The AI Industry Is Moving Money in Circles and Calling It Growth

Let me explain what is happening in the AI industry right now using the simplest possible terms, because the people profiting from this are counting on you not understanding it.

NVIDIA invests in AI startups. Those startups use the money to buy NVIDIA chips. NVIDIA books the purchases as revenue. The revenue makes NVIDIA's stock go up. The higher stock price gives NVIDIA more capital. NVIDIA invests in more startups. Those startups buy more chips. For every $10 billion NVIDIA invests in OpenAI, they get $35 billion in GPU purchases back. Nobel economist Paul Krugman called it a "financial ouroboros," which is a polite academic way of saying a snake eating its own tail and calling it dinner. Bernstein analyst Stacy Rasgon said the deals "will clearly fuel circular concerns." Jay Goldberg of Seaport Global compared it to asking your parents to co-sign your mortgage. NVIDIA has now pumped $30 billion or more into AI startups including OpenAI, Anthropic, xAI, CoreWeave, Mistral, and dozens of others. They also hold a 7% stake in CoreWeave, which is one of NVIDIA's largest customers. Lambda, another portfolio company, buys NVIDIA chips with borrowed money collateralized by the GPUs themselves, then leases compute time back to NVIDIA. Read that sentence again if you need to.

I first wrote about this circular financing dynamic four months ago, when NVIDIA committed $100 billion to OpenAI. What has happened since then has only confirmed the thesis: the circle is getting wider, the numbers are getting larger, and the gap between what is being spent and what is being earned continues to grow.

And NVIDIA is not even the most brazen one. Microsoft invested $13 billion in OpenAI. That sounds like the biggest bet in tech history. Except OpenAI turned around and spent $12.4 billion of it on Microsoft Azure compute. That number exceeded OpenAI's total revenue for the same period. Microsoft also collects a 20% revenue share on all ChatGPT and API revenue, which pulled in $865 million through Q3 2025. And OpenAI has committed to $250 billion in future Azure purchases. So Microsoft created a customer, called it a partnership, and is now sitting back collecting rent on both sides of the transaction. Amazon put $8 billion into Anthropic across three rounds, and Morgan Stanley estimates AWS will generate $1.28 billion from Anthropic this year alone, growing to $5.6 billion by 2027. Then Amazon booked a $9.5 billion pretax paper gain in a single quarter from Anthropic's rising valuation. Google did the same thing: $3 billion invested in Anthropic, $10.7 billion in "equity securities gains" in Q3 2025. Unrealized. Paper. Not money anyone can spend or touch or eat for dinner. But it shows up on the income statement looking exactly like profit, and nobody on the earnings call is asking follow-up questions about it.

The total mapped commitments are staggering. OpenAI alone has disclosed $1.09 trillion in infrastructure commitments across Microsoft Azure, Oracle Cloud, NVIDIA, AMD, and Amazon AWS. In 2025, 70% of all VC funding went to AI. One out of every two venture dollars in the past two years flowed into AI startups. And most of it flowed right back to the same cloud providers and chip makers that invested it in the first place.

Nobody is making money. Everyone is moving money in a circle and calling it growth.

The Most Expensive Vaporware in Human History

And if you think I am being hyperbolic, let me tell you what happened this week. The Information reported that Stargate, the $500 billion AI data center project that Donald Trump announced at the White House in January 2025, the one he called "the largest AI infrastructure project in history," has not staffed up and is not developing any of OpenAI's data centers. Thirteen months after the announcement. Zero employees hired for the joint venture. OpenAI, Oracle, and SoftBank spent the entire year fighting over who would own what and who would control what and who would pay for what, and the whole thing devolved into what The Information describes as a three-way tug of war that went nowhere. OpenAI wanted to build its own data centers, which would have been the smart move for their long-term independence, but they could not get financing because lenders looked at a cash-burning company without a proven long-term business model and said no. Analysts project OpenAI could run out of cash by mid-2027. So they went back to the negotiating table with their Stargate partners, tail between their legs, and spent months in marathon sessions in Tokyo trying to work it out with SoftBank's Masayoshi Son. They eventually settled on bilateral deals that bypass the Stargate framework entirely. The $500 billion project announced at the White House by the President of the United States was, functionally, a press release.

And this week, OpenAI also revised its cash burn projections upward by $111 billion. They now expect to burn $665 billion cumulatively through 2030. That is $218 billion between 2026 and 2029, which Sherwood News helpfully noted is 23 times what Tesla burned during its entire cash-incinerating phase from 2007 to 2018. Their adjusted gross margins dropped to 33%, well below their 46% target. They spend $1.69 for every dollar of revenue they bring in. Training costs alone are projected to hit $440 billion through 2030. Sam Altman had publicly floated $1.4 trillion in total infrastructure spending; that has now been quietly walked down to about $600 billion, which is still a number so large it does not mean anything to a normal human being but represents a 57% reduction from what the CEO was saying out loud just months ago. And the main beneficiaries of all that spending? Microsoft, Amazon, and Oracle. The money goes out OpenAI's front door and comes right back in through the side doors of the companies that invested in it.

Creative Writing for Capitalists

Now layer the valuation scheme on top of this, because the circular capital is just the engine. The Wall Street Journal documented how AI startups are manufacturing billion-dollar valuations through tiered investment rounds where different investors pay wildly different prices for the same company, sometimes days apart. About 20 of these deals have happened in the past year, and the mechanics are the kind of thing that sounds illegal until you find out it is technically not, which should tell you something about our securities laws.

A prestigious VC "preempts" a round by investing at a lower valuation. Days later, other investors come in at a dramatically higher price. The startup announces the higher number. The lead investor's stake doubles on paper overnight. A company called Serval raised its Series A at $220 million with about $1 million in annual recurring revenue. Sequoia invested in early December at under $400 million. Days later, December 11, Serval announced a $75 million Series B at $1 billion. A $600 million increase in less than a week. Nothing changed about the business. The deal was structured that way. Another company, Aaru, literally had two tiers in the same round: some investors paid roughly $450 million, others paid $1 billion. The blended valuation fell below $1 billion, but the company reported the higher number. Their revenue was under $10 million. That is a 100x revenue multiple at the headline price, which is not a valuation. It is a creative writing exercise.

Chris Douvos of AHOY Capital said this practice "absolutely does inflate valuations" and described it as a way to "weaponize a startup's balance sheet to try to anoint a winner and suck all the air out of the room." Peter Wendell of Sierra Ventures called it a method to "make it look like you have a high valuation" while the lead gets a sweetheart deal underneath. Stanford professor Ilya Strebulaev studied 135 unicorns using actual legal filings and found reported valuations average 48% above fair value. Nearly half the companies valued above $1 billion are not actually worth that. Common shares, the kind employees hold, were overvalued by 56%. But PitchBook reports the headline number. CB Insights reports the headline number. TechCrunch reports the headline number. None of them verify whether the transaction was arms-length. The formula is just the latest share price times total shares, which applies the most expensive preferred class to all shares including the common stock held by employees that has none of those protections. And then VCs mark their portfolios to that number, and use those marks to raise bigger funds, and deploy that capital at even higher prices. That is not a market. That is a closed system manufacturing its own evidence of value.

The Startups Are Already Dying

And the startups being valued this way? They are dying. Eighty-five percent of AI startups are expected to be out of business within three years. The cohort launched in 2022 burned through $100 million in three years, double the cash-burn speed of earlier generations. AI startups have a median lifespan of about 18 months before shutting down or pivoting. Funding dropped 23% in Q1 2025, the sharpest quarterly drop since the 2018 crypto winter. The "AI Hype" has been replaced by "AI Exhaustion," and companies that were valued at hundreds of millions are quietly folding because they did not build businesses, they built features on top of someone else's infrastructure. Noogata had marquee customers and solid backing but deals stayed small and stuck in pilot mode until they shut down. Wuri went through Y Combinator and pivoted three times without finding product-market fit. Builder.ai's AI turned out to be fake. The median gap between funding rounds has stretched to 696 days. The golden age of "raise money, figure it out later" is over and the entire class of startups these valuations were applied to is evaporating in real time.

The People at the Top Always Get Liquid First

Be honest with yourself for a second about what this means for the people who are not in the room when these deals are structured.

When a startup announces a $1 billion round, that number becomes the anchor for everything downstream. It triggers a mandatory 409A revaluation. Higher preferred stock prices pull up the fair market value of common stock, which directly sets the exercise price for employee options. Employees granted options after an inflated round need the company to achieve even more growth just to break even. And the tax trap is real. When employees exercise incentive stock options, the spread between strike price and fair market value triggers Alternative Minimum Tax even if the shares cannot be sold. If the company craters, the employee has paid real taxes on phantom income. A WeWork employee took out a loan to exercise at roughly $50 per share before the IPO. When the IPO collapsed and shares repriced to about $4, they still owed a six-figure tax bill with no liquidity to pay it. Adam Neumann had cashed out $700 million and walked away with a $1.7 billion exit package.

This is the pattern. Every single time. The people at the top get liquid before the music stops. The people at the bottom hold paper.

Hopin peaked at $7.75 billion before selling its core assets for $15 million. 99.8% decline. The founder sold $140 to $200 million in personal shares before the crash. 1,300 employees watched their equity evaporate. Fast paid employees $200 to $240K base plus equity that looked like golden tickets. The company was generating $600,000 in revenue against a $10 million monthly burn. It shut down entirely. IRL was valued at $1.17 billion after SoftBank led a $170 million round. An internal investigation found 95% of its 20 million users were bots. The SEC charged the CEO with fraud. Most option plans give employees 90 days to exercise after departure. If the 409A has ballooned, exercising might cost hundreds of thousands of dollars for illiquid shares that cannot be sold. One tech employee described options with $3 million in paper value that would require $300,000 cash to claim. So you either come up with the money for shares you cannot sell in a company whose valuation may be fiction, or you walk away from what you were told was your upside for years of work.

Heads I Win, Tails You Hired McKinsey

And the venture capital business model is designed to reward all of this. The standard 2-and-20 structure means a $1 billion fund generates $200 million in fees over its life before producing a single dollar of returns. Carry calculations are performed "as if the fund had realized all assets at their reported fair value," meaning paper markups flow directly into carry before anyone can actually sell anything. The median TVPI for VC funds raised 2018 to 2020 is 1.8x, but median DPI, actual cash returned, is only 0.4x. Seventy-eight percent of reported returns are unrealized. Only 9% of 2021-vintage funds returned any capital within three years. More than 60% of 2019-vintage funds returned nothing after five years. The VCs inflate, the startups cooperate, the trackers and press have no mechanism to verify, and the people whose money is at risk are making allocation decisions based on numbers that are 48% higher than reality. CalPERS posted venture returns of 0.49% annualized from 2000 to 2020. Negative 24.8% in fiscal year 2023. And they are expanding their venture allocation sixfold to up to $5 billion. Right now. At the top. Howard Marks put it best: "You can't eat IRR."

Everyone Knows. Nobody Cares. The Story Is Too Good.

Now zoom out even further, because the circular capital, the fake valuations, and the dying startups are only part of it. The hyperscalers funding this whole operation are themselves spending money they do not have. The top five (Amazon, Microsoft, Google, Meta, and Oracle) are projected to spend $600 to $690 billion on infrastructure in 2026. That is a 36% increase over 2025. Capital intensity has hit 45 to 57% of revenue, levels that look more like utilities or industrials than tech companies. And they can no longer fund it from cash flow. Aggregate capex, after buybacks and dividends, now exceeds projected cash flows. Morgan Stanley expects hyperscaler borrowing to top $400 billion this year alone, more than double 2025. JP Morgan projects $1.5 trillion in total tech sector debt issuance over the coming years. These companies went from cash-funded software businesses to leveraged infrastructure companies in about 18 months, and nobody seems to have noticed.

Against that $600 billion in spending, AI-related services are expected to generate about $25 billion in revenue in 2025. That is a 24-to-1 ratio of infrastructure spending to revenue. Only 25% of AI initiatives have delivered expected ROI. Fewer than 20% have been scaled across enterprises. MIT's NANDA initiative found that only 5% of generative AI pilot programs achieve rapid revenue acceleration. Five percent. The vast majority stall, delivering little to no measurable impact on the income statement. Bain calculated the AI economy needs $2 trillion in annual revenue by decade's end to justify current capex. Best-case forecasts say $1.2 trillion. That is an $800 billion gap. Everyone knows. Nobody cares. The story is too good.

Same Scam, Different Logos

I know people get tired of hearing "this is like 2008" or "this is like the dot-com bubble." But I need you to sit with the specifics because these are not vibes. They are structural replicas.

Cisco extended $2.4 billion in credit to telecom startups in the late 1990s so they could buy Cisco equipment. Cisco booked it as revenue. By 2001, $900 million written off, $2.2 billion inventory write-down, stock fell 88%. Twenty-five years later still has not recovered. Lucent committed $8.1 billion in vendor financing. The SEC charged them with manipulating over $1 billion in revenue. Over 20 telecoms failed. More than $1 trillion in unrecoverable debt. The capacity swaps between Global Crossing, Qwest, and Enron Broadband used the same trick: record the "sale" as revenue immediately, capitalize the "purchase" over years. Qwest restated more than $2 billion. CEO got six years. WorldCom CEO got 25. Gary Winnick sold $734 million in stock before the collapse and paid a $55 million settlement with no criminal charges.

CDO tranching did the same thing these tiered rounds are doing now. Same pool of subprime mortgages, different tranches, different ratings depending on payment priority. BBB-rated tranches re-pooled into CDOs where two-thirds were rated AAA. By December 2008, 80% of those AAA CDOs were junk. CDOs accounted for $542 billion of the nearly $1 trillion in institutional losses. In tiered AI rounds, the same underlying company gets priced at dramatically different levels depending on which slice you buy, with the headline number used to validate the entire thing.

SoftBank invested in WeWork at $20 billion, then $20 billion again, then $42 billion, then $47 billion. Each time the only investor willing to pay. WeWork was losing $219,000 every hour. SoftBank's cumulative losses hit $14.2 billion. Across the Vision Fund, $53 billion in losses over two years, wiping out every dollar since launch. And now SoftBank is back, co-leading Stargate, the $500 billion project that has not hired anyone, while booking a $4.2 billion gain on its OpenAI stake. Masayoshi Son is running the same playbook with different logos.

The historical pattern has a 100% hit rate over 30 years: these schemes run three to five years before exposure, total collapse within 12 to 24 months of the first cracks, criminal charges two to five years later. Jim Chanos: "The fraud cycle always follows the financial cycle with a lag." Bill Gurley warned that AI burn rates are "bigger than they've ever been in the history of venture capital." Aswath Damodaran said the AI market must generate $4 trillion in revenues to justify current investment, then fully exited NVIDIA. Howard Marks noted that Thinking Machines Lab raised $2 billion at $10 billion without a product or even telling investors what it was building. Lisa Shalett at Morgan Stanley warned of a "Cisco moment" within 24 months. Over 50 securities class actions alleging false AI statements have been filed in the past five years. The SEC created a dedicated unit and listed AI as a top examination priority for 2026. And the company at the center of the entire thesis, OpenAI, just revised its cash burn upward by $111 billion, cannot get its flagship infrastructure project off the ground, might run out of cash by mid-2027, and cut its total spending projection from $1.4 trillion to $600 billion without anyone in the press asking why the number dropped by 57%.

These Are the Cracks

We are in the exposure phase. The Stargate stall. The cash burn revisions. The pilot failure data. The 48% valuation overstatement. The growing chorus of the most credible names in finance all saying the same thing at the same time. These are the cracks.

I write about this because neurocollective exists at the intersection of AI hype and AI reality. The technology is real. AI works. We see it every day in the organizations we work with. But 95% of enterprise AI pilots are failing to deliver measurable ROI, not because the technology does not work but because the entire information environment has been corrupted by the financial machinery sitting on top of it. When every startup claims to be a unicorn, when every vendor claims 10x ROI, when every investment round is reported as validation of the technology rather than validation of a financial structuring trick, leaders cannot make rational decisions. The urgency you feel to "not get left behind" is partly manufactured by companies that need you to spend money to justify their valuations. The gap between what AI can genuinely do for your organization and what the financial machinery says it is worth has never been wider. Navigating that gap is the entire game. And the people manufacturing these valuations have every incentive to make sure you cannot tell the difference.