Over the past quarter, CoreWeave's backlog ballooned to $104 billion—a headline that would make any bull roar. But the silence of the incremental growth told a different story. Net new orders added only $70 billion, a modest 7% increase from the prior quarter's $994 billion, while the company recognized $2.58 billion in revenue. The market cheered the beast, but I heard the whisper of the bear: demand is strong, but not exploding. And in that whisper, I found the truth about the AI compute economy.
I've been watching this industry since the days of ICOs, when trust was a whitepaper and value was a promise. Now, trust is a GPU cluster, and the promise is compute. CoreWeave is not a tech company in the traditional sense—it's a broker of silicon, a middleman between Nvidia's factories and the insatiable appetite of AI labs. Its IPO in early 2025 at $40 per share was a bet on the narrative that AI infrastructure is the new oil. But oil has a history of boom and bust, and the custodians of that oil often forget the covenant between the resource and the community.
Context: The Machine Behind the Machine
CoreWeave's business model is deceptively simple: buy massive quantities of Nvidia GPUs, build data centers, and lease the compute at a premium. Its moat is not in algorithms or software but in the ability to secure power, deploy clusters faster than competitors, and maintain high utilization rates. The company's Q2 2025 earnings report, released on August 12, revealed a revenue of $2.58 billion—a 112% year-over-year increase—and an adjusted EBITDA of $1.51 billion, implying a margin of roughly 58.5%. That's a staggering number for a cloud-like service, especially when compared to AWS's typical 30% operating margin. But the GAAP net loss of $626 million tells a different story: under the hood, the machine is bleeding cash.
The key to understanding CoreWeave lies in its backlog. The $104 billion figure represents the total value of contracts signed but not yet recognized as revenue. These are long-term commitments from AI companies—likely including names like Microsoft, OpenAI, and others—that guarantee a certain amount of compute over time. The backlog increased by only $70 billion net of the $2.58 billion recognized, meaning new signings were roughly $72.6 billion. That's a slowdown from the previous quarter's pace, suggesting that the buying frenzy may be stabilizing. The CEO boasted that new contracts in Q2 had margins 5 to 10 percentage points higher than recent quarters, which implies that CoreWeave is finally pricing its service at a level that reflects true demand. But every broken token taught me how to hold value—and here, the token is the GPU hour, and its value is only as stable as the contracts that bind it.
Core: The Theology of the Numbers
Let me walk through the gospel of the income statement. Revenue of $2.58 billion is impressive, but it's the cost side that reveals the soul of the business. Cost of revenue likely includes depreciation of GPUs, power costs, and data center operations. With a net loss of $626 million, the company is spending roughly $3.2 billion in total expenses, including interest, taxes, and depreciation. The adjusted EBITDA of $1.51 billion excludes these non-cash charges, giving a rosy picture. But the cash flow statement—which I had to infer from the report's details—shows a capital expenditure guidance of $35 to $39 billion for the full year. That's a massive reinvestment, far exceeding revenue. CoreWeave is essentially a machine that burns cash to build more machines, hoping that future revenue will cover the debt.
This is not unlike the early days of Bitcoin mining, where miners reinvested every dollar into ASICs, betting on a rising price. The difference is that CoreWeave's output is not a decentralized asset but a service rented to centralized clients. The company's guidance raised revenue expectations to $12.4-13.2 billion for the full year, but capex is set at $35-39 billion. That means for every dollar of revenue, they spend nearly three dollars on building infrastructure. The math works only if the revenue grows exponentially in the coming years. But the backlog growth of 7% suggests a linear trajectory, not an exponential one.
In the silence of the bear, we heard the truth: the AI compute market is not a bottomless pit. The hype cycle of 2024-2025, fueled by the ChatGPT boom and the fear of missing out, drove companies to sign massive contracts. But as the novelty wears off, the utilization rates of these GPU clusters may fall. CoreWeave's contracts lock in customers, but they also lock in the cost structure. If demand softens, the company is left with empty racks and depreciating assets. The covenant of compute is that the provider must serve the community, not just the shareholders.
Contrarian: The Centralization of Trust
Every article about CoreWeave frames it as a success story of AI infrastructure. But I see a different narrative: the centralization of compute power into a single point of failure. The company's reliance on Nvidia's GPU supply chain is a vulnerability. If Nvidia shifts its allocation, or if a competitor like AMD gains traction, CoreWeave's moat evaporates. Moreover, the company's data centers are concentrated in certain regions, making them susceptible to energy price shocks, geopolitical risks, or even natural disasters. My code was the covenant, not just the contract—and here, the code is the hardware, and the contract is the financial engineering.
The contrarian angle is that CoreWeave's high margins are a mirage. The adjusted EBITDA margin of 58.5% is achieved by excluding the cost of the very assets that generate revenue. If you include depreciation (which is a real cost, as GPUs wear out and become obsolete), the margin drops significantly. The net loss of $626 million is a red flag that the company is not yet self-sustaining. The market is pricing CoreWeave on future expectations, not current reality. This is a bet on the continuation of the AI boom, not a reflection of a profitable business.
Furthermore, the backlog of $104 billion is not all created equal. Some contracts may be cancelable or subject to renegotiation. The CEO's comment about improved margins suggests that earlier contracts were underpriced, meaning that the backlog might include lower-margin deals that will be fulfilled at a loss. The company's guidance of $35-39 billion in capex implies that they are building capacity for demand that may not materialize. In the world of crypto, we call this "over-leveraging on sentiment."
Takeaway: The Future of Compute as a Commons
CoreWeave's Q2 results are a testament to the power of narrative. The numbers are strong, but the cracks are visible to those who look beyond the headlines. The real question is not whether CoreWeave can grow, but whether the AI compute infrastructure should be centrally owned by a few players. The blockchain community has long advocated for decentralized compute networks—projects like Akash Network, Golem, and Render Network—that distribute the resource across many nodes, reducing single points of failure and aligning incentives with users. In the bear market of 2022, we learned that centralized services can collapse under their own weight. The same principle applies to AI compute.
In the silence of the bear, we heard the truth: the covenant between compute provider and consumer must be built on transparency, not just contracts. CoreWeave's success is a mirror reflecting our collective desire for speed, but it also reflects our neglect of resilience. As we head into a sideways market, the true value will be in the protocols that survive the crash, not the ones that merely ride the wave. The next cycle will reward those who built with the community, not just for the shareholders.
My code was the covenant, not just the contract. And the covenant demands that we question the silence of the bear.