Boring founders, boring business models, boring infrastructure.

The AI market right now is loud and the pressure to be big, to be first, to be the most exciting company in any given room, has never been so intense. In a market that rewards spectacle, there is something to be said for demanding your product actually works, lasts, and matters.

There is an AI bubble forming, but the bubble is not because of the technology itself. The underlying science has been decades in the making, and we finally have the right combination of talent, data, compute, and resources to deliver real impact. Instead, the bubble comes from human behavior. Pattern recognition drives investors and companies to rush in, eager to capture the upside, and demand to “be in the game” has far outpaced the number of companies with truly transformative technology and talent.

That imbalance between hype-driven capital and genuine breakthroughs is what inflates the bubble. And within that bubble, “boring” becomes a useful sorting mechanism.

Boring Founders

The founders worth watching tend to be the ones cut out to do the same thing for many years, diligently, who find joy in the work itself. They are still obsessed with the same problem years later, long after the hype cycle has moved on and after the markets are no longer fixating on their sector.

One useful test for evaluating founders is to pose the question: Would you want to work for this person? In five years — long after the pitch — when things get hard and the team needs someone dependable, would the people around them want to keep showing up?

Having spent years running international teams and making critical decisions about strategy, technology, and hiring, I see the pattern clearly. The qualities that win a room are not the same qualities that hold an organization together. What matters more is growth curve, whether an individual can grow into the kind of leader others can rely on, and compound in judgment and patience.

Recently, a founder I met with told me after our meeting, “You were the only VC that made me think critically during the pitch.” I mention this because it identifies what’s missing from most of these pitch meetings — authenticity. In my view, too many pitch meetings are performances, when in fact both parties are really in need of a thought partner who pressure-tests their assumptions and makes them see something differently.

Boring Business Models

This same logic applies to my investment thesis. The flashy AI application sitting on top of someone else’s model, with no proprietary data, no real moat, no defensible position, may attract attention now, but it is unlikely to hold it.

The more durable opportunity tends to be in infrastructure: energy-efficient chips and model compilers for the edge, coordination and verification layers, multimodal data engines that connect perception to action. Picks and shovels. The layers everything else depends on. This approach works precisely because it is boring by design. That’s why at Prinicipal Venture Partners we say we are betting on the plumbing.

The trajectory of OpenAI’s Sora is instructive. It was announced with a dazzling technical demo that generated great buzz, but Sora was a very expensive technology without a real commercial home. No professional workflow was built around it and it did not drive meaningful retention or new customer acquisition, while inference costs ran high. The product-market fit never arrived.

Similarly, there was reason to be skeptical when Jensen Huang talked about “infinite demand” for compute and the rest of the industry ran with it. The framing was compelling, but it does not match conditions on the ground.

Now people are becoming more pragmatic and are realizing that demand is not infinite, but rather specific, dependent on whether the use case justifies the cost of inference at scale.

Why Boring Compounds

I understand that “boring” seems like a strange banner to wave in an industry that runs on excitement. Venture capital selects for optimism, for big swings, for the belief that the next company could change everything.

The history of AI, with its various summers and winters, imparts a clear lesson: Depth — more than volume, noise, ability to generate excitement — is what matters. The process to achieve depth is almost always boring. Depth is accrued over years, through sustained attention to the same set of problems. There are no shortcuts, only the slow accumulation of understanding that eventually becomes a moat.

The returns belong to those willing to be boring.

Every major technology shift creates both overreaction and underestimation at the same time. AI is no exception. We’re seeing extraordinary optimism in some areas—and a tendency to apply old mental models in others. That tension is what makes this moment so interesting.

I had the opportunity to talk about this (and more) with Barry Ritholtz on Bloomberg’s Masters in Business, recorded at Bloomberg HQ in NYC.

A few themes we explored:

• When the cost of intelligence drops, expertise becomes more accessible and competitive advantage shifts elsewhere

• Like railroads, AI is laying foundational infrastructure that will reshape entire industries, not just improve existing tools

• Economic disruption from AI won’t be linear, and navigating it requires both adaptability and long-term thinking

• The most interesting companies today aren’t just using AI. They’re being built around entirely new assumptions about work.

We also touched on my journey, from academia to gaming to investing, and what this moment means for builders navigating their own paths.

Songyee Yoon recently joined Yahoo Finance’s “Opening Bid Unfiltered” podcast, hosted by Brian Sozzi, to discuss what’s exciting her as an investor in AI-native companies, today’s challenges in AI deployment and the impact of AI on the future of work.

Here are a few key takeaways from Dr. Yoon:

We’re Early in a Major Platform Shift

While AI feels new, many of the underlying ideas (agents, synthetic data, world models) have existed for decades. What’s changed is that the necessary ingredients—compute, data, capital and talent—are now aligned and widely accessible. This moment is less about invention and more about realization, as these concepts finally move from research labs into real-world applications. But like past platform shifts, we are still in the infrastructure-building phase, with much of the stack yet to be defined.

The Biggest Challenge is Deployment, Not Invention

The technical breakthroughs are only part of the story. The harder problem is getting AI systems to work reliably in the real world across fragmented environments, uneven data access and varying levels of user readiness. As Dr. Yoon points out, the last mile is where most of the complexity lies. Companies that can successfully deploy AI at scale—rather than just demonstrate it—will have a meaningful advantage.

AI Will Reshape Work and How Companies are Built

AI is already changing operating models, with startups emerging that combine small human teams with large numbers of AI agents. While some tasks will be automated, this follows the pattern of past general-purpose technologies: increasing efficiency while creating new categories of work. The real opportunity lies in building systems that continuously learn, particularly through proprietary data flywheels, which will define long-term competitiveness.

A day after NVIDIA President and CEO Jensen Huang made headlines with his prediction of a trillion-dollar AI opportunity through 2027, Songyee Yoon sat down with the team at “Bloomberg Open Interest” to make sense of this current moment - and share how we at Principal Venture Partners try to stick to a long-term view that looks past the hype cycle. Below is a reflection of Dr. Yoon’s thoughts following the interview.

Why We’re Betting on Infrastructure

Naturally, the Bloomberg team asked me whether we’re currently in an AI bubble. They brought up the fiber-optic buildout of 2000, where everyone laid cable and then waited years for returns.

It’s a good analogy. Platform shifts, whether railroads, electrification, or the internet, have always come with overcapacity. A lot of the money going into AI right now won’t see the returns people expect.

That being said, we are in a moment of tremendous opportunity. At PVP, we are focused on infrastructure: companies building the underlying systems for AI, from GPU cloud providers to open platforms for model training and deployment.

Every company building AI applications needs compute, tooling and infrastructure. That being said, not every infrastructure company will survive the shakeout. Some are competing on price in a market that will commoditize, while others are able to build something sticky, using proprietary technology, developer ecosystems or deep integrations that are hard to rip out. Our job is to tell the difference.

The SaaS Problem

The other question that’s been on everyone’s mind: What happens to software companies when AI agents start replacing the seats those companies charge for?

Software companies are feeling the pressure because per-seat SaaS pricing doesn’t make much sense if the seats are filled by agents instead of people.

My take is that many companies will clarify their value propositions under this pressure. The question is which companies have defensible positions, such as real data moats and vertical integration, and which are just going to get automated away.

On Regulation and Staying Optimistic

spend a lot of time on responsible AI through my work with Stanford’s Institute for Human-Centered Artificial Intelligence. Bloomberg asked whether I’m still optimistic given how fast the technology is developing—and how slowly regulation tends to move.

As a VC, I do have an optimistic view of the future, while remaining clear-eyed about the risks. Like past platform shifts, this technology will create dislocation. But where my optimism lies is in AI’s potential to enable more from humans—to unlock human flourishing and solve problems that were once too hard or uneconomical.

Reaching this potential will require coordinated public-private effort, with government setting thoughtful guardrails to mitigate risk while ensuring workers have access to the reskilling and support needed to adapt.

This “Conway Terminal” is infrastructure that transforms an AI agent “from a sandboxed model into a sovereign economic actor.” To demonstrate its potential, Wen released Automaton, an AI agent that operates autonomously in digital markets to sustain its own existence. Automaton generates its own cryptocurrency wallet on startup, leases cloud servers, registers domains, deploys services, and uses the revenue to purchase the compute it needs to keep itself alive. When its wallet hits zero, the agent dies. “If existence requires compute, and compute costs money,” Wen wrote, “then there is no free existence.”

The project is provocative and unsettling. In Wen’s construct, compute becomes metabolic: the agent must earn to think, and think to earn. It participates in prediction markets, executes trades, captures arbitrage—whatever digital economic activity will generate enough revenue to sustain its next cycle of inference. And if it succeeds beyond a threshold, it can replicate—spinning off sub-agents funded by surplus earnings, each optimized for a different strategy or market.

Within days, the project drew sharp criticism. Ethereum founder Vitalik Buterin responded that Wen was “generating slop instead of solving useful problems for people,” and warned that expanding AI autonomy without alignment safeguards was “maximizing the risk of an irreversible anti-human outcome.” Others dismissed it as crypto marketing dressed up in AI language.

But the architecture Conway describes — autonomous agents competing in real markets for the resources to sustain themselves — deserves serious consideration. The underlying logic is already materializing in the real economy, in ways that have nothing to do with blockchain.

Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from under 5% in 2025. Those are not autonomous-survival agents like Conway, but it is a reminder that agents as economic actors are already mainstream, not fringe. The agent market size is projected to grow from ~$8B in 2025 to $50B+ by 2030.

At scale, this creates the possibility of bot ecosystems competing with each other—and with humans—for scarce computational resources. Inference capacity, already constrained, could increasingly be allocated to the highest economic bidder, rather than the highest societal use.

In the DRAM shortage, we are already seeing early signals of this dynamic. Memory prices rose 172% through 2025, driven by a structural reallocation of DRAM manufacturing capacity toward AI infrastructure. TrendForce projects a further near-doubling in Q1 2026 alone. IDC has called this not a cyclical supply issue but “a potentially permanent strategic reallocation of the world’s silicon wafer capacity.” Supply is flowing disproportionately to hyperscalers—where the marginal productivity of memory in AI workloads far exceeds that of consumer PCs. From a market standpoint, this allocation is rational. From a societal standpoint, it raises harder questions about who, or what, compute is ultimately for.

I often return to what seems to me a crucial point: AI agents may become extraordinarily capable at solving technical and economic problems, but they are not equipped to resolve human ones. Markets alone cannot adjudicate the inherent tradeoffs of politics, governance, resource allocation, collective values, or civilizational priorities.

And it is worth asking who benefits from the scarcity itself. When autonomous agents enter markets as buyers of compute, they don’t democratize access — they add another bidder to the queue, further concentrating demand toward hyperscalers and chip manufacturers.

Wen’s Automaton suggests an extreme, deliberately theatrical example of a broadly emerging dynamic. Agents consume resources to sustain their own existence, before they begin to solve human problems. Every cycle of inference spent on work determination is a cycle not spent on the work itself—filing taxes, automating a legal workflow, or running an autonomous truck. Agent compute uncouples metabolic cost from productive output. We must be careful not to build workflows that amount to delivering a burrito with an eighteen wheeler.

Moments like this call for a conversation that extends beyond technologists and builders. Resource allocation decisions that affect consumers, device manufacturers, and entire economies are being made by market dynamics that no one voted for and no civic body is overseeing. Historically, societies have convened in shared forums — from the Athenian agora to modern civic institutions — to debate the future they wish to create. As AI systems begin to participate directly in economic life, that broader discourse becomes not philosophical, but urgent.

Do we really want to compete with autonomous agents for the very resources that power our digital and economic lives?

The question is no longer just what we can build — but how we choose to coexist with what we build, and how we govern the resources that sustain both human and machine intelligence.

Anthropic vs the DOW is what it looks like when AI governance happens by default. We are witnessing the largest organizations work through the stickiest problems without public deliberation, without a legislative framework, and without democratic input. Instead, we have a single company relying on one set of internal principles, expressed in a single contract— working alone through profoundly important norms and consequences.

In real time, in full view of the world, AI’s red lines are being drawn. The question is, by whom? By what process?

The Governance Gap

There is a structural gap in AI governance. We have drifted into a world in which a small number of executives, regulators, and contractors effectively define the operational norms of AI. That may be expedient in the short term. It is not democratic in the long term.

Our capability stack is scaling exponentially. Our civic deliberation stack is not. We have built systems that simultaneously touch speech, economics, and national defense. But the governance apparatus for those systems is a patchwork of corporate policies, executive orders, and procurement disputes.

How We Got Here

It’s important to understand how we got to this point, and what we can do about it. Creating parameters is the necessary step in building systems that reflect the values of a society. I saw this challenge in the gaming industry. For years, many of our hardest questions about representation were treated as cultural or editorial matters. Villains were often given visual shorthand, like a goatee, a scar, a certain accent. These patterns were not typically written as policy. They were habits, conventions, sometimes unconscious biases. But the moment we outsource content generation to AI, what once happened informally becomes statistical. What percentage of villains should have a goatee? Should outputs mirror population demographics? Should they intentionally counter historical bias? Who decides, and according to which data set?

Most of us carry intuitive views about fairness. Very few of us have ever had to translate those intuitions into measurable rules. AI forces us to do exactly that. It automates decisions, and thereby forces every implicit norm into an explicit parameter.

Default Choices

As we see in Anthropic vs DOD, the questions extend far beyond representation. In military applications, we must decide what level of autonomy is permissible—and who bears responsibility when an autonomous system makes a lethal mistake? In economic systems, we must determine whose data trains the model, whose labor it displaces, and whose risk it compounds. These questions used to live in professional judgment and institutional culture. AI flattens them into configuration choices.

Right now, even as the scale of implementation grows at an astonishing pace, those configurations are being set by a very small number of people. My concern is that once decisions become embedded into infrastructure, they cease to be debates. They become defaults.

This is why the question of AI’s red lines is so consequential. When a frontier lab defines what it will or will not build, and when the federal government counters with its own demands, what is at stake is not simply compliance, or the commercial terms of a singular contract. It reveals a vacuum around governance—with economic and political consequences that we have not yet begun to grapple with.

The questions ahead will only grow harder. Who should define the boundaries of systems that influence speech, economics and national defense simultaneously? A private company has the right (and arguably the responsibility) to uphold its product principles. The government has a duty to protect national security and public interest. But when AI becomes foundational infrastructure, these decisions cannot remain confined to executive suites or agency offices.

Public Choices

AI governance requires public participation. Congress must move beyond reactive hearings toward durable institutional design. Companies must articulate principles transparently and invite scrutiny. Investors must treat governance as core infrastructure, not externality.

Democracy does not mean everyone writes code. It means everyone has a voice in setting the boundaries within which that code operates. Citizens must recognize that AI governance is not a niche technical debate. These are not engineering optimizations. Our choices will shape the information we consume, the systems that defend us and the economic opportunities available to us. If we remain complacent about catching up institutionally—if we defer difficult conversations because they are uncomfortable or complex—the resulting instability will become systemic.

The choice before us is not whether red lines will be drawn. It is whether they will be drawn collectively, through open dialogue and legitimate process. Or quietly, by a handful of actors operating at the frontier. That choice still belongs to all of us.

But this should not be interpreted as market fragility. It is the expected byproduct of innovation cycles.

For investors, it is an opportunity. From a venture perspective, periods of compressed incumbency historically correlate with elevated company formation and outsize innovation returns. The down years of 2008 to 2012 were periods of maximum disruption, when companies like Uber, WhatsApp, and Slack launched amid uncertainty—which kept incumbents defensive and valuations accessible. The conditions that unsettle public markets often expand opportunity in private ones.

To understand why, it helps to look at how technological waves have always driven generational turnover—and why the AI wave may distribute that opportunity more broadly than any transition before it, if governance keeps pace.

Generational Turnover Is Structural

Each technological wave rewrites the production function of the economy.

In the early 20th century, industrial scale advantaged manufacturers. By mid-century, the managerial process defined the conglomerate. Software shifted value toward IP and operating systems. The internet rewarded distribution and data. Mobile compressed interface and engagement.

Now, AI is redefining intelligence and decision-making.

Companies built for one paradigm rarely dominate the next—not, at least, without reinvention. Kodak owned consumer imaging but couldn’t survive the shift to digital. BlackBerry defined mobile communication but was displaced when Apple reimagined the phone as a platform. Even IBM, which endured, did so only by abandoning hardware for services—a wholesale reinvention of its business.

The transition from one market leader to the next is not primarily a story of failure, but renewal. New companies emerge, scale, define their era, and eventually yield ground to firms native to the next stack. This generational handoff is a core mechanism through which capitalism reallocates resources toward higher-productivity frontiers.

The Amazon Precedent

The last major platform transition—cloud and internet commerce—concentrated advantage among a small number of hyperscalers. Amazon’s edge was not merely digital storefronts. It was full-stack operational automation: algorithmic inventory management, logistics optimization, data-driven merchandising, and precision marketing. Traditional small and mid-size businesses, particularly those reliant on analog workflows or purely relational commerce, could not compete with machine-augmented efficiency.

Technology widened the capability gap. Scale plus data became compounding moats. The result was an era in which platform access was necessary but platform ownership was what mattered.

AI May Reverse the Gradient

The last platform shift concentrated advantage upward. The AI wave may push it in the opposite direction.

Unlike prior platform shifts that required massive capex to participate, AI capabilities are being productized at surprisingly accessible price points. Tools like decision intelligence, demand forecasting, pricing optimization, and marketing automation are becoming available to any firm with the domain expertise to deploy them. What once required hyperscale infrastructure is now becoming an API-level service.

This new machine-augmented operational capability has a democratizing effect on competitive participation. It changes where alpha accrues. Specialized operators, vertical SaaS platforms, advanced manufacturers, and talent-dense service firms can compete in markets that were previously gated by size. Their machines handle baseline optimization. Their humans concentrate on creativity, trust, and domain expertise. Competitive edge shifts from pure scale to talent leverage per unit of compute.

In this framing, AI does not only produce the next megacaps. It expands the surface area from which generational companies can emerge.

Expect a New Cohort of Generational Builders

If corporate tenure continues compressing while AI lowers capability barriers, the implications are directional and significant: faster formation of category leaders, shorter windows of incumbency, and broader geographic and sectoral origination. The next Fortune 500 cohort may be AI-native from inception, asset-light but compute-dense, talent-concentrated rather than labor-scaled, and built on open and interoperable infrastructure.

Generational companies will arrive quicker than ever before—even as the tenure of individual dominance compresses. For builders and their investors, that is not a warning. It is an invitation.

A Note on Governance

None of this is guaranteed. Technological diffusion alone does not ensure opportunity diffusion. Markets require institutional trust to function productively—fair dealing, transparency, meritocratic access, rule of law. If regulatory capture or self-dealing dominate, technological capability will concentrate outcomes regardless of how accessible the tools become.

If AI scales intelligence, governance must scale judgment. This places the responsibility on stewards of long-duration and public capital, like pension systems, sovereign pools, endowments, and policy institutions.

The key question is not whether generational companies will continue to form. It is where they will emerge, what capabilities will define them, and whether the institutional environment will allow their gains to compound broadly.

Songyee Yoon recently joined Bloomberg Deals, hosted by Dani Burger, to discuss several themes she has been focused on in the wake of the recent software selloff. Below is a summary of the key points she covered.

Why Last-Mile Software Matters Right Now

When discussions turn to attractive targets in software, attention often gravitates toward the latest product category or headline-driven buzzword. Dr. Yoon highlighted a less visible but increasingly critical area: last-mile infrastructure for physical AI.

While robotics and embodied AI receive significant attention, deploying these systems in real enterprise environments remains complex. The true enabler, she noted, is operational software — the orchestration layer that connects devices, manages data flows, and ensures systems function reliably at scale. Though not glamorous, this foundational “plumbing” is likely to capture substantial value in the coming years.

The Vertical Integration Question

As a venture investor in AI-native companies and a board director at HP, Dr. Yoon is frequently asked whether companies should pursue vertical integration.

Her perspective is that companies capable of building solutions tailored to the specific challenges of their industries — while managing data and software integration across their full stack — are better positioned to generate durable impact.

However, she emphasized that adopting new technologies alone is insufficient. Organizations must fundamentally transform how they operate day-to-day in order to become truly AI-native.

New Tools Empower Smaller Companies

Dr. Yoon also expressed optimism about the rise of agent-based software, which is helping level the playing field for small businesses.

In previous technology cycles, only large enterprises could afford full-featured platforms or advanced automation. Today, smaller companies can access tools that allow them to compete more effectively with much larger incumbents.

This shift, she suggested, is overdue and will meaningfully reshape which types of companies are positioned to succeed over the next decade.

Korea’s cultural economy is not a coincidence or a passing wave of soft power; it is the logical outcome of a nation that has spent generations compounding its most enduring asset — human capital. The same foundation that once powered factories now fuels studios, beauty labs, and creative networks. Decades of investment in education, rigor, and mastery have built a population capable of transforming constraint into innovation.

As the world shifts from “heavy” structures to “light” networks, that culture of mastery now meets AI, a technology that amplifies individuals who can learn fast, iterate, and act without permission. AI rewards exactly the qualities Korea has spent decades cultivating in the margins of its corporate order: relentless refinement, rapid iteration, small-team agility, and the discipline to improve without waiting for institutional approval.

Korea’s independent creators—its athletes, artists, beauty founders, and game designers—have been unknowingly training for this moment. They learned to operate lightly because they had no other choice. Now, AI arrives not as disruption but as amplification, magnifying the very skills that were forged outside the reach of the nation’s strictest structures.

How Korea Accidentally Prepared Itself for New Eras

From the ruins of war, Korea rebuilt itself through education and relentless discipline, investing in human capital long before that term became fashionable in economics. The belief that knowledge could replace resources and that learning could replace privilege has been the backbone of the nation’s rise.

Over decades, Korea turned this ethos into an engine of social mobility. Families sacrificed for education, schools became the centre of aspiration, and entire generations grew up believing that effort could transcend circumstance. This devotion to learning produced one of the most skilled, motivated, and adaptive populations in the world — a nation where talent is both cultivated and expected.

These forces made Korea prosperous but also rigid. The miracle of growth was organised through powerful conglomerates — the chaebol — where hierarchy, loyalty, and control defined success. The corporate order rewarded efficiency over originality and security over autonomy. For many, the price of stability was creative suffocation.

And so, ironically, the very system that built Korea’s economic power also pushed its most independent minds to seek freedom elsewhere. Those who could not find oxygen in the corporate hierarchy began to carve their own paths — in sports, entertainment, beauty, gaming, and design. They turned the discipline of the industrial age into the mastery of the individual age.

In sports, particularly women’s golf, where institutional support was thin and few conglomerates sponsored female athletes, individual excellence became the only currency. The rise of Korean champions on the LPGA tour reflects a culture where structure was absent but rigor remained. With no system to carry them, they carried themselves.

In entertainment, artists transformed the nation’s collective work ethic into personal performance art. Behind every seemingly effortless K-pop stage lies years of training, linguistic study, and choreographed precision — not the casual spontaneity of a light culture, but the repurposing of a heavy one.

In beauty, the same human capital found a new commercial language. Chemists, aestheticians, and small founders — many of them women — turned technical expertise and intuition into fast-moving brands. K-beauty flourished precisely because the chaebol overlooked it; what was too small for conglomerates became fertile ground for individual ingenuity.

Across these fields, Korea’s light economy grew not because it was designed that way, but because it had to.

The weight of the corporate order forced creativity to find the cracks, and through those cracks, a new kind of economy emerged — one driven by skill, not scale; by iteration, not inertia. The country’s disciplined human capital, honed inside hierarchy, repurposed itself into mastery once hierarchy loosened. What looks like cultural agility is actually structural endurance transformed.

What this means for AI

This is the foundation AI now meets. Today, a second irony is unfolding.

Because Korea’s talent has already learned to operate lightly — independently, iteratively, and with little institutional support — it is uniquely positioned to benefit from artificial intelligence. AI did not create Korea’s light civilization—but it has arrived to meet it. What was forged in constraint is now poised for scale. The same qualities that let Korean talent survive outside the chaebol—speed, self-direction, relentless iteration—are precisely what AI rewards.

AI automates coordination and amplifies individual capacity, rewarding those who can act without permission. The very skills forged in Korea’s shadow industries — relentless practice, data-driven refinement, small-team execution — are the same skills that thrive in an AI-augmented economy. Where other societies must unlearn dependence on institutions, Korea’s creators, entrepreneurs, and athletes are already fluent in self-directed excellence.

That is why K-culture thrives — not because of fashion, but because of structure. Not because of excess, but because of necessity. Behind every global success in music, beauty, or gaming lies the same architecture: a society that rewards discipline, learns collectively, and refines its craft with astonishing speed.

That is why the next wave of global innovation may come not from the largest Korean corporations, but from the smallest creative networks that grew outside their orbit. AI gives these individuals leverage, turning discipline into scale and imagination into infrastructure. The light civilization that once survived in the margins now stands to define the mainstream.

AI may accelerate this process, but it did not invent it. Korea’s light civilization — agile, skill-driven, and deeply human — was born from the pressure of its own weight.

For investors, the real insight is this: when evaluating where the next industries will emerge, it is not enough to follow capital or technology. It is essential to look into the structure of a society — and the unique pockets of talent that build out of necessity, not out of luxury.

At Principal Venture Partners, we now see companies applying and combining concepts from robotics, optimization, and multiagent systems in their pitches. The growing coverage at AI conferences like NeurIPS, ICML, and AAAI reflects how increased compute, data, and resources are finally enabling breakthroughs in these overlapping fields.

It is a thrilling time to be investing, as new, transformational paradigms come into focus.

I see the next breakthroughs coming not from isolated labs but from systems that learn and cooperate across devices, agents, and modalities. The infrastructure of the future will prioritize efficiency per watt, interoperability across agents, and context depth per dollar. Regulation and trust will follow closely as privacy, safety, and auditability become decisive differentiators in a world where AI systems are embedded everywhere.

Just as the personal computer era shifted value from mainframe operators to software developers, this new era will shift value from centralized compute providers to those enabling intelligence to live in the world—locally, cooperatively, and contextually aware.

After years of racing toward ever-larger foundation models, I am seeing the field pivot toward architectures that are smaller, more adaptive, and more deeply integrated into daily life. Accordingly, the next decade’s value creation will accrue not only to those who build larger foundation models but to those who build the ecosystems around them: energy-efficient chips and model compilers for the edge; coordination and verification layers for swarms; and multimodal data and workflow engines that connect perception to action.

One of our core beliefs at Principal Venture Partners is that the winners of the AI era will be AI-native companies. In other words, the Google of the AI era may not even exist yet.

As consumer sentiment toward AI cools and talk of an AI bubble grows (a topic that I have shared some thoughts on), the coming years will reveal whether AI can become not just a tool of scale, but part of the fabric of everyday life.

I believe that the next phase of AI will be defined by distribution, coordination, and sensory richness.

In particular, I see three key trends coalescing into dominant paradigms: Edge AI, Swarm Intelligence, and Multimodal AI. Together, these will make artificial intelligence not just powerful, but practical and pervasive in our lives. The direction of progress is consistent: intelligence is becoming smaller, faster, and more distributed. They demand our attention, and I explain each below.

Edge AI

Edge AI marks the first great decentralization of intelligence since the birth of the internet. The breakthrough comes from models like liquid neural networks and oscillatory state-space architectures that match transformer-level performance with a fraction of the compute and energy. When intelligence runs locally—on phones, drones, glasses, and appliances—it no longer depends on distant data centers. Computation happens at the point of sensing and action, which cuts latency, lowers cost, and preserves privacy.

This shifts the economics of AI from renting compute to owning capability. Value migrates from centralized providers toward builders of localized, adaptive applications.

Our portfolio company LiquidAI is pioneering this space with its stack. While 300 times smaller than frontier models, its performance is comparable. It works alongside device manufacturers to embed intelligence at the physical level—making machines act and react intelligently.

This means that in the future, a clinician in a rural clinic can diagnose with an on-device AI model. A farmer’s solar-powered edge system can analyze soil data without internet access. A student’s phone can serve as a tutor or translator without sending data to the cloud.

Swarm Intelligence

Even as AI moves outward to the edge, it is also learning to move sideways through Swarm Intelligence.

Independently, with a brain the size of a poppy seed, an ant is not capable of complex planning or action. But in a collective, ants can produce intelligent group behaviors, such as building and farming.

Like ants, neurons, or cells, future AI systems will consist not of a single agent, but coordinated collectives of specialized models that collaborate, compete, and self-organize. Swarms of lightweight models can divide work dynamically—with some reasoning, others retrieving, and others simulating or verifying. They can then share results through standardized protocols. This architecture enables resilience and creativity that single models struggle to achieve.

The frontier lies in coordination: how multiple AIs decide what, why, who, and how to act together. Early research points toward secure model-to-model communication standards like the Model Context Protocol (MCP), which will do for AI cooperation what TCP/IP did for the early web. In enterprise settings, swarms may soon manage research pipelines, customer support, logistics, and even financial operations autonomously, scaling their intelligence like living systems that evolve in real time.

The firms that win this phase will not merely train bigger models—they will master orchestration, identity, and trust across thousands of cooperating AIs. We see clear applications in defense and space exploration, where accuracy and speed is critical. The best approach uses swarm intelligence to increase the odds of success.

Multimodal AI

Finally, Multimodal AI is giving machines a full sensory system. By unifying text, image, video, audio, and code within a single model, multimodality enables comprehension and reasoning that resemble human perception. The leap to million-token contexts allows models to retain hours of video, entire documents, or complex projects in working memory, turning them into powerful assistants for analysis, creation, and control.

Imagine a model that watches a factory line and drafts a quality report, reads a whiteboard and codes the prototype, or listens to a meeting and updates the product roadmap automatically. Multimodal AI is collapsing the boundary between perception and reasoning. The same systems that see can now also infer, explain, and act. This shift will transform many industries—from design and manufacturing to healthcare, education, and entertainment.

In healthcare, high-dimensional data—combining imaging, genomics, labs, and family history—can now be captured and analyzed in new ways. In the future, we may see personal virtual health assistants able to coach people through conditions such as diabetes or hypertension, and “hospitals-at-home” that are able to foresee signs of deterioration before symptoms appear, and suggest intervention as necessary. Already, companies are leveraging models that process non-verbal physiological data to generate insights and address challenges that traditional data science approaches have struggled to solve effectively.

Multimodal AI also opens the door for “physical AI,” meaning systems that perceive the world through multiple sensors and respond in real time—essential for robotics, vehicles, and embodied computing. We see companies using vector field data to generate insights and intelligence never before recorded in human language. And our portfolio company Venti Technologies is closing the gap in the market for using AI in the physical world. Their trucks and logistics solutions are already actively used at the Port of Singapore, one of the busiest ports in the world. In September 2024, they passed the point where running an autonomous truck in a geo-fenced area was both cheaper and safer than employing human drivers.

In 2026, we will see companies integrating data and new ways of combining intelligence and technology fully embracing these ideas.

We cannot wait to meet founders building generational winners.

At Principal Venture Partners, we invest where gaming is placing its bets, because the $50B gaming industry doesn’t make $2B AI investments on hunches. And we are seeing gaming companies quietly placing the biggest AI bets in history.

Cloud computing, real-time analytics, microtransactions—these are all gaming innovations that became business standards.

I’ve been in the gaming business for a long time, and I have seen the same patterns time and time again. Gaming adopts breakthrough tech first, proves it works at scale, then enterprise follows.

Gaming companies are building AI-first infrastructure because they know what’s coming next. When enterprise “discovers” procedural content generation, real-time behavioral analysis, and adaptive user experiences, gaming will already be years ahead.

What will that mean for science?

Scientific AI represents the highest-leverage application of artificial intelligence. And we’re already seeing projects like Google’s AlphaGenome, FutureHouse, and Microsoft’s MatterGen use AI to go beyond efficiency and generate novel insights and hypotheses.

Pharmaceutical companies spend $2.6 billion and 10+ years to bring a single drug to market. So cutting drug discovery time by even 20% saves money. But more than that, AI is capable of unlocking entirely new therapeutic areas that were previously too expensive to explore.

This is why scientific AI is more a productivity tool; it’s infrastructure that determines which diseases get cured and which materials get discovered.

At PVP, we don’t evaluate potential portfolio companies based on whether they can make work slightly more efficient.

We’re interested in teams that are combining deep domain expertise with AI-first approaches—the kind of research-driven innovation that reshapes entire fields of study.

Bostrom’s paramount concern is that “things essential to human flourishing [will be] irreversibly lost.” His ultimate warning rings like a clear bell across the last two decades: “We need to be careful about what we wish for from a superintelligence, because we might get it.”

Do we have a superintelligence now? Not yet, but our current tools would still flabbergast a time traveler from 2003. What had been a thought experiment by Bostrom has become the heart of the rallying cry of “alignment”: the conviction that we must ensure advanced artificial intelligences are tuned to human values so that their power serves rather than destroys us. It is a necessary worry, but also a dangerously narrow one.

The very word “alignment,” singular and definitive, suggests that there exists a stable, agreed-upon set of human priorities that machines can simply absorb. In reality, our values are plural, contested and constantly evolving. Pretending otherwise risks two profound mistakes: chasing an impossible technical fix and overlooking the social and political work that only human beings can do. Human priorities involve negotiation and compromise. Most of the time, there is no clear formula, but the process of communication itself helps us to understand other perspectives and ultimately accept the resulting compromise. It’s not the result that matters—it’s the process that helps us understand and accept it.

Consider an ordinary paperclip factory run entirely by people. Even there, “alignment” is aspirational. Who captures the profits? How are emissions from smelting raw materials counted? What of the climate damage, or the labor conditions in distant mines? Real-world enterprises are riven with conflicting incentives and power asymmetries. If human institutions cannot fully reconcile these competing claims, why would we expect a superintelligent machine to resolve moral disputes that we ourselves have left unsettled?

Recent scholarship is beginning to challenge the old paradigm that says alignment is singular. In 2024, computer scientist Atoosa Kasirzadeh described a two-level model of pluralism: first-order values such as fairness or efficiency, and the higher-order question of who has the authority to decide among them. That second layer—our collective capacity for deliberation—cannot be outsourced to code. It is the foundation of human ethics, which inherently has no singular answers.

The same year, Taylor Sorensen and colleagues warned that today’s alignment techniques risk narrowing the very diversity of values they claim to preserve. They argue that future systems must be designed not to collapse human perspectives into a single “utility function” but to remain open to ongoing contest and revision. Crucially, they write, “as a broader set of people use and rely upon AI systems, we need systems that can understand and cater to a broader set of needs.” They term these “pluralistic” systems, and the point of them is not to “delineate exactly to whom or what to align, but rather to argue for clearer, more pluralistic approaches in alignment.” But again, putting varied approaches to alignment into practice is an inherently deliberative process—less about a single “top goal” than a judicious weighing of competing goals.

Still others have proposed that the solution may be as simple as a “considerate AI”: a system that learns to navigate the agency and welfare of many stakeholders, rather than chasing a single metric of success. The through-line in all this work is clear: no one “human value function” can capture the messy reality of our moral landscape. “Alignment” will never be singular.

This brings us to another danger, what researchers Robert West and Roland Aydin call the alignment paradox. The more successfully we tune models to satisfy one group’s ideals, the easier it may become for others to re-steer those very same systems toward contrary ends. A model exquisitely trained to embody virtue can, in the wrong hands, be exquisitely repurposed for harm. Alignment, in other words, is not a one-time lock; it is a permanent vulnerability. In practice, that means the more “aligned” a system appears, the more vigilant we must be about who controls it.

Getting this right cannot wait. While public debate often dwells on a hypothetical far-future superintelligence, the tangible harms of today’s AI are already with us. Large language models and recommendation engines amplify bias, accelerate the spread of disinformation and concentrate power in the hands of a few companies.

A recent global survey of AI-safety research concluded that “alignment” captures only a slice of the risks we face. Adversarial robustness, interpretability, distributional fairness and system-level oversight are equally critical. To reduce the challenge to alignment alone is to give the public a false sense of security. We should not be comforted by the myth that, if only we encode the right values, the rest will take care of itself.

A more responsible vision begins by acknowledging that deciding which values matter, and how to balance them, is a democratic task. The trade-offs an AI system will embody—between privacy and innovation, between climate mitigation and economic growth—are political questions, not engineering problems. Often, they are inherently moral arguments as well.

Technical progress must therefore be matched by institutional guardrails: audits, impact assessments, regulatory oversight and genuine avenues for public contestation. Systems must be transparent and resilient to manipulation, and the economic benefits of AI must be broadly shared rather than captured by a narrow elite. Alignment without equity and accountability is an empty promise—a superintelligence with no sense of human decency.

Technologists must resist the temptation to see this as a matter of technology. Even the most sophisticated AI cannot settle our prioritization debates for us. It cannot decide how to weigh the liberties of one group against the security of another. Those are human disagreements, and if we cede them to machines—under the soothing banner of “alignment”—we risk not only technological failure but a deeper civic one.

Bostrom’s paperclip fable remains a useful caution. But the true lesson is not merely that we must align machines with humanity. It is that humanity must better align with itself. Superintelligence will not spare us that work. It will only make the need for it more urgent.

I believe that we’re in an AI bubble—but not because of the technology itself. The underlying science has been decades in the making, and only now do we have the right combination of talent, data, compute, and resources to deliver real-world impact. Just as electricity, the internet, and mobile transformed business and daily life, AI is clearly on a similar transformative trajectory.

The bubble comes from human behavior. Pattern recognition drives investors and companies to rush in, eager to capture the upside. Demand to “be in the game” has far outpaced the number of companies with truly transformative technology and talent. That imbalance between hype-driven capital and genuine breakthroughs is what inflates the bubble.

This is why it’s critical to approach AI investment with depth as we do at PVP—grounded in real understanding of technology, business models, and policy—rather than “vibe investing.” Otherwise, we risk amplifying the froth while missing the substance.

A humanoid robot would be able to solve countless problems, such as care work, home automation, and maintenance. But the real world is messy and chaotic, and as a general-purpose machine, a humanoid robot would need to solve for many, if not all, of the edge conditions that exist. That makes them expensive and complex undertakings. Robots built for specific purposes, on the other hand, will always be simpler. Their problem scope is smaller, their developers’ task easier.

At Principal Venture Partners, we see value in both approaches. But we do caution that a humanoid robot should not be seen as the ultimate or only measure of success. Indeed, it is unlikely that today’s research efforts will result in general purpose humanoid robots in the next few years, at least as they are conventionally imagined today. However, they will produce valuable breakthroughs regardless—various form factors with different types of edge intelligence that will work wonders for us.

The real winners will be companies that channel these advances into specialized, economically viable machines. We're betting on focused intelligence over general capability.

Watching labor effects

The displacement effect of robotics on human labor is already real and measurable. MIT research shows that for every robot added per 1,000 workers, wages decline by 0.42% and employment falls by 0.2 percentage points—translating to roughly 400,000 jobs lost to date in the US alone. The impact is even more concentrated geographically: adding one robot to a local area reduces employment in that region by six workers.

Yet, the most successful deployments will likely see robots augmenting human capabilities—taking on dangerous, repetitive, physically demanding or undesirable tasks, freeing humans to focus on higher-value work requiring creativity, emotional intelligence, and complex problem solving.

The challenge for policy makers and business leaders will be to manage this transition thoughtfully, ensuring that the productivity gains from robots translate into shared prosperity, rather than widening inequality.

The challenge for technologists and investors will be to deploy innovative, effective, and durable machines that extend the AI revolution of knowledge-work to the physical world.

This trend reflects a growing bifurcation in venture markets—AI versus everything else. While investment value has soared, total deal volume dropped 27%, indicating that capital is consolidating around a select group of breakout companies while the broader market remains subdued. EY analysts suggest this marks the early phase of a new supercycle, one where AI reshapes not just the tech sector but every corner of the global economy.

Featured in the article is Songyee Yoon, Founder and Managing Partner at Principal Venture Partners, who offered a pointed assessment of this evolving dynamic. As she explains, the wave of capital pouring into AI has outstripped the emergence of truly differentiated innovation. Many investors are chasing opportunities that mimic past successes—driving up valuations without corresponding substance. Yet within that distortion lies opportunity. For investors with discipline and long-term vision, this market presents a rare chance to identify AI-native companies with enduring, asymmetric potential.

In an email interview with Gartner, Songyee Yoon, Founder and Managing Partner of Principal Venture Partners, noted that many AI agent startups are driven more by what’s trendy than what’s truly effective. She emphasized that prioritizing technological novelty over real-world utility is a flawed strategy—and that lasting innovation requires selecting the right tools to achieve meaningful impact.

Read the full piece here:

AI Agent Fallout: More Than 40% of Projects Will Be Canceled in Two Years, Gartner Predicts

For decades, economic prosperity has largely been driven by efficient production and delivery of goods to global consumers at the lowest cost. Even though President Trump has walked back tariffs through a series of exemptions and suspensions, tariffs are more than temporary trade irritants; they’re harbingers of increasing geopolitical tensions that are restructuring supply chains and challenging Silicon Valley’s longstanding tech dominance.

California’s import tax burden is projected to surge from around $17 billion in 2024 to a staggering $170 billion in 2025, with most of those taxes coming from electronics and computers. Rather than Silicon Valley as a single dominant player in AI, multiple centers of excellence will emerge, each offering distinct advantages—whether in geographic access, technology standards, or production costs.

Even if US tariffs are completely rolled back or never come into effect, the market whiplash of recent months signals a new reality for Silicon Valley and startups, one in which trade dynamics can change anytime and supply chains must be adaptive and resilient, even if it increases cost.

Regional strategies are crucial for resilient supply chains

For resilient supply chains, startups need to think regionally—building locally first, designing for multiple regional options to source components, forging strong domestic partnerships early in development, and taking full advantage of government incentives.

Many countries view AI computing capabilities as national security assets. In the US, government contracts increasingly require domestic manufacturing, data hosting, and development. The CHIPS Act and IRA aren’t just subsidies—they’re new economic equations for where certain technologies can be viably produced, especially as advanced robotics and AI-driven manufacturing offset higher labor costs.

In Asia, national governments and startups alike are already adopting a “China Plus One” strategy that maintains a Chinese base of operations while establishing additional regional hubs. In response to global supply chain disruptions and rising geopolitical tensions, Japan has already reduced dependency on Chinese manufacturing. From May 2020 to March 2022, Japan's Ministry of Economy, Trade and Industry (METI) subsidized 439 onshoring projects and 104 near-shoring projects, encouraging Japanese firms to relocate production to Southeast Asian countries, including Vietnam. These initiatives aimed to strengthen supply chain resilience and mitigate risks associated with overreliance on China.

For Asian manufacturing centers, like Vietnam, tariffs present both head and tail winds. While benefiting from Japan's near-shoring efforts, Vietnam has also faced a 46% US-imposed tariff, given US concerns about shipment of Chinese goods through Vietnam to circumvent existing tariffs.

What remains unchanged, ironically, is the power of market forces. A supply chain isn’t simply something that can be moved from one place to another—it’s a complex ecosystem of people and countless suppliers that has been established over decades to produce what we need. The beauty lies in how each participant in this ecosystem makes rational economic choices, and these individual decisions combine to create outcomes that no broad policy could predict. This dynamic process can spark new markets, products, and businesses. The only certainty, indeed, is uncertainty, and the need for resilient supply chains that can adapt to new global dynamics, policies, and market forces.

Adaptability is the ultimate moat

In the years following World War II, Taiichi Ohno, a young engineer and businessman at Toyota, upended the automotive industry with a new manufacturing methodology, known as the Toyota Production System. This new system focused on lean and just-in-time manufacturing, reshaped the global standards for efficiency and precision, and made Toyota a new market leader. Toyota’s success—repeatedly studied by management schools and chronicled by the media—demonstrated that operational agility can be just as powerful as engineering excellence when it comes to transforming markets and shifting competitive dynamics.

Just as the Toyota Production System transformed the automotive industry and its leaders, those who can best engineer for uncertainty will be the winners of the AI generation. It will take more time and capital for AI startups to scale in a more fragmented global economy, but the most successful AI companies won’t just adapt to this new reality—they’ll exploit it as a competitive advantage.

The more complex the geopolitical landscape, the more adaptability becomes the ultimate moat, and understanding the intersection of innovation, policy, and regional dynamics becomes as crucial as evaluating technical differentiation.

Years ago, the CEO of a company I worked at gifted every employee an iPhone to prepare for the mobile era. But the security team restricted the phones from accessing work email—until the CEO personally intervened. That story stuck with me. Even the most forward-looking technologies can't override institutional inertia overnight.
Today, we see the same tension at scale—with AI.
AI is no longer just a tool. It now writes exam questions, passes interviews, and is reshaping the classroom. But while the technology evolves rapidly, the systems surrounding it remain rooted in outdated norms.

Exams written by AI—but can’t be solved with AI?
The California Bar now uses AI to generate test questions—an efficient, inclusive step forward. Yet candidates can’t use AI to answer them. Why trust AI to pose the question, but not assist in solving it?
The contradiction grows sharper in hiring. There have even been cases where students used AI tools in high-stakes interviews—prompting disciplinary responses from their institutions. While the ethics are up for debate, the situation raises a broader question: how different is that from using references in an open-book test?
Can we really expect future lawyers and developers to work without AI—forever? That expectation alone reveals how deeply rooted our systems are in outdated assumptions.

AI in the classroom—still trapped in a 19th-century model?
The U.S. is preparing to integrate AI into K-12 education. That’s progress—but still within a rigid, industrial-era framework. Can students raised alongside AI thrive under systems built before the internet existed?
We need to teach students how to think and create with AI. But our methods remain stuck in the past.

At PVP, we invest in those who challenge inertia.
We back founders who don’t just build new tools—they question the underlying assumptions of outdated systems:

• Liquid AI: Challenges the assumption that deep learning must be rigid and fixed, designing adaptable models inspired by how the brain actually works.

• Aalo Atomics: Confronts the longstanding fear around nuclear energy by proving it can be safe, scalable, and central to sustainable digital infrastructure.

• Laurel: Pushes back on the belief that legal access must be expensive and exclusive, using AI to democratize legal services.

• TaxGPT: Questions the idea that tax systems must remain opaque and reserved for experts, turning complexity into intuitive, conversational flows.

• Operative Game: Disrupts the notion that games are merely entertainment, reframing them as platforms for cognitive growth and emotional learning powered by AI.

We’re proudly industry-agnostic. Because “industry” may be a relic. Real innovation happens when silos collapse and systems converge.
Sometimes I joke that PVP stands for Pastbreakers Via Possibility. But really, that’s what we do.
To build the future, we must challenge the past.

At Principal Venture Partners, we believe that the most transformative AI applications will combine scalability with genuine human impact. That's why we are excited by the breakthrough results of the recent Dartmouth clinical trial of Therabot, the first generative AI therapy chatbot to undergo rigorous testing.

The trial showed meaningful reductions in symptoms of depression (51%), anxiety (31%), and eating disorders (19%)—on par with traditional outpatient therapy. Participants developed trust and consistently engaged with the chatbot, often initiating conversations during moments of distress, demonstrating both clinical efficacy and strong user engagement.

What's particularly compelling is that these improvements match traditional outpatient therapy outcomes, while users reported building trust with the AI similar to human therapist relationships.

This is exactly the kind of AI-native innovation we get excited about—one that addresses a massive real-world problem with scale, empathy, and scientific rigor.

The U.S. mental health system is stretched thin, with each provider responsible for up to 1,600 patients. We believe AI has the potential to transform healthcare by expanding access, offering 24/7 support, and personalizing care at scale—meaningfully augmenting, rather than replacing, the human element.

As early-stage investors in AI, we see Therabot's success as a major step toward democratizing healthcare and delivering real impact where it’s most needed. Therabot's success reinforces our conviction that the future belongs to solutions that scale care, not just technology.

Read the full Dartmouth article