"Industry Research" · AI Compute Stack Series, Part 3. The first two parts looked at chips (Nvidia) and value migration (to power). This one looks at a fascinating bunch of new players—they don't make chips, just buy them and rent them out—yet they've become the sexiest and most dangerous type of company in this AI narrative.
Neocloud 是什么:AI 算力的「二房东」
Let me explain this new species.
Neocloud (new cloud, also called cloud, AI-specific cloud) refers to companies like CoreWeave, Nebius (NBIS), Lambda, Crusoe. Their business model is simple enough to state in one sentence—they buy Nvidia GPUs at scale, build data centers, and rent out compute power by the hour or by contract to whoever needs it.
They are the "sublessors" of AI compute: they don't produce the source of compute (chips), and they're not necessarily the end users. They stand in the middle, renting scarce GPU compute wholesale and retail.
You might ask: didn't we already have AWS, Azure, Google Cloud? Why do we need these new players?
Because the traditional cloud giants are general-purpose clouds—they do everything, and GPUs are just one piece of their vast operations. Neoclouds, on the other hand, are born specifically for AI compute: from data center design, network architecture, to scheduling software, everything is optimized around "how to run GPU clusters most efficiently." The result is that in this niche of GPU compute, they often deploy faster, are more specialized, and sometimes cheaper. During a period of extreme AI compute scarcity—when even the traditional cloud giants themselves are queuing for GPUs—this kind of specialized sublessor finds its niche.
一个耐人寻味的细节:英伟达为什么扶持它们
The rise of Neocloud has one behind-the-scenes push worth noting separately—Nvidia itself.
Nvidia not only sells GPUs to these Neoclouds, but also prioritizes them for scarce GPU allocations, and even invests directly in them (Nvidia has investments or deep ties with CoreWeave and Nebius). This seems unusual: why would Nvidia prop up a bunch of "sublessors"?
My take: Nvidia is creating "customer diversity" for itself, hedging its dependence on hyperscale customers.
Remember the biggest risk from the previous article: Nvidia's top customers (Microsoft, Google, Amazon) are all building their own chips—they are potential gravediggers. One of Nvidia's countermeasures is to support a group of allies that are Nvidia-only, don't do their own chips, and help it sell GPUs to a broader market. Neoclouds are exactly that: they are 100% loyal to Nvidia (no ability or incentive to self-develop), and they distribute Nvidia's compute to smaller customers who can't afford or don't want to build their own infrastructure. For Nvidia, Neoclouds are a diversification line of defense against "customer self-development" and a channel to reach the long tail. That's why Nvidia is willing to allocate them chips and give them money.
Understanding this relationship is important, because it is both the Neocloud's biggest source of confidence (Nvidia has their back) and its biggest hidden risk (the circular financing we'll discuss next).
商业模式的另一面:这是一门杠杆生意
After the sexy side, the dangerous side. The most dangerous part of Neocloud is hidden in its balance sheet—this is a business that relies heavily on leverage.
The logic: buying GPUs and building data centers requires massive upfront capital. These Neoclouds mostly lack the free cash of the giants. They rely on heavy debt to finance GPU purchases—using future rental income and the GPUs themselves as collateral to borrow money and expand. This leads to a structural vulnerability I must highlight:
First, GPUs are rapidly depreciating assets. GPUs aren't real estate; they're technology. Nvidia iterates yearly—today's most advanced GPU will be crushed by the next generation in two to three years, with rental income shrinking sharply. Neoclouds borrow long-term debt to buy rapidly depreciating assets—that is a mismatch in both maturity and value. If they can't earn enough rent before the GPU depreciates, the math is ugly.
Second, debt is fixed; rental income is cyclical. The debt and interest are fixed, must be paid. But rental income depends on AI compute demand and pricing—which is cyclical. In good times, rent > debt costs, leverage amplifies profit, it's a money printer. Once AI capex slows or GPU rents fall, rent may not cover debt, and leverage amplifies losses. This is exactly what I wrote about leverage in asset allocation: leverage doesn't change whether you're right or wrong, it magnifies the consequences of being right or wrong. And when you're wrong, it can kill you.
Third, customer concentration. Take CoreWeave: its revenue is heavily dependent on a few large customers like Microsoft and OpenAI. This means those "locked-in long-term contract revenues" on its books are tied to a tiny group of customers' continued performance. If a major customer cuts orders or builds its own capacity, the "certain cash flow" of the Neocloud will suffer a big hole—while its debt doesn't shrink accordingly.
Fourth, the worry of circular financing. Remember Nvidia investing in Neoclouds? Connect the chain: Nvidia gives money to Neocloud → Neocloud uses it to buy Nvidia GPUs → Nvidia books revenue. There is a circular flavor—an upstream supplier funds downstream buyers to buy its own products. In good times, this is a virtuous cycle (everyone grows together). Critics worry, however, that this cycle can artificially inflate real demand, and if one link breaks, the cycle can collapse in reverse. I don't call this a problem per se, but it's a structure worth watching for health concerns.
Combine these four points—long-term debt, short-lived assets, cyclical rent, few customers, embedded in a circular financing arrangement—and you see: Neoclouds are the most leveraged, most cycle-sensitive, thinnest-margin-of-safety link in the entire AI compute chain.
我的判断:它是 AI 周期的「煤矿里的金丝雀」
So how to view Neocloud? My judgment: They are the best "canary in the coal mine" for the health of the entire AI capex cycle.
Miners take canaries into the mine because canaries are most sensitive to poison gas—when they fall, the miners know it's time to evacuate. In the AI compute chain, Neoclouds are that canary—because they are most leveraged and most fragile, any trouble in the AI capex mine will hit them first and hardest.
This means two things:
First, as long as the boom lasts, Neoclouds offer the highest beta. Their leverage amplifies upside in growth; their purity (100% bet on AI compute demand) makes them the sharpest long tool. If you believe AI capex is a multi-year trend, Neoclouds are the highest-beta expression of that belief.
Second, they will be the first and most severe to collapse. Once AI capex peaks—even just slows—their leverage, asset depreciation, and customer concentration will all reverse simultaneously. They won't suffer a "gradual margin compression" like Nvidia; they could face a "cash flow break, debt default" hard landing. They have none of Nvidia's moat, none of power companies' demand safety, none of memory giants' net cash—they only have leverage and a bet.
So my attitude toward Neoclouds matches my attitude toward leverage itself (remember that piece—for most people, the answer on leverage is zero): They are a high-odds, high-fragility tool; their very existence is the most direct thermometer for the AI capex party. I will watch them closely (as a cycle signal), but I know clearly that they are the link with the worst anti-fragility, the least able to survive "that one" crash.
写在最后
Neoclouds are a fascinating specimen in this AI narrative. They concentrate the AI compute boom into its purest—and riskiest—business: borrow money, buy shovels, rent them to the gold diggers.
Their fascination lies in their purity: unlike Nvidia with its complex moat story, unlike power with its cross-narrative safety margin, Neoclouds are a high-leverage pure long on "AI compute demand persists." In good times, they run the fastest on this chain.
Their danger also lies in that purity: when a business bets its entire existence on the two assumptions "demand stays strong forever" and "money stays cheap forever," and amplifies it with leverage, it turns itself into a fuse—quiet normally, but the moment the cycle turns, it ignites before anyone else.
I'm not saying Neoclouds will blow up. In a world where AI capex continues, they may keep charging for years. What I'm saying is: the value of understanding Neoclouds is not in predicting whether they go up or down, but in the fact that they are a mirror—they reflect how deep the leverage has been used in this AI boom, how heavily the belief in "demand eternity" has been wagered. When you want to know how long this party can last, don't just look at Nvidia's earnings—look at whether the most sensitive canary is still singing.
If I had to leave just one line—
Neoclouds are the most leveraged link in the AI compute chain. In a boom, they are the highest-beta money printers; in a downturn, they are the first to break, a house of cards. Their value is less as an investment, and more as a signal—they tell you just how much people believe "this time is different" in the grandest bet of our era.
The AI Compute Stack series concludes here: from chips (Nvidia), to value migration (power), to leveraged sublessors (Neocloud). Next track, we leave the hard world of "compute" and step into a softer, even more imaginative layer—AI applications and software: why the winners in this layer may not yet exist today.
——
Risk disclosure: This article is for industry research purposes only. The companies mentioned are for analytical illustration and do not constitute any investment advice. Markets are risky; invest with caution.
On investment analysis, market insight, and asset allocation — not chasing short-term noise, but understanding what truly drives long-term returns.
