Table of Contents
The OpenAI Nvidia AMD deals announced this week have sent shockwaves through Silicon Valley, not for their innovation, but for what they reveal about the precarious financial architecture propping up the AI boom. OpenAI will purchase $500 million in AI servers from both Nvidia and AMD, funded in part by investments those very chipmakers made in OpenAI’s recent $40 billion funding round. It’s a circular arrangement that has analysts drawing uncomfortable parallels to the vendor financing schemes that preceded the dot-com crash.
The artificial intelligence industry’s explosive growth is starting to look less like a revolution and more like a hall of mirrors. OpenAI’s recent blockbuster partnerships with Nvidia and AMD, worth a combined $160 billion, have ignited a debate that Wall Street can no longer ignore: Is the AI boom real, or are we watching the same money circle through the same handful of companies, inflating valuations to unsustainable heights?
The deals themselves are staggering in scale. In late September, Nvidia announced it would invest up to $100 billion in OpenAI to fund a massive data center buildout, with OpenAI committing to fill those facilities with millions of Nvidia chips. Two weeks later, OpenAI struck a similar arrangement with AMD, agreeing to deploy 6 gigawatts of AMD processors while securing the option to acquire up to 10% of the chipmaker.
On the surface, these partnerships signal confidence in AI’s transformative potential. Look closer, and the circular nature of the financing becomes impossible to ignore.
The OpenAI Nvidia AMD Deals: A Circular Financing Model
Here’s how the money flows: Nvidia invests $100 billion in OpenAI. OpenAI uses that capital to buy Nvidia chips. Nvidia books the revenue. OpenAI also purchases cloud computing from Oracle, which buys chips from Nvidia. Nvidia holds a stake in CoreWeave, a data center operator that supplies infrastructure to OpenAI and has spent $7.5 billion on Nvidia GPUs. SoftBank, which owns a $3 billion stake in Nvidia, is partnering with OpenAI and Oracle on the $500 billion Stargate data center project, where Nvidia serves as a “core technology partner.”
It’s a tightly wound ecosystem where the same dollars appear to be counted multiple times, flowing between a small circle of companies that are simultaneously customers, suppliers, and investors in one another.
“You don’t have to be a skeptic about AI technology’s promise in general to see this announcement as a troubling signal about how self-referential the entire space has become,” wrote analysts at Bespoke Investment Group. “If Nvidia has to provide the capital that becomes its revenues in order to maintain growth, the whole ecosystem may be unsustainable.”
The parallels to the late 1990s dotcom bubble are hard to miss. Back then, telecom equipment giants like Lucent Technologies and Nortel Networks extended billions in loans to cash-strapped telecom startups so they could purchase equipment. Lucent committed $8.1 billion in vendor financing, Nortel extended $3.1 billion, and Cisco promised $2.4 billion in customer loans. When the bubble burst in 2000, 47 competitive local exchange carriers went bankrupt, and the equipment makers were left holding worthless debt. The Nasdaq fell 77%, wiping out trillions in market value.
Vendor Financing, 2025 Edition
Leading British tech investor James Anderson told The Guardian he sees uncomfortable echoes of that era. “It’s not quite like what many of the telecom suppliers were up to in 1999-2000, but it has certain rhymes to it,” Anderson said. “I don’t think it makes me feel entirely comfortable from that point of view.”
The scale of Nvidia’s exposure dwarfs what Lucent faced. Nvidia’s direct investments in AI companies total $110 billion, representing 67% of its annual revenue. By comparison, Lucent’s vendor financing commitments peaked at 24% of revenue. Neil Wilson, UK investor strategist at Saxo, an investment bank, put it bluntly: the situation “looks, smells and talks like a bubble.”
What makes this moment particularly precarious is the emergence of a new financial instrument: GPU-backed debt. CoreWeave alone carries $10.45 billion in debt collateralized by the value of Nvidia’s graphics processing units. Lambda Labs secured a $500 million GPU-backed loan. The entire “neocloud” sector has accumulated more than $15 billion in debt backed by chips that may depreciate faster than anyone wants to admit.
The bet underlying this debt is that GPUs will hold their value over four to six years. But the evidence suggests otherwise. While companies have extended depreciation schedules to six years for accounting purposes, Google architects report that GPUs running at 60-70% utilization in AI data centers typically survive one to two years, with three years as the maximum. Meta’s Llama 3 training experienced 9% annual GPU failure rates, suggesting 27% of chips could fail within three years.
If GPU values collapse, the entire debt structure crumbles, and the companies holding that debt face catastrophic losses.
The Concentration Problem
Nvidia’s customer base is far more concentrated than Lucent’s ever was. Lucent’s top two customers, AT&T and Verizon, accounted for 23% of revenue in 2000. Nvidia derives 39% of its revenue from just two customers and 46% from four customers. A total of 88% of Nvidia’s revenue comes from data centers.
This concentration creates systemic risk. If any of these major customers, particularly OpenAI, stumbles, the ripple effects could be devastating. OpenAI reported $4.3 billion in revenue in the first half of 2025, but posted an operating loss of $7.8 billion, according to The Information. Nearly half of that loss was stock-based compensation, but the company remains deeply unprofitable while taking on massive infrastructure commitments.
OpenAI’s valuation has skyrocketed from $157 billion in October 2024 to $500 billion today. Anthropic, another AI lab, nearly tripled its valuation from $60 billion in March to $170 billion last month. These valuations are predicated on the belief that AI will generate transformative economic returns. But the evidence for that transformation remains elusive.
The Productivity Paradox
In August, MIT published research showing that 95% of organizations are getting zero return from their investments in generative AI. The issue wasn’t the quality of the models but how they were being deployed. McKinsey reported that while eight out of 10 companies use generative AI, the same proportion report no significant impact on their bottom line. AI tools are being used for broad, low-value tasks like producing meeting summaries rather than specific, high-value applications like identifying supply chain risks or generating strategic insights.
This productivity gap is the Achilles’ heel of the AI boom. The entire investment thesis rests on the promise that AI will dramatically improve economic efficiency, allowing a single employee to produce far more value in a working day. If that promise fails to materialize, or if it takes longer than expected, the trillions being poured into AI infrastructure will look like a colossal misallocation of capital.
“The experience of a quarter of a century ago won’t necessarily be repeated, but the scale of recent investment increases by tech firms already indicates that they are taking significant risks,” analysts at Oxford Economics wrote in a recent note. “If it starts to become clear that AI productivity gains, and thus the return on investment, may be limited or delayed, a sharp correction in tech stocks, with negative knock-ons for the real economy, would be very likely.”
The Off-Balance-Sheet Gambit
Adding another layer of complexity, tech companies are increasingly using Special Purpose Vehicles (SPVs) to finance AI data center construction. A hyperscaler like Meta partners with a private equity firm like Apollo, contributing capital to a separate legal entity that builds and owns the data center. The hyperscaler maintains operational control through long-term lease agreements, but because it doesn’t directly own the SPV, the debt remains off its balance sheet.
The appeal is straightforward: companies can pursue massive infrastructure buildouts without alarming credit rating agencies or spooking investors with ballooning debt levels. But it also obscures the true scale of financial exposure. American tech companies are projected to spend $300 billion to $400 billion on AI infrastructure in 2025. How much of that is hidden in off-balance-sheet vehicles is anyone’s guess.
Data center assets now represent 10% to 22% of major REIT portfolios, up from near zero two years ago. The thin equity layer in these SPVs, typically 10% to 30%, means that if data center utilization falls short of projections or if GPUs depreciate faster than expected, equity holders face steep losses before debt holders experience impairment.
The Counterargument
Not everyone is sounding the alarm. Nvidia’s defenders point out that the company is fundamentally different from Lucent. Nvidia generates more than $50 billion in annual operating cash flow and holds $46.2 billion in net cash. Its credit rating was upgraded to Aa3 by Moody’s in March 2024, while Lucent was downgraded to A3 in December 2000, months before its collapse.
Nvidia’s top customers, Microsoft, Alphabet, Amazon, and Meta, generated a combined $451 billion in operating cash flow in 2024. These are not cash-strapped startups burning through venture capital. They are among the most profitable companies in history, and they are betting their futures on AI.
There are also clear signs of genuine demand. Microsoft and AWS both report AI capacity constraints. OpenAI’s ChatGPT is now used by 800 million people a week, up from 500 million in March. Labor market data shows wages rising twice as fast in AI-exposed industries, and workers using AI report performance improvements of up to 40%.
Unlike the telecom bubble, where demand was speculative and customers were burning cash, the AI boom is being driven by companies with real revenues and real customers. The question is whether those revenues will grow fast enough to justify the trillions being invested.
What Comes Next
The AI industry is at a crossroads. The infrastructure buildout is unprecedented in scale, the financial arrangements are increasingly complex and circular, and the productivity gains that would justify it all remain frustratingly out of reach for most organizations.
Sam Altman, OpenAI’s CEO, has sought to calm fears, suggesting that booms and busts are part of every industry’s life cycle. “Between the ten years we’ve already been operating and the many decades ahead of us, there will be booms and busts,” Altman said during a tour of OpenAI’s massive data center complex in Abilene, Texas. “People will over-invest and lose money, and underinvest and lose a lot of revenue.”
That may be true. But for investors, employees, and the broader economy, the difference between a healthy correction and a catastrophic collapse is everything. The dotcom bubble didn’t just wipe out overvalued startups. It destroyed trillions in wealth, triggered a recession, and left a generation of investors scarred.
The AI boom may yet deliver on its promise. The technology is real, the applications are expanding, and the potential is enormous. But the circular financing, the soaring valuations, the off-balance-sheet debt, and the productivity paradox are all flashing warning signs that this boom is built on shakier ground than its champions want to admit.
As more money flows through the same small circle of companies, the line between genuine growth and financial engineering grows harder to discern. And when that line disappears entirely, history suggests, the reckoning is swift and brutal. The question isn’t whether AI will change the world. It’s whether the companies building it can survive long enough to see it happen. For a deeper look at how hardware innovation is reshaping the tech landscape, see our coverage of Qualcomm’s acquisition of Arduino.