Written by: Maxence Visseau | Arkevium Capital & Arkevium Research
Forecasting when the peak of an economic and market cycle will arrive is a very difficult business. Anticipating how it will end is slightly easier. For the past few years our stance on the AI boom has been consistent. We want to participate in the boom while continuously testing the assumptions embedded in prices. We do not know exactly when the top comes. Neither do you. But we have much stronger conviction about the risks that will shape the next phase.
The technology can transform the economy while delivering poor returns to many of the investors financing that transformation. Railways did it. Telecommunications did it. AI can do it as well. Recent data suggest that the cycle is changing. The first phase was driven by rapid adoption. The next phase, in our opinion, will depend on usage intensity.
We believe AI adoption growth has already passed its most explosive point in the United States. Usage and investment can continue accelerating. The economic impact may still be ahead of us. But the investment framework must now change.
Anthropic Reveals the Shape of the Adoption Curve
Anthropic is on track to generate annualized revenue of more than $65 billion based on current performance, up more than sevenfold from its pace at the end of last year. Extraordinary in isolation. Almost mechanical in context.
The surge should not have shocked anyone, because in substance it is a slightly lagged version of the adoption curve that preceded it (Chart 1). According to Ramp's AI-adoption metrics, Anthropic's penetration rate within the US business community has risen nearly threefold over the past year.
This data is tracked in quasi real time by AI-specialist investors. To a large extent it was already priced. The market did not rerate on the news because the news was, functionally, six months old when it arrived.
The mathematics of future adoption are also becoming less favorable. Anthropic’s penetration reached 43.5% of the businesses in Ramp’s sample in July. It cannot triple again within the same market. Across providers, more than half of businesses in Ramp’s US sample are already using AI (Chart 2).
Adoption can continue rising. The remaining pool of new corporate customers is simply smaller. Customer acquisition will contribute less to growth as penetration approaches saturation. We believe the United States has passed the point of peak acceleration in adoption breadth. And that conclusion tells us where future growth must come from. The next phase will be driven by deeper integration inside companies that have already adopted the technology.
Usage Depth Is Replacing Customer Growth
Breadth measures how many companies use AI. Depth measures how much intelligence each company consumes. The evidence on depth remains extremely strong.
Monthly AI spending per employee is rising across every corporate cohort. The median adopter now spends more than $10 per employee each month. Companies in the top 10% spend several hundred dollars. The top 1% spend several thousand dollars. Year over year growth is above 100% across all three groups, with the strongest acceleration concentrated among the most intensive users (Chart 3).
In the US at least, the growth of AI revenue will come from extracting more from the existing client base rather than from signing up new customers. Companies are expanding from occasional chatbot use toward coding, customer support, research, data analysis and automated operating workflows. Agentic systems could increase consumption further. An agent can run continuously, call several models and execute hundreds of tasks. The compute intensity of these workflows can be many times greater than that of a human asking a chatbot a single question.
International markets also provide room for expansion. The competitive environment will be more difficult. Chinese open source models can exert greater pressure on prices, especially in markets where cost carries more weight than concerns around security, governance and data sovereignty.
Lower prices can still increase aggregate demand. As intelligence becomes cheaper, more workflows become economically viable. This is the Jevons paradox applied to compute. The variable is therefore the interaction between price and volume. Falling token prices provide limited information in isolation. Spending per employee tells us whether lower costs are generating enough incremental consumption to expand the total market. So far, they are. But Jevons is ultimately a statement about quantities. It offers no guarantee about which part of the value chain will capture the resulting profits.
Intelligence Is Becoming a Routed Commodity
At its simplest, revenue at the model layer can be decomposed into four variables:
Number of customers × workflows per customer × compute per workflow × price of compute
The first variable is approaching saturation in the United States. The second and third remain highly expandable. The fourth is falling rapidly. That combination supports explosive token growth. It creates a much less certain outlook for margins. Anthropic’s own usage data provide an early warning. Its newest and most capable model represented only 6% of the tokens purchased from the company and 11.4% of spending. This occurred despite the model costing roughly twice as much as OpenAI’s flagship product. Corporate buyers are already discovering that the most capable model is often unnecessary for the task.
Once a model crosses the capability threshold required for a workflow, customers begin optimizing for cost, latency, reliability, security and integration. Further improvements in benchmark performance have declining economic value for that specific use. Large companies are also adopting several providers. Different workloads require different levels of intelligence. A basic classification task does not require the same model as advanced research or software engineering. As routing systems improve, model selection will increasingly resemble best execution in financial markets. Each task will be sent to the cheapest model capable of completing it with an acceptable level of accuracy.
This will increase price competition. Open source adoption will accelerate the process. Companies will still require chips, cloud capacity and electricity to operate those models. Aggregate compute demand may remain strong. The economic rent can migrate away from the model developer.
The divergence between semiconductor equities and the price of intelligence illustrates this tension. The SOX Index remains around 70% above its January base, while the usage weighted price paid per million tokens has fallen sharply from its midyear peak. The infrastructure trade remains priced for scarcity even as the product generated by that infrastructure becomes cheaper (Chart 4).
A coherent equilibrium can emerge in which token volumes grow exponentially. AI becomes embedded across the economy and returns at the model layer deteriorate. If usage rises fivefold while prices fall tenfold, total spending declines by half. Even when spending rises, producer margins can compress if efficiency gains are transferred to customers through competition. The hurdle for shareholders is also much higher than revenue growth. Incremental gross profit must exceed depreciation, electricity, financing costs and the research spending required to maintain competitive performance. Jevons can protect compute demand. It cannot protect AI rents.
AI Has Become an Industrial Capital Cycle
Investors should stop analyzing the AI buildout as an asset light software cycle. The infrastructure requirements resemble an industrial expansion. Data centers require semiconductors, memory, networking equipment, cooling systems, power generation, transformers and new grid connections. The physical intensity is enormous.
Estimates suggest that the AI capital cycle is already delivering the largest investment driven boost to GDP in modern history. Its impact approaches 3.6%, compared with roughly 2.25% for railways and close to 1.1% for highways and telecommunications (Chart 5).
The scale of future spending is even more important. Capital expenditure across Amazon, Alphabet, Meta, Microsoft and Oracle is projected to exceed $1 trillion annually from 2027. Combined spending remains above that threshold through 2030 under current estimates (Chart 6).
The latest full year capital spending plans from Amazon, Alphabet, Meta and Microsoft alone exceed $700 billion. Fiscal periods differ and part of this expenditure supports businesses outside AI. The direction remains clear. The technology sector is committing capital on a scale normally associated with national infrastructure programs.
The physical data center may remain useful for decades. The chips inside it may become economically outdated within a few years. This creates a maturity mismatch. The industry is deploying long duration capital against demand whose composition, pricing and competitive structure can change every few months.
Demand is currently strong enough to justify continued investment. Hyperscalers still report capacity constraints, growing backlogs and rapid cloud expansion. These observations confirm scarcity today. They do not establish the return on the next $700 billion of capital. The dangerous point will arrive when incremental gross profit begins growing more slowly than the capital base required to produce it. Revenues can still rise during this transition. Earnings can still grow. Economic returns will deteriorate if invested capital grows faster.
Capital expenditure is also partly a lagging indicator. Data centers ordered today may enter service next year. Equity markets will price their expected utilization before the capacity becomes operational. The first signal of a turn will probably appear in the second derivative of orders, bookings and utilization expectations. It may arrive well before reported infrastructure revenues peak.
History Warns Against Confusing Utility With Returns
History offers several versions of this story, and the lesson is more specific than the usual bubble analogy.
The railway buildout transformed the economic geography of the nineteenth century. It reduced transportation costs, connected markets and enabled entirely new industries. The investment returns were much less impressive. Following Britain’s Railway Mania, the return on total railway capital was estimated at 2.8% in 1849. It averaged only 3.6% between 1849 and 1858.
The railway network created enormous value for the economy. That value was distributed across landowners, manufacturers, consumers and new businesses. Many investors who financed the network captured very little of it.
Telecommunications provide an even closer analogy. Between 1996 and 2000, US investment in communications equipment increased from approximately $62 billion to more than $135 billion per year in constant dollars. The Nasdaq Telecommunications Index rose from 198 in April 1997 to 1,230 in March 2000. By May 2003, the index had fallen to 136. The infrastructure survived. Internet traffic continued growing. But much of the original equity was destroyed.
AI can follow a similar path. Falling inference costs improve the technology’s economic usefulness. They can simultaneously undermine the returns earned by the companies producing that intelligence. This buildout differs from telecommunications in important ways. The leading builders are among the most cash generative companies in history. They entered the cycle with strong balance sheets and have funded most investment internally. The infrastructure can also support a wide range of workloads. These differences reduce the probability of a systemic financing crisis. However, they do not remove the possibility of overcapacity, margin compression or poor marginal returns.
Corporate Transformation Will Follow a Productivity J Curve
Buying subscriptions is adoption. Redesigning a company around AI is transformation. The second requires investment in data architecture, employee training, compliance, workflow redesign and new product development. Most of that organizational capital is expensed rather than capitalized, which means the cost lands on the income statement before the productivity gain shows up in it. This is the productivity J-curve associated with every general-purpose technology, and it is why measured returns to adoption will look disappointing for another year or two before they look obvious.
This delay matters for investors because the eventual beneficiaries may sit outside the conventional AI sector. The firms buying the technology could capture more value than the firms producing it. Their ability to do so will depend on market structure. The ideal adopter has a large wage bill, repetitive information intensive processes, digitized proprietary data and enough pricing power to retain the productivity gain. A company operating in a highly competitive commodity market may pass most of its savings to customers. A business with differentiated products, strong distribution or concentrated market share can retain more of the surplus through higher margins.
The result will be substantial dispersion within sectors. Investors who classify every company as either an AI winner or an AI loser will miss the more important distinction. From our perspective, the question is who can convert cheap intelligence into retained economic profit.
We Want to Own the Bottleneck While It Remains a Bottleneck
The standard recommendation to own picks and shovels has become too simplistic. Picks and shovels can be overproduced. Our principle is to own the bottleneck for as long as it remains scarce.
Today those bottlenecks include advanced semiconductors, high-bandwidth memory, packaging, power generation, grid connections, transformers and cooling systems. These businesses can keep earning excess returns even as model prices fall, because cheaper intelligence stimulates more compute consumption. But scarcity is a cycle, not a permanent moat. When supply catches up, the equities will reprice before their revenues peak. The second derivative of orders matters more than the absolute level of the backlog, and lead times are a better leading indicator than either.
At the model layer, explosive revenue growth must be assessed against multi vendor adoption, falling inference costs and the speed at which competitors approach frontier performance. Faster diffusion increases the economic value of AI. But it can also shorten the duration of excess returns.
At the application layer, the defensible assets are distribution, proprietary data, deep workflow integration and accountability for an outcome. A generic interface is easy to reproduce. A product embedded in a company’s operating system, trained on internal data and responsible for completing a regulated workflow is much harder to displace.
Outside the AI sector, we favor adopters that can retain productivity gains. Pricing power is essential. A large addressable productivity improvement has limited investment value when competition transfers the entire benefit to customers.
Three Scenarios From Here
1) Abundance Without a Demand Collapse
Our base case, at roughly 50% probability, is abundance without a collapse in demand. Usage continues to grow rapidly and Jevons remains operative. Agentic workflows consume far more inference than chatbots. Total spending rises, but more slowly than token volumes as prices fall. Model-provider revenues stay impressive while gross margins and valuation multiples compress. Returns migrate toward applications with distribution and toward companies using cheap intelligence to expand margins.
This regime favors selective infrastructure exposure, differentiated applications and high quality adopters. It creates a difficult environment for model valuations that require sustained premium pricing.
2) An Extended Scarcity Regime
The second scenario, at approximately 35%, is an extended scarcity regime. Agents unlock enough genuinely new tasks that compute consumption outruns both efficiency gains and new capacity. Frontier performance stays economically differentiated and the leading labs hold premium pricing. The current infrastructure winners keep outperforming, particularly where supply cannot respond quickly. Electricity, grid equipment, advanced packaging and leading edge memory remain the strongest parts of the value chain.
The main portfolio risk in this scenario comes from reducing infrastructure exposure too early.
3) The Fiber Moment
We assign a probability of 15% to a sharper capacity correction. New infrastructure arrives as corporate AI budgets become more disciplined. Many pilot projects fail to generate acceptable returns. Reserved compute goes underutilized. Model competition intensifies. Prices fall faster than workloads expand. Depreciation and financing costs continue rising while new orders slow.
The first casualties would be leveraged capacity owners and late-cycle suppliers whose valuations assume current scarcity is permanent.
The technology would keep spreading, exactly as internet traffic kept growing after the telecom bust, and cheaper infrastructure would eventually seed a second wave of applications with an entirely different set of shareholders capturing the returns.
Bottom Line
Our conclusion is therefore less bearish than it may initially appear. We may have passed the peak in adoption breadth. We have almost certainly not passed the peak in usage, investment or economic impact. The technology can succeed spectacularly while many of the securities financing it disappoint. Railways, electrification and fibre all demonstrate that economic transformation and investor returns are separate questions with separate answers.
None of this tells us when. We said at the outset that forecasting the peak is very hard and that anticipating the shape of the ending is somewhat easier, and we still believe it. The correct posture from here is to keep riding it while systematically rotating risk out of the things that will be obsolete in three years and into the things that will still be scarce in thirty.
