This article first appeared in The Edge Malaysia Weekly on August 3, 2026 - August 9, 2026

FOR over a year, there has been concern that a burgeoning circular financing loop in the artificial intelligence (AI) ecosystem and increasing use of off-balance sheet vehicles could undermine the multi-trillion dollar growth engine that has been powering the global economy from AI frontier labs in the US and China, Taiwan’s chip foundries, South Korea’s memory chip makers, and chip packaging and testing firms in Southeast Asia. Until now, investors and the tech industry as well as policymakers believed that funding issues were unlikely to derail the global economy. That changed this past week as these concerns collided with quarterly earnings reports from Big Tech firms, mounting geopolitical risks and macroeconomic and policy pressures.

How serious is the circular financing and off-balance sheet debt? Where are the fault lines? Who is the most vulnerable? What do investors need to watch out for now? Could this finally trigger the long-anticipated bursting of the AI bubble? What can the industry and regulators do to prevent a full-blown crisis?

On July 26, The Wall Street Journal reported that chip behemoth Nvidia was in talks with AI frontier lab OpenAI to provide US$250 billion (RM1.02 trillion) backstop to lease a massive US$500 billion data centre project in Ohio. Nvidia’s guarantee would cover the data centre lease and debt needed to fund its build-out, but not the AI chips that go inside it. There is a separate financing deal between Nvidia and OpenAI for US$350 billion of graphic processing units or GPU chips. Nvidia invested US$30 billion in OpenAI in April. On July 27, Nvidia inked another deal with South Korean chip maker SK Hynix to lock down high bandwidth memory supplies over the next five years. Hynix affiliate SK Telecom will build a cloud business using Nvidia’s Vera Rubin chips.

Nvidia is not alone in such circular deals. Google agreed to backstop lease payments for Anthropic’s data centres in May, effectively helping it secure US$35 billion. And, oh, there is also Advanced Micro Devices’ (AMD) circular deal with OpenAI, which involves a massive infrastructure agreement paired with large equity warrants, drawing comparisons to dotcom-era vendor financing. Under that partnership, OpenAI agreed to deploy 6 gigawatts of AMD chips. In return, AMD issued warrants to the frontier lab for up to 160 million shares, or nearly 10% of its total capital. Because it owns 10% of AMD, OpenAI is incentivised to buy chips from AMD. Yet, as I mentioned earlier, AMD’s biggest competitor, Nvidia, owns around 4% of OpenAI and has another circular loop that forces OpenAI to buy some Nvidia GPUs.

Another chipmaker, Broadcom, is involved in its own circular financing loop, helping backstop and arrange vendor financing for major AI labs buying its custom silicon. Private credit providers like Apollo Global Management and Blackstone Inc structure financing vehicles that help fund frontier model developers like OpenAI and Anthropic to buy Broadcom chips. It’s hard to put a precise value on total intertwined deals but estimates vary from as low as US$800 billion to as high as US$1.2 trillion.

The OpenAI’s backstop, the deal between OpenAI and Oracle, and the SK Hynix arrangement are all classic vendor financing loops. Nvidia invests US$30 billion in OpenAI, which in turn uses the money to secure a contract with Oracle and Japan’s SoftBank Group, who then sign a deal with Nvidia to buy GPU chips. Nvidia then turns around and gets SK Hynix to supply high bandwidth memory for those chips. Nvidia books the chip sales as revenue, its shares rise and it is then able to boast that it has more capital to invest in the next AI lab. It is called circular financing because the money stays within the small club of cloud service providers, GPU makers, AI labs and memory chip firms. Everybody benefits and all boats rise. It is only when the tide goes out that you see who has been swimming naked.

To be sure, circular financing loops are a normal feature of fast-growing infrastructure booms. If AI products generate enough genuine external demand, the loop can be viewed as just aggressive capital formation. If they don’t, it can be seen as a debt-fuelled bubble. “The scale and circular nature of these commitments — where suppliers, customers and investors overlap — has prompted comparisons to the late-1990s tech bubble,” Stephanie Aliaga, a strategist for JP Morgan said in a recent note. “While some caution is warranted, the better question is not whether today’s deals resemble the dotcom era, but whether the underlying fundamentals do.” Here’s the thing: AI infrastructure is currently running near full capacity, unlike the idle fibre-optic networks of late 1990s and early 2000s, which helped trigger the bankruptcies of telecom companies like Worldcom and Global Crossing in 2001 and 2002.

So, should we be worried about circular financing loops? Yes, we should. For one thing, vendor financing can fake demand. When Nvidia funds its own customers, sales can look more robust than genuine end-market appetite for the product. Another issue is concentration risk in two or three closed loops. The same firms sit on different sides of the deals. If AI revenue disappoints, losses can cascade through all or most of the players immediately. Finally, these circular loops echo dotcom bubble era telecom vendor financing where suppliers such as Cisco Systems, Nortel Network and Lucent Technologies lent customers money to buy their networking gear. When actual usage fell short of their inflated numbers, the entire telecom and networking equipment sector crashed. Nortel went out of business. Lucent was rescued by Alcatel, which in turn was bought by Nokia.

A bigger issue is off-balance sheet debts or debts hidden away in special purpose vehicles (SPV). Until early last year, most of the hyperscalers were funding all their AI infrastructure spending from the cash on their own balance sheets. With AI infrastructure spending ballooning to over US$770 billion this year, the once cash-rich hyperscalers now rely on borrowing from banks, issuing bonds and using joint ventures with private credit firms. A recent Nikkei Asia study pegged combined off-balance-sheet AI debt across Google’s parent Alphabet Inc, Amazon.com, Meta Platform, Microsoft and Oracle at US$1.65 trillion on top of the US$1.35 trillion debt they actually had on their balance sheets. Bloomberg estimates off-balance sheet debts of hyperscalers at over US$1.8 trillion.

Meta had the biggest portion (US$420 billion) of off-balance sheet debt, or nearly three times the debt on its own balance sheet. The Bank for International Settlements recently dubbed it as “shadow borrowing” and noted these arrangements are “economically equivalent to debt” even though they are off the books. Off-balance sheet debts are not just limited to SPVs. They also include finance leases, residual-value guarantees, credit-wrap derivatives, as well as leases signed but “not yet commenced”, which sit off-balance sheet until the facility goes live. So Meta or Oracle might borrow today and if their massive data centres don’t go live until 2030, that debt remains off the balance sheet for another four years.

Last year, Blue Owl Capital formed a US$27 billion joint venture deal with Meta to fund and develop the massive Hyperion data centre campus in rural Louisiana. Blue Owl took an 80% interest while Meta has 20%. You won’t find any mention of US$27 billion data centre in Meta’s balance sheet. Some analysts have described off-balance sheet debts of AI labs as credit default swaps (CDS) of the AI era.

Here is what is going on. Instead of building data centres and buying Nvidia, AMD or Broadcom’s chips themselves, hyperscalers like Google and Meta sign long-term lease agreements with data centre operators. That allows them to offload risk and avoid recording the debt on their own balance sheets. These data centre operators are often shell companies that are created and funded by private credit providers like Blue Owl, Apollo Global and Blackstone.

Google creates an SPV, or a fund, with a private credit firm, which then rents compute capacity to Google. The income generated by that rental agreement is paid to the private credit fund. Essentially, Google is offloading the risks to some of the riskiest private credit firms like Blue Owl, which recently have been weighed down by heavy withdrawal demands and forced to gate redemptions and limit cash payouts. The private credit industry, which thrived during the zero interest era, has had a tougher time in the current higher-for-longer interest rates environment. While the US Federal Reserve last week held interest rates steady amid sticky inflation concerns, it signalled higher rates.

If huge off-balance sheet debts and circular financing were not enough, two weeks ago China’s frontier lab Moonshot AI unveiled its Kimi K3 model that is almost as good as the models being released by its American peers. The K3 release wiped out US$392 billion off OpenAI and Anthropic’s valuations. The two labs were expected to launch their IPO later this year. IPOs are important because the firms would have raised USS$70 billion to US$90 billion each to pay for their AI spending. Now with the IPOs delayed, they would need to work with private credit players and try to hide more of their debts through off-balance sheet vehicles.

The problem is that AI labs’ bet on circular financing had assumed that profits will remain concentrated at the “model layer”. That means frontier models need to stay scarce and expensive, which helps to justify their high valuations and the huge compute spend. If open-weight models from China keep matching American frontier peers at a small portion of the cost, then that assumption breaks. It also means profits move away from frontier labs to chip design firms like Nvidia and AMD, chipmakers like TSMC and Intel, cloud service providers like Amazon’s AWS and Microsoft’s Azure, as well as new AI applications. Open source doesn’t threaten the chips as much because model makers would still need GPUs to run the next Kimi K3. Still, it will hurt the economics of frontier labs, whose future revenue is supposed to justify the whole AI structure.

Will OpenAI or Anthropic go bust at some point soon? OpenAI posted a US$6.9 billion non-generally accepted accounting principles (GAAP) on US$5.7 billion in revenue for the first quarter of 2026. Some analysts have mentioned bankruptcy within a year at this rate. Yet, OpenAI continues to raise funds. Its previous US$40 billion funding round lasted a year. It now has about US$100 billion  in its coffers. A US$14 billion cash burn a year will buy it seven years unless costs balloon. The problem is costs are soaring. Anthropic’s cash burn is projected to fall to a third of its revenue in 2026 and 9% by 2027. The real risk isn’t an AI lab suddenly running out of cash but a lab failing to raise the next round because investors start to lose faith. The more likely outcome for a weaker lab like OpenAI is a distressed acquisition or a forced merger.

Most vulnerable are neo-cloud firms such as CoreWeave, Nebius, Lambda and Crusoe. They buy Nvidia chips with borrowed money, often using the GPUs themselves as collateral, and then rent out capacity. The math only works at high utilisation. CoreWeave carries a junk-tier Ba3 rating, and its credit-default swaps recently blew out to 855 basis points. The market is implying a 50% chance of default over five years. Tens of billions in GPU-collateralised loans across these firms come due between now and end-2028. Ratings agency Moody’s recently flagged Oracle and CoreWeave as AI’s weakest credit link. Oracle’s debt is rated Baa2 with a negative outlook or only two notches above junk. Finally, there is social media powerhouse Meta, which guided 2026 capex to between US$125 billion and US$145 billion. Unlike Alphabet or Amazon, which operate public cloud business, it needs external customers paying to rent its capacity. The entire return has to come back through Meta’s own ad engine and its Llama models. Meta is making the biggest relative bet with the least direct monetisation path and and as such is seen among the more vulnerable. Whether it will be able to create new growth drivers to steady itself remains to be seen.

Assif Shameen is a technology writer based in North America

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.