The intelligence margin: Can frontier AI ever acquire ‘software economics? 

For decades, software possessed an economic superpower. Build the product once. Sell it repeatedly. Allow revenue to scale much faster than the physical infrastructure required to produce each additional unit. Microsoft did not manufacture another factory every time somebody opened Excel. Adobe did not construct another data center every time somebody edited a photograph. Software economics became attractive precisely because replication approached zero marginal cost. Artificial intelligence (AI) may be rewriting that equation. 

The emerging frontier-AI industry confronts a paradox: What happens when the world’s most sophisticated software begins acquiring the economics of heavy industry? 

Anthropic’s prospective IPO may provide one of the clearest tests yet. According to an IPO prospectus reviewed by Reuters, Anthropic generated nearly $4.6 billion in revenue in 2025, approximately twelve times the previous year’s level. Yet it recorded an operating loss of approximately $8.06 billion. Its operating expenses reached about $12.65 billion, of which $7.33 billion, roughly 58 percent, went toward computing and infrastructure. Even more strikingly, the company disclosed approximately $518 billion in future cloud, computing, and infrastructure obligations. 

The widely reported $42 billion net loss requires an important qualification. Roughly $34 billion represented an accounting charge associated with financial instruments potentially convertible into shares. Therefore, the $8.06 billion operating loss is considerably more useful for understanding Anthropic’s underlying operating economics. But the fundamental question remains. Can intelligence eventually become cheap enough to reproduce that frontier AI acquires software economics? 

When software begins to look like a steel mill 

Traditional software economics rests upon an extraordinary asymmetry: High initial development cost + negligible replication cost = enormous operating leverage. Frontier AI complicates this equation. 

Building a better model requires training compute. Serving millions of users requires inference compute. Reasoning models may consume substantially more computation per query. Autonomous agents can execute long chains of actions rather than answering a single prompt. 

The product being sold is software. But the factory producing it increasingly resembles industrial infrastructure. Data centers become factories. GPUs become machinery. Electricity becomes feedstock. Tokens become units of production. This is why the conventional obsession with AI revenue growth is insufficient. Revenue can increase spectacularly while the economic machinery necessary to produce that revenue expands almost as rapidly. 

A company can become technologically more powerful and commercially larger without necessarily becoming economically more efficient. That distinction may become one of the defining investment questions of the AI age. 

The intelligence operating leverage 

We need a metric beyond revenue growth. I propose what might be called Intelligence Operating Leverage (IOL): IOL = Percentage Growth in Revenue ÷ Percentage Growth in Compute and Infrastructure Cost 

The interpretation is straightforward. If IOL > 1, revenue is expanding faster than the infrastructure cost required to produce intelligence. If IOL = 1, revenue and intelligence-production costs are essentially scaling together. If IOL < 1, infrastructure requirements are growing faster than monetization. The ratio should ideally be calculated over several years and supplemented by cost per unit of useful intelligence — such as cost per million tokens, cost per completed agentic task, or compute cost per dollar of revenue. One year’s number can be distorted by advance capacity purchases or temporary underutilization. The destination matters more than the snapshot. 

Can frontier-AI companies progressively move their IOL sustainably above one? If they can, AI could gradually inherit the economics of software. If they cannot, frontier AI may look less like SaaS and more like telecommunications, cloud infrastructure or utilities — enormously important businesses, but businesses in which continuous capital expenditure is inseparable from growth. 

There is, however, a powerful counterargument 

The bearish interpretation of AI economics overlooks something remarkable: The cost of intelligence is collapsing. Stanford’s 2025 AI Index found that the inference cost of obtaining performance equivalent to GPT-3.5 on the MMLU benchmark fell from approximately $20 per million tokens in November 2022 to $0.07 by October 2024 — a decline exceeding 280-fold in roughly 18 months. AI hardware price-performance was improving by around 30 percent annually, while energy efficiency was improving approximately 40 percent annually. 

The improvement has continued. Alphabet reported that it reduced Gemini serving unit costs by 78 percent during 2025 through model optimization, efficiency improvements and higher infrastructure utilization. 

This changes the equation dramatically. Suppose an AI company requires twice as many tokens next year but manages to halve its effective cost per token. Usage has doubled without doubling the corresponding production cost. 

That is precisely how operating leverage can emerge. The race is not simply: Revenue vs. compute. It is: Demand growth × monetization vs. Compute consumption × cost per unit of compute. And that is a far more interesting race. 

The ‘Jevons Paradox’ of intelligence 

There is another complication. Making intelligence cheaper may not reduce total spending on intelligence. It may increase it. When steam engines became more efficient, coal consumption did not necessarily fall because cheaper mechanical power encouraged vastly greater usage. Economists know this phenomenon as the Jevons Paradox. 

AI may experience something similar. A chatbot answering one question might require one inference sequence. An autonomous agent researching a market, examining hundreds of documents, writing code, interrogating databases, running simulations, communicating with other agents and revising its work could require thousands. Thus, cost per unit of intelligence decreases while simultaneously units of intelligence consumed increases. 

Efficiency improvements could therefore be swallowed by exploding demand. We may make intelligence one hundred times cheaper, and consequently consume one thousand times more of it. That would be wonderful technologically while remaining extraordinarily capital intensive economically. 

Microsoft offers an early warning 

Even diversified technology giants are encountering this pressure. Microsoft reported $281.7 billion in FY2025 revenue and $128.5 billion in operating income, demonstrating economics very different from those of frontier-model startups. Yet Microsoft Cloud’s gross margin fell to 69 percent, with the company explicitly attributing part of that decline to scaling AI infrastructure. Intelligent Cloud revenue grew 21 percent, while its cost of revenue increased 36 percent. That is revealing. 

AI can simultaneously create enormous demand and compress margins. The technology may be revolutionary without every layer of its value chain possessing revolutionary economics. 

The competitive advantage may move 

This raises another uncomfortable question. What exactly is the competitive advantage of a frontier-AI company? If competing models converge in capability while inference costs collapse, intelligence itself may become increasingly commoditized. The competitive advantage could migrate elsewhere: proprietary data → distribution → enterprise integration → agent ecosystems → switching costs → trusted workflows → specialized domain knowledge → infrastructure efficiency. 

In that world, possessing the smartest model for six months may matter less than possessing the most economically efficient system for converting computation into valuable customer outcomes. 

The winner may not be the company producing the most intelligence. It may be the company producing the highest-value intelligence per dollar of capital consumed. That is the Intelligence Margin. 

Safety, competition, & capital 

Anthropic’s situation also exposes a deeper governance contradiction. Frontier laboratories operate under four simultaneous imperatives:  

Safety says: slow down. 

Competition says: accelerate. 

Capital markets say: grow. 

Infrastructure says: spend. 

Public ownership will not eliminate these tensions. It will make them measurable every quarter. Anthropic’s proposed governance structure illustrates the difficulty. Reuters reports that its founders would retain 50.1 percent voting control through a special Founder LLC, explicitly designed in part to protect the company’s public-benefit mission from ordinary market pressures. 

That may prove to be one of the most fascinating corporate-governance experiments of the AI era: a capital-hungry enterprise seeking enormous public-market financing while simultaneously attempting to insulate crucial technological decisions from those same capital-market pressures. 

From ‘Software Margin’ to ‘Intelligence Margin’ 

Perhaps we are asking the wrong question when we ask whether AI companies will become profitable. The more consequential question is: What kind of profitability will frontier intelligence ultimately permit? 

If model efficiency, specialized chips, utilization rates, software optimization and falling inference costs allow revenue to grow persistently faster than compute expenditure, frontier AI could eventually acquire something resembling traditional software economics. 

But if every increase in capability generates another wave of larger models, longer reasoning chains, autonomous agents, larger data centers and greater electricity consumption, AI may become something entirely different. Not software, or manufacturing, or utilities. But a new economic category: industrialized intelligence. 

For investors, managers and policymakers, revenue growth alone will consequently become an increasingly inadequate measure of success. We should start watching another number: How many dollars of economically valuable intelligence does each additional dollar of infrastructure create? Because the great AI race may ultimately be decided not by who builds the most intelligent machine, but by who discovers how to make intelligence scale faster than the capital required to manufacture it. 

And when that happens, the most valuable margin in the twenty-first-century economy may no longer be the ‘software margin.’ It will be the ‘Intelligence Margin.’