As promised, it is time for a follow-up on my personal all-time-favorite Substack article: “Investing in the Age of AI,” published on April 7, 2025. At its heart was the key insight that AI had created a self-reinforcing flywheel. Better applications attract more users, generate more data and create greater demand for cloud computing. That cloud demand drives investment in increasingly powerful chips, which in turn enables better models and entirely new applications. Each revolution of the wheel makes the next one faster.Almost 1.5 years later, that flywheel could hardly be more intact. If anything, it is spinning considerably faster than I imagined. Models have become vastly more capable, AI applications are beginning to perform real work, cloud investment has exploded and demand for leading-edge compute remains insatiable. Yet the share prices of the companies I identified as its principal beneficiaries have also risen dramatically:Five of the six more than doubled. SK hynix increased more than ninefold. The average gain was 254%.Viewed this way, the natural conclusion is that the boat has sailed. It would be tempting to take a victory lap and move on to the next theme. However, share price is only the numerator. What happened to the denominator?To answer how badly the market underestimated what these businesses were becoming, we can rerun the valuations using the prices investors could have paid in April 2025, but the earnings that analysts now expect these companies to generate in 2027.The result is nothing short of astonishing. At its April 2025 share price, Nvidia would have been trading at roughly 6.6 times the earnings now expected for its January 2028 fiscal year. TSMC would have been available at around six times 2027 earnings. Alphabet at less than ten times. Even Amazon, which looked relatively expensive at the time, would have traded at approximately 16.5 times today’s 2027 earnings estimate.SK hynix is the most extreme example. Its share price has risen more than ninefold, yet its current 2027 earnings multiple remains below the 4.9 times forward earnings at which it traded when I wrote the original article. The numbers reflect exceptionally strong (and potentially cyclical) memory-market conditions, so they should not be extrapolated blindly. But they demonstrate the scale of what the market failed to anticipate.The performance was not merely multiple expansion, but these companies became substantially larger earnings machines.Recent operating results help explain why. Alphabet reported 30% operating-income growth and 82% growth in Google Cloud revenue. Amazon’s AWS sales grew 37%. Nvidia’s quarterly revenue more than doubled, while SK hynix reported a 557% increase in operating profit. The earnings estimates followed the businesses upward, rather than simply chasing their share prices.That leaves us with the central question for this follow-up.Did we already experience the great AI trade? Or are we attempting to value a fundamentally different kind of industrial transformation using frameworks built for an earlier era?Over the past seventeen months, AI has begun moving from answering questions to performing work. Coding agents can now operate across meaningful portions of the software-development loop. Models can use tools, navigate digital environments and pursue increasingly long-horizon tasks. OpenAI’s forthcoming Astra model is being discussed not merely as a better chatbot, but as a potential AI research intern capable of implementing and running experiments and pursuing follow-up work from scientific papers.The next transition may be even larger.Until now, most economically useful AI has remained confined to digital environments. It can write, analyse, code, design and communicate, but it cannot reliably manipulate the physical world. Robotics gives intelligence a body. In a world constructed around human proportions, tools and infrastructure, humanoid robots could allow AI to move from digital work into factories, warehouses, laboratories, hospitals and eventually homes.If that transition occurs, the addressable market expands from software expenditure to work itself.This article is therefore an attempt to establish where we now stand, how much the underlying technology has changed, and whether the companies enabling that transformation still offer attractive asymmetry, even after some of the most extraordinary share-price gains in modern market history.Preview: I still own four of the six companies from the original article: Nvidia, SK hynix, Amazon and SoftBank. In fact, today Amazon, SK Hynix, and Nvidia are even cheaper on a forward PE basis than they were at the time of my original writeup. Moreover, I have since added three new AI-related investments.Before diving into earnings growth and valuations, however, it is worth stepping back to consider the bigger picture. If AI is progressing toward increasingly autonomous intelligence capable of performing meaningful digital and eventually physical work, then the central question is not merely whether next year’s earnings will beat expectations. It is where we stand in the early innings of a technological transformation that could reshape the structure of the global economy over the next decadesThe intelligence curve steepensWhen I wrote the original article, OpenAI’s leading general-purpose model was GPT‑4.5, supplemented by its o1 reasoning model, while Anthropic’s frontier model was Claude 3.7 Sonnet. They were already impressive, but nearly the entire current generation of reasoning and agentic models still lay ahead.The improvement since then is visible in the Artificial Analysis Intelligence Index, which combines performance across a range of evaluations covering mathematics, science, coding, reasoning and agentic work. No composite score can capture something as broad as intelligence perfectly, but the direction is clear: the frontier has continued rising rapidly. Frontier-model capabilities have continued rising rapidly. The Epoch Capabilities Index combines performance across numerous benchmarks into a single longitudinal measure of general model capability. Source: Artificialanalysis.ai Notice how frontier models keep pushing the boundaries of intelligence with every new release. This suggests AI capabilities - however impressive today- have not reached their limits and we can expect more.Raw intelligence, however, is only one part of the story. A highly intelligent model that produces a single answer and then stops remains a tool. The more consequential development is that models can increasingly apply that intelligence across an extended sequence of actions.From chatbot to coworkerSeventeen months ago, most people still experienced AI primarily as a chatbot. A human asked a question, the model produced an answer and the interaction ended. The model could generate impressive text or code, but the human remained responsible for defining the task, supplying the relevant context, checking the result and deciding what to do next.The decisive development since then has been the emergence of agency.An AI agent does not merely respond. It can break a larger objective into smaller tasks, decide which tools it needs, search for information, inspect files, write and execute code, evaluate the result, correct mistakes and continue working until it has produced something useful. Increasingly, it can also delegate parts of a problem to other agents and combine their work.Coding became the first major proving ground because software provides almost ideal conditions for machine work. The environment is digital, actions can be executed immediately and the results can often be verified automatically. If the code fails, the agent receives an error message and can try again. This creates a rapid feedback loop that is much harder to reproduce in less structured knowledge work, and vastly harder in the physical world.Nonetheless the same pattern is now spreading beyond coding. Models can navigate websites and desktop software, analyse documents, create financial models, produce presentations, conduct research and interact with the applications in which people already work. GPT‑5.6 can convert unstructured information from documents and workplace applications into finished professional outputs.Its internal adoption may be even more revealing. Over six months, OpenAI says the share of its research compute devoted to coding inference increased 100-fold, while internal agentic token use rose approximately 22-fold. Average daily output tokens per active researcher using GPT‑5.6 were more than twice the previous high recorded with GPT‑5.5.This does not mean that the fully autonomous AI coworker has arrived. Agents remain unreliable. They can pursue the wrong objective, lose track of important context or complete nine steps correctly before failing on the tenth. OpenAI’s own AutomationBench result for GPT‑5.6 Sol remains only 18.1%. Impressive demonstrations should therefore not be confused with dependable, unsupervised deployment across an entire company.Nevertheless, the direction of travel matters. The important change is not merely that AI can solve more difficult problems, but that it can remain useful across progressively longer sequences of work.METR’s task-completion horizon attempts to measure the duration of work, expressed in the time required by a human expert, that an AI agent can complete with a given probability of success. Historically, the frontier’s 50%-reliability task horizon has doubled approximately every seven months.AI agents are becoming capable of completing progressively longer tasks. METR’s task horizon measures the duration of work, expressed in human-expert time, that a frontier agent can complete with 50% reliability. Source: METR.Again, note how the trend is intact, clearly suggesting that AI will keep improving considerably.That is the transition from chatbot to coworker. The human increasingly defines the objective and reviews the output, while the machine performs more of the work in between.From coworker to researcherThe next step is considerably more consequential.A coworker applies existing knowledge. A researcher creates new knowledge. If AI begins contributing meaningfully to scientific and technological discovery, it does not merely increase productivity within the existing economy. It could accelerate the rate at which the technological frontier itself advances.OpenAI’s forthcoming Astra family of models is important in this context. According to reporting by Alex Heath for TIME, Astra has demonstrated the ability to operate computer software at high speed and coordinate multiple agents on difficult problems. In one demonstration, sixteen agents divided a research-level mathematics problem into subproblems and assembled a proposed proof. OpenAI chief scientist Jakub Pachocki says Astra has already passed the company’s internal benchmark for an automated AI “research intern”: given an experimental idea, it can implement the idea inside OpenAI’s codebase, run the experiment and return the results. It can also take a scientific paper and perform follow-up work that previously occupied a human researcher for a week.Source of the Screenshot: Alex Heath on X.comOpenAI reported on August 1, 2026 that an internal version of Astra generated ten advances in mathematics and theoretical computer science, each resolving or making substantial progress on a long-standing open problem. The problems covered fields including geometry, coding theory, quantum complexity, cryptography and extremal combinatorics. The model generated the mathematical arguments, humans helped prepare them for publication, and the results were subsequently formalised into machine-verifiable Lean certificates. OpenAI estimates that the total inference required to discover the solutions would have cost approximately $2,000 at its current API rates.The subsequent response has been striking. Fields Medalist Timothy Gowers described the results as “extraordinarily impressive,” calling the construction of a non-sofic group one of the most important open questions in group theory and the multicolour Ramsey result a major problem he had not expected to see resolved in his lifetime. Independent mathematicians have already begun building on the work. Gábor Kun and Andreas Thom extended the mechanism to construct additional non-sofic groups, while Francesco Fournier-Facio used Astra’s central criterion to produce a finitely presented, torsion-free example.The strongest results appear mathematically genuine, machine-verifiable and important enough that leading researchers are already extending them. While Astra did not create the underlying mathematics from nothing; it combined existing ideas in ways that humans had not previously discovered, which is also a reasonable description of a great deal of genuine mathematical progress. AI is beginning to move from retrieving and recombining human knowledge toward making independently useful contributions to its creation.This potentially introduces a second flywheel inside the original AI flywheel.Today, human researchers design better algorithms, architectures and training methods, which produce more capable models. Those models are now beginning to assist the researchers building their successors. Better AI therefore contributes to the creation of even better AI.The result need not to be an instantaneous “intelligence explosion” to matter economically. Even a moderate acceleration in AI research could compress development cycles, improve chip utilisation, automate experimentation and increase the number of promising ideas that can be tested. A research team assisted by hundreds of persistent agents could explore a much larger possibility space than the same number of humans working alone.How close are we to AGI?This brings us to the question that previously belonged primarily to science fiction: how close are we to artificial general intelligence?OpenAI defines AGI as highly autonomous systems that outperform humans at most economically valuable work. In his recent interviews with TIME, OpenAI chief research officer Mark Chen estimated that the company is “80% of the way” there. Sam Altman stopped short of saying that AGI had already been achieved, but said he expects OpenAI to have an internal system that he would call AGI by the end of 2026. Greg Brockman suggested that, viewed retrospectively from two years in the future, the present moment might be remembered as the point at which AGI emerged. These are extraordinary statements. They are also statements made by the executives of a company raising and spending unprecedented amounts of money to build AGI. Their definition is not universally accepted, their assessment is not an independent one and the incentives for generating excitement are obvious.The term itself creates additional confusion. Does AGI require a system to outperform every individual human at almost everything? Must it possess common sense, creativity and social understanding? Does it need to operate independently for months, or is it sufficient for a human to direct it? Different definitions can place the same model either far beyond the threshold or nowhere close to it.However, others argue that the AGI threshold has already been crossed. In a 2026 commentary published in Nature, Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks contend that the human-level machine intelligence imagined by Alan Turing is now a reality. Their argument is that current models can reason, communicate, create and solve problems across a wider range of domains than any individual human. Demanding consciousness, embodiment, perfect reliability or competence at every possible task, they argue, applies a standard that no human could satisfy either.Jensen Huang has reached a similar conclusion from a more economic perspective. On Nvidia’s latest earnings call, he said that “for many tasks” we could already claim to have achieved AGI. He immediately added that such milestones have become almost meaningless. What matters, in his view, is that AI is already performing productive work, generating economically valuable output and improving through increasingly agentic workflows.These claims do not settle the debate. The Nature authors use a functional definition of intelligence, while Huang explicitly limits his claim to “many tasks.” Neither demonstrates that current systems can autonomously replace a human across most economically valuable work—the higher bar contained in OpenAI’s own definition. But the fact that serious researchers and industry leaders can plausibly argue that AGI is already here tells us that the disagreement has shifted from whether machines can display general intelligence to how much autonomy, reliability and breadth we should demand before acknowledging it.Perhaps AGI will not arrive as a single model accompanied by a press release. It may instead become a threshold that we recognise only retrospectively after increasingly general systems have already begun transforming the economy around us.For investors, there is a more useful way to frame the question.What proportion of economically valuable tasks can an AI system perform, at what level of reliability, at what cost and with how much human supervision?Under this definition, AGI is less like a single finish line and more like an expanding economic frontier. A model does not need to become universally superhuman before it begins to alter labour markets and corporate cost structures. It only needs to become good enough, cheap enough and reliable enough to automate particular tasks. As those task categories widen, a growing share of knowledge work becomes addressable.Robotics: When Intelligence Enters the Physical WorldSo far, almost all economically meaningful AI activity has taken place inside computers.AI already has something resembling eyes through multimodal vision. It can interpret images, video, documents and the contents of a computer screen. It can also interact with the digital world through software tools. What it does not yet have is a reliable body.The physical world is far less forgiving than a software environment. Objects deform, lighting changes, floors become slippery and humans behave unpredictably. A digital agent can retry a failed action at almost no cost. A robot dropping a glass, damaging a machine or injuring a person faces a very different standard. Physical intelligence requires perception, mobility, dexterity, real-time control, safety and an understanding of cause and effect - all operating together.This is why progress in robotics has historically appeared slower than progress in software. Each new physical task required extensive programming, engineering and task-specific training. Data was scarce because collecting it involved real machines operating in real environments.Foundation models, simulation and synthetic data may begin to change that. Instead of programming a separate robot for every task, developers are attempting to build general vision-language-action models that can interpret instructions, understand their surroundings and transfer learned behaviours across machines and environments. Robots can practise millions of actions in simulation before those behaviours are deployed in the physical world.Nvidia’s Isaac GR00T platform is one example of this emerging stack. It combines robot foundation models, simulation, synthetic-data generation and onboard computing, with the aim of allowing developers to train and deploy more general humanoid capabilities. Nvidia is making its humanoid reference architecture available to researchers and supporting platforms such as the Unitree G1. (Nvidia)Humanoids are particularly interesting because the world has already been constructed for the human body. Our doors, stairs, shelves, tools, vehicles, factories and workplaces assume human proportions and capabilities. A sufficiently capable humanoid could theoretically enter these environments without requiring the entire physical infrastructure to be redesigned.That does not mean every useful robot will look like a human. Specialized machines will often be cheaper, safer and more efficient. Autonomous vehicles, warehouse robots, industrial arms and purpose-built medical machines may create enormous value long before a general-purpose household humanoid becomes economical.But humanoids illustrate the outer boundary of the opportunity. If intelligence can move from the computer into a general-purpose body, the addressable market expands from information work to physical work. The speed of progress is perhaps most visible at Beijing’s annual World Humanoid Robot Games, also known as the “Robot Olympics.”At the inaugural competition in August 2025, the limitations were difficult to miss. Robots repeatedly collided during football matches, collapsed while running and often required humans to carry them away. Tiangong Ultra won the 100-metre race in 21.5 seconds—respectable for a bipedal machine, but more than twice Usain Bolt’s world-record time. The spectacle was entertaining precisely because the robots appeared so fragile and awkward.One year later, the same robot completed the distance in 8.64 seconds, almost a full second faster than Bolt’s 9.58-second human record. More than 2,000 robots from 666 teams competed in 51 categories, ranging from sprinting, football, boxing and choreographed dance to industrial assembly, hotel services, housework and emergency response. The video below shows some of the highlights.The Robot Olympics illustrate that the machines remain brittle, specialized and often dependent on controlled conditions. Yet the slope of progress increasingly mirrors what we have already seen in AI models: initially amusing demonstrations improve with surprising speed until, almost without warning, they cross a threshold that previously appeared years away.The compute billThis transition also changes the amount of computing power the world will require.A chatbot is a relatively contained workload. The user enters a prompt, the model generates an answer and the process stops. An agent may work for minutes, hours, days or eventually weeks. It reads and rereads large amounts of context, calls tools, writes code, evaluates results, corrects itself and communicates with other agents. A single human request can therefore trigger dozens or hundreds of separate model interactions.Multi-agent systems multiply that demand further. The Astra mathematics demonstration did not involve a single model producing a single answer. Sixteen agents divided the problem, pursued different approaches, exchanged information and assembled the final result. Persistent agents do not wait for a human to provide the next prompt. They continue generating and consuming tokens while the work remains unfinished.This helps explain the token-usage chart circulated by Hedgeye. According to the data shown, AI agents are already consuming approximately five times as many tokens as human-directed interactions, with agentic token use increasing fourteenfold since February.Source: Hedgeye on X.comThe precise figures depend on how agentic and human usage are classified, so I would not treat the chart as a universal measurement. But its direction is corroborated by OpenAI’s internal figures: a 22-fold increase in agentic token use and a 100-fold increase in research compute devoted to coding inference in only six months. Goldman Sachs, meanwhile, estimates that broader adoption of AI agents could drive a 24-fold increase in monthly token consumption between 2026 and 2030, reaching 120 quadrillion tokens per month.Personally, I increasingly hit the usage limits on both my ChatGPT and Claude Pro subscriptions. As the systems become capable of tackling more complex tasks, my own consumption rises with them.This is the agentic version of Jevons’ paradox. Chips become faster, models become more efficient and the cost of producing an individual token declines. But cheaper intelligence unlocks more use cases, longer tasks and more agents. Consumption can therefore increase far faster than the cost per unit falls.The move from training to inference reinforces this dynamic. Training a model is an enormous but discrete event. Deploying millions (or eventually billions) of agents creates a continuous workload. Every task completed, experiment performed and robot controlled requires additional inference. If AI becomes embedded in everyday work, compute demand becomes tied not merely to the periodic construction of new models, but to the ongoing production of economic output.No one expresses the outer edge of this thesis more boldly than Masayoshi Son.Son expects a world containing 100 trillion autonomous AI agents by 2040. He believes supporting them will require approximately three terawatts of AI data-centre capacity and $5 trillion of annual investment. His economic justification is equally audacious: he predicts that AI-related revenue will eventually become equivalent to 20% of global GDP.What would 20% of global GDP actually mean?Masa Son’s forecast becomes easier to appreciate once translated into dollars. Global nominal GDP was approximately $117 trillion in 2025. If it grows at 5% annually in nominal terms, the world economy would more than double to roughly $243 trillion by 2040. Son’s assumption that AI revenue could become equivalent to 20% of global GDP would therefore imply an annual AI revenue pool approaching $49 trillion.Sources: Author’s calculations based on approximately $117 trillion of global nominal GDP in 2025, as reported in the IMF World Economic Outlook, and Masa Son’s 2040 assumptions reported by Reuters.The resulting number is almost difficult to comprehend. Annual AI revenue of $48.6 trillion would be equivalent to more than 40% of the entire world economy today. From that perspective, Son’s projected $5 trillion of annual investment would amount to only around 10% of the revenue generated by the resulting AI economy.Of course, these numbers should not be mistaken for a carefully documented base-case forecast. Son has not publicly provided the detailed assumptions behind them. But I think it is a useful framework to quantify the investment opportunity ahead of us and it is a framework worth updating and refining.If intelligence becomes an abundant input into digital work, scientific discovery and eventually physical labour, the corresponding compute market need not resemble today’s cloud market. Compute becomes a productive input analogous to energy, but one that produces cognition, decisions and actions. The amount demanded will depend not only on how many people use AI, but on how many agents each person or company deploys, how long those agents operate and how much work they perform.This brings us back to the flywheel from the original article.Better AI enables more useful applications. More useful applications create greater demand for cloud infrastructure. Greater cloud demand drives investment in chips, memory, networking and energy. Better infrastructure enables larger and more capable models. Those models become agents, researchers and eventually the intelligence inside robots, opening entirely new categories of applications and beginning the cycle again.The AI flywheel is in full force and has arguably even expanded since my original article.From technological revolution to investment returnsEven if the direction of travel is right, the investment conclusions are not trivial. Technological revolutions often attract excessive capital. Competition can destroy margins, value may migrate from one layer of the ecosystem to another, and even a remarkably accurate prediction about the future can produce poor returns when the entry price is too high.The AI flywheel tells us where value is being created. It does not tell us which companies will ultimately capture it, or how much investors should pay for that possibility.That is the dividing line between a technological thesis and an investment thesis. The relevant investment question is which businesses possess durable claims on the resulting cash flows, how much growth is already embedded in their valuations, and where the market may still be underestimating what comes next.It is therefore time to return from the technological north star to the numbers.I will do so through a series of company-level analyses covering the AI-linked businesses I own. And there is no more fitting place to begin than with the man whose vision concluded this first part and the company through which he is attempting to make it reality.Part I: SoftBank — Buying Masa Son’s AI vision at a 47% discount to NAV?Masayoshi Son’s track record combines extraordinary foresight with extraordinary volatility. It includes some of the greatest investments in modern history, most notably Alibaba and Arm, but also spectacular failures (WeWork) and periods of severe capital destruction. That combination makes him both one of technology’s most visionary investors and one of its most controversial capital allocators.Yet the pattern across more than three decades is remarkable. Son has repeatedly positioned SoftBank around major technology shifts: the commercial internet, mobile data, e-commerce, semiconductor architecture and now artificial intelligence. Source: Softbank Earnings results presentation from August 6, 2026From technology investor to ASI platform providerDuring earlier cycles, SoftBank primarily participated by investing in companies positioned to benefit from the next platform: Yahoo during the commercial internet era, Alibaba in e-commerce and mobile operators during the transition to smartphones. The businesses were connected by a broad technological vision, but they did not necessarily form an integrated operating system.Son’s ambition today goes considerably further. SoftBank no longer wants merely to own shares in companies that benefit from AI. It wants to become what Son calls the world’s leading ASI platform provider—assembling the models, chips, infrastructure and physical machines required to create and deploy intelligence at scale.Before considering that vision, however, it is important to establish what SoftBank actually owns today.