The Circular Economy of AI

The most important thing to understand about the AI boom in 2026 is that it is not fake in the simple sense. The money is real. Nvidia sells chips. Amazon, Google, Microsoft, Oracle, and CoreWeave sell compute. Anthropic and OpenAI sell models and coding agents. Enterprises are paying for access. Developers are using the tools. The question is not whether economic activity exists.

The question is whether the activity closes the loop with external productivity: customers outside the AI infrastructure complex becoming meaningfully more valuable businesses because they use AI. If that happens, the circularity looks like an early-stage buildout. If it does not, the circularity starts to look like a machine that books revenue and valuation gains from its own financing arrangements.

I. The Loop

The capital cycle now has four visible steps. Hyperscalers and chipmakers fund AI labs. The labs spend huge amounts on chips and cloud. New funding rounds lift private valuations. Those valuations can flow back into public-company earnings through mark-to-market accounting, even when no stake has been sold.

Alphabet and Amazon's Q1 2026 earnings made the loop unusually visible. Fortune reported that Alphabet's record quarterly profit included roughly $28.7 billion in pre-tax fair-value gains on its private equity stakes, primarily Anthropic, while Amazon disclosed a $16.8 billion pre-tax gain in non-operating income from its Anthropic investment.^[1] Both are pre-tax markups on stakes that have not been sold, not operating profits from search, ads, AWS, or retail. They are valuation gains.

The accounting is not a trick in the illegal sense. ProMarket traces the mechanics to ASU 2016-01, a 2016 accounting rule that pushed many equity investments through fair-value changes in net income rather than hiding the movement elsewhere in the financial statements.^[2] But the economic interpretation matters: when a private AI lab raises at a higher valuation, part of that increase can appear as profit for the companies that funded it.

Then the lab spends. Anthropic's latest round pushed the company to a reported $965 billion post-money valuation, with AP reporting $47 billion of annualized revenue by May 2026.^[3] TechCrunch reported that part of the round included previously committed hyperscaler investments; the $5 billion Amazon tranche reported in April, paired with a pledge of $100 billion of cloud spending back to AWS, is part of that committed amount rather than a separate additional infusion.^[4] Axios reported that Anthropic also agreed to secure up to 5 gigawatts of compute from Amazon.^[5]

This is why "circular" is the right word. It is not that every dollar returns to its origin. It is that the same small set of firms increasingly appear as investors, suppliers, customers, and accounting beneficiaries in the same transaction web.

II. Circular Does Not Automatically Mean Fake

There is a lazy version of the critique that says circular financing proves the AI boom is fraudulent. That is too strong. Capital-intensive technology buildouts often have vendor financing, pre-purchase commitments, and strategic investments. Telecom networks, aircraft manufacturing, semiconductor fabs, and cloud infrastructure all use variations of this pattern because supply is scarce and capacity must be reserved years ahead of demand.

The better critique is narrower: circular financing becomes dangerous when it lets the system recognize the appearance of demand faster than real external demand materializes. A chipmaker investing in a cloud company that buys its chips can be rational. A cloud provider funding a model lab that buys its compute can be rational. A model lab selling tools to enterprises can be valuable. But each step should eventually be justified by customers outside the loop earning more money, not just by the next repricing round.

III. How We Would Actually See Productivity

The cleanest macro definition of labor productivity is output per hour worked. The Bureau of Labor Statistics measures nonfarm business productivity that way. At the firm level, the equivalent questions are more practical: is revenue per employee rising faster than peers? Are gross margins or operating margins expanding? Is headcount growing more slowly than output? Are sales cycles shortening? Are claims, underwriting, support, engineering, or content workflows producing more billable output rather than just more internal drafts?

That is why "hours saved" is not enough. A knowledge worker saving six hours a week is useful, but it is not automatically economic value. The saved time has to be converted into something a customer pays for, a cost line that disappears, or a product cycle that accelerates enough to raise revenue. Otherwise the gain is absorbed as more polish, more meetings, more optionality, or higher expectations.

IV. The Current Evidence Is Mixed

At the aggregate level, there is not yet a broad AI productivity breakout. BLS revised U.S. nonfarm business labor productivity down to a 0.3% annual rate in Q1 2026 (output up 1.0%, hours up 0.7%), while total factor productivity in the private nonfarm business sector increased 0.8% for 2025.^[8] That is not a collapse, but it is even further from the visible step-change implied by the scale of AI capital spending.

Enterprise surveys and studies tell the same divided story. MIT's NANDA initiative, as reported by Fortune and other outlets, found that about 5% of enterprise generative AI pilots achieved rapid revenue acceleration while the vast majority produced little to no measurable P&L impact.^[9] PwC's 2026 AI Performance study found that 74% of AI's economic value is captured by only 20% of organizations, with leaders using AI for growth and business-model change rather than only isolated productivity pilots.^[10]

Goldman Sachs reached a similar conclusion from public-company earnings commentary: no meaningful relationship yet between AI adoption and productivity at the economy-wide level, even while some specific use cases show much stronger gains.^[11] This is exactly what an early general-purpose technology should look like if the technology is real but the organizational adaptation is lagging. It is also exactly what a bubble can look like before the external demand fails to arrive.

V. The Difference Between Selling AI and Using AI

There are many examples of companies growing revenue because they sell AI. Anthropic, OpenAI, Nvidia, ServiceNow, Salesforce, and cloud providers can all point to AI as a top-line driver. That is real, but it is not the same as proving that AI buyers are becoming more productive.

The harder case is a non-AI company saying: we used AI internally and it increased revenue by X% or expanded margins by Y%, net of implementation costs. Those examples exist, but they are thinner and less auditable than the vendor revenue stories. Klarna's widely cited revenue-per-employee improvement is a useful case, but it is mostly a headcount and operating leverage story, and the company later softened its AI-only customer-service posture after quality concerns.^[12] Insurance underwriting, software development, customer support, and financial operations all have credible workflow wins, but most are reported as throughput, time saved, or cost avoidance rather than clean revenue lift.

VI. The Bull Case

The strongest defense of the AI cycle is historical. General-purpose technologies usually require complementary investment before they show up in productivity statistics. Electricity did not transform factories until layouts, motors, labor practices, and management systems changed. Computers were visible everywhere before they were visible in productivity. AI may be in the same pre-reorganization phase.

On this view, the current circularity is not a red flag by itself. It is the financing structure of a supply-constrained buildout. The labs need compute before customers can rely on the products. The cloud providers need anchor tenants before building gigawatts of capacity. Enterprises need time to move from copilots to redesigned workflows. The productivity proof comes late because the first generation of spending buys experimentation and infrastructure, not immediate transformation.

There is evidence for this more optimistic story. Individual software developers, analysts, consultants, and support workers do report task acceleration. Coding agents are now one of the clearest paid use cases. Anthropic's reported revenue growth suggests real customer willingness to pay for Claude and Claude Code.^[3][4] PwC's data does not say nobody is getting value; it says value is concentrated among firms that reorganize around AI rather than sprinkling tools on top of old processes.^[10]

VII. The Bear Case

The bear case is that the system is confusing internal demand for final demand. A model lab buys compute because investors fund the lab. The cloud provider recognizes revenue from the lab and a valuation gain on its stake. The lab's valuation rises because revenue is growing and strategic investors are competing for exposure. The resulting earnings and market capitalization then support more infrastructure spending.

This can work for a while even if end customers are not earning the productivity gains needed to justify the full buildout. The danger is not that there is no demand. The danger is that the demand is partly endogenous to the financing structure.

The warning sign would be a widening gap between AI vendor revenue and buyer outcomes. If enterprises keep spending more on AI while most still cannot identify revenue growth, margin improvement, or headcount leverage, then the system is pulling future productivity forward into current valuations without enough proof that the productivity will arrive.

VIII. What Would Change the Verdict

The AI circular economy becomes much less concerning if three things happen over the next several years.

First, revenue per employee should rise outside AI-native companies. Not just at Nvidia, Anthropic, Microsoft, Amazon, and Google. The stronger proof would come from banks, insurers, retailers, manufacturers, logistics firms, healthcare providers, and professional-services companies showing higher output per worker relative to non-adopters.

Second, AI spend should move from pilot budgets into redesigned operating models. The productivity proof will not come from giving every employee a chatbot subscription. It will come from changing workflows: fewer handoffs, shorter queues, automated exception handling, faster product cycles, and pricing models that convert speed into revenue.

Third, reported AI gains should reconcile to financial statements. Dashboards showing hours saved should be treated as leading indicators, not final evidence. The final evidence is revenue, margin, working-capital efficiency, churn reduction, or lower cost per transaction.

Conclusion

The clean version of the AI productivity story is still possible: AI makes workers faster, companies ship more, customers receive more value, revenue rises, margins expand, and the massive infrastructure buildout is justified after the fact. That is the durable circular economy: infrastructure enables productivity, productivity funds demand, demand supports infrastructure.

But the current version is not yet that clean. The strongest evidence sits inside the AI supply chain itself. The chipmakers, cloud providers, and model labs are generating revenue from each other. Private-market valuations are producing public-company gains. Enterprises are buying the tools. What remains weak is the broad, auditable proof that the buyers are becoming more productive businesses.

That is the whole tension of AI in 2026. The loop is real. The productivity dividend may be real. But the dividend has not yet shown up clearly enough to close the loop.