
CoreWeave's $87B Balance Sheet: The Middleman That Hyperscalers Don't Need
A bear thesis argues CoreWeave is a replicable middleman: revenue tripled to $3.66B in Q3 2026, but net losses widened to $3.80 per share amid $87B in liabilities.
- By
- Rebecca Stone
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
CoreWeave carries roughly $87 billion of liabilities on behalf of customers who could, by a bear analyst's reckoning, carry that debt themselves. That figure sits at the heart of a new short thesis arguing the AI cloud provider is an unnecessary middleman — one whose thin margins and extreme leverage "portend substantial loss of value in a sizable portion of scenarios."
The numbers behind the argument are stark. Revenue nearly tripled year over year, hitting $3.661 billion in the quarter ended September 30, 2026, up from $1.365 billion a year earlier. But operating expenses almost tripled too, climbing from $1.313 billion to $3.454 billion. Net losses widened from $1.87 per share over the first nine months of 2025 to $3.80 over the same period in 2026.
What does CoreWeave actually do?
Founded in 2017 as a crypto mining operation, CoreWeave pivoted to supplying AI-ready compute. Its playbook: buy or rent data centers, fill them with high-end NVIDIA chips, and lease that capacity to hyperscalers and AI labs under long-term contracts.
The bear case sees a structural flaw. NVIDIA already sells chips directly to hyperscalers. Data center developers already build for them. CoreWeave inserts itself between parties who don't need a go-between.
The analyst draws direct parallels to two companies that leaned on the same "reinventing" marketing language and struggled: WeWork, which leased offices and re-leased them to tenants, and Safehold, which bought land under buildings at a 3%-5% cap rate — far more than that land was worth to the sellers. Both middlemen found their suppliers and customers could do the deal without them. WeWork went bankrupt; Safehold keeps struggling.
Why do hyperscalers use CoreWeave at all?
Amazon, Meta, OpenAI and Anthropic clearly have the sophistication to build their own AI data centers, the thesis notes. So why let CoreWeave spend the capital?
The answer offered is blunt: it's a good deal for everyone except CoreWeave. NVIDIA "doesn't care whether they sell to AMZN or CRWV," and sells to the highest bidder either way. The hyperscalers, meanwhile, "would rather have the $87B of liabilities sit on CoreWeave's balance sheet than their own." A stronger model, the analyst argues, would serve small customers incapable of building their own infrastructure. Instead, per CoreWeave's own 10-Q filing, concentration runs deep: "We recognized approximately 30% and 24% of our revenue from our top two customers for the three months ended September 30, 2026."
Can scale fix the margins?
Usually, revenue growth spreads overhead across a larger base and margins improve. Not here, the bear argues, because CoreWeave's two largest expenses are variable, not fixed:
- Depreciation scales linearly with more equipment owned.
- Interest scales linearly with more debt.
The tenth data center costs just as much as the ninth. Growth required "extraordinary amounts of debt to buy rapidly depreciating equipment."
What could prove the bear wrong?
The thesis flags one genuine escape hatch. Depreciation is a non-cash expense, so CoreWeave can generate significant cashflow despite paper losses. The open question is how much of that cashflow must flow back into replacing GPUs.
CoreWeave depreciates its technology equipment over six years — roughly in line with hyperscaler schedules. If chips keep most of their value well beyond that window, a larger share of cashflow becomes true earnings. If they don't, essentially all cashflow goes to replacement hardware just to hold revenue flat.
The bottom line, in the analyst's words: if your counterparties can replicate your function, the only way to keep their business is to compete on price. Absent a scenario where equipment outlives its depreciation schedule, the thesis predicts CoreWeave follows the WeWork path — and it's bearish on the rest of the Neoclouds too.
Source: seekingalpha.com
More from Rebecca Stone
Show full bio
Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,092 articles
Related articles
coreweave-s-70-slide-looks-cheap-but-the-debt-isn-t-analyst-keeps-hold
CoreWeave's 70% Slide Looks Cheap, But the Debt Isn't: Analyst Keeps Hold
coreweave-and-applied-digital-one-lease-two-fragile-balance-sheets
CoreWeave and Applied Digital: One Lease, Two Fragile Balance Sheets
coreweave-doubled-revenue-and-still-lost-more-money
CoreWeave Doubled Revenue and Still Lost More Money
coreweave-funds-ai-buildout-with-prepayments-and-2-6-billion-loan
CoreWeave Funds AI Buildout With Prepayments and $2.6 Billion Loan


