Hook: The EBITDA Mirage
CoreWeave just reported Q2 2025 numbers. Revenue hit $2.58 billion, up 112% year-over-year. Adjusted EBITDA margin? 58.5%. That is a number that would make any hyperscaler jealous. AWS runs at roughly 30% operating margin. Azure is lower. Google Cloud is barely breakeven. Yet here is a company that went public eight months ago at $40 per share, already printing margins that would make a DeFi protocol blush. But the GAAP net loss was $626 million. The market cheered. The stock popped. The narrative locked in: AI compute demand is infinite, and CoreWeave is the only pure-play pick and shovel.
But data demands respect, not reverence. A 58.5% EBITDA margin paired with a net loss is not a contradiction. It is a signal. Adjusted EBITDA is a construction. It strips out depreciation, interest, stock-based compensation, and restructuring costs. In a capital-intensive business like GPU cloud, depreciation is the single largest real cost. CoreWeave spent $1.5 billion on depreciation and amortization in Q2 alone. That is not a one-time adjustment. That is the cost of doing business. The question is not whether the company is growing. It is whether the growth is sustainable enough to absorb the structural costs before the debt comes due.
Context: The AI Infrastructure Wholesaler
CoreWeave is not a hyperscaler. It is not a model provider. It is a GPU wholesaler. The company buys NVIDIA’s latest chips, builds out data centers with liquid cooling and high-bandwidth networking, and leases the compute capacity to the largest AI labs and enterprises. Its competitive advantage is speed: it can stand up a GPU cluster faster than AWS or Azure because it is not burdened by legacy infrastructure or enterprise compliance layers. It also has a close relationship with NVIDIA, securing early access to B200 and next-gen Blackwell chips.
But the business model is simple: borrow money, buy GPUs, get customers to sign long-term contracts, and hope utilization stays high. The revenue is visible because of those contracts. The backlog at the end of Q2 was $104 billion, up from $99.4 billion at the end of Q1. That is a 4.6% sequential increase. But the revenue recognized in Q2 was $2.58 billion. Subtract that from the backlog change, and net new orders were roughly $7 billion. That is a strong number, but it is not accelerating. The Q1 backlog was $99.4 billion, up from $90 billion? We do not have the exact figure, but the trend suggests that the rate of new bookings is not keeping pace with the hype. The $104 billion backlog is a massive number, but it is a nominal number. Those are contracts, not cash. They are subject to cancellation, renegotiation, or delivery delays.
In my 2017 ICO audit, I saw $14 billion in token presale commitments that never materialized into product. The nominal value was a marketing tool. The same risk applies here. The backlog is a snapshot of signed agreements, but the terms matter. Are there penalties for early termination? Is the contract a binding commitment or a framework for future orders? The article does not specify. The market assumes it is guaranteed revenue. Data demands a closer look.
Core: The On-Chain Evidence Chain
Let us walk through the numbers as if they were transaction data on a blockchain. Every line item has a counterparty, a timestamp, and a context.
Revenue: $2.58 billion. That is a 112% YoY jump. But compare to Q1: CoreWeave reported $2.1 billion in revenue. So sequential growth was 23%. Annualized, that is around 92% sequential growth, which is impressive but not exponential. The Q2 revenue beat the raised guidance by a small margin. The company raised full-year revenue guidance to $12.4-$13.2 billion. That implies Q3 and Q4 average revenue of $3.5-$3.8 billion each. That is a significant ramp from Q2’s $2.58 billion. The company is betting on accelerating demand. But the backlog growth of only 4.6% suggests that the existing contracts do not support that ramp. The company must sign new contracts at a faster rate in Q3 and Q4. If they fail, the guidance will be missed.
Adjusted EBITDA: $1.51 billion. The margin of 58.5% is high. But we need to decompose it. The largest adjustments are depreciation and stock-based compensation. In Q2, depreciation was $1.5 billion. Stock-based compensation was likely a few hundred million. The GAAP net loss of $626 million means that after interest expense and taxes, the company is still deep in the red. The interest expense is a growing burden. CoreWeave has taken on significant debt to finance its capex. The company raised $1.5 billion in debt in Q1 2025, and likely more in Q2. The interest on that debt is not included in adjusted EBITDA. As rates stay high, the interest expense will eat into the cash flow.
Capital expenditure: The company raised its capex guidance to $35-$39 billion for the full year. That is roughly 3x the revenue midpoint of $12.8 billion. This is a capital-intensive model. The company is spending $3 for every $1 of revenue. That is not sustainable without external financing. The IPO raised $2.5 billion. The debt raises have added another $3-$4 billion. But the capex is accelerating. If the company spends $35 billion in 2025, it will need to raise another $10-$15 billion in debt or equity. That will dilute shareholders or increase leverage.
Now, the new contract margins: CEO Matt McDivitt stated that new contracts signed in Q2 have margins 5-10 percentage points higher than the new contracts signed in recent quarters. That is a positive signal. It suggests that CoreWeave is gaining pricing power. But it also implies that the earlier contracts were underpriced. The company may have offered discounts to win anchor customers. As those contracts roll off, the margin profile improves. But the timing is critical. The backlog is weighted towards the future. If the high-margin contracts start later, the near-term EBITDA might be lower than the average.
Contrarian: The Correlation-Causation Trap
The market is buying the narrative: AI demand is infinite, GPU supply is constrained, CoreWeave is the conduit. The data supports the correlation: revenue up, backlog up, margins up. But correlation is not causation. There are three blind spots.
First, the demand for AI compute is not homogeneous. The initial wave was training large models. Training requires massive compute clusters, but it is a one-time or periodic cost. Inference is recurring and cheaper per token. The demand for training is peaking as the largest models are already trained. GPT-5, Gemini 2, and Claude 4 are in production. The next wave of AI models may be more efficient, requiring less compute. CoreWeave’s revenue is heavily weighted towards training workloads. If the industry shifts to inference, CoreWeave’s advantage in speed of deployment matters less. Hyperscalers already have inference infrastructure in place.
Second, the competitive landscape is shifting. NVIDIA is reportedly developing its own cloud service, similar to what AWS does with its own chips. NVIDIA’s DGX Cloud is a direct competitor. If NVIDIA decides to prioritize its own cloud, it could starve CoreWeave of GPU supply. CoreWeave’s relationship with NVIDIA is a double-edged sword. It is the company’s biggest asset and its biggest liability. In my 2022 Terra/Luna analysis, I saw a similar dependency: Terra’s UST relied on the Luna token. When the relationship broke, the entire system collapsed. CoreWeave is not Terra, but the dependency is real.
Third, the macro environment: interest rates are still high. CoreWeave’s debt carries a floating rate component. If rates stay high, the interest expense will grow. The company’s net loss will widen. The EBITDA margin can only hide the interest cost for so long. The market will eventually focus on free cash flow. In Q2, free cash flow was negative by a wide margin. The capex outflow was $8.8 billion (quarterly level). Even if we assume some capex is for growth, the maintenance capex alone is significant. The company is not generating cash. It is a cash-burning machine that relies on the capital markets to survive.
Takeaway: The Next Signal to Watch
CoreWeave’s Q2 report is a study in selective disclosure. The adjusted EBITDA is a mirage. The real story is the cash flow and the backlog quality. The next signal to watch is the Q3 backlog growth. If it accelerates to 10%+ sequential growth, then the demand is real. If it stays below 5%, the market will reprice the stock. Also, watch for any customer concentration disclosures. The top 10 customers likely account for over 80% of revenue. If one of them signs a competing deal with AWS or Google, the stock will drop.
Gravity always wins when leverage exceeds logic. CoreWeave is flying high on the AI tailwind, but the altitude is dangerous. The data shows a company with strong revenue growth but negative cash flow, high debt, and a single-supplier dependency. The market is pricing in perfection. One miss, and the volatility will be a tax you pay for uncertainty.
Data demands respect, not reverence. The numbers are public. The analysis is reproducible. The question is whether you believe the narrative or the evidence. In my 2017 ICO audit, I learned that promises are cheap. The blockchain never lies, but the people who write the contracts do. CoreWeave’s backlog is a promise. The cash flow is the truth. Watch the cash.
Volatility is the tax you pay for uncertainty. The market is certain about CoreWeave. I am not. The data tells me to wait for the next on-chain confirmation.
