A backlog of 100 billion dollars but deep losses: the life and death of the Big Three in Xinyun's capital efficiency

According to Woofun AI, the computing power hunger caused by generative AI is reshaping the infrastructure landscape, and the industry's focus has changed from a simple shortage of GPUs to who can turn electricity, chips, and data centers into usable computing power at lower capital costs. By dismantling the latest financial reports from CoreWeave, Nebius (NBIS.US), and Cerebras, App Economy Insights's “How They Make Money” revealed the core conflict that Neocloud (new cloud vendor) faces in the process of building next-generation AI infrastructure between debt overdraft and revenue fulfillment.

As the purest representative of the new cloud model, CoreWeave's business logic is based on purchasing Nvidia (NVDA.US) GPUs, deploying them in data centers, and then leasing them to leading customers such as OpenAI, Microsoft (MSFT.US), and Meta (META.US). Long-term commitment contracts contributed 98% of the company's revenue in the second quarter, with on-demand usage accounting for only 2%. Total revenue increased 112% year over year to $2.6 billion.

However, high data center rents, electricity, and other expansion costs are growing faster than revenue, causing gross margins to drop 8 percentage points to 66 percent. The company recorded an operating loss of $49 million and a net loss of $626 million, of which $6.4 billion in interest expenses directly related to GPU-secured debt financing became a heavy burden. Worse still, the profit and loss statement did not fully reflect the size of its expenses. Capital expenditure for the second quarter reached 9.4 billion US dollars, more than three times the revenue for the quarter.

Although demand far exceeded production capacity, the backlog of orders reached $104 billion, up 246% year over year, and a $25 billion customer commitment was signed at the beginning of the third quarter, the inflection point in profit margins has yet to be fully established. Although the adjusted operating margin rose to 5% from 1% in the previous quarter, the revenue structure is improving as a contribution margin of 5 to 10 percentage points higher than recent contracts in the second quarter. Coupled with the price increase of about 25% in July. Annual recurring revenue from the storage, CPU, networking, and software business has surpassed $400 million, and the contracted annual recurring revenue of the managed inference business jumped from $1 million to over $100 million in a quarter.

However, the cost of growth is huge, and CoreWeave raised the 2026 capital expenditure forecast to $35 billion to $39 billion, with the goal of increasing active power capacity to over 1.85 GW by the end of the year. Compared to the target revenue annualized operating rate of $19 billion at the end of 2026, the $104 billion backlog of orders appears huge, but order conversion is still limited by physical production capacity. Its investment logic ultimately depends on the ability to quickly convert power and GPUs into revenue while preventing financing costs from eating up profit margin improvements.

Nebius (NBIS.US)'s rise path is quite different, stemming from the 2024 split of Russian tech giant Yandex. Its Nasdaq-listed Dutch holding company sold the Russian business for $5.4 billion, and the retained international business then formed Nebius (NBIS.US), and Yandex co-founder Arkady Volozh took back the helm. Unlike the Rebuilding Internet Integrated Group, Nebius (NBIS.US) uses existing engineering talents and cloud computing experience to create a cloud platform designed specifically for AI. In the second quarter, AI Cloud's business revenue reached $575 million, accounting for 98% of total revenue, and the company's total revenue increased 454% year over year to $582 million. Gross margin increased 6 percentage points to 77%, and adjusted EBITDA reached US$236 million, corresponding to a profit margin of 41%, but the company still recorded an operating loss of US$176 million, mainly due to the depreciation and amortization of US$260 million brought about by the new infrastructure included in the income statement.

Notably, the price of computing power is rising. The average value of the four new AI cloud contracts is over $1 billion, with a contract value of 20 million to 25 million US dollars per megawatt, and short-term contracts of up to 40 million to 50 million US dollars per megawatt.

According to data compiled by Woofun AI, management expects the contracts signed in the second quarter to recover capital expenses and operating costs in about 22 months, which is significantly shorter than the previous two to three year cycle. Although capital expenditure for the second quarter was as high as $5.7 billion, nearly 10 times the revenue for the quarter, and capital expenditure is expected to reach $20 billion to $25 billion for the full year, customer advances provided strong support. Nebius (NBIS.US) expects to receive more than $9 billion in advance payments from customers in 2026, covering approximately 50% to 60% of related capital expenses.

This increase in capital efficiency has significantly increased the economic value of each additional megawatt of production capacity. In the long run, this is more strategically significant than the 454% revenue growth rate.

As an alternative among new cloud vendors, Cerebras did not buy Nvidia (NVDA.US) GPUs, but instead designed its own wafer-level processors and commercialized them through sales systems and Cerebras Cloud rental computing power. In the second quarter, the company's revenue increased 74% year over year to $180 million, with cloud and other services revenue growing 281% to $126 million and hardware revenue falling 23% to $54 million. The company recorded an operating loss of US$477 million, which was mainly affected by large equity incentive costs arising from the May listing. Excluding relevant factors, the core operating loss was only $34 million.

The gross margin under the report was 14%, but the core gross margin reached 41%, an increase of about 9 percentage points over the previous year, down from 46.5% in the first quarter. This is partly due to paying to rent back previously sold systems to meet cloud service requirements. Core cloud business revenue nearly tripled to $128 million, surpassing hardware business for the first time. Cerebras raised its core revenue guidance for fiscal year 2026 to $8.8 billion to $890 million and raised gross margin and operating margin expectations. As of 2027, the company has operated or contracted more than 600MW of data center capacity, and core gross margin is expected to pick up after bottoming out in the third quarter.

Although remaining fulfillment obligations amount to $25.4 billion and OpenAI is still its main customer, turning the backlog into revenue still requires significant infrastructure investment. The real test is whether the company can effectively convert the huge backlog of orders while repairing profit margins as new production capacity is launched.

Together, these three new cloud vendors face a structural contradiction: market demand has exceeded available computing power, but to meet demand, huge amounts of capital must be invested before revenue arrives. This upfront capital investment often results in negative free cash flow, and debt, leasing, depreciation, and customer concentration are almost as important as revenue growth. A huge backlog of orders is not equal to revenue, much less cash flow. Production capacity must be built ahead of schedule, causing capital expenditure, debt, and depreciation to rise at the same time, while revenue delivery is lagging behind.

This timing mismatch makes it necessary for new cloud vendors to continuously balance the security of the capital chain and the speed of scale expansion during the expansion process.

At the same time, the trend of tech giants building their own production capacity poses a long-term competitive threat. Meta (META.US) is expanding its self-developed chips and multi-GW GPU clusters, and SpaceX is also starting to sell the right to use its Colossus cluster. These giants have stronger financial strength and vertical integration capabilities. Once their AI production capacity is fully launched, new cloud vendors may face the risk of becoming indispensable infrastructure partners to temporary exports to fill short-term computing power gaps.

This change in the competitive landscape requires new cloud vendors not only to maintain technological leadership, but also to prove their long-term value in terms of business models.

The next stage of competition will depend on three core variables: production capacity, profit margin, and ability to finance. New cloud vendors need to turn contract requirements into infrastructure that actually runs on electricity, and improve return on investment as utilization increases. At the same time, they must finance the next round of expansion to avoid debt or equity dilution undermining the economy of the entire business model. The speed of capacity conversion determines the pace of revenue realization. The ability to repair profit margins reflects an increase in operational efficiency, while financing capacity is related to the bottom line of survival. These three constrain each other and jointly determine the ultimate fate of new cloud vendors in the midst of fierce competition.

Demand is locked in, and capital efficiency will determine who wins. In a context where the shortage of computing power has become a consensus, the winner and loser of new cloud vendors is not only technological advancement or order size, but how to achieve efficient capital turnover and profit conversion under the pressure of huge capital expenditure. Companies that can optimize capital structures, shorten recovery cycles, and effectively control financing costs will have an advantage in this infrastructure race. Conversely, if capital efficiency issues cannot be solved, even if there is a 100 billion backlog of orders, it may fall into debt. This is the key to a new round of reshuffle in the AI infrastructure sector following the monopoly of traditional cloud computing giants.