NVIDIA, AMD expected early movers)

• ~2030: Full market maturity

The standardization effort under OCP is significant. It signals that the major platform architects (Meta, Microsoft, Google) are already engaged in defining the spec. This isn’t a science project — it’s a product roadmap.

Korea’s Structural Advantage

This is where it gets especially interesting for anyone tracking Korean semiconductors.

Korea has a unique structural advantage in HBF that no other geography can match:

1. SK hynix has world-leading HBM packaging expertise AND is now co-developing HBF specifications

2. Samsung possesses both advanced NAND flash manufacturing AND HBM packaging capabilities

3. SanDisk (Western Digital) only has NAND — they need a partner for the full stack

Korean companies can integrate both memory tiers. They can build the complete dual-tier solution in-house. No one else can.

SK hynix is explicitly repeating its HBM playbook: secure early ecosystem leadership through standardization, build the manufacturing moat, and lock in design wins before competitors catch up. It worked spectacularly with HBM. They’re running the same play.

Professor Kim at KAIST frames this as a paradigm shift from GPU-centric to memory-centric AI architectures. If he’s right — and HBM’s trajectory suggests he often is — the center of gravity in AI hardware is shifting toward memory makers. That favors Korea.

Investment Implications

For venture and growth investors, the HBF opportunity clusters around four areas:

1. Supply Chain Enablers

TSV (through-silicon via) technology, advanced bonding equipment, and thermal management solutions. These suppliers will receive orders before the main HBF products ship. The pick-and-shovel play.

2. System Integration

Memory controllers, interface IP, and architecture tools for the dual HBM+HBF tier. Whoever builds the “glue logic” between these memory layers captures significant value.

3. Inference-Optimized Accelerators

Startups designing AI accelerators specifically architected around the new memory hierarchy. The chip design assumptions change when you have 10x more on-package memory available.

4. TCO Optimization

HBF dramatically reduces data center total cost of ownership for inference workloads. Companies helping operators transition to this architecture — software-defined memory tiering, workload orchestration, capacity planning tools — have a clear value proposition.

Risks Worth Watching

This isn’t a risk-free thesis:

• Latency gap: NAND is fundamentally slower than DRAM at random access. HBF mitigates this with parallelism, but the gap exists.

• Thermal challenges: Placing NAND flash adjacent to hot GPUs/accelerators requires careful thermal design. Heat degrades flash reliability.

• Durability: Continuous AI inference workloads put sustained stress on flash cells. Long-term endurance under these patterns needs validation.

• Adoption timeline risk: 2030 full maturity means a long runway before peak revenue. HBM4/5 improvements could narrow the gap.

That said, the standardization momentum and multi-company investment suggest this is a question of when, not if.

Bottom Line

HBF represents the same kind of structural shift that HBM created a decade ago:

• A new memory tier emerges to address an AI bottleneck

• Korean companies are positioned at the center

• Early movers in standardization gain outsized advantages

• The supply chain builds out 2-3 years before peak demand

For investors tracking AI infrastructure, this is the next supply chain to map. The window to build conviction — and find the picks-and-shovels opportunities — is now.

Ethan Cho is Partner and CIO at TheVentures, a Seoul-based seed-stage VC firm focused on AI, mobility, and deep tech. He writes about AI infrastructure, venture capital, and the Korean tech ecosystem.

On May 19, 2026, Anthropic hired Andrej Karpathy onto its pre-training team. Four days earlier, Peter Steinberger had posted a screenshot showing his employer OpenAI paid $1,305,088 in API charges over thirty days so his three-person team could run roughly 100 Codex instances. Between these two events, Anthropic users on the $200 per month Max plan continued reporting that their five-hour usage windows were exhausting in ninety minutes.

These are not three unrelated facts. They form the visible surface of a monetization sequence the gaming industry mapped a decade ago and that frontier AI labs now appear to be re-running, with one structural difference that matters for investors.

The Gacha Sequence

Free-to-play games follow a known four-stage monetization arc.

• Stage one: generous onboarding to acquire users at near-zero unit cost.

• Stage two: continuous release of new characters, modes, and servers to drive average revenue per user upward.

• Stage three: tightening of free resources while premium gating intensifies.

• Stage four: fatigue and churn among the heaviest spenders, who originally subsidized the rest of the user base.

The pattern has held across Genshin Impact, Honkai Star Rail, Fate/Grand Order, and most Korean MMOs from Lineage onward. The compression cycle from stage one to stage four typically runs eighteen to thirty-six months for a single title, though publishers extend it through new IP releases and cross-promotion.

Anthropic’s last twelve months map onto this sequence with uncomfortable precision.

• The company released Claude for Excel, Claude for Chrome, Cowork, and Claude Code in rapid succession.

• It tightened OAuth token policies in February 2026 to prevent third-party tools from running on Pro and Max subscription credentials.

• It cut effective rate limits on its highest tier plan. And on May 19, it hired the most recognizable researcher in the field, a move that pattern-matches to peak-cycle vanity acquisition rather than capability gap.

Inverse Gacha

If this were simply a gacha cycle, the investment conclusion would be straightforward: short the cycle, wait for fatigue, exit. But there is a structural difference between game tokens and AI tokens that flips the failure mode entirely.

A Genshin Impact five-star character has zero utility outside the game. A Claude completion produces code, contracts, research, and revenue. The first is pure dopamine. The second is replaceable input to actual production.

This matters because in gacha games, fatigue equals churn equals total customer loss. The player either stays in the ecosystem or leaves it entirely.

In AI, fatigue equals substitution. The Anthropic power user who hits rate limits at ninety minutes does not stop using LLMs. They route the next request to Kimi K2.5, or to Codex, or to a local model.

The Perel Web Studio case from February documented exactly this pattern: Claude Max as the orchestrator, Kimi K2.5 as the overflow worker at roughly ninety cents per million tokens, with monthly Kimi spend of $150 to $300 covering work that would have run several times higher on direct Anthropic API rates.

The investment-relevant claim is therefore narrower than the gacha analogy suggests. Anthropic does not face an existential cliff. It faces what is best called the inverse gacha problem: the harder it squeezes power users to defend ARPU, the faster those power users discover that substitution is cheap and quality loss is bounded. The risk is margin compression, not customer collapse.

What This Is Not

Three counterarguments deserve direct engagement.

First, Anthropic’s ARR growth remains exceptional. Public reporting through May 2026 suggests revenue grew several multiples year over year, which is not the trajectory of a company in late-stage decline. This is correct. The inverse gacha thesis does not predict revenue collapse in the next twelve months. It predicts that the next phase of growth comes with structurally lower gross margins as power users distribute workloads across substitutes.

Second, Anthropic’s enterprise distribution moat is real and growing. Embedding Claude into Excel, Chrome, Salesforce, and Slack creates switching costs at the workflow layer that consumer-grade substitution cannot replicate. This is also correct, and it is the primary reason this is a margin thesis rather than a collapse thesis. Enterprise contracts protect the floor. Power user behavior compresses the ceiling.

Third, the Karpathy hire signals research depth, not desperation. Pre-training acceleration using Claude itself is a recursive-improvement direction that could pay off in capability terms. This is the steelmanned version, and it may prove correct. The counterview is that pre-training moved past peak frontier importance during 2025, and that the marginal return on hiring a pre-training specialist is lower than the marginal return on product velocity or enterprise distribution. Reasonable investors will disagree on this point.

Three Signals Worth Tracking

A thesis becomes investable when it specifies falsifiable signals. Three are worth watching over the next twelve to eighteen months.

Signal one: Opus to Sonnet downgrade trends among existing Max subscribers. A clear migration from Opus to Sonnet for power users would confirm that quality loss from downgrade is bounded enough to drive substitution. Observable through pricing tier mix disclosed in enterprise reporting, or through community usage surveys.

Signal two: Claude Code daily active session length. If sessions shorten while subscription counts grow, that is the gacha pattern where casual users replace power users at the top of the funnel. Observable through third-party usage trackers and developer community reports.

Signal three: enterprise multi-model declarations. The frequency with which large enterprise customers publicly announce multi-model architectures (Claude plus Kimi, Claude plus Llama, Claude plus DeepSeek) is a leading indicator of margin pressure. A doubling of such announcements in any six-month window would constitute confirmation.

Implications for Allocation

For LPs evaluating exposure to the frontier lab category, the inverse gacha frame suggests three positioning principles.

1. Direct equity exposure to Anthropic at current valuations prices in continued ARPU expansion. This thesis implies the next leg of revenue growth comes at lower gross margin. That is not a short thesis. It is a “buy the revenue, fade the multiple” thesis, with different implications for entry timing and structure.

2. Substitution-layer exposure becomes interesting on the same logic. Kimi K2.5 by Moonshot, DeepSeek, and the open-weight inference ecosystem benefit directly from power user substitution behavior. This includes inference providers, routing layers, and prompt-portability tooling.

3. Application-layer companies that pre-commit to single-model architectures carry hidden margin risk not yet priced into their fundraising rounds. Founders who design their products to be model-agnostic from day one will outperform those who optimize for any single frontier lab, regardless of which lab wins on raw capability.

The frontier lab competition is not OpenAI versus Anthropic versus Google. It is a question of which lab adapts its monetization architecture to a category where users have utility-equivalent substitutes one API call away.

The lab that solves this earliest will define the next phase.

The lab that runs the gacha playbook unchanged will discover that AI users do not behave like Genshin players.

Ethan Cho is Partner and CIO at TheVentures, a Seoul-based early-stage venture capital firm founded by the Viki co-founders. Prior roles span Google Korea, Qualcomm Ventures, Samsung Strategy & Innovation Center (SSIC), KB Investment, and Fast Ventures. Early investor in Toss and Dunamu. Seoul National University BBA, Columbia MBA, Korean CPA.

The circular financing diagram has been on the front page of every financial publication since November. Nvidia funds OpenAI, OpenAI buys Nvidia chips, Oracle leases compute, CoreWeave issues debt against those leases, Microsoft books Azure as revenue, SoftBank borrows against its tech holdings to keep Stargate moving. Everyone has seen the chart. Most readers have seen it five times this month.

The chart is not the story. The chart is the evidence. The story is what you believe happens next, and the existing commentary has split into two camps that are both wrong.

The bear camp says it is Lucent 2.0. The loop breaks, equity goes to zero, the buried fiber gets bought a decade later by the next platform generation.

The bull camp says AGI justifies the capex. Revenue catches up. Stargate pays itself back. Both camps price this as a binary: bubble or no bubble.

That is the wrong frame. Private companies that grow this large, this fast, with this much politically aligned capital wired through them do not resolve into either outcome. They get nationalized at a haircut. They become Fannie Mae.

I’ll call the threshold that triggers this outcome the Implicit Sovereign Floor: the point at which a private company becomes structurally inseparable from national industrial policy, after which its valuation is no longer set by markets but by the cost of letting it fail. OpenAI has crossed that line. Anthropic, despite better economics, is six quarters behind.

That is the actual trade. Not whether OpenAI’s loop holds. Whether you are positioned for the moment a private AI company stops being priced as equity and starts being priced as a sovereign credit instrument, with a state-implicit floor and a state-imposed ceiling.

What Lucent missed

In 1999, Lucent financed equipment purchases by newly formed telecom carriers, then booked those purchases as revenue. The carriers used Lucent’s loans to buy Lucent’s switches. When the dollar stopped circling, Lucent collapsed first. Cisco lost 86 percent of its market cap. The fiber buried during that cycle eventually became the backbone of Google and Netflix. The equity holders who funded it never recovered.

Lucent had no political constituency. OpenAI does. That single difference inverts the entire historical analogy.

How the floor activates

Fannie Mae began as a private corporation. By 2008 it held or guaranteed roughly half of the US mortgage market. Treasury did not nationalize Fannie because shareholders deserved protection. Treasury acted because allowing it to fail would have repriced every mortgage-backed security held by every pension fund, sovereign wallet, and bank in the system simultaneously.

OpenAI is approaching that role for AI infrastructure. The $852 billion private valuation is propped against $1.4 trillion in compute commitments held by Oracle, CoreWeave, Microsoft, and SoftBank. A 40 percent reset of OpenAI’s implied IPO valuation forces write-downs across all four counterparties at once. Oracle’s stock has already fallen roughly one-third since November on this concern. CoreWeave is down further.

The US government cannot allow that simultaneous repricing. The CHIPS Act framed semiconductors as strategic national infrastructure. The next chapter of that legislation, whatever it ends up being called, will frame AI compute the same way. When it arrives, the implicit floor under OpenAI becomes explicit.

This is what “too big to fail” actually means. Not that the equity is safe. The system around the equity gets backstopped, and equity holders take a sharp haircut on the way to that floor.

Why Anthropic is six quarters behind, not safe

The standard counterargument: Anthropic is winning. The numbers support this. Anthropic’s annualized revenue went from $9 billion at the end of 2025 to $30 billion by the end of March 2026. In the combined enterprise spend of OpenAI plus Anthropic, Anthropic’s share moved from roughly 10 percent at the start of 2025 to over 65 percent by February 2026. Claude Code alone produces $2.5 billion in annualized revenue and is responsible for 4 percent of all public GitHub commits.

These are real numbers backed by real customers. Anthropic’s economics are genuinely better than OpenAI’s.

The trap is the cost curve, not the revenue curve. Frontier model training requires clusters that scale faster than any individual company’s operating cash flow can fund. Anthropic’s compute commitments to Amazon and Google have already begun to mirror the structure of OpenAI’s commitments to Microsoft and Oracle. The vendor-financed loop forms as soon as cluster size passes a threshold, regardless of how good the underlying business is.

By the time Anthropic crosses $50 billion in annualized revenue, somewhere in late 2026 or early 2027, its compute commitments will likely sit between $400 billion and $600 billion. At that point, Amazon’s quarterly earnings, Google’s TPU monetization narrative, and a portion of US export control policy will all run through Anthropic’s continued operation. The Implicit Sovereign Floor activates not because of bad management, but because of physics. Frontier AI cannot be funded out of operating cash flow at any conceivable scale.

What this is not

This is not a prediction that OpenAI or Anthropic will collapse. The opposite. Both will survive, because both will be backstopped. The question is whether their equity holders capture the returns, or whether the return profile gets compressed by an eventual Treasury or Department of Energy intervention that prioritizes operational continuity over shareholder upside.

This is not a claim that AI is a bubble. The technology works. Claude Code writing 4 percent of global commits is not vaporware. The underlying productivity shift is real and will compound for a decade. The question is not whether AI changes the world. It is who captures the economic surface area of that change.

This is not a claim that infrastructure investors are foolish. They are correctly pricing the implicit floor. Oracle’s downside is bounded by the same logic that bounds its upside. The trade has become a sovereign credit position dressed up as growth equity.

Where the actual returns sit

If the infrastructure layer is a sovereign credit trade with capped upside, the real returns sit in three places the consensus is currently mispricing.

First, the application layer that runs on top of frontier models without owning the compute. The companies that win here look like Bloomberg Terminal in financial data: thin software margins on top of expensive infrastructure they do not pay for. Cursor, Granola, and a small number of vertical AI companies fit this shape. Most “AI startups” do not.

Second, the efficiency and distillation layer. As the cost of frontier capability becomes politically constrained, the value of running 90 percent of that capability at 5 percent of the cost rises sharply. This is where Korean, Chinese, and French open-weight labs have a structural advantage US capital is currently dismissing.

Third, the post-reset entry into infrastructure itself. The Lucent fiber cycle ended with Google and Netflix buying capacity for cents on the dollar. The AI compute cycle will end the same way. The investors who buy GPU capacity from a distressed hyperscaler subsidiary in 2028 will compound at a rate that the investors buying OpenAI secondaries at $1 trillion in 2026 mathematically cannot.

The trade

The most consequential AI investments of this decade will not look like AI investments at the moment they are made. They will look like distressed infrastructure assets, mid-cap vertical software, and weight-efficient research labs in jurisdictions the consensus considers peripheral. The companies on the front pages right now are building the rails. The returns will sit with whoever buys those rails after the first repricing, likely between 2028 and 2030.

Ethan Cho (조여준) is Partner & CIO at TheVentures, a Seoul-based early-stage venture firm. His prior investments include Toss and Dunamu, made during his tenure at Qualcomm Ventures Korea. Earlier roles include Google Korea and Samsung Strategy & Innovation Center (SSIC). Columbia MBA.

The Seoul press covered it. The English-language press did not.

The ruling is easy to miss because the mechanism is administrative rather than legislative. Korea’s Financial Services Commission convened its Immunity Review Committee (면책심의위원회) and granted regulatory immunity, called myeonchaek, to Korean financial institutions participating in the ₩150 trillion National Growth Fund. The scope covers four activities: co-investing alongside the Fund’s direct equity commitments, committing as LPs in policy-backed funds, participating in infrastructure investment and lending, and joining low-interest co-lending facilities. Willful misconduct and gross negligence remain sanctionable. Everything else does not.

This is what I call the Myeonchaek Unlock. It is the thesis that Korea’s binding constraint on domestic institutional venture participation was never capital scarcity, expertise, or regulatory capacity. It was compliance-department loss aversion. Remove that, and Korean bank, insurance, and pension money moves into venture at a scale foreign LPs have not yet priced into their Korea exposure.

The constraint that was

Korea’s 2026 National Growth Fund totals 150 trillion won, roughly $110 billion at current rates, split evenly between government sources and private, public, and financial-institution sources over a five-year horizon. It was announced in March 2025, scaled up from 100 trillion won to 150 trillion won in September 2025, and formally adopted on September 9, 2025.

The arithmetic worked. The behavior did not. Korean financial institutions with the balance-sheet capacity to anchor venture-adjacent strategies, what a Silicon Valley LP would call insurance general accounts, bank-affiliated asset managers, and retirement-funds allocators, have historically been governed by a sanction regime that penalizes loss-making decisions in ex-post examination. The result was predictable. Compliance officers disallowed exposures that would have been routine in Boston or Menlo Park. The National Growth Fund was at risk of becoming a government-only vehicle dressed as a co-invest structure.

The March 6 ruling addresses that directly. In the FSC’s own framing, the goal is to reduce the burden of ex-post examination and sanctions on unpredictable losses, so participating institutions can move more aggressively into what the commission calls “productive finance.” Coverage applies specifically to the National Growth Fund and the four delineated activities. It is not a blanket rewrite of Korean financial oversight.

Why foreign LPs should price it

Three second-order effects are worth tracking.

First, the composition of Korean domestic LP capital in venture is about to shift. Prior to this ruling, the largest domestic LPs in Korean venture funds were the Korea Fund of Funds (KVIC), Growth Ladder Fund, and a handful of corporate pension accounts. Bank and insurance balance sheets were structurally absent from riskier vintages. The Myeonchaek Unlock changes that. Foreign LPs assessing Korean funds should now expect the LP base of any National Growth Fund sub-fund to include a meaningfully broader slate of Korean financial institutions than prior vintages, with implications for co-investment flow, follow-on capacity, and exit-timing alignment.

Second, GP selection criteria are changing. The 2026 National Growth Fund framework rewards, among other things, technical specialization in advanced strategic industries and a demonstrated willingness to deploy into long-gestation deep-tech categories. KDB has indicated that specialist GPs will receive scoring advantages. Fund managers selected in this vintage will not look like generalist 2019 Korean VCs. Foreign LPs co-investing into sub-funds should expect a materially different investment personality than the last cycle. This is the “who gets picked” consequence of the “who can underwrite” unlock.

Third, the window is time-bound. The National Growth Fund operates on a five-year horizon. The myeonchaek is tied to that program. The cultural effect, domestic Korean financial institutions actually building competency in venture underwriting, will outlast the program only if the first vintage performs. Foreign LPs who co-invest in the 2026-2028 vintage get domestic partner exposure under conditions Korean private LPs will not see again.

What this is not

Three counterarguments deserve explicit treatment.

It is not a blanket loss guarantee for the National Growth Fund. The myeonchaek is regulatory, not civil. Fund managers remain fully subject to their fiduciary duties under the Financial Investment Services and Capital Markets Act, and GPs must still commit one percent of sub-fund size as mandatory subordinated capital. Skin-in-the-game has not disappeared.

It is not a clean replica of the 1979 ERISA reinterpretation. ERISA’s rule change applied to the full U.S. pension universe. The myeonchaek applies only to National Growth Fund participation. The structural logic, removing loss-aversion as a binding constraint on institutional risk capital, is parallel. The scope is narrower.

It is not evidence that Korea has become a state-capitalist venture ecosystem. The popular narrative frames the ₩150 trillion commitment as a sovereign-scale intervention. The Myeonchaek Unlock pushes in the opposite direction. It is a concrete policy effort to draw private institutional capital into risk-on positions by removing the regulatory cost of occasional failure. Whether it succeeds is an empirical question. Calling it a failure in advance, or a state takeover in advance, is lazy.

What to watch

Four markers will tell foreign LPs whether the Myeonchaek Unlock holds.

Sub-fund GP selection is scheduled for completion by mid-2026. Watch the composition, specifically the share of AUM flowing to specialist deep-tech managers versus generalist PE/VC houses. A specialist-tilt validates the policy intent.

LP commitment rosters for the first sub-funds will reveal whether domestic financial institutions actually move. If insurance and bank balance sheets show up in size, the unlock is real. If they do not, the ruling was cosmetic.

Co-invest terms for foreign LPs entering alongside National Growth Fund sub-funds will be negotiated in the second half of 2026. This is the operational window to price the unlock into Korea exposure.

Finally, the 2027 extension decision. If FSC extends or broadens myeonchaek scope, the cultural reset is durable. If not, 2026 was a pilot.

Ethan Cho is Partner & CIO at TheVentures, a Seoul-based early-stage venture firm founded by the Viki co-founders. Prior roles include Qualcomm Ventures, Samsung Strategy & Innovation Center (SSIC, Silicon Valley), KB Investment, and Fast Ventures, with early investments in Toss and Dunamu. Columbia MBA, Seoul National University BBA, Korean CPA. He is building the K-AI Fund, a cross-border AI investment vehicle. Writes “The Signal” for LPs tracking Korean venture from outside Korea.

Italy follows the same pattern. Renaissance workshops became Gucci, Prada, Ferragamo.

Japan’s Edo craft culture became Muji, Issey Miyake, Comme des Garçons.

Luxury industries do not create cultural depth. They export it.

This is the part most allocators miss when they look at Korea.

The question I get from non-Korean LPs is always some version of: Is K-pop a fad? The question assumes Korean content is a recent invention, a marketing phenomenon, a government subsidy story. It is not. Korea has 5,000 years of accumulated aesthetic inventory sitting in warehouses nobody outside the peninsula has ever walked through.

Visit the National Museum of Korea in Seoul. It drew 4.2 million visitors in 2024, more than the Met, more than the British Museum. Walk the Goryeo celadon hall. Walk the Joseon white porcelain hall. Walk the calligraphy wing. You are not looking at a museum. You are looking at a distribution problem that just got solved.

The distribution layer is what changed, not the inventory.

For most of the last century, Korean content could not leave Korea. Language barrier. No global platforms. No translation infrastructure. No payment rails. An aesthetic tradition older than most European nations sat on the shelf because there was no way to ship it.

Then the stack arrived. Netflix globalized scripted content. YouTube globalized music. TikTok globalized short-form. Spotify globalized audio. Shopify, Coupang, Oliveyoung globalized physical goods. Papago and ChatGPT collapsed the language moat. Within roughly fifteen years, every logistical reason Korean culture stayed domestic disappeared.

This is why Squid Game, Parasite, BTS, Blackpink, Stray Kids, K-beauty, and Korean cuisine all broke globally inside the same narrow window. It was not a wave. It was a backlog clearing.

Paris took 300 years to build what Korea is compressing into 30.

The compression is possible because Korea skipped the infrastructure build. It inherited the distribution layer from American platforms the week it was ready to sell.

For LPs thinking about Korea as an investment geography, this reframes what you are actually underwriting. You are not betting on one hit drama or one cosmetics brand. You are betting on a country with a 5,000-year inventory and a 15-year-old distribution channel, where the ratio of unshipped to shipped content is still heavily tilted toward unshipped.

I have a term for this: Cultural Inventory Arbitrage. The gap between what a country has accumulated and what it has exported. Korea’s gap is the widest in the developed world. France closed its gap over three centuries of industrial buildout. Korea is closing its gap in a single generation, because it does not need to build the pipes. The pipes came free.

Founders who win in this environment share one trait. They understand the inventory natively because they grew up inside it, and they understand the distribution layer because they grew up on it. I call this Founder Intelligence: high-context Korean cultural fluency combined with American-scale ambition and execution. It is not a soft quality. It is a measurable selection filter. Every Korean founder we back clears this bar.

If you want to see what we are investing in before it ships, visit the National Museum of Korea. The ticket costs five dollars.

The thesis it illustrates is worth considerably more.

Ethan Cho is CIO of TheVentures, a Seoul-based early-stage VC firm founded by the co-founders of Viki (acquired by Rakuten). His frameworks include Founder Intelligence, Cultural Inventory Arbitrage, and the E/D/R framework for AI investment. Previously at Google Korea, Qualcomm Ventures, Samsung Strategy & Innovation Center, and KB Investment. Early backer of Toss and Dunamu. Columbia MBA, Seoul National BBA. Ranked #1 Korean LinkedIn influencer by Favikon.

The KOSPI rose more than 75% in 2025, its best year in over two decades and the strongest performance of any major global index.

Over the same twelve months, Korean private capital deal value collapsed from $24 billion to $16 billion. Venture fundraising fell from $9.3 billion to $1.1 billion.

That gap between public market euphoria and private market paralysis is the widest I have seen in fifteen years covering this market across Google Korea, Qualcomm Ventures, Samsung’s Strategy & Innovation Center, KB Investment, and now TheVentures.

It will not stay this wide.

LPs who wait for confirmation will pay the recovery price, not the reset price.

Competition is for losers. Korea is where competition left.

The best venture returns in history came from markets where one or two investors had privileged access and everyone else was absent. Sequoia in 1970s semiconductors. Benchmark in 1990s internet. Founders Fund in 2000s enterprise. The pattern repeats. Outsized returns do not come from outbidding ten other smart funds for the same deal. They come from being one of the few investors paying attention when a market is structurally mispriced.

Korea in 2026 is that market.

The sharpest LPs push back with the same question: if Korean venture is such a great opportunity, why did zero VC funds above $250 million close in 2025? Why did sub-$50 million funds collapse from 297 in 2022 to just 32?

The answer is the entire thesis.

2025 fundraising reflects underwriting decisions made in 2023 and 2024, during peak NPS commitment freeze, the Homeplus fallout, and regulatory uncertainty around PE leverage limits. Fundraising is the lagging indicator. It tells you what domestic LPs believed 18 months ago, not what the market looks like now. What the market looks like now is a competition vacuum.

The emerging-manager layer that normally bids up seed and Series A deals is gone. The mid-size funds that normally consolidate Series B rounds are gone. Into that vacuum walked foreign capital and corporate venture, which together took 40% and 63% of deal value respectively. Arm led Rebellions at a $1.4 billion valuation. FuriosaAI turned down an $800 million offer from Meta. The bids that cleared in 2025 came from investors with proprietary conviction and strategic reasons to be there. Everyone else was absent.

This is the shape of a market bottom. Not the absence of quality. The absence of competition for it. Every other major Asian venture market is either crowded (Japan, India), politically constrained (China), or structurally shallow (Southeast Asia).

Korea is the only developed Asian market where a dedicated manager can still enter quality deals without five other funds at the table.

The structural tailwind LPs are underweighting

While domestic sentiment was retreating, Korean policy was loading the next cycle.

December 2025: the government launched the KRW 150 trillion National Growth Fund to back semiconductors, AI, robotics, biotech, and future mobility through direct equity, fund of funds, infrastructure financing, and lending. The Ministry of SMEs and Startups launched the “Four Venture Powerhouses” strategy with 2030 targets of KRW 40 trillion in annual venture investment, 10,000 AI and deep-tech startups funded, and 50 unicorns and decacorns created. M&A guarantees expand from KRW 30 billion to KRW 200 billion by 2030. The Korea Fund of Funds is being extended beyond its 2035 sunset in ten-year increments. A dedicated KRW 333.3 billion early-stage vehicle is being established. A Startup Venture Campus opened in Silicon Valley in January 2026, with additional hubs planned in Singapore, Tokyo, London, and New York.

Corporate acquirers are also pre-committed. Samsung, SK, Hyundai, and LG have collectively pledged KRW 800 trillion, about $580 billion, to AI, semiconductors, energy transition, and advanced manufacturing over five years. January 2026 reforms abolished the mandatory CVC divestment rule and raised the corporate contribution cap for accelerator-led funds from 30% to 40%. Strategic buyers for the 2026-2028 vintage are policy-enabled and capital-committed.

Pair these two facts. Entry prices are the cleanest they have been in six years because competition left the market. Exit demand is the most structurally loaded I have seen in fifteen years because policy and corporate capital both converged on the same destination.

That configuration does not exist anywhere else in Asia today.

Founder Intelligence is the actual asset class

The narrow bet worth making is not “Korean VC.”
It is a specific founder profile.

Korea is a high-context society where founders grow up reading rooms, managing hierarchy, and surviving compressed social feedback. The best of them then train inside American-style execution cultures, usually at a US tech company, a US-educated startup, or a cross-border subsidiary. The combination produces a founder who can sell into Bentonville on Monday and run a factory floor in Gumi on Friday.

I call this Founder Intelligence.

The Korean Diaspora cohort, founders operating natively across Seoul, the Bay Area, New York, and Tokyo, is where the outsized returns sit. They are unreachable through a standard pan-Asia fund. They do not show up in Seoul-only networks either. The manager needs feet in both worlds and a career arc that proves it.

Why most LPs will still miss this

I have watched this movie before. Every cycle, foreign LPs who want Korea exposure end up in one of three dead ends.

1. They buy into a large Seoul-based generalist fund and get diluted into domestic-focused deal flow.

2. They take a pan-Asia fund where Korea becomes a 10% sleeve with no sourcing edge.

3. They co-invest opportunistically and only see deals at the price everyone else already bid up.

Each path feels like Korea exposure. None of them capture the actual trade. Each path also puts the LP back into a competitive market rather than the non-competitive one Korea currently offers.

The Korean Diaspora thesis requires a manager who sources domestically in Korean, underwrites in English, and whose LP base, portfolio logic, and career arc all point outward from Seoul by design rather than by accommodation. That description fits a small handful of firms globally.

My track record across earlier roles includes leading early investments in Toss and Dunamu during my time at Qualcomm Ventures, now two of Korea’s most valuable private companies.

TheVentures was founded by the Viki co-founders after their sale to Rakuten, and the firm was built from day one around cross-border founders rather than around the domestic Korean market.

Cross-border is the origin, not the marketing.

What you do in the next 90 days or leave on the table

Three actions.

1. Ask your pan-Asia managers what share of their 2025 Korea allocation went into sub-$500 million pre-money rounds. The honest answer is almost always near zero. That is the gap the emerging-manager collapse just opened and the gap foreign LPs are currently mispricing as absence of opportunity rather than absence of competition.

2. Screen Korea-dedicated managers on one question: show me a portfolio of founders who scaled into the US market within 24 months of seed. If the answer is theoretical, pass. If the answer is a list of names with revenue, that is the Founder Intelligence signal. Most will fail this test.

3. Move before the National Growth Fund closes its first vintage and the KVIC Fund of Funds extension is ratified. The public market already priced in Korea’s governance reforms and added 75% in twelve months. Private markets are 18 months behind the same signal. LPs who enter now are buying the reset. LPs who wait are buying the recovery, against ten other funds that finally noticed.

The capital is already flowing. Arm. Meta. KKR. Air Liquide. The question is not whether foreign capital is moving into Korea.

It is whether you are moving with it now, while the market is still empty, or after everyone else arrives.

Ethan Cho (조여준) is Chief Investment Officer at TheVentures, a Seoul-based early-stage venture firm founded by the Viki co-founders. His 20-plus year career spans Google Korea, Qualcomm Ventures (where he led early investments in Toss and Dunamu), Samsung Strategy & Innovation Center, KB Investment, and Fast Ventures. BBA, Seoul National University. MBA, Columbia Business School. Ranked #1 Korean LinkedIn influencer (Favikon, 2026).

Now run the same scenario with one variable changed. The same script, the same animation studio, the same voice cast, the same music. But the commissioning platform is TVing, a Korean domestic streaming service. Under that scenario, the project ends as a children’s program on a weekend slot, closes after three weeks of middling ratings, and never leaves the peninsula. Same content. Same creative team. Different outcome by two orders of magnitude.

This thought experiment is the cleanest way to explain why Korea is the most mispriced startup market in Asia.

This is not a story about content.

The counterfactual works because it isolates the variable that actually matters. The creative input was globally competitive. The distribution channel was the bottleneck. When the content was routed through a global channel, it cleared global prices. When it would have been routed through a domestic channel, it would have cleared domestic prices. The gap between those two prices is the arbitrage.

The same gap exists across Korean technology companies right now. Korean founders are building products that compete credibly against US and Chinese peers. Korean capital markets, Korean M&A activity, and Korean IPO valuations are not. When an analyst in Boston or New Haven looks at Korean VC returns on a spreadsheet, what they see is the domestic distribution channel, not the underlying content.

They are watching TVing, not Netflix.

What US endowments are actually measuring

The standard objections to Korean VC allocation are familiar to anyone who has sat across from a CIO. Fund sizes are too small for meaningful ticket deployment. Exit comparables come back weak because the IPO market discounts Korean listings by thirty to forty percent against global peers. Prior experience with Korean funds produced disappointing net multiples. Governance and reporting quality trails what an endowment expects from a 2018 vintage European or Israeli fund. Currency hedging eats into returns.

Every one of these objections is defensible. None of them measure founder quality. They measure the channel that Korean founders historically had no choice but to use.

That is the mispricing. An endowment is not refusing to pay for Korean founders because Korean founders are weak. It is refusing to pay because the exit channel is weak, and it has been pricing the two together as if they were inseparable. They are no longer inseparable. Samsung is not the only distribution path anymore. Kosdaq is not the only liquidity venue anymore. A Korean founder with the right background can now build a company that routes entirely through US capital markets, US customers, and US acquirers, while keeping engineering and operating leverage in Seoul at a fraction of the cost.

Founder Intelligence and the arbitrage pattern

At TheVentures, I have spent the past several years formalizing what I call Founder Intelligence, the investment thesis that the durable edge in Korean venture is not sector, not technology, and not timing, but a specific founder profile that combines high-context Korean social fluency with American-style execution velocity.

The Korean Diaspora founder sits at the intersection of two operating systems that rarely coexist in one person. One operating system is Korean: dense network access, engineering talent inside the country, cost discipline inherited from growing up in a brutally competitive domestic market, and the social capital to recruit senior Samsung or Naver engineers on trust alone. The other operating system is American: comfort with venture-scale ambition, willingness to burn capital against a clear plan, native English go-to-market fluency, and a board-management instinct shaped by US norms.

Almost every globally scaled Korean-origin company of the last decade has been built by a founder with both operating systems active. Coupang. Sendbird. Moloco. Lunit. The ones that stayed inside one operating system stayed inside the domestic channel, and their cap tables priced accordingly. The ones that bridged both scaled onto global channels, and their cap tables priced accordingly.

The investment thesis is narrower than “back Korean startups” and sharper than “back Korean-Americans.” It is this: back Korean-origin founders who can operate in both contexts simultaneously, and enter their cap tables at Korean seed and Series A prices before the global channel repricing happens.

Why the gap has not closed

The reasonable question from any endowment CIO is this. If the mispricing is real, why has it not already been arbitraged away by smarter local capital?

Three reasons.

First, Korean LPs, who fund the majority of Korean VC, optimize for policy and stability mandates rather than global outlier returns, which means most Korean fund strategies are designed to look defensible to a domestic LP committee, not to catch a Coupang.

Second, US-based funds that could play this arbitrage do not have sourcing density in Seoul and treat Korea as an occasional flyer rather than a primary hunting ground.

Third, the founder pool this thesis depends on, Korean-origin operators with both operating systems active, was too small a decade ago and has only recently reached the density required to support a dedicated strategy.

The window that produced the mispricing is the same window that is closing it. An endowment that takes a position in the next twenty four months is participating in the repricing. An endowment that waits five years is buying after it.

What this means for a US endowment allocator

The ask is narrow and specific. Not a blanket Korea allocation. Not a fund of Korean funds. A direct exposure to the Korean-origin founder cohort that is already routing through global channels, sourced by a GP with full context on both operating systems and a portfolio concentrated on the founder profile described above.

If that framing is interesting enough to pressure test against your current Asia allocation, I will make the case directly. The data I would bring to that conversation is not public, and it should not be. It is the part of the thesis that does not belong in a newsletter.

Ethan Cho (조여준) is Chief Investment Officer at TheVentures, a Seoul-based early-stage venture capital firm founded by the co-founders of Viki (acquired by Rakuten). His investment career spans Google Korea, Qualcomm Ventures, Samsung Strategy and Innovation Center, KB Investment, and Fast Ventures, with early investments in Toss and Dunamu. He holds a BBA from Seoul National University and an MBA from Columbia Business School, and is a licensed Korean CPA.

Walk into a Sand Hill Road partner meeting and the question is: who left Stripe last quarter, and how do we get into their seed round before anyone else hears the name. Walk into a Yeouido investment committee and the question is: how many times is this Pre-IPO secured against downside, and what is the redemption clause if the listing slips past 2027.

These are not two cultures investing in the same asset class with different styles. They are two financial systems that have placed their bets at opposite ends of the risk curve. American venture capital concentrates its conviction at the earliest, most uncertain moment. Korean institutional capital concentrates its conviction at the latest, most certain moment. The inversion is structural, and once you see the cause, the behavior of every deal in both markets becomes predictable.

The cause is the LP, not the GP.

In the United States, the dominant venture LP is the university endowment, the family office, and the pension fund willing to lock capital for ten to fifteen years against the explicit expectation that most funds will return modestly and a few will return everything. The power law is not a belief in Silicon Valley. It is a budgeted assumption. Yale does not call Sequoia after a down year and ask for a redemption schedule.

In Korea, the dominant LP is 모태펀드 (the Korea Fund of Funds) and its policy-fund cousins, evaluated by the National Assembly and the Board of Audit. Their fiduciary frame is not “did you generate alpha across the portfolio.” It is “did this specific fund return its principal with a multiple, and if not, who is responsible.” A Korean GP whose fund liquidates below 1.0x does not get to raise the next one. Often the firm itself does not survive. Variance is not rewarded here. Variance is treated as negligence.

That single fact reorganizes everything downstream.

If your LP measures you on every fund’s liquidation multiple, you cannot run a portfolio where eight of ten investments go to zero. You need eight of ten to return two times. The only way to engineer that distribution at scale is to invest after product-market fit is proven, after revenue is contracted, after the IPO underwriter is selected, and to wrap each check in redeemable convertible preferred shares with put options, ratchet clauses, and IPO-failure penalties. That is not risk capital. That is structured private credit dressed in venture clothing.

If your LP is Yale, you can do the opposite. You can pay forty million dollars for eight percent of a six-person company because you know that one such position every five years pays for the entire fund. You will lose money on most of these checks. Yale knows. Yale signed up for it in writing.

I call this the Risk Capital Inversion, and it is the single most important thing a non-Korean LP needs to understand before allocating to a Korean fund. The Korean venture market is not undercapitalized. It is miscapitalized. The capital that exists is structurally prohibited from doing what venture capital is supposed to do, which is buy uncertainty cheaply and hold it long enough for the math to work.

This is also the opportunity. A Korean fund backed by LPs who do not require quarterly downside-protected reporting, who accept the power law in writing, who measure returns at the portfolio level rather than the fund level, can do in Korea what American funds did in Silicon Valley a generation ago.

Not because the founders are absent. The founders are there. I have spent twenty years watching them. The capital structure to back them properly is what has been missing.

That is the gap TheVentures was built into, and it is the gap our AI Fund is being raised to fill.

Ethan Cho (조여준) is Chief Investment Officer at TheVentures, a Seoul-based early-stage venture capital firm founded by the co-founders of Viki (acquired by Rakuten). His twenty-year career includes Google Korea, Qualcomm Ventures, Samsung Strategy & Innovation Center, KB Investment, and Fast Ventures, with early investments in Toss and Dunamu. He holds an MBA from Columbia Business School and is ranked the #1 Korean voice on LinkedIn by Favikon.