(This article is authored by Tech Business; published on Titanium Media with authorization.)
Article | Tech Business
In the frenzy surrounding large AI models, computing power has long been regarded as the “oil” of the new era.
To seize computing power, Silicon Valley giants have poured tens of billions of dollars into the race; meanwhile, a host of “Neocloud” service providers without ties to major tech firms have been borrowing heavily to acquire hundreds of thousands of NVIDIA GPUs, aiming to play the role of “shovel sellers” in this gold rush.
However, this game of passing the buck is approaching its end.
Recently, Elon Musk engaged in a fiercely candid conversation on X with renowned investor Chamath Palihapitiya. Musk sharply exposed the current artificial intelligence computing market’s façade of prosperity: over the past few weeks, the debt markets have virtually cut off funding to companies that lack revenue streams independent of their AI data centers.
Consequently, Musk made a starkly grim prediction: “This means that those more than 200 ‘new cloud’ service providers will undergo a fate akin to that of guppies—gradual extinction and consolidation. A few may survive, but their numbers will be very limited.”
Meanwhile, Chamath, the investor who accurately shorted the subprime mortgage crisis, added another blow, pointing out that by 2027 there will be a large number of “stranded compute” resources and data centers abandoned halfway through.
This conversation not only strips away the façade of AI computing power hype but also signals that a major reshuffling in the AI infrastructure sector has already begun.
A sudden brake in the debt market: 200 new cloud companies exit
To understand this crisis, you must first grasp what “Neoclouds” are.
Over the past two years, companies like CoreWeave have emerged as major players. Unlike Amazon AWS or Microsoft Azure, which boast vast, general-purpose cloud computing ecosystems, they focus on a single task: purchasing GPUs from NVIDIA and then renting them at premium rates to large‑model firms or AI startup teams that urgently need computational power.
This is, at its core, an extremely extreme “high-leverage + heavy-asset” financial game. These companies lack stable cash flow to support themselves and must rely on debt financing from Wall Street to purchase GPUs that often cost tens of thousands of dollars apiece. Their business model is exceedingly fragile: as soon as the price of computing power rentals drops, or the cycle of borrowing new funds to repay old ones breaks down, the companies can instantly find themselves in dire straits.
In the conversation, Musk pinpointed with remarkable accuracy the current critical issue: the funding tap in the debt market has been turned off.
After a prolonged period of euphoria, Wall Street financial institutions have finally begun to scrutinize risks. When they realized that many AI startups’ revenues could not even cover their exorbitant compute costs, and that “new cloud” companies not only lacked collateral but were heavily reliant on high-interest debt, investors’ risk-averse instincts were fully triggered.
Musk cited his own company as an example: “SpaceX raised $40 billion in debt to finance GPUs, yet we maintain an investment-grade credit rating and generate revenue from Starlink and launches. If you don’t have solid cash flow, don’t expect to secure a single penny from the debt markets.”
This is the cruelest reality: more than 200 new cloud companies that lack “substantial revenue independent of AI data centers” are on the brink of a funding crunch. Musk described this impending wave of bankruptcies as a “massive die-off, like guppies.” Once the tide of capital recedes, all that will be left on the beach is a pile of discarded scraps and secondhand GPUs that can’t be monetized.
“Stranded Computing Power” in 2027: The Ruins After a Construction Boom
In this conversation, Chamath raised an even more critical industry pain point: the “stranded computing power” crisis of 2027.
Chamath stated bluntly that he is witnessing a disaster caused by bureaucracy. Many tech giants (hyperscalers) face extremely sluggish processes when building data centers. “For NVIDIA, just signing a non-disclosure agreement (NDA) takes three months, and the engagement period even requires 18 months.”
This sluggish supply chain and bureaucratic red tape, coupled with physical constraints such as grid capacity and building permits, have prevented many data centers from being completed on schedule. Meanwhile, major tech companies, in pursuit of their own large‑model businesses, keep all self‑built computing power tightly under their control, prioritizing internal needs before allocating any surplus to third parties.
This has resulted in an utterly absurd waste of capital: many companies have spent billions of dollars to acquire top-of-the-line GPUs and build massive data centers, only to find that by 2027, due to a lack of power, network connectivity, or viable use cases, this computing capacity has become “half‑finished projects,” gathering dust in warehouses with no revenue-generating potential.
These “stranded computing resources” are not only a financial disaster for tech giants but also a severe blow to small and medium-sized AI companies that rely on renting computing power. The money has already been spent, yet the equipment has become useless scrap metal.
The squeeze from the tech giants: Who’s smiling, and who’s bleeding? In this AI‑computing winter, not everyone is shivering.
As Chamas pointed out, this mismatch will only benefit two types of companies: First, non‑bureaucratic giants that can rapidly scale to industrial levels—for example, Musk’s xAI, which, backed by a robust engineering culture, can swiftly build a supercomputing cluster with 100,000 H100 GPUs. Second, players with extremely deep pockets and sufficient agility—such as Microsoft and Meta, which boast vast cash reserves, allowing them to ignore short‑term debt‑market volatility and continue stockpiling resources.
As for the “new cloud” companies and large‑model startups caught in the middle, they face not only the inability to rent computing power but also the extremely harsh “digital exploitation” imposed by the tech giants.
Chamath revealed that major tech companies now even prohibit third-party model firms from direct access, requiring them to route their models through the companies’ “model routers” and imposing exorbitant interconnection fees. He even hinted that Anthropic’s previously leaked S‑1 confidential filing mentioned that “the major players informed them they had no spare compute capacity.” This confirms that the shortage of computing power and the resulting oligopolistic dominance have reached an extremely dangerous level.
The cloud of short-selling: A “duet” between Musk and Chamath?
The most interesting detail of this tweet unfolded in the comments section.
While Musk and Chamath were pontificating about the AI computing power bubble and voicing pessimistic forecasts, Jack Gee, founder of the short-selling firm Hedgeye, fired back directly below: “You two are really working well together. This certainly looks like a clever way to diversify risk and take on competitors.”
Jack Gee’s sarcasm hit the nail on the head. Back in 2021, during the SPAC (special purpose acquisition company) bubble, Chamath leveraged wild hype and false marketing to swindle countless retail investors before cashing out and exiting. Now, both he and Musk are heavily invested in AI infrastructure while publicly downplaying those “new‑cloud” companies that piled up debt to hoard GPUs.
Their statements have, in effect, served to suppress competitors. Once market panic intensifies and debt interest rates soar, those rivals that relied on high leverage to acquire assets will find themselves cornered, while industry giants like SpaceX and xAI—backed by stable cash flows and diversified financing channels—can take advantage of the situation to scoop up undervalued assets, thereby cementing an oligopolistic structure in the sector.
Conclusion: The “Coming-of-Age Ceremony” of AI’s Grand Infrastructure Project
The relentless surge in AI computing power has finally run into the brick wall of financial common sense.
Over the past two years, many believed that simply purchasing an NVIDIA graphics card would guarantee profits. Yet this debt-fueled frenzy masked a fundamental flaw in the business model. AI infrastructure is, at its core, a long‑cycle, capital‑intensive, low‑tolerance‑to‑error traditional industry. It requires grid‑level power support, supply‑chain stability, and an enormous pool of capital reserves.
2027 is set to be an extremely turbulent year for the AI computing power industry. A large number of substandard “new cloud” service providers will be weeded out, non-compliant AI startups will be starved of resources, and “stranded computing power” lacking electricity and supporting infrastructure will become the most tragic remnants of this bubble.
In the tide of the technological revolution, bubbles inflated solely by PPTs and borrowed money will ultimately be mercilessly shattered by gravity. In the end, only a select few—with the strongest cash flow, the most robust engineering capabilities, and the coolest business acumen—will be able to laugh last.
For more exciting content, follow TMTPost on WeChat (ID: taimeiti), or download the TMTPost App.