Silence speaks louder than the algorithmic hum.
On August 12, 2025, CoreWeave released its Q2 earnings. The headline numbers screamed velocity: $2.58 billion in revenue, up 112% year-over-year. Adjusted EBITDA hit $1.51 billion, a margin of 58.5%. The market nodded, the stock stirred. But beneath the surface, a quieter pattern emerged—a divergence between the polished metrics and the raw cash flow. The ledger remembers what eyes forget.
Context: The AI Infrastructure Wholesaler
CoreWeave is not an AI model company. It is a GPU cloud provider, a wholesale reseller of Nvidia H100 and B200 clusters. Its business model is simple: front-load massive capital expenditure to build data centers, lock in long-term contracts with hyperscalers and AI labs, and earn a spread between the cost of hardware and the price of compute. It went public in early 2025 at around $40 per share, riding the wave of AI inference demand. The company’s moat is not code—it is power procurement, supply chain relationships with Nvidia, and the ability to deploy capacity faster than traditional cloud providers.
But the Q2 report reveals a structural tension. The adjusted EBITDA margin is best-in-class, yet the GAAP net loss stands at $626 million. This is not a temporary mismatch. It is a clue to the underlying physics of the business.
Core: The On-Chain Evidence of a Dual Narrative
Let me walk through the data points as I would a set of validator logs.
First, revenue and EBITDA. Revenue of $2.58B is impressive, but the sequential growth from Q1 2025 (which was around $1.9B) is about 36%. That is still strong, but decelerating from the 150%+ year-over-year pace of earlier quarters. More importantly, the adjusted EBITDA margin of 58.5% is extraordinary for a hardware-heavy business. Traditional cloud providers like AWS or Azure operate at 30-35% margins. CoreWeave’s margin suggests one of two things: either they are pricing with extreme discipline, or they are capitalizing a significant portion of costs as capital expenditures rather than expensing them. The latter is common in early-stage infrastructure plays.
Second, the backlog. As of June 30, 2025, contracted backlog stood at approximately $104 billion, up from $99.4 billion at the end of Q1. That is a sequential increase of only 4.6%. In a quarter where they recognized $2.58B in revenue, the net new bookings were roughly $7.0B (assuming no cancellations). This implies a book-to-bill ratio of about 2.7x, which is healthy but not explosive. The CEO mentioned that new contracts signed in Q2 have 5-10 percentage points higher margins than recent quarters. This is a positive signal of pricing power, but it also suggests earlier contracts were underpriced—a potential drag on future margins when those contracts ramp.
Third, the guidance. Full-year 2025 revenue guidance was raised to $12.4B-$13.2B, implying H2 revenue of roughly $6.3B-$7.1B, up from H1’s $5.9B (approximately). Capital expenditure guidance was raised to $35B-$39B, up from a prior range that was likely around $30B. This is a huge number—capital spend is roughly 3x revenue. The company is burning cash to build capacity, but the GAAP net loss of $626M in Q2 reveals that depreciation and interest costs are eating into the reported profitability.
I traced the ghost in the validator’s code: the cash flow statement. CoreWeave’s operating cash flow in Q2 was likely negative due to the massive capex, but the adjusted EBITDA completely ignores depreciation and interest. The 58.5% EBITDA margin is a starkly different picture from the negative GAAP net income. This asymmetry is not a bug—it is a feature of capital-intensive infrastructure businesses. But the market often rewards EBITDA over net income, especially in high-growth sectors. The real question is whether the company can generate positive free cash flow within a reasonable timeframe as the installed base starts to generate cash faster than the new builds.
Beauty hides in the candle’s wick.
Digging deeper, I examined the capex-to-revenue relationship. In Q2, capex was likely around $9B-$10B (annualized rate of $35B-$39B). Revenue was $2.58B. That means for every dollar of revenue, the company spent about $3.5 on capex. This is unsustainable unless revenue growth accelerates dramatically or the capex intensity declines. The company’s own guidance implies H2 revenue will be roughly 4-5% higher than H1, but capex is staying high. The wick is burning fast.
Another hidden signal: the backlog growth of only 4.6% sequential. If the market is as hot as headlines suggest, why isn’t backlog growing faster? One explanation is that the company is deliberately limiting new signings to avoid overcommitting capacity. Another is that competition from other GPU cloud providers (e.g., Lambda, Vultr, or even AWS’s own GPU instances) is squeezing the deal flow. The CEO’s comment about higher margins on new contracts may indicate that CoreWeave is now able to price more aggressively because they have proven uptime and reliability—but it could also mean they are moving away from loss-leader pricing that was previously used to land anchor customers.
Contrarian: The Calculation Error in the Market’s Enthusiasm
Symmetry is a liar; asymmetry tells the truth. The market sees a 112% revenue growth, a 58.5% EBITDA margin, and a $100B backlog, and it extrapolates a straight line to cloud dominance. But the asymmetry between GAAP and non-GAAP, between capex and revenue, and between backlog growth and revenue recognition, suggests a different story.
First, the GAAP loss of $626 million is not just accounting noise. It reflects real cash costs: interest on debt to finance GPU purchases, depreciation of hardware that loses value quickly, and stock-based compensation that dilutes existing shareholders. The adjusted EBITDA margin is a smoothed version of the business, but it hides the cost of capital. In a rising interest rate environment (or even a stable one), the cost of debt for a company with negative net income is significant.
Second, the backlog of $104 billion is not a guarantee of future revenue. It includes contracts that may span 5-7 years, and many are subject to cancellation or renegotiation. The company’s own prospectus mentioned that a significant portion of its backlog is concentrated among a small number of customers. If one of those customers (e.g., Microsoft, OpenAI, or a Chinese AI lab) pulls back on spending, the revenue could evaporate quickly.
Third, the technology moat is thin. CoreWeave’s competitive advantage is not proprietary software or algorithms; it is access to Nvidia’s latest GPUs and the ability to build data centers faster. But Nvidia is investing in its own GPU cloud (DGX Cloud) and is also selling to competitors. The hyperscalers (AWS, Azure, GCP) are all building their own AI infrastructure. CoreWeave’s edge is temporary—it is the first mover in a segment that will become commoditized as GPU supply increases. The capital expenditure cycle is a double-edged sword: if demand growth slows, the company will be left with underutilized assets and massive depreciation drag.
Tracing the ghost in the validator’s code: the unit economics.
Let me do a quick back-of-the-envelope calculation based on publicly available data. Suppose CoreWeave has about 500,000 GPUs deployed (a rough estimate from its disclosure of $35B+ capex and current GPU prices around $30,000 per H100). At $2.58B quarterly revenue, that implies revenue per GPU per quarter of about $5,160, or $1,720 per month. This is at the high end of market rates, but it requires near-100% utilization. If utilization drops even 10%, the margin compression is severe. The EBITDA margin of 58.5% leaves little room for error.
Takeaway: The Next Signal
The next quarter will reveal whether the backlog growth accelerates or stagnates. If new bookings in Q3 exceed $8B (sequential growth of 15%+), the market’s enthusiasm will be validated. If they fall below $5B, the story changes. Pay attention to the ratio of capex to revenue—if it drops below 3x, the company is becoming more efficient. If it stays above 4x, the cash burn is accelerating.
Color coded, not just counted. The beauty of this data is that it shows two realities: one of high growth and high margins, and another of heavy cash consumption and customer concentration. The market will eventually reconcile them. The ledger remembers what eyes forget.
Painting with private keys—the asymmetry is the truth. CoreWeave is a bet on the continued scarcity of AI compute and the company’s ability to maintain its supply chain advantage. But the data whispers that the window is narrow. Watch the backlog growth rate. Watch the GAAP net income trajectory. For now, the silence between the block beats louder than the revenue hum.