This article first appeared in The Edge Malaysia Weekly on October 26, 2026 - November 1, 2026

FOUR years ago, OpenAI unveiled ChatGPT and unleashed an unprecedented artificial intelligence (AI) boom. For a lack of a better yardstick, the barometer S&P500 Index is up over 115% inclusive of reinvested dividends while the XLK, the State Street Technology Select Sector SPDR ETF, the largest pure play tech barometer, is up 210%. For nearly 18 months now, there has been persistent talk of an imminent AI bubble burst akin to the dotcom crash of March 2000. Since then, the S&P500 is up 32% and XLK is up 60%. Over the past fortnight, however, a key segment of the AI ecosystem known as neoclouds, such as CoreWeave, Australia-based Iren and Amsterdam-based, Nasdaq-listed Nebius Group, has been reeling after another Australian company, Firmus Grid, was forced to shelve its US$5 billion (RM20.4 billion) initial public offering (IPO) on Sydney’s Australian Securities Exchange earlier this month. 

Firmus, which sells Nvidia-powered computing capacity to AI frontier labs such as such as OpenAI and hyperscalers like Meta Platforms, pulled its listing plans at the last minute after investors balked at its proposed US$30.4 billion valuation, up from US$10 billion just two months ago. Surging bond yields had raised concerns about Firmus’ soaring cost of debt. Chip behemoth Nvidia is a supplier of graphics processing units, or GPU chips, and has a small equity stake in Firmus. The Sydney-based neocloud would have had US$35 billion in debt upon completion of its data centres in Malaysia, Indonesia and Australia, but just US$5 billion in annual revenue. Investors were concerned about how Firmus would ever pay back its loans were told that it would keep on building more expensive data centres. Unfortunately, that would only be possible by doubling or tripling its total debt.

As Firmus launched its IPO roadshow three weeks ago, it was clear that its business model just didn’t make sense to investors. It was betting on borrowing more massive sums against hypothetical future customer contracts to buy expensive Nvidia chips, hoping the future rent would cover its mounting debt load. While it had a bunch of “memorandums of understanding” (MOUs) with Meta Platforms and OpenAI, they were merely “options to acquire capacity”, not actual contracts. 

There was also the controversy around Firmus’ co-CEO and co-founder Oliver Curtis, who had served jail time for insider trading. Investors buying into the vastly overvalued IPO of a debt-laden firm run and founded by a convicted felon and insider trader were understandably worried. After Firmus yanked its IPO, other neoclouds around the globe that were readying their own listing saw investors and regulators scrutinise their plans and valuations. Among them: Britain’s Nscale, which has faced pushback over its US$35 billion target IPO valuation. On Oct 20, another neocloud, South Korea’s Elice Group, dubbed the “CoreWeave of Korea”, raced to its own listing. Elice’s IPO was valued at US$842 million, but its stock dropped 32% upon listing. It has since rebounded but is still below its listing price. 

What in the world is a neocloud? Companies such as CoreWeave, Nebius, Iren, Lambda and Crusoe buy AI GPUs and then lease them to customers such as top frontier labs OpenAI and Anthropic which need compute, or computing power, to build and run their latest AI models. GPUs, of course, are parallel processors that are widely used for general-purpose AI and heavy math workloads. Neoclouds do one thing that hyperscalers were too slow or too big and diversified to do: rack up Nvidia GPUs at scale and rent them by the hour. Neoclouds were also offering GPU capacity at a significant discount to hyperscaler pricing. Early chip access and a thinner service layer helped them offer on-demand rates that were 50% to 70% below hyperscaler list prices.

So how big is the fast growing neocloud space? There are, by some counts, over 530 firms around the world renting or leasing GPUs that are not hyperscalers. Neocloud sector revenues are forecast to exceed US$42 billion this year from US$25 billion last year. Research firm Synergy estimates that total neocloud revenues could exceed US$400 billion by 2031. Many neocloud firms such as Iren, Core Scientific and Cipher were once Bitcoin miners who pivoted to the more lucrative GPU rental business when cryptocurrency prices collapsed. While AWS, Azure and Google Cloud were seen as general-purpose and expensive, neoclouds were offering access to the newest chips, faster, cheaper, with InfiniBand networking tuned for AI training. Not everybody loved neoclouds but demand for GPUs was so high that anybody who wanted access to Nvidia’s GPUs had to do business with them on their terms.

Here’s how it works: Neoclouds, most of which are not well capitalised, buy GPUs with borrowed money, and sign rental contracts with firms that need compute. They basically earn the spread between rental revenue and the cost of capital plus cost of power as well as the cost of hosting the GPUs or running the data centre. Their problem is a structural mismatch. The debt used to buy a GPU fleet is amortised over five to six years, but as I have explained in my tech columns before, the economic life of a GPU is much shorter, more like four years than six, because Nvidia ships a new generation of GPUs every year. Over the past two years, we have gone from Nvidia’s Hopper to Blackwell to, more recently, Rubin GPU chips. Each new generation of chips puts pressure on the rental price of the previous one. The Hopper H100 that once fetched US$8 hour is now available for US$2.79; Blackwell (or B200) which was US$7.46 two weeks ago fell 8% in a single week. So the asset backing the loan loses value faster than the loan gets paid down. 

Think of how a bank, which primarily makes money from selling mortgages, would handle its loan book if the houses being mortgaged were going to be close to worthless in four years. For starters, the bank will only have three-year mortgages because the house could be worth nothing at the end of the fourth year. Moreover, the bank will also charge a much higher interest for its mortgages. But back to neocloud financing — those loans carry 60% to 70% loan-to-value against GPU collateral. A fleet of GPUs can go underwater fairly quickly on its financing long before it is paid off. Surely, you might wonder, ageing GPUs might retain some residual value. They do, but the market is now pricing them way below it used to last year, or the year before. The used-GPU resale index was down about 22% year on year in August 2026, a reflection of collateral erosion showing up in real prices.

Over the last two weeks, neocloud woes have engulfed the sector’s biggest player, CoreWeave, whose debts are huge and front-loaded. CoreWeave carries US$35 billion in total debt with a quarterly net interest expense of around US$640 million, almost twice as much as it had a year earlier. Yet CoreWeave had an audacious capital expenditure plan of around US$39 billion for this year. Much of the neocloud’s GPU-backed debt set to mature between this year and 2028 will be rolled over in an environment of higher interest rates that is likely to weigh on the sector for years.

CoreWeave stock plunged 34% between Oct 6 and 21, amid growing credit market fears over the heavy debt-funded buildout of AI infrastructure. While the stock was being hammered, CoreWeave’s debts were pounded even more. The yield on the company’s notes due 2031 touched almost 16% while the cost of insuring its debt via five-year credit default swaps, or CDS, spiked to 11.4%. “The market is demanding substantially more compensation to own CoreWeave’s unsecured credit,” notes Madison Rezaei, an analyst for Bernstein. “CoreWeave does not just need customers willing to buy the compute. It needs lenders willing to fund the hardware on terms that leave something attractive for shareholders.”

Will the neocloud business model withstand challenges like rising interest rates, falling GPU prices and the advent of open-source models from China? Neoclouds is the most capital-intensive and leveraged corner of tech. Gartner predicts that over 80% of neocloud firms will cease independent operations by 2030 due to intense cost pressures and a lack of differentiated business models. While lower GPU prices lower the cost to buy a new fleet of chips, they also pressure the rental rate and the resale value of the fleet that the banks have financed. Falling chip prices will more likely punish the incumbents that levered up at the top of the cycle  while slightly helping late entrants.

Lower compute costs and cheap Chinese open-source models produced by labs like DeepSeek and Moonshot AI are another long-term threat. Yet cheaper AI doesn’t mean less demand for GPUs. Indeed, this is a Jevons Paradox situation: When inference gets cheaper and good-enough open models proliferate, total inference volume tends to explode. Moreover, inference is a better business for neoclouds than training. Training can be lumpy, winner-take-all and concentrated in a few labs while inference is steady, distributed, latency-sensitive and closer to a real utility with recurring revenue. The operators that reposition from “training capacity for frontier labs” towards “inference capacity for everyone” are more likely to have an edge. 

In Nvidia’s most recent quarterly report, three direct customers made up 21%, 17% and 16% of the chip giant’s total sales. Here is another way to look at things: About 75% of all GPUs go to hyperscalers — Amazon’s AWS, Google Cloud, Microsoft’s Azure, Meta Platforms and Oracle. The sixth player is SpaceX, which is growing rapidly as a provider of compute. Another 15% of GPUs is sold to neoclouds. Tech research firm Gartner expects them to account for over 20% of all GPUs sold by 2030. Hyperscalers and neoclouds use GPUs made by Nvidia and AMD in their cloud data centres and lease, or rent, the computing power or compute out to frontier AI labs such as Anthropic and OpenAI. Barclays estimates that 73% of hyperscalers’ AI revenues, or nearly a third of their total cloud revenues, now come from the two big frontier labs. 

Nvidia relies on the hyperscalers to buy its GPU chips. AWS, Azure and others count on OpenAI and Anthropic to buy compute. The smallest, most vulnerable and highly indebted hyperscaler, Oracle, depends on OpenAI to buy most of its compute. If OpenAI lands in hot water and faces consequences, Oracle will be in a jam almost immediately, the hyperscalers will lose almost quarter to a third of their total cloud revenue, and Nvidia’s biggest customers will stop ordering new GPUs. To hedge his bets, Nvidia CEO Jensen Huang has diversified GPU sales away from hyperscalers by promoting sovereign AI or selling state-of-the-art chips to governments in the Middle East, Asia and Europe. 

As competition mounts, subscale operators dependent on bank loans and vendor financing are unlikely to make it through a higher-rate, lower-price cycle, which in turn would lead to a consolidation wave. Huang told journalists this past week that demand for neoclouds is “off the charts”, though their growth is currently limited by a shortage of working capital rather than a lack of customers. Nvidia, which is an investor in Coreweave and several other neoclouds, is looking to shore them up. 

There will likely be a few survivors with genuine scale, diversified customers and ideally their own power and data centre real estate. Increasingly, power, not chips, is becoming the actual scarce asset. Surviving neoclouds would need to adapt and look beyond raw GPU rental by offering complete AI integration, developer tools and managed services. They will undergo a transformation and an extreme makeover over the next few few years. The AI boom still has a long way to go. 

Assif Shameen is a technology and business 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.