In my inverted-bubble argument, I predicted that AI providers would eventually have to raise prices to support the economics of what they were building. Rising prices would expose a harder question: how much demand would remain commercially viable at the prices providers needed to charge? I expected this to be resolved by a lengthy market, driven back-and-forth between OpenAI and it’s competitors as I saw them at the time which were a small group of American frontier labs Attempting to gain dominance over each other.
What I didn’t expect was that the supply side would get so saturated so quickly, and CoreWeave is becoming an important test (and unwilling proof) of that argument. It has committed customers, valuable infrastructure and significant barriers to switching. It also has enormous financing requirements at a time when customers have more ways to distribute their computing expenditure. Higher prices can improve the return on individual contracts while leaving the company exposed to insufficient growth in the volume of business it wins. And it is exposed.
The financial numbers emerging from the industry make that tension harder to ignore. Microsoft disclosed $10 billion in quarterly revenue from commercial arrangements with OpenAI, including revenue-sharing payments, and $7.2 billion in outstanding receivables from the company at September 30. Those figures are not a measure of OpenAI’s losses. They reveal the extraordinary scale of the financial commitments connecting model developers and their infrastructure suppliers. (sec.gov) With so much instability and so much in motion, it’s very hard to predict what will happen. Will co panics slow down on training? Will demand for inference become so diffuse it spreads too far for stability? We don’t know. CoreWeave doesn’t know. But it needs the numbers to start to stabilize.
The question is how much profitable customer demand can sustain those commitments, and how that demand will be divided. AI adoption can continue expanding while particular providers lose market share, pricing power or access to affordable capital. There will be losers within a growing industry. CoreWeave’s combination of substantial financing obligations and dependence on future compute spending makes it particularly exposed.
CoreWeave’s own pricing has been moving upward. In September, the company said it was contracting new capacity at higher prices, including three-to-six-month agreements priced at approximately $40 million per megawatt on an annualized revenue basis. That is evidence of pricing strength in those agreements. It does not establish the economics of the entire fleet or the durability of those prices over several years. (investors.coreweave.com)
A provider’s need to charge more and its ability to secure higher prices can coexist. The test is whether enough customers will continue buying enough capacity at those prices to support the investment. We cannot establish from the announcement alone how much of the increase reflects scarcity, improved service, higher costs or the mix of contracts being signed.
CoreWeave’s latest full-year guidance anticipates $35 billion to $39 billion in capital expenditure against revenue of $12.4 billion to $13.2 billion. Infrastructure spending is intended to produce revenue over several years, so the comparison does not measure profitability. It shows how much investment depends on future receipts. All figures are in US dollars. (marketscreener.com)
Meanwhile, the routes through which customers can obtain compute are multiplying. Anthropic confirmed an agreement to use all the capacity at SpaceX’s Colossus 1 data centre, providing access to more than 300 megawatts and over 220,000 NVIDIA GPUs. Anthropic also describes training and running its models across Amazon Trainium, Google TPUs and NVIDIA GPUs. Large-scale demand can reach multiple suppliers and hardware platforms. (anthropic.com)
Meta has explored becoming a supplier as well. Reuters reported in July that it was discussing a potential two-year, $10-billion compute agreement with Anthropic. The strategic possibility is consequential: companies that purchase infrastructure from specialist providers can also build businesses reselling capacity to other customers. (reuters.com)
This creates a limit to CoreWeave’s switching moat. Moving an existing deployment can be difficult, expensive and disruptive. A functioning cluster combines business knowledge, a working relationship, chips, networking, storage, power and operational expertise. Customers have reasons to keep existing workloads where they are. Their next deployment is a separate decision.
A customer can maintain its CoreWeave relationship while placing additional training or inference workloads elsewhere. It could pay CoreWeave more per unit while directing a smaller proportion of its expanding budget to the company. Total AI spending could rise, customer retention could remain strong, and CoreWeave could still win less incremental business than its expansion anticipates. Technical capabilities and switching costs continue to protect installed workloads. Their protection is weaker around spending that has not yet been committed.
Existing agreements provide a substantial cushion. CoreWeave reported that 98% of second-quarter revenue came from committed contracts, which it describes as take-or-pay. Reduced usage would generally leave customers responsible for contracted payments, subject to contractual terms and CoreWeave delivering the service. But that protection has a duration, and it depends on that duration not stretching out, and customers remaining willing to pay. (sec.gov)
CoreWeave’s August $2.6-billion loan facility has an approximately five-year maturity, while the underlying customer contracts average approximately three years. The company explicitly described lenders as accepting renewal risk. Some of the economics therefore depend on the terms available when that capacity must be contracted again. (investors.coreweave.com)
A slowdown in training or inference spending would intensify the pressure. Even continued growth could disappoint if capacity arrived faster than paying demand. Customers could delay current projects, delay expansion, seek shorter commitments or negotiate more aggressively. A provider requiring higher prices would then face a market offering customers greater choice.
Inference is a source of potential demand, and CoreWeave already serves it. But the relevant measure is the revenue and cash generated by that work. More users, requests and tokens do not automatically translate into proportionately more infrastructure spending. Efficiency, utilization, hardware selection and competition all affect what customers must purchase. And there are ramifications beyond just the business: CoreWeave is a critical part of Canadian AI sovereignty infrastructure and a massive tenant to our biggest projects, both partnering with Bell and planned upcoming data centres from others.
The concern is the collision between what financial structure requires and how customer spending is evolving. Higher prices are necessary, but may be insufficient. Existing customers may remain loyal, yet distribute their next investments more widely. Infrastructure may remain useful, yet generate inadequate returns for the capital committed to it.
That is the next test of the inverted bubble: whether providers can sell enough at the prices their economics require. CoreWeave’s customers don’t have to abandon it for that equation to stop working. So far the market doesn’t seem to be absorbing that uncertainty well.
