The belief in computing power is about to crumble: Elon Musk holds a pessimistic view on the sector, and 200 emerging cloud enterprises are facing the "depreciation dead end".

During the gold rush of AI large models, the "water seller" used to be the most lucrative business with guaranteed profits. As long as you could get NVIDIA GPUs and build a computer room, you would never worry about large model companies waving checks to rent computing power.

But now, the belief of "getting rich by hoarding GPUs" is collapsing in an all-round way.

Recently, the US stock computing power leasing sector has encountered a historic valuation crash. From crypto mining enterprises Soluna and Bit Digital, to star giants CoreWeave and Applied Digital, a group of companies collectively referred to as "Neoclouds" have seen their stock prices halved, with their technical patterns completely broken down.

Faced with this miserable situation, Elon Musk delivered the most fatal blow on social media: "In the past few weeks, the debt market has almost completely cut off funding channels for companies that have no substantial income independent of AI data centers. More than 200 new cloud service providers will face guppy-like extinction and consolidation."

Many people attribute this crash to the tightening of macro capital. But peeling off the coat of financial leverage, what is really strangling these more than 200 Neoclouds are two major drastic changes in the underlying logic of the AI industrial chain: the rapid depreciation of hardware, and the shrinking demand brought by the algorithm efficiency revolution.

The First Crisis: GPU Is Not Real Estate, But a "Depreciation Black Hole"

The biggest illusion of Neocloud service providers is to regard NVIDIA's GPUs as "digital real estate" that can preserve and appreciate in value.

In the traditional commercial real estate model, you borrow money to build a building, which can collect rent for decades, and the asset itself will appreciate slowly. But GPUs are completely different. The underlying law of technology hardware determines that it is a cruel war of attrition.

First of all, its hardware iteration speed is extremely fast. From H100 to B200, and then to the currently planned Rubin architecture, NVIDIA launches new products almost every year. Once a new generation of chips hits the market, the computing power leasing price of the previous generation of chips will plummet. In the second-hand market in Silicon Valley, the H100, which was once hard to get, has seen its leasing price begin to loosen with the expectation of mass production of new chips.

Secondly, large model companies are extremely "picky" customers. To train larger models, they always need the latest clusters with the largest interconnection bandwidth. Those old graphics cards hoarded in small and medium-sized Neocloud computer rooms will soon face the embarrassing situation of no one showing interest.

This forms a fatal "depreciation deadlock": Neocloud service providers borrow huge three-year debts from Wall Street to buy expensive graphics cards. But in less than three years, the rent of these graphics cards in the market may not even be enough to cover the interest and electricity bills. What's more, the physical loss of the cluster, the aging of network cables, and the maintenance of the liquid cooling system are all eroding profits all the time. When old cards cannot be rented out and new cards cannot be afforded, the break of the capital chain becomes inevitable.

The Second Crisis: Draining the Demand Side of Its Lifeblood

Apart from asset depreciation, Neoclouds are facing an even more terrifying enemy: increasingly intelligent algorithms are reducing the absolute thirst for computing power.

In the past year, the consensus in Silicon Valley was that "computing power is justice", and great efforts create miracles. But a series of recent technological breakthroughs (represented by open-source forces such as DeepSeek) have sounded the alarm for all companies that have invested heavily in computing power.

Through extreme MoE (Mixture of Experts) architecture optimization, better data ratio and more efficient inference frameworks, large model companies have found that they do not need 100,000 H100s, and only thousands of cards are enough to train models comparable to GPT-4 level.

This has directly changed the computing power consumption view of the entire industry. When "the intelligent output per unit of computing power" becomes the core indicator, those Neocloud companies that used to make easy money by "selling high-priced computing power" immediately feel the piercing chill. Large customers start to budget carefully, build their own clusters, or directly sign long-term low-price agreements with large manufacturers; while small and medium customers cannot afford the high rent at all. Worse still, high-quality human corpora are drying up, and the marginal benefit of large model pre-training is decreasing. The demand side has instantly shifted from "computing power shortage" to "computing power surplus", and the hundred-billion-level blue ocean that was originally highly anticipated has turned into a red ocean in the blink of an eye.

The Tragic Folding of the Industrial Chain: "Digital Contractors" Trapped in the Cracks

In this reshuffle, the reason why Neocloud service providers fall first is that they are in the link with the thinnest profit and the highest risk in the AI industrial chain.

Looking upward, there is the monopoly of NVIDIA and hyperscale cloud vendors. NVIDIA takes the most generous hardware profits in the industrial chain, while giants such as Microsoft, Google and Amazon control the training of high-end large models with their huge cash flow and self-built computing power. The giants even begin to charge high "model router" tolls, requiring third-party model companies to access through their gateways. Neocloud companies can only take over scattered inference orders that the giants are unwilling to do.

Looking downward, there is extreme involution among large model startups. As the price war of large model APIs (such as the ultra-low-priced large models from ByteDance and Alibaba) enters a white-hot stage, model companies themselves are struggling on the verge of life and death, let alone paying the high computing power rent of Neocloud companies in cash.

Neocloud service providers are essentially "digital contractors" trapped in the middle. They borrow high-interest loans to purchase equipment, depend on the charity of large manufacturers and the ups and downs of model companies, and have no pricing power at all. Once the macro environment fluctuates, they are destined to be the first to be sacrificed.

The Physical Dilemma in 2027: Stranded Super Infrastructure

Apart from the fatal flaw in business logic, Neoclouds are also facing extremely cruel physical constraints.

Renowned investor Chamath Palihapitiya pointed out that many giants are facing the dilemma of a two-year delivery cycle for transformers and no clear timeline for grid connection approval when building data centers.

This has led to extremely absurd capital waste: many companies have spent billions of dollars buying top GPUs and building huge computer rooms, but due to lack of power and grid access, these computing powers have become "half-finished projects" piling up in warehouses and gathering dust. For Neocloud service providers with tight capital chains, they often cannot survive until the day the grid is connected, and have already vanished in the debt meat grinder.

Who Is Reaping the Final Fruits?

Musk and Chamath's pessimistic remarks, though every word makes sense, also carry the unique coldness of Silicon Valley.

As short-selling agencies taunted in the comment section, Musk and Chamath themselves are players with heavy positions in computing power infrastructure. While publicly pointing out the "deadlock" of small and medium-sized Neoclouds, they are relying on SpaceX, xAI and huge cash flow to quickly build 100,000-card level super clusters at extremely low cost.

This is a dimensionality reduction strike. When the new computing power cycle starts, these giants with sufficient capital, strong engineering capabilities and their own business support (such as Starlink, autonomous driving, internal large models) will acquire the second-hand hardware of bankrupt Neoclouds at extremely low prices, completing the major reshuffle and oligopoly of the industry.

Conclusion: Getting Rid of Bubbles and Returning to Real Hardcore Strength

The end of more than 200 Neoclouds has taught a lesson to all those who blindly add leverage in the AI craze.

The AI large model is a giant beast that devours resources, but it also follows common business sense: intermediaries without core technical barriers, diversified profit-generating capabilities, and only relying on borrowing to "resell" hardware resources are doomed to not go far.

The future AI computing power market will no longer be a wild era where "hoarding GPUs can print money". It will belong to those hardcore players who can achieve the ultimate combination of software and hardware, push the Token cost to the lowest, and truly convert large models into actual productivity. Those Neoclouds swept up by the computing power myth and fallen on the ruins of "stranded computing power" in 2027 will become the most easily forgotten cost in this technological revolution.

This article is from the WeChat official account "Tech Business", written by Tech Business, and authorized for release by 36Kr.