Month: June 2026

At Code Class 2026, Zhang Xun, Managing Director of Source Code Capital, delivered a talk titled “Accelerating Exponential Growth”. In his view, AI now sits on an exponential curve that is about to turn nearly vertical, while the human brain, shaped by two million years of evolution, is hardwired to underestimate the explosive effects of compounding. At the same time, value is concentrating in an ever-smaller number of companies as the power-law distribution becomes increasingly extreme.

Zhang moved from the persistence of Moore’s Law to a recurring industry pattern: general methods combined with more compute repeatedly outperform elaborate human engineering. He then turned to a power-law world in which fewer than 1% of companies account for net value creation, before returning to what compounding and long-termism really mean. The question he ultimately sought to answer was not whether AI is coming, but what capability is truly scarce in an era whose potential is being consistently underestimated. His answer: seeing the pattern earlier than others, and having the patience to back and stay with that tiny handful of exceptional companies over the long term.

Key Takeaways:

• At any given point in time, it can look as though growth has reached its limit. But zoom out, and the exponential curve has never stopped, with each successive stretch being steeper than the last.
• Two million years of evolution have made the human brain highly sensitive to relative change but poor at intuiting absolute magnitude, so we tend to underestimate exponential growth precisely when it matters most.
• More general methods, combined with greater compute, have repeatedly and decisively outperformed elaborate human engineering.
• This is a world with fewer and fewer winners—perhaps less than 0.5% of companies account for net wealth creation.
• Until a pattern is falsified, act as though it holds—remain skeptical, but place your bet.
• Compounding follows an exponential curve too, and like Moore’s Law, it is easy to underestimate in its early stages.
• The AI era may have given us the longest hill and the wettest snow yet. The only question is whether we have the stamina to keep the snowball rolling.

The following is an edited and abridged transcript of Zhang Xun’s talk:

The title of today’s talk is a bit of a mouthful—“Accelerating Exponential Growth”. But what I want to share comes down to something very simple, yet deeply counterintuitive: we may be approaching the point where an exponential curve begins to look almost vertical, while our brains are almost hardwired to underestimate it.

Beyond Exponential: The Underestimated AI Era

Take a look at this chart. It’s a visualization of Moore’s Law that I photographed at the Computer History Museum in Silicon Valley. The x-axis represents time, and the y-axis indicates the number of transistors on a single chip—a reasonable proxy for the computing power of that chip.

Let’s walk through a few milestones. In 1971, a single chip contained roughly 2,300 transistors. By 1994—23 years later—that figure had increased about 900-fold, equivalent to a compound annual growth rate of roughly 35%. If you had been a seasoned investor or semiconductor veteran in 1994, you would almost certainly have concluded that the previous 23 years had already been an extraordinary period of growth for semiconductors, especially leading-edge process technology, and that it would be difficult for the industry to accelerate much further.

But here’s what actually happened next. Over the 16 years from 1994 to 2010, transistor counts on a single chip ballooned to roughly 2 billion—about the level of Intel’s flagship processors at the time, when Intel was the world’s largest semiconductor company. Rather than slowing down, the compound annual growth rate actually accelerated to roughly 54%.

What is most striking is the chart itself. For visual effect, its creator placed the 2010 data point near the upper-right corner, almost as if growth had reached the limit of what the chart could contain. But extend the same chart to 2026, through the transition from serial to parallel and then accelerated computing, and the picture changes dramatically. The 2010 point that once looked almost “off the chart” ends up near the bottom of the expanded curve. From 2010 to 2026, the industry achieved another 168-fold increase, equivalent to roughly 37.75% annualized growth.

This is what makes exponential growth so counterintuitive: at any given point in time, it can look as though we’ve reached the limit. But extend the horizon, and the exponential trend continues, with the latter part of the curve often steeper than the previous part.

Anthropic CEO Dario Amodei has described AI as a tsunami, arguing that the public has very little intuition for how close we may be to “the end of the exponential”. By that, he does not mean the exponential curve literally ends. He means the point where a long, seemingly gradual phase gives way to an extremely steep rise. We may now be close to that inflection point, while most people still have very little sense of how near it is.

Why do people systematically underestimate this? The answer lies in how the human brain evolved.

Over the past two million years, our ancestors looked up and saw the sun, and looked around at prey such as pigs, cattle, and sheep. Our brains evolved to handle differences on the scale of a few units or perhaps a dozen–more than enough for hunting and gathering. Later, the German scientists Weber and Fechner formulated what became known as the Weber–Fechner law, and Daniel Kahneman extended this line of thinking in his own work, and was awarded the Nobel Prize in Economics. The basic idea is that the human brain is far more sensitive to changes in relative ratios than to absolute changes.

A change from 100 to 200 is immediately recognized as a “doubling”, which feels dramatic. However, a change from 10,100 to 10,200 is perceived as “only 1%” and barely feels significant, even though the absolute increase is the same.

This perceptual system helped our ancestors survive in hunting and agrarian societies, but today it can mislead us. Technology does not progress linearly; it compounds, especially in the AI era. Once compounding reaches a later stage, the incremental growth added in each successive period can quickly exceed everything accumulated before it. Our intuition is simply not built to read that curve, so we repeatedly underestimate exponential growth precisely when we should take it most seriously.

The same pattern of underestimation is visible at the company level. Plot today’s largest companies by market capitalization, with years since founding on the x-axis and revenue or ARR on the y-axis, and a clear generational pattern emerges. At the same stage of maturity, the growth curves of the mobile internet generation, represented by Google and Meta, are markedly steeper than those of the PC and software generation, represented by Apple and Microsoft. Each successive generation has a steeper slope.

The AI generation has taken another step up in speed. Anthropic’s last publicly disclosed ARR was around $4.5–4.6 billion, and by May or June it could reach $5–6 billion. OpenAI’s current ARR is reportedly around $3–4 billion. And these curves are still accelerating, with both companies moving toward tens of billions of dollars in annualized revenue, potentially even approaching $100 billion, at a pace that is difficult to comprehend. If we use them as benchmarks for the AI era, the second derivative—“super-exponential” acceleration—appears even more extreme than in the mobile internet era. The speed at which this generation of companies goes from zero to scale would have been almost unimaginable in the previous generation.

Behind this acceleration, the evolution of the AI industry reveals another recurring pattern. Let me connect the dots through a series of milestones.

In 1997, Deep Blue used massive computing power to overpower generations of accumulated chess knowledge and carefully developed heuristics to defeat the best human chess player. The contrast between AlphaGo and AlphaZero is even more revealing. AlphaGo still learned from human game records; AlphaZero started from scratch, relied on no human game data, and learned entirely through self-play—yet ultimately became even stronger.

The same is true for machine translation. Earlier systems relied heavily on linguists parsing sentence structures and building phrase tables, and they often produced awkward or even absurd results. After the Transformer was introduced, neural networks that were not explicitly programmed with grammatical rules dramatically outperformed earlier linguistics-driven approaches. From ImageNet and AlexNet to speech recognition and then AlphaFold—which made a breakthrough on the protein-folding problem that had challenged biology for decades—these developments share a common underlying pattern.

Kai Yu, the founder of Horizon Robotics, has made a similar point from another angle. He argues that the point of autonomous driving is not to imitate human drivers because 99% of human drivers are not worth imitating. They corner too aggressively and brake too hard. The goal is to discard human driving habits and move toward the genuinely most efficient way of driving. At a fundamental level, that is exactly what AlphaZero and Tesla FSD are doing.

This entire sequence points to the same conclusion: at least in AI, more general methods combined with greater compute have repeatedly and decisively outperformed elaborate human engineering.

This idea may come as a shock to people working inside the industry. I remember an interview with Yang Zhilin, the founder of Moonshot AI, in which he was asked what he had gained most from his time at Google. He answered that as a CEO, especially in the AI era, you have to learn to step back from individual details and resist the temptation to over-engineer locally. This era rewards bets on general methods and scale, not endless refinement at the margins.

The Power Law: Fewer and Fewer Winners in the AI Era

If the previous section was about how fast growth happens, the next question is who ultimately captures that growth, and the answer is just as counterintuitive.

Value creation follows a power-law distribution, and that distribution is becoming increasingly concentrated.

Over the past 30 years, the top 1% of companies in the U.S. market have accounted for roughly 80% of public-market value creation; in China, the figure is about 76%. There is an even more striking way to visualize this. Put the number of companies on the x-axis and the total wealth created by listed companies worldwide on the y-axis, using U.S. Treasury returns as the benchmark. According to empirical research by Bessembinder and others, listed companies around the world created roughly $45 trillion in wealth between 1990 and 2018. But break that down and the distribution is startling: 60.9% of companies failed to outperform U.S. Treasuries and effectively destroyed value relative to that benchmark; 37.8% created just enough wealth to offset those losses; and only about 1.3% of companies accounted for virtually all net wealth creation.

In other words, virtually all net wealth creation came from an extraordinarily thin tail of companies. That was already true in the mobile-internet era.

In the AI era, this tail may become even thinner. This next part is a hypothesis rather than an established fact. If we take 2023—the first full year of the ChatGPT era—as the starting point, and the world goes on to create another $200 trillion, $300 trillion, or even $500 trillion in wealth, the blue distribution curve could become an even steeper red one: perhaps 80% of companies fail to outperform Treasuries and destroy value relative to that benchmark, another 19.5% merely offset those losses, and fewer than 0.5%, perhaps even 0.3%, account for almost all net wealth creation.

There is already some evidence of this playing out in front of us. At the close of U.S. trading on May 29, the evening before this talk, all ten of the largest companies by market capitalization in the U.S. were technology companies. In addition to the familiar Magnificent Seven, the list included TSMC, Micron, and Broadcom. Look back at snapshots from 1880, 1920, 1940, or 2000, and it is hard to find another moment when the world’s ten largest companies were so uniformly concentrated in one sector.

Put simply, this is a world with fewer and fewer winners: a tiny handful of companies account for the vast majority of incremental wealth creation.

Of course, no matter how powerful a pattern appears, we have to be honest about its limits. In The Logic of Scientific Discovery, Karl Popper argued that scientific theories cannot ultimately be proven; they can only be falsified. Moore’s Law is fundamentally an empirical pattern fitted to historical experience, not a scientific axiom. Newtonian mechanics works within certain bounds, but ceases to apply at sufficiently small scales. Likewise, the power law, this apparent pattern of super-exponential growth, is likely subject to the same limitation: it cannot be proven once and for all, only falsified.

That is precisely what makes this interesting: until the pattern is falsified, acting as though it holds may be the more rational way to make decisions.

As Dario has put it, we believe that scaling laws will continue to hold. We question them every time, yet every time training scale increases by another order of magnitude, they continue to work. So, we “bet on scaling” with a healthy respect for uncertainty: remain skeptical, but keep placing the bet. This may be a better mental model for navigating a power-law era.

Long-Termism: Investing Through AI’s Steep Curve

Put these forces together and the implication for long-term investors is clear: technology is accelerating at a super-exponential pace that human intuition tends to underestimate, while value is becoming increasingly concentrated in a very small number of companies.

What can be deduced from these facts? This means that the truly scarce capability is identifying those few companies on the thin tail early, while most people are still underestimating them, and having the patience to stay with them for the long term as they climb that steep curve.

The power law tells a cold, unforgiving story: most companies destroy value. But viewed from another angle, that concentration is also where the greatest opportunity lies. Precisely because so much value accrues to so few companies, the potential returns from identifying the right winners early and staying with them for the long term may be greater than in any previous era. The steeper the curve and the fewer the winners, the greater the potential reward for those who get in early and have the patience to stay.

Time itself is another severely underestimated force, and Warren Buffett is perhaps the clearest illustration of this. There is a widely circulated set of figures: 99.6% of Buffett’s wealth was created after age 52, 99.976% after age 40, and 99.9993% after age 30. In other words, almost all of his wealth came later in life, not because he suddenly became smarter, but because compounding takes time. A snowball needs a sufficiently long hill before it can grow to extraordinary size. Compounding is itself an exponential process, and like Moore’s Law, it is easy for our intuition to underestimate it in the early stages.

This is exactly what Source Code Capital has been doing for the past twelve years. Since our founding in the spring of 2014, we have focused on the same basic task: identifying exceptional companies early and staying with them as they grow over the long term. The AI era has not changed the nature of that work; it has simply increased both the potential payoff and the difficulty. The curve is steeper, so seeing it early matters more. There are fewer winners, so the cost of being wrong is higher. And the cycle is longer, so patience is more valuable than ever.

In an era like this, we increasingly come back to a few simple principles. Respect exponential growth; do not use linear intuition to measure an accelerating world. Bet on general methods and long-term trends rather than becoming absorbed in local optimization. Accept the reality of power laws and focus your attention on the very small number of companies that truly matter, rather than spreading your effort evenly across mediocre opportunities. And always remain humble in the face of any apparent pattern, because even the strongest trend may eventually be falsified.

In the end, this is a discipline of patience. Buffett has a famous metaphor: investing is like rolling a small snowball down a very, very long hill. The secret is to find a hill long enough and snow wet enough.

The AI era may have provided us with the longest hill and the wettest snow yet. The only question is whether we have the patience to keep the snowball rolling all the way down the hill.

At Code Class 2026, Han Guang, Managing Director at Source Code Capital, who leads the firm’s growth-stage and AI-focused investments, delivered a talk titled “Standing at 1769 of the Intelligence Revolution”. In his view, this wave of AI is not an extension of the previous revolution in communications, but another revolution in productivity—one in which “intelligence itself” is being industrialized. Scaling laws continue to hold; “one year in AI is like three years in the human world.” Given that 68% of the global population of 8.3 billion has yet to use AI, we are not at the peak of this wave—we are in its earliest days.

From Silicon Valley’s “three months is too long” mindset and “wish-driven work”, to new frameworks like “tokens are not equal” and “to-human versus to-agent”, and finally to the organizational questions facing CEOs and Source Code Capital’s own investment thinking, Han Guang’s point is not to ask whether AI is coming, but to recognize that it is already here. The flagship of this new era set sail three years ago. For entrepreneurs, the choice is whether to stay on the old continent and watch it sail away, or jump aboard.

Key Takeaways:

• Believers and skeptics of scaling laws have made very different choices over the past few years—and seen very different returns.
• The CEOs of frontier labs aren’t hyping things up; they’re presenting facts.
• The Industrial Revolution commoditized mechanical power; this time, we may be industrializing intelligence itself.
• We must build a new world designed for intelligence.
• Not all tokens are created equal—some are as valuable as diamonds, while others are worth little more than water.
• We used to categorize companies as B2C or B2B. Now that boundary is blurring: are you building for humans or for agents?
• The flagship of this new era set sail three years ago. Do we stay on the old continent and watch it sail away, or jump aboard?

The following is an edited and abridged transcript of Han Guang’s talk:

The theme of my talk today is “Standing at 1769 of the Intelligence Revolution”. I’d like to share a few thoughts with you. AI is not a continuation of the previous communications revolution; it represents a fundamentally different kind of productivity revolution. For the first time since the invention of the steam engine in 1769, humanity is once again thoroughly industrializing a primary factor of production. The Industrial Revolution industrialized mechanical power; this time, we are industrializing intelligence itself.

Over the past six months, chatter about AI “hitting a wall”, “being in a bubble”, or “needing a cool-down” has resurfaced regularly. But my view is that scaling laws continue to hold. One year of AI progress is equivalent to three years of human development. We are not at the peak of this wave—we are in its earliest days. So, what I want to discuss today is not whether AI is coming. It’s already here. For entrepreneurs, the question is: do we stand on the shore and watch the ship sail away, or do we jump aboard?

A World with Scaling Laws Is a Very Different World

Let’s start with the “g-force” we’ve all felt over the past six months.

I imagine many of you have felt the same way I have over the past six months: as if we’re trapped on a bullet train accelerating wildly, pinned to our seats and holding our breath. All we can do is hold on tight to the handrails so we don’t get thrown off.

Looking back over the past few years, if you asked me what mattered most, I would undoubtedly say scaling laws.

A world in which scaling laws hold and one in which they don’t are two entirely different worlds. Believers and skeptics of scaling laws made completely different decisions and saw very different returns.

While scaling laws have a formal academic definition, there’s also a simpler way to understand them: an AI model’s intelligence scales predictably with the logarithm of the compute and data resources invested in training and running it. There are two key points: first, the resources being deployed grow exponentially over time; second, the gain in model intelligence is strictly predictable. This “predictability” is paramount. When you listen to what Sam Altman or Dario Amodei have been saying lately, it essentially boils down to one phrase: “We could already see this coming several years ago.”

Standing here today, a natural question arises: Have scaling laws hit a wall? How can we tell? The best way, of course, would be to be inside a frontier lab. But for those of us on the outside, there are still a few ways to judge.

First, listen directly to those who see the reality firsthand—people who are “one degree of separation from the truth”. Information degrades as it is disseminated; by the time you’re two degrees removed, the signal gets drowned out by noise. The individuals closest to ground truth are the CEOs and leading researchers at frontier labs. Over the past six months, they have repeatedly stated that scaling laws continue to hold in various forms of communication, including essays, blog posts, interviews, and talks. In January, Dario wrote “The Adolescence of Technology”, arguing that every few months, public sentiment either becomes convinced that AI is “hitting a wall” or becomes excited about a new breakthrough while AI’s capabilities continue to improve steadily and resolutely. In an interview last month, he reiterated that current progress was broadly in line with what he had expected as early as 2017.

Second, look at the effective compute going into frontier models. If scaling laws had hit a wall, that compute would no longer be growing exponentially. If it is still expanding exponentially, that suggests the people closest to the frontier still don’t see a wall ahead. There’s a piece of work called “Situational Awareness” that plotted a chart in 2024 tracking the effective compute invested in frontier models, with time on the horizontal axis and effective compute on a logarithmic vertical axis, each tick representing an order of magnitude (OOM). The chart extended through GPT-4 and made a number of predictions that looked crazy at the time and that almost no one believed. Two years later, these predictions have proved broadly accurate, which is striking. With the help of AI, we updated the chart and were surprised to find that the trend line was still climbing linearly on the log scale. After the release of OpenAI o1, there was even a slight inflection toward faster growth. There is a good chance that we are still firmly on the scaling-law trajectory.

There’s another interesting observation: AI progress is proceeding at roughly three times the pace of a child’s cognitive development—one year of AI progress is roughly equivalent to three years of human intellectual development. In 2019, GPT-2 was like a preschooler. In 2020, GPT-3 was like an elementary schooler. By 2023, we had a smart high schooler; and by the second half of 2025, it suddenly seemed more like a PhD student. The ratio is roughly 1:3. Extrapolate that trajectory forward. What kind of intelligence might we have by 2030? Perhaps we only began to grasp how formidable AI had become by the end of last year. This was not because its progress suddenly accelerated—it had been advancing steadily all along—but because we finally crossed our perceptual threshold.

A year or two ago, I thought Dario, Sam Altman, and Demis Hassabis were exaggerating because the world they described sounded like science fiction. Looking at it now, however, I realize that they weren’t hyping it at all. They were describing facts.

Dario has said that “powerful AI”, smarter than a Nobel laureate across most relevant fields, could emerge within one or two years, and is highly likely to arrive within the next few years. He has spoken of “a country of geniuses in a data center”, predicted that half the world’s population will be talking about AI by the end of this year, and recommended The Making of the Atomic Bomb because he believes advanced AI could prove comparable in historical significance. Sam Altman has said that we may be only a couple of years away from an early form of true superintelligence, by which he means an AI capable of serving as the CEO of a large company or conducting research beyond the level of the world’s best scientists. Demis is the most conservative of the three. In January, he estimated a 50% chance that AGI would arrive by the end of the decade, but his bar is much higher: train an AI on everything we knew by 1911 and see if it can independently derive general relativity. Even a 50% chance of that is both alarming and exciting.

Are we already witnessing the first glimmers of an AI scientist? A few days ago, a general-purpose reasoning model from OpenAI, with no specialized training, independently produced a proof concerning the unit-distance conjecture first posed in 1946. The result was reviewed and validated by four mathematicians. The proof itself is not exceptionally difficult; it is constructive in nature, but it offers an early glimmer of the AI scientist. It suggests that the world they have been describing may indeed be possible.

The 1769 Revolution Industrialized Mechanical Power. This Time, It’s Intelligence Itself

1769 was the year James Watt improved the steam engine, marking a turning point in human history. Before then, nearly all the mechanical power behind the goods around us came from human and animal labor, supplemented by a small amount of wind and water power. After that, machines began to supply most of it—first with steam engines, then with internal combustion engines, and today, increasingly, with electric motors. Over the course of two centuries, the percentage of mechanical power supplied by humans and animals decreased from 98% to less than 1%.

So, what if we are facing a computing revolution this time?

Today, most cognitive activity is still performed by the human brain, supplemented marginally by mechanical and electronic computation (PCs, smartphones) and now, neural networks. If this really is a computing revolution, could the human brain’s share of general cognitive work shrink dramatically over time? What will supply tomorrow’s general intelligence? Perhaps neural networks, some other architecture, or something beyond our current imagination. But the direction is unmistakable.

What is remarkable is that the Industrial Revolution commoditized and industrialized mechanical power; this time, we may be doing the same to general human cognition, or intelligence itself.

This may be the fastest wave in history. It took Apple and Microsoft about two decades to reach $10–20 billion in revenue. The mobile internet moved much faster, with leading companies reaching $100–200 billion over a similar time span. Yet the two breakout AI companies have taken only three or four years to reach annualized revenue in the tens of billions, with a path toward $100 billion. This pace is unprecedented in business history and far exceeds that of earlier technological generations.

So how large is this market, really?

One heuristic is labor substitution. Global GDP stands at roughly $110 trillion, with around $60 trillion allocated to wages. Suppose that AI could replace 50% to 80% of that labor and that machine replacement is priced at a discount to human labor. For simplicity, let’s use $10 trillion as a round number. How much of that opportunity have we captured so far? Only about $80 billion—a tiny fraction.

But is substitution really the right way to size the market?

Back in 1700, before the Industrial Revolution, total global mechanical power from humans and animals was just 10 gigawatts. Two hundred years later, that surged 50-fold to 500 gigawatts. When the power loom was invented, the market it created was far larger than the hand-weaving activity it displaced; it dramatically expanded the market itself. Any forecast bounded by “substitution” is likely to be constrained by our lack of imagination.

And we are still in the earliest days. Across the global population of 8.3 billion, 68% have never touched AI. Paying users make up only 1.2%, and coding users just 0.3%. Yet this tiny fraction drives staggering consumption: a developer consumes 300 times more tokens than a free user, and 10 times more than a standard paying user, and that trajectory is still climbing. Over the past two years, as we shifted from chatbots to AI agents, token consumption per task exploded by 1,000 times.

We need to build a new world for intelligence, structured in two layers.

The first layer is enabling intelligence generation, which immediately hits physical walls in energy and hardware. If you chart the AI industry’s compute demand as Dario describes it, you find that in three to four years, meeting all of that projected demand could require power equivalent to roughly half of average U.S. electricity generation. This is a systemic physical-infrastructure bottleneck across power, chips, storage, interconnects, long-distance networking, power electronics, liquid cooling, and even construction capacity, all facing full-stack shortages. Fundamentally, growth in compute and scaling in intelligence demands exponential investment, but physical infrastructure can only expand linearly.

The second layer is building a native ecosystem for agents. Today, virtually all of the infrastructure of the digital world—identity systems, payment rails, security checks, authentication protocols—was designed for humans. Some of that infrastructure already exists for agents; much of it does not. In the future, agents may have their own communication protocols, authentication systems, and payment tools, and perhaps one day even their own economic systems and markets. It will be an entirely new and exciting world, though perhaps a bit unsettling as well.

We’re All Novices: Did You Spend Time at the Keyboard Yourself Today?

In March, we spent some time on the ground in Silicon Valley. Being there in person gave us a completely different sense of what was happening, and of the magnitude of the change. I’d like to share three key phrases that left a lasting impact:

First, “Three months is too long.” I asked a frontier researcher what to expect by year-end. He retorted: “Year-end is nine months away. That’s far too long; so much could happen. How could I possibly predict that?” Inside frontier labs, people no longer spend much time discussing what might happen four months out, because time itself feels compressed. The very time horizons we rely on today must be challenged.

Second, “Make wishes.” I asked another researcher, “You’re as close to AI as anyone. How do you direct its work?” He said, “Direct it? I’m learning from it. Every morning I come to work with three wishes: ‘Dear AI, please grant these three wishes today.’ And by the end of the day, it delivers.” “Making wishes” has become one of my favorite phrases lately.

Third, “Results in two hours.” One researcher noted a shift in how his manager assigns tasks: “This is very promising—I want results in two hours.” Not “when you finish”, and certainly not tomorrow. Two hours. So, what should be the new baseline for the speed at which an organization moves and iterates?

So how do we become super-individuals in the AI era? We came away with four words for ourselves.

First, dive in—Don’t stand idly on the shore—jump into the water.

Second, unlearn—Let go of old standards, old metrics, and familiar tools and processes. Start again from facts and first principles.

Third, be open—In this era, we’re all beginners. No one has an inherent head start, and mindset shapes action.

Fourth, enjoy—This is an age of abundance, full of new things to explore and play with. The whole world feels like one enormous playground. We should enjoy it.

I often ask myself: Is AI always on? We’re all managers, spending our days managing people and organizations. But have you spent time at the keyboard yourself today? When you run into a problem, is your first instinct to ask AI, or to ask another person? Do you reach for the tool you already know, or first see if a new tool can solve the problem better? There is a downside, of course. If you become obsessed with coding or with trying every new tool, you may end up neglecting the things that matter most—and exhausting yourself in the process. Are you falling into a counterproductive loop?

We need to prepare for a new world in which humans and agents work side by side. The duration of tasks agents can complete autonomously is doubling roughly every four months. Two years ago, in the era of chatbots and tab completion, the model was essentially one human working with one AI assistant. Today, some people already work with ten agents at once. Before long, perhaps one of those ten will be “promoted” to “CEO” and manage the others for you. You could then be running an agent-native company, with one person overseeing fifty agents. Eventually, we may not need to manage individual agents at all. They could operate autonomously like a “lights-out factory”—a workforce running inside a data center. You give it an objective, and it gives you the result. In 1937, Ronald Coase wrote “The Nature of the Firm”. In the agent era, might we see the emergence of a new “nature of the firm”? I’m looking forward to finding out.

So, I’d ask the CEOs in the room to think about a few questions. If, three years from now, every employee is working with five to ten agents, does your current organizational structure still make sense? Which roles in your company are fundamentally about shuttling information around and moving workflows forward? What kind of senior person will you hire next? If AI takes over most junior-level work, how will you develop the next generation of senior talent? Of your ten most important workflows, which can AI already handle end to end, or close to it, and where is the remaining bottleneck? And if the cost and difficulty of execution fall dramatically, what becomes your moat?

Not All Tokens Are Created Equal: From Monetizing Attention to Commodifying Intelligence

Times are changing, and so is investing. Jensen Huang has described AI as a five-layer cake: energy, chips, infrastructure, models, and applications. Startups founded over the past three years rushed to fill out every layer of that stack, yet the total cake remains vast. This is a multi-trillion-dollar frontier, and we should be competing for our piece.

At the same time, we are convinced this represents a fundamentally distinct kind of revolution. Most of us in this room, including those of my generation, have never really lived through a productivity revolution. What we have experienced has largely been a revolution in communications. Communications accelerated information distribution. The AI revolution crafts entirely new “intelligence products”, transforming intelligence from something scarce, precious, and artisanal into something industrialized. We need a different lens to understand it.

In the communications era, we monetized attention. Whether I’m a billionaire or an ordinary knowledge worker, an hour of our attention is broadly comparable—“attention is more or less equal.” Business models were built around advertising and transaction fees, with the goal of keeping users on the platform for as long as possible. But today, “tokens are not equal”. Some tokens carry enormous value, like diamonds, while others are highly commoditized and worth little more than water. Some generate high margins; some create strong retention; others are purely transactional.

That probably means our evaluation metrics need to change as well. In the communications era, we focused on DAU, time spent, and cohort retention. Today, we track token volume. Tomorrow, will we track “token value density”—or metrics we haven’t even conceived yet? I think that shift is inevitable.

We used to classify companies as B2C or B2B, but that boundary is starting to blur. In both cases, a human may be the one using the product—so where exactly does B2B end and B2C begin? Increasingly, people are asking a different question: are you a “to-human” company or a “to-agent” company?

There’s another difficult question we don’t yet have an answer to: if model companies become this powerful, where does that leave application-layer companies? Here are a few ways to think about it.

First, AI capabilities are advancing almost vertically, unlocking new capabilities at extraordinary speed, while real-world adoption lags behind. It is this gap that presents an opportunity for application-layer companies. Speed is everything.

Second, good products speak for themselves. It’s an old principle, but it still holds. Even my two-year-old daughter instinctively knows to put four fingers through a mug handle and steady the mug with her thumb. A well-designed product requires almost no explanation. Turn the handle ninety degrees and anyone can tell that it’s poorly designed. While many products may look similar on paper, users can intuitively sense quality.

Third, the last mile matters. Model companies will capture some vertical use cases, but they will not capture all of them. In a market measured in the tens of trillions of dollars, there will always be verticals where intelligence has to be embedded deeply into workflows before it can create real value. For serial entrepreneurs, the winning playbook is to go deep before going wide.

Fourth, building a moat is especially difficult when the market is moving this fast; every part of the stack is in motion. It may therefore be more important to think from the user’s perspective than from the technology’s perspective. How deeply can you embed your product into customers’ workflows? How do you accumulate domain knowledge? And what softer organizational capabilities do you have, especially the organizational learning velocity? Those may matter more than ever.

Fifth, be imaginative. When electricity first emerged in the late nineteenth century, the most obvious early applications were the light bulb (1879) and the telephone (1876). Who could have imagined that the next hundred years would bring a world filled with electrical appliances of every kind? Today, we can readily imagine chatbots and coding agents. Why shouldn’t the next fifty years bring an explosion of intelligent applications we cannot yet conceive of? This is an era for pioneers, and for visionary product managers.

So what are we investing in? First, intelligence itself—one of the largest value pools in this cycle, with the competitive landscape in the U.S. already beginning to take shape. Second, hardware: because of the tension between exponential demand and the linear expansion of physical capacity, hardware may be the most important near-term theme. Beyond that will come the infrastructure needed to produce, distribute, and support intelligence—to build this new world for AI. And finally, applications. They will emerge more gradually. This is not a two- or three-year cycle; it is a ten- or twenty-year opportunity. We will continue looking for brilliant product managers who know how to put intelligence to work and turn it into high-value solutions for customers.

The ship of this new era set sail three years ago. That is a fact. Do we remain in the old continent and watch it fade into the horizon, or do we jump aboard? The choice is entirely in our hands.

At Code Class 2026, Sun Weijie, founder of DP Technology, an early portfolio company of Source Code Capital, delivered a talk on “AI for Science.” From systematically building the field’s underlying infrastructure from the outset to seeing AI for Science become a national strategic priority shared by China, the United States, and Europe in 2025, DP Technology has remained focused on one difficult but necessary task: building a common scientific foundation for research and discovery worldwide.

The following is an edited and abridged selection from Sun Weijie’s talk.

Thank you very much for inviting me to Code Class. Every time I attend, I feel an immense sense of pride and warmth. DP Technology was one of Source Code Capital’s early investments, and our core mission is AI for Science.

The ultimate goal of AI for Science is to create AI scientists that surpass humans across the board. Just as AI may surpass humans at writing code this year or next, the core problem AI for Science aims to solve has now achieved broad global recognition.

From Early Pioneer to Global Consensus

Founded in 2018, DP Technology was an early pioneer in AI for Science. We began building the field’s underlying infrastructure systematically from day one. The early path of exploration was exceptionally difficult, which is why we are especially grateful for Source Code Capital’s sustained support and partnership from the beginning. It was not until 2025 that AI for Science finally achieved global consensus as an industry direction.

In 2025, China, the United States, and Europe all introduced policies that elevated AI for Science to the level of national strategy. China’s “Artificial Intelligence Plus” action plan, released in July, placed AI for Science at the top of its priority areas. Europe’s Horizon plan, released in August, ranked it as the second strategic priority. In November, the United States announced the Genesis Mission, elevating the field directly to the level of a fundamental national policy.

It is worth remembering that the United States has launched only two comparable top-tier basic-science initiatives in its history. The first was the Manhattan Project, which supported the development of the atomic bomb. The second was the Apollo Program, which helped the United States gain a technological advantage during the Cold War. The Genesis Mission is the third, and its strategic intent is unmistakable.

Why does AI for Science occupy such a high strategic position? Put simply, it will be the primary gateway through which humanity acquires scientific knowledge and conducts research. It will also become the underlying engine of technological innovation across industries and a critical lever in competition among major powers. Notably, the six dimensions outlined by the U.S. Genesis Mission are broadly aligned with the “four beams and N pillars” infrastructure framework for AI for Science that we proposed in 2022.

The Limits of Traditional Research and the Answer from AI Scientists

What fundamental problem is AI for Science meant to solve? Our answer is that it is not merely about enabling breakthroughs in one or two subfields. It is about creating a common scientific foundation for research and discovery across the world.

For a long time, the return on investment of traditional human research has been extremely low—far lower than in factories, data centers, or internet platforms. The numbers are striking: global annual research spending is approximately $2.8 trillion, while China alone spends RMB 3.6 trillion each year, equivalent to roughly 2.7% of GDP. Yet the scientific output generated by this enormous investment remains limited.

The root cause is a fundamental weakness in the traditional research paradigm. Traditional research places human scientists at the center and relies on tools to explore the world. Its core activities can be reduced to three words: “read, compute, and do”—read the literature, run calculations, and conduct experiments. This model has two major pain points:

First, the tools are outdated and inefficient. Even today, the way we read the literature is not fundamentally different from how scholars leafed through printed papers in the seventeenth century. Research software and laboratory instruments have also failed to fully carry forward the productivity gains of successive technological revolutions.

Second, talent cannot be replicated. Human beings are the agents of research, but their scientific judgment cannot be copied, scaled, or compounded, directly creating a severe shortage of top talent. There are an estimated 60 to 80 million frontline research professionals worldwide, yet fewer than two million top scientists are believed to be capable of making core breakthroughs in new drugs, new materials, and basic science.

AI for Science will fundamentally change this situation. On the one hand, it can upgrade and replace traditional research tools with intelligent systems. On the other, AI agents can already simulate human researchers, take on research tasks independently, and complete the full loop from question to result. This is the core vision of DP Technology: to build AI scientists that surpass humans across the board and solve the global shortage of top scientific talent.

Based on the trajectory of technological progress and our accumulated experience, we believe AI agents capable of surpassing human scientists across the board may take shape within the next two or three years. An equally clear opportunity today is to make traditional research infrastructure intelligent—upgrading it while equipping AI scientists with the knowledge, compute, and experimental capabilities they need.

Simply increasing compute and execution capacity cannot solve complex scientific problems. It requires a mature tool system for “reading, computing, and doing.” With this goal in mind, we have built a closed-loop research system centered on scientific agents and integrating all three capabilities. An AI scientist cannot work without any one of these four elements.

We have already built AI scientists with the equivalent of five to eight years of experienced research expertise in multiple subfields. One example is MatMaster, an AI materials scientist developed jointly with the Suzhou Materials Laboratory. These AI scientists are built on SciMaster, our general-purpose scientific intelligence foundation. Internal testing shows that SciMaster reaches postdoctoral-level performance across disciplines and already surpasses human postdocs in core tasks including literature review, scientific computing, experimental operations, and scientific writing.

Read, Compute, Do: Three Research Infrastructure Pillars and Commercial Deployment

Around “reading, computing, and doing,” we have systematically rebuilt the three pillars of research infrastructure.

Knowledge systems (read). We have fully structured a vast body of high-quality literature, patents, and data from around the world. AI agents can search across these scientific resources and produce literature reviews with precise source tracing and low hallucination rates. A deep research report that takes the system ten minutes is equivalent to a month of work by a core R&D professional—a tenfold gain in efficiency. The challenge lies in data cleaning: manually cleaning and annotating a single PDF is expensive, and hundreds of millions of papers worldwide cannot be processed by humans alone. Through human-machine collaborative annotation, we have reduced the cost to nearly one-thousandth of the original level. Since the launch of Bohrium Science Navigator, its registrations and visits have remained the highest in the industry, and 4.35 million researchers worldwide now use our products.

Computing systems (compute). Compute tools are the core foundation for iterating on scientific research. We have completed the intelligent adaptation and MCP integration of scientific computing tools from open-source communities and public platforms, and our platform now hosts more than 50,000 categories of scientific software and tools. Where public tools cannot deliver high-precision computation, we develop specialized scientific foundation models, including Uni-SMART for multimodal scientific literature, Uni-AIMS for molecular representation, Uni-Fold for proteins, Uni-RNA for genes, DPA for atoms, and Uni-MOL for molecular conformations. Together, they support high-precision computation in materials, chemistry, biology, and other fields. Notably, in May 2026, DP Technology and its partners jointly released DPA4, a new model architecture to which DP Technology was a core contributor. DPA4 ranked first worldwide on the CPS composite metric in Matbench Discovery, a leading international benchmark for materials discovery, making it the latest SOTA model. On the authoritative molecular benchmark SPICE-MACE-OFF, DPA4 also achieved a new SOTA result with fewer parameters, surpassing the previous leader eSEN to take the top spot.

Experimental systems (do). Every frontier discovery ultimately has to be synthesized, tested, and validated in the laboratory. The central pain point today is that agents cannot interact efficiently with physical laboratory hardware because there is no unified underlying operating system. We were among the first to enter this field and developed UniLab OS, an intelligent laboratory operating system. It now connects and standardizes more than 150 broad categories and nearly 2,000 types of laboratory instruments, enabling AI to take control of the laboratory and conduct experiments autonomously.

The efficiency gains from this system are substantial. In polymer-materials R&D, it used to take a PhD student an entire morning to organize 100 sets of reagents and tune the equipment. Our intelligent system can now deliver the equivalent of a full day’s output from dozens of PhD students in a single day. In Shanghai, the intelligent discovery platform for peptide drug molecules alone can match the around-the-clock output of 500 to 800 chemists. In Yibin, our intelligent R&D system for electrolytes and solid-state electrolytes delivers output equivalent to that of nearly 100 specialists.

Once the infrastructure is in place, how do we close the commercial loop? Our capabilities can be packaged as standardized software services for customers, while large enterprises and research institutions can receive customized R&D solutions. Here I want to highlight the FDE, or forward-deployed engineer, model. It is particularly well suited to scientific research because it addresses three major pain points in the industry:

First, R&D needs are highly fragmented and long-tailed. For a cathode-materials customer, we mapped more than 160 R&D steps across the full process—far too many for any single-point tool to cover. But saving 30% to 40% of the time at each step compounds into a transformative gain in end-to-end efficiency.

Second, the industry has three separate barriers: AI technology, scientific expertise, and engineering implementation. People who understand algorithms may not understand materials chemistry, while domain experts may lack an understanding of AI. A product alone cannot bridge all three gaps.

Third, R&D data is a company’s core intellectual property, and companies will never share it externally. A model that depends on acquiring customer data to iterate is therefore not sustainable.

These three factors point to an optimal commercial model: use 70% to 80% standardized platform capabilities as the foundation, then add 20% to 30% customized development to fit each company’s R&D workflow deeply. To deliver this, we assembled a “three-in-one FDE task force” whose members combine the attributes of domain scientists, AI algorithm experts, and product-delivery specialists. They can translate customer needs into problems AI can solve in practice and deliver quickly.

DP Technology’s core strengths can be summed up in three words: earliest, deepest, and fastest. We were among the first companies in the world to define AI for Science and systematically build its research infrastructure. We have gone deep on foundational infrastructure with exceptional generality and fundamental reach, while remaining among the industry leaders in market coverage and commercial deployment. Today, millions of researchers worldwide and more than 100 universities in China use our platform, and hundreds of leading companies in fields including life sciences and materials science work with us closely.

The Next Five Years and 2040: A Transformative Shift

The direction of scientific research is fully aligned with the national strategies set out in China’s 15th and 16th Five-Year Plans. China’s core goals are to achieve basic self-reliance and strength in science and technology and address key technology bottlenecks by 2030, reach a high level of self-reliance and strength by 2035, and share Chinese research infrastructure and approaches with the world. AI for Science will play a central supporting role in this process.

Over the next five years, the world will see a wave of investment in AI for Science infrastructure, with data, compute, and experimental infrastructure all undergoing comprehensive upgrades. Traditional research-literature databases and scientific software will be fully replaced by AI agents. Scientific instruments will not disappear, but they will be upgraded into AI-native intelligent devices. The industry has reached a consensus: within five years, every R&D-intensive company in biomedicine, new materials, chemicals, and related fields will have to become AI-driven, or fall irreversibly behind in R&D efficiency.

Looking further ahead, we may witness a transformative shift in scientific research before 2040. Scientific discovery could become as simple as using a search engine, organized into a standardized industrial pipeline. Long-standing scientific problems that humanity has pursued for generations may finally be solved. The barrier to research could fall dramatically: a high-school student might be able to distill and define a scientific question, while the population able to participate in scientific discovery could grow from 245 million today to more than 500 million by 2040. The traditional logic governing the distribution of scientific value would be restructured.

The goal of achieving a high level of self-reliance and strength in science and technology will be realized by our generation of researchers. Our core vision is to create AI scientists that surpass humans and solve the global shortage of research talent—to build the Anthropic of scientific research.