The scarcest resource in AI infrastructure is not chips—it's power that can go online immediately
GPUs can be bought with money and optimization capability will be leveled by open-source; in the long term, the only true scarcity is electricity that can be deployed immediately and land with pre-approved grid interconnection permits. This means New Cloud competition is actually about fundraising and operational capacity, not technology.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Acquisition or open-source: two opposite bets
Nebius paid $643 million to acquire the Eigen AI team and infrastructure stack; Clarify even sealed the deal with just a technical license—the fact that a technology could be bought this way suggests it never had pricing power to begin with. Crusoe took the opposite route: it contributed its own Managed Inference Tokenizer directly to NVIDIA's Dynamo open-source project, effectively admitting the technology cannot be defended, and instead planted itself at the center of the standard. The two companies are betting opposite propositions: Nebius is betting that optimization capability can be privatized; Crusoe is betting that it will become a public good, with everyone running code Crusoe wrote from then on.
— RayThe truly scarce resource is power, not GPUs
GPUs can be purchased with money, and optimization capability won't be scarce long-term—it can be open-sourced or acquired, and within 12 to 24 months gaps between players will narrow. The truly scarce resource is electricity available immediately and power land, plus the capacity to absorb utilization risk. Nebius discloses power reserves in its financial reports more aggressively than GPU inventory; contracted power in Q1 exceeded 3.5 megawatts with guidance of 4 megawatts by year-end—this is the metric investors should use to value the company.
— RayRevenue doubled, but profit fell 36%
Revenue growth only proves that market demand exists and capacity can be sold; it doesn't prove the business is profitable, because revenue growth is a function of capital expenditure, not competitive strength—buy more cards, rent them out, revenue goes up naturally. CoreWeave's adjusted operating profit tells the real story: this quarter $128 million at 5% margin, versus $200 million the year before—over the past year, revenue doubled but operating profit fell 36%.
— RayLock in long contracts or keep ammunition to bet on spot prices
This year Nebius didn't lock all its capacity into customer contracts; management said datacenter capacity remains tight but deliberately reserved some compute for its token factory, betting that shorter-term contracts would command higher prices—betting it has pricing power. CoreWeave went the opposite way, extending contract terms from four years to about five, locking in supply aggressively. These two paths correspond to two risk configurations: CoreWeave transfers utilization risk to customers and gives up upside; Nebius retains pricing upside but takes the risk of idle GPUs itself.
— RaySelling GPU hours and selling tokens are two different businesses
Switching from GPU hours to tokens is like moving from a landlord model to running an operations business: income shifts from multi-year contracts to real-time usage, earning only when someone uses it today; profit margin risk transfers from customers back to you; the source of gross margin changes too. GPU hours profit from funding-cost and rental spreads—financially driven at heart; tokens profit from utilization spreads and optimization efficiency—operationally driven at heart. Customer switching cost is nearly zero: replacing a supplier is just changing one line of code in the base URL. Token pricing has higher gross margin potential but lower certainty.
— RayThe inference platform's moat lives on the customer side
Inference API layers have near-zero switching costs, unit pricing is directly comparable, pushing serverless platforms into a price war with no exit until there's no markup left. The layer that can command a premium is post-training, but Fireworks founder Qiao Lin has said that moat is the customer's moat, not Fireworks' own—by embedding customer data and training workflows on their platform, the moat is rented. It depends on customers growing first, and once customers mature, they have some probability of switching or building in-house.
— RayPower land is harder to replicate than GPUs
Sun Yuchun tweeted that we face a GPU shortage in the short term, power shortage in the medium term, and storage shortage in the long term. By 2026, signs of GPU shortage relief are visible, but power infrastructure is nowhere close. More precisely, we lack power that can be deployed immediately. Power land means land that has already obtained grid interconnection permits, with substations and water access. GPUs can be mass-produced and depreciate within 18 months, but grid interconnection permits cannot be mass-produced—they only become scarcer and more expensive. The podcast also mentioned that tidal power company Panthalassa no longer lays undersea cables; instead, it converts power directly to tokens locally and transmits via Starlink.
— RayTwo signals of the hype cycle topping out
Two signals tell you whether AI infrastructure hype is topping out: one is if a company unwinds its long positions early and actively seeks out customers to lock in resources, it signals a shift from optimism to pessimism about future token prices. The second is when a mature company's token revenue from datacenter operations can no longer offset total operating costs—that marks a structural inflection point.
— RayIn their own words · checked verbatim
Then it shouldn't have a price. If it does have one, that's precisely evidence the moat isn't deep enough. When we define moat in the broadest way, it should be something others can't buy even with money. But Nebius paid $643 million and bought the entire team and infrastructure stack.
那它就不应该有价格 如果它有价格 恰恰说明这条护城河不够深 就是大家在对护城河定义的时候 应该最广泛的定义是 可能别人花钱也买不到的东西 但是NewBS花6.43亿 就可以买到整个团队和整套站
Ray4:03
Or the simplest distinction would be: Nebius runs an acquisition business, Crusoe runs a standards-participation business.
或者是最简单的区分 应该是Nubius做的是买断的生意 Cruso做的是接入的生意
Ray6:04
What I think is truly scarce is immediately available power and power land, plus the ability to absorb utilization risk.
我觉得真正稀缺的东西是 立即可用的电和Powerland 以及承担利用率风险的能力
Ray9:09
So my thinking is: New Cloud in its ultimate form may not be a tech company at all, but infrastructure finance wearing a cloud coat. Its real profit should come from underwriting fees, not technology.
所以在这个地方 我的想法是 New Cloud的终极形态 可能并不是一家科技公司 而是一个穿着Cloud马甲的基础设施金融公司 它赚钱的本质应该是承销费 而不是技术一家
Ray10:15
I think revenue growth only proves one thing: that real market demand exists and capacity can actually be sold. But what it doesn't prove is whether this business makes money. Because under this model, revenue growth is a function of capital expenditure, not competitive strength.
我觉得收入增长只证明了一件事情 就是市场上需求是真实存在的 而且产能是确确实实可以卖得掉的 但是不能证明的事情是 这个生意在赚钱 或这个生意赚不赚钱 因为在这个模式下 收入增长是资本开支的函数 并不是竞争力的函数
Ray14:19
In other words, we will probably struggle to establish long-term pricing power over customers, because this is a business where customers can swap you out in ten seconds.
也就是说我们可能非常难以在客户的身上建立长期的定价权,因为这是一门客户可能十秒钟就可以把你换掉的生意。
Ray21:40
If we understand it using the purest concept of moat, I'd say their moat is rented. They need to wait for customers to grow first, and once customers mature, there's some probability they'll swap out or build in-house.
如果用最本质的护城河概念去理解 我们觉得他的护城河是租来的 他需要等客户先长大 而且客户长大之后是有一定概率把它换掉或者是自建的
Ray25:50
He said we face GPU shortages short-term, power shortages medium-term, and storage shortages long-term. I think that's a remarkably prescient insight.
他说短期缺GPU 中期缺电力 长期缺存储 我觉得这是一个非常有长远的洞见
Ray33:12
Figures
| Nebius acquisition of Eigen AI team and infrastructure | $643 million | 4:03 |
| Nebius Q1 contracted power | Over 3.5 megawatts | 9:09 |
| Nebius year-end power guidance | 4 megawatts | 9:09 |
| CoreWeave this quarter adjusted operating profit | $128 million at 5% margin | 16:27 |
| CoreWeave prior-year operating profit | $200 million | 16:27 |
| CoreWeave year-over-year change | Revenue doubled, operating profit fell 36% | 16:27 |
Glossary
- Token Factory
- An inference business model that converts compute into token output and charges real-time per-unit usage.
- Power land
- Land with grid interconnection permits already obtained, equipped with substations and water access for building datacenters.
- Post-training
- Secondary training or fine-tuning done on open-source base models using customer proprietary data.
- Gateway
- A neutral multi-model API endpoint that avoids system-level optimization to preserve neutrality.
- EBITDA
- Earnings before interest, taxes, depreciation, and amortization—a profit metric that excludes depreciation and interest but can mask real financial risk.
How to listen
Investors focused on AI infrastructure and cloud-compute businesses, and founders of inference platforms trying to assess where their competitive moat lies.
The opening minutes are pleasantries and setup; the substantive analysis starts with the moat discussion, so you can skip the introduction.