The AI Bubble’s Hidden Feedback Loop: Circular Financing, Options Trading and Why the Boom Keeps Feeding Itself

The AI boom is generating extraordinary revenue. Investors should ask how much represents independent end demand, and what happens when that financing system meets a public market where record options activity, leverage and mechanical hedging increasingly influence short-term price formation.

There is a number buried in corporate spending data that deserves far more attention than another benchmark score or model launch.

In July, the most intensive 1% of corporate AI adopters in Ramp's data spent almost $8,000 per employee per month on artificial intelligence. By August the figure had fallen to $7,205, still hundreds of times the $11.95 spent by the median AI adopting company. Ramp's lead economist has gone further: roughly 1% of customers account for about 80% of the OpenAI and Anthropic enterprise revenue visible in its data, an extraordinary degree of concentration compared with other software categories Ramp tracks. Those largest customers skew heavily towards technology companies and businesses selling AI products and services themselves.

That concentration deserves a different interpretation from the standard story of explosive enterprise adoption.

It suggests that the marginal dollar of AI revenue may be much more financially reflexive than investors assume.

A normal technology adoption cycle begins with capital, produces a product, finds an independent customer and earns revenue from that customer. Parts of the AI economy increasingly look different. Capital providers finance AI companies. Cloud and semiconductor suppliers invest in model companies. Model companies commit enormous sums back to those suppliers for compute. Venture backed application companies receive subsidised API access and spend their funding on inference. Infrastructure providers borrow against long term contracts with the same small group of AI customers.

The money moves. Revenue appears. Capacity gets built. Valuations rise.

The unresolved question is how much genuinely independent economic demand sits at the end of the chain.

There may also be a second feedback loop above it. The public market valuing this infrastructure has become increasingly derivative intensive. Options positions create hedging requirements, those hedges generate transactions in the underlying market, and movements in the underlying market change the required hedge again. Financing can therefore reinforce AI demand while market structure can simultaneously reinforce, suppress or reverse price movements.

Follow the money, not the tokens

The first clue is the extraordinary dispersion in corporate AI spending.

Ramp's Summer 2026 data found that the top 1% of AI adopters spent more than $7,400 per employee per month, compared with $650 for the top 10% and just $11.95 for the median adopter. Ramp subsequently revised July's top 1% figure to approximately $8,000.

Spending of that magnitude can certainly be economically rational for an AI company whose product sends millions of customer requests through a model API. Per employee spending is therefore an imperfect measure of excess. But that observation leads to a more important point: much of the industry's enterprise demand appears concentrated among companies whose own economics depend on the continued expansion of AI.

That is very different from broad, diversified demand across unrelated industries.

If an insurance company buys database software, the database vendor is exposed primarily to the insurer's need for databases. If an AI application company raises venture capital, spends the proceeds on model inference, sells another AI product into an ecosystem financed by the same capital markets and then raises again at a higher valuation, several supposedly independent links in the revenue chain share the same funding source.

Revenue remains revenue. Its resilience can still vary dramatically according to where the money originated.

This distinction should matter to investors as much as ARR growth itself.

The subsidy is already visible

There is also direct evidence that venture backing and model consumption have been deliberately connected.

OpenAI's startup programme tells founders that portfolio companies backed by participating venture firms including Thrive Capital, Sequoia, Andreessen Horowitz, Kleiner Perkins and Conviction may receive free API credits, higher usage limits and technical support. OpenAI explicitly tells startups seeking API credits to contact their venture investor.

Anthropic operates a similar programme. Its VC partner programme allows investment firms and accelerators to provide portfolio companies with API credits, while its startup programme says qualifying founders can receive free Claude credits and priority limits.

These programmes do not prove that venture firms are secretly manufacturing model revenue. They establish something narrower and more useful: AI consumption is deliberately subsidised through the venture ecosystem.

That changes the economics of early demand.

Give thousands of venture backed companies subsidised compute and they will experiment more aggressively, build products with higher inference intensity and become customers earlier than they otherwise would. When the credit expires, the relevant question is whether the resulting product generates enough outside cash flow to support the same consumption at market prices.

The venture industry has always subsidised customer acquisition indirectly by financing losses. AI adds another layer because one of the subsidised inputs is itself becoming a major reported revenue stream elsewhere in the same investment ecosystem.

The distinction sounds technical. Financial bubbles are often built from technical distinctions investors initially decide do not matter.

The much bigger circularity sits above venture capital

The strongest evidence for circular AI finance no longer comes from startup credits. It comes from the largest companies in the industry themselves.

Consider Microsoft and OpenAI.

Microsoft's 2026 annual filing says it had funded $11.9 billion of a $13 billion investment commitment to OpenAI and held roughly a quarter of the company on an as converted basis. The same filing says Microsoft recorded $24.1 billion of fiscal 2026 revenue from commercial arrangements with OpenAI, including revenue sharing payments.

Separately, OpenAI committed to purchase an incremental $250 billion of Azure services. Microsoft therefore sits simultaneously in several positions: investor, infrastructure supplier, technology partner, revenue participant and major economic beneficiary of OpenAI's growth.

Now look at Amazon.

In February 2026 Amazon agreed to invest $50 billion in OpenAI. At the same time, OpenAI expanded an existing $38 billion AWS agreement by another $100 billion over eight years and committed to consuming roughly two gigawatts of Amazon Trainium capacity. Capital moves from Amazon into OpenAI while contractual compute spending moves from OpenAI towards Amazon.

Nvidia supplies another striking example. In September 2025 it announced plans to invest as much as $100 billion in OpenAI as OpenAI deployed up to 10 gigawatts of Nvidia systems. In OpenAI's February 2026 financing, Nvidia separately committed another $30 billion while expanding dedicated inference and training capacity for OpenAI.

Anthropic follows a similar architecture.

Amazon invested a total of $8 billion in Anthropic while becoming its primary cloud and training partner. Anthropic committed to AWS Trainium and Inferentia infrastructure.

Then, in November 2025, Microsoft and Nvidia agreed to invest up to $5 billion and $10 billion respectively in Anthropic. In the same announcement, Anthropic committed to purchase $30 billion of Azure compute and contract up to another gigawatt of Nvidia based capacity.

The accounting can be entirely legitimate while the economics remain reflexive.

A supplier invests in its customer. The customer signs a large contract with the supplier. The contract supports the supplier's revenue expectations. The customer's access to capital improves because investors see enormous capacity, strategic backing and rapid growth. Rising valuations increase the economic value of existing stakes. Fresh capital finances another round of infrastructure commitments.

Each transaction can make sense individually. Viewed together, they form a financing machine.

Sona Asset Management recently attempted to map that machine. Its analysis, reported by the Financial Times, identified 176 AI related transactions involving 202 entities with a headline value of roughly $3.6 trillion over three and a half years. Around 120 received what Sona called a high circularity flag because companies were connected through multiple investment, financing, supply and customer relationships. The $3.6 trillion figure includes future commitments and financing arrangements, so it should not be confused with cash already spent. The network structure is the important finding.

AI increasingly finances itself through its own value chain.

Revenue quantity is obscuring revenue quality

Investors have spent three years asking how quickly AI revenue can grow. They should start asking what kind of revenue is growing.

There is a considerable difference between a dollar ultimately funded by a hospital automating administrative work and a dollar funded by a venture round raised by an AI startup whose largest operating expense is inference from another venture backed AI company.

Both dollars can appear as model revenue.

Only one conclusively demonstrates that an external customer has found economic value worth paying for from internally generated cash flow.

This suggests a set of metrics the industry rarely reports.

How much model revenue comes from profitable customers rather than externally financed customers? How concentrated is inference spending among AI companies? How much infrastructure backlog ultimately depends on OpenAI, Anthropic and a small set of hyperscalers? What percentage of demand would survive if venture funding and strategic investment stopped growing?

Revenue growth without answers to those questions tells investors less than it appears to.

The relevant metric may eventually become something resembling "capital independent revenue": spending ultimately supported by cash generated outside the AI financing ecosystem.

That denominator would be uncomfortable. It would also be useful.

The financial leverage is moving into infrastructure

The circularity becomes more consequential once debt enters the system.

CoreWeave illustrates the mechanism. The AI cloud provider says it finances infrastructure primarily with asset level debt supported by take or pay customer contracts. At the end of 2025 it reported $60.7 billion of remaining performance obligations. By June 2026, total indebtedness had reached $35.6 billion.

OpenAI alone committed as much as $11.9 billion to CoreWeave through 2030 under a capacity agreement. CoreWeave disclosed that a special purpose vehicle would hold infrastructure and incur debt to finance equipment supporting that contract.

The structure turns expected future AI demand into collateral for present infrastructure spending.

That works beautifully while the customer remains well financed and demand keeps expanding. It becomes much less forgiving when utilisation disappoints, pricing declines or capital markets demand higher returns.

The financing engineering is spreading. The Financial Times reported in September that technology companies are using residual value guarantees and special purpose structures to support as much as $300 billion of AI data centre and chip financing, allowing significant exposure to sit outside conventional corporate debt measures.

Physical constraints make the timing more dangerous.

The International Energy Agency says new transmission lines can require four to eight years to build in advanced economies, while waiting times for critical equipment such as transformers and cables have roughly doubled in recent years. Half of US data centres under development are being built inside existing large clusters, increasing the risk of local power bottlenecks.

Meanwhile, the IMF has identified another mismatch. Its 2026 financial stability work estimates that reported depreciation policies across major hyperscalers imply average property and equipment lives of roughly seven years, while advanced GPUs may become economically obsolete within roughly two years.

That gap matters because the AI boom is converting a software narrative into a capital intensive industrial system.

Software can be rewritten. Debt must be serviced.

The loop eventually meets a price

Ramp's latest numbers provide an early warning about what happens when economics begin to intrude.

Spending among its top 1% of AI adopters fell 9.7% in August, from $7,976 to $7,205 per employee. Ramp also found that the effective price per million tokens had fallen 41% from its 2026 peak by early September. Businesses were increasingly directing workloads towards cheaper standard models when the additional performance of frontier models did not justify the extra cost.

This is how technology markets normally mature. Performance diffuses. Prices decline. Buyers optimise.

For AI finance, however, falling unit economics interact with an infrastructure base being built under assumptions of enormous future consumption.

The labs need cheaper inference to stimulate adoption. Infrastructure investors need enough revenue to earn a return on extraordinary capital expenditure. Model companies need enormous compute commitments to remain technologically competitive. Their suppliers increasingly finance them so those commitments can be made.

Every participant has a rational reason to keep the machine accelerating.

The system becomes fragile when acceleration itself becomes an assumption embedded in financing.

The market has built its own feedback loop

The financing structure is only half of the reflexivity story.

Options trading has moved from the edge of US equity markets towards their centre. In July 2026, OCC cleared 1.55 billion options contracts, 25.7% more than a year earlier. Average daily options volume for 2026 through July reached 70.8 million contracts, almost 24% above the comparable 2025 period. Index option volume in July was up 33.6% year on year, while ETF option volume rose 59.1%.

The shift is even more striking in same-day contracts.

Cboe reported that zero-days-to-expiry options accounted for a record 66.2% of all SPX options volume in July 2026. A product that was once peripheral now represents most trading in the world's most important equity index options franchise.

That does not mean two-thirds of the stock market is speculative gambling. Options volume and economic exposure are different things. It does mean that any analysis of modern market behaviour which treats derivatives as a side market is increasingly incomplete.

Short-dated options are unusually sensitive to movements in the underlying index as expiration approaches. That sensitivity matters because dealers generally hedge their options exposure dynamically.

When dealers are long gamma, they tend to sell into rallies and buy into declines to rebalance their hedges. Those flows can resist the prevailing move and encourage mean reversion. When dealers are short gamma, the mechanics reverse: hedging can require buying as prices rise and selling as prices fall, adding momentum to the original move.

The underlying price changes the hedge. The hedge transaction then affects the underlying price.

That is reflexivity expressed through market plumbing.

Research increasingly finds evidence of this mechanism, although the effect is subtler than the popular claim that "0DTE controls the market." Work using proprietary options data has found that positive dealer gamma is associated with stronger intraday price reversal, while negative gamma is associated with greater momentum. More recent research finds that 0DTE positioning can reduce volatility and strengthen order-flow reversals because dealer hedging absorbs part of a market move.

This is important because reflexivity does not always create volatility. It can suppress it.

That distinction helps explain one of the stranger features of the modern market: a violent decline can encounter mechanical buying pressure without any corresponding improvement in the news that caused the decline. A market can fall sharply, find what looks like an inexplicable floor and reverse as hedging flows, positioning and liquidity interact.

The rebound does not prove that fundamentals suddenly improved. Sometimes the transmission mechanism changed.

The reverse can also happen. When positioning leaves dealers short gamma, the same hedging process can add to a decline instead of absorbing it. A market structure that dampens movement under one positioning regime can accelerate it under another.

This is why the explosion of short-dated options matters even if one rejects the claim that they dictate market direction.

Cboe's own work offers an important boundary. Its studies have generally found aggregate net 0DTE dealer hedging small relative to the enormous liquidity of S&P futures. Research published in 2025 estimated such hedging at roughly 0.2% of daily SPX liquidity under typical conditions. That makes it difficult to argue that 0DTE alone explains every rally, sell-off or V-shaped recovery.

The stronger argument is more interesting anyway.

Market prices increasingly emerge from an interaction between information and endogenous flows. Investors respond to prices. Options exposure responds to prices. Dealers respond to changing options exposure. Systematic portfolios respond to volatility. Those transactions feed back into prices, which can trigger another round of responses.

Fundamentals remain the anchor over long horizons. The path between one fundamental anchor and the next has become increasingly financialised.

That matters enormously for the AI trade.

Suppose disappointing AI economics cause a cluster of highly valued technology shares to fall. The first-order question is what the earnings news implies. The second is how one of the most heavily optioned parts of the equity market transmits that shock.

If positioning is stabilising, derivative hedging can absorb part of the decline and produce a sharp reversal. If positioning is destabilising, the same market machinery can accelerate the move. In either case, observing the index rebound tells investors less about the original fundamental problem than it once might have.

A V-shaped recovery is a price pattern. It is not an economic argument.

The danger comes from confusing market resilience with fundamental validation.

The larger post-2008 lesson

This structure did not emerge in a financial vacuum.

After the 2008 crisis and again in 2020, the Federal Reserve used large scale asset purchases to suppress longer term interest rates and ease financial conditions. The policy response was defensible in both emergencies. It also helped establish a generation of investors and executives for whom extreme financial disruption became associated with equally extreme policy intervention.

Fiscal capacity has since deteriorated. The Congressional Budget Office expects US federal debt held by the public to equal 101% of GDP in 2026, with a deficit of $1.9 trillion, or 5.8% of GDP. Net interest expense is projected to exceed $1 trillion this year and rise to $2.1 trillion by 2036.

The significance for AI is simple. The sector is attempting one of the largest private capital formation programmes in modern history at the same moment that governments are entering the next downturn with heavier debt burdens and the private AI infrastructure stack itself is becoming more leveraged.

The assumption that another layer of capital will always arrive deserves a higher discount rate than markets have historically assigned to it.

The bubble can contain a revolution

Calling all AI revenue fake would miss the point.

AI can become one of the most important technologies of the century and still generate one of the great capital misallocations of the century. Railways transformed economies and bankrupted investors. Fibre optic networks became indispensable after the telecom bubble destroyed enormous amounts of capital. The internet justified the technological thesis without validating every price paid in 1999.

The distinction between technological truth and financial truth is where serious investors should focus now.

There is real AI demand. There is also extraordinary customer concentration, subsidised startup consumption, supplier financed customers, enormous reciprocal compute commitments, growing infrastructure leverage and an equity market whose short-term behaviour increasingly reflects derivative positioning alongside economic information.

These are two forms of reflexivity sitting on top of one another.

Inside AI, capital finances spending that becomes somebody else's revenue, supporting valuations that attract more capital. In public markets, prices alter derivative exposures, those exposures generate hedging flows and those flows can alter prices again.

Neither mechanism makes the underlying activity fictitious.

Both can make apparent strength more self-reinforcing than it first appears.

The question is therefore larger than whether OpenAI or Anthropic can keep growing revenue, or whether the Nasdaq can keep recovering from every sell-off.

It is whether AI demand and AI valuations can survive a world in which fresh capital becomes scarcer while the financial mechanisms that have cushioned, recycled and amplified that capital begin working in reverse.

As long as capital keeps arriving, circularity can resemble growth. As long as derivative positioning absorbs shocks, reflexivity can resemble resilience. When those mechanisms reverse, investors discover how much strength came from final demand and how much came from the machinery surrounding it.