As generative AI shifts from model training to larger-scale inference deployment, tech companies are increasingly allocating budgets towards data centers, cloud capacity, and chip commitments, rather than just funding the models themselves. On Thursday, CoreWeave announced it had signed an expanded agreement valued at $21 billion with Meta to provide cloud computing capacity through the end of 2032; this builds on a previous $14 billion agreement between the two parties and signifies Meta's ongoing expansion of AI infrastructure investment in a dual approach of self-building and outsourcing. Reuters reported on the same day that Meta's AI infrastructure investment could reach $135 billion this year.
This trend is not limited to Meta. According to Reuters, giant deals surrounding AI infrastructure over the past few months have involved companies such as OpenAI, Oracle, AMD, Nvidia, Google, Anthropic, Amazon, CoreWeave, and SoftBank. This includes OpenAI's reported long-term agreement to purchase around $300 billion worth of computing power from Oracle, the Stargate data center project's announced investment framework of up to $500 billion, and AMD's multi-year agreement to supply Meta with AI chips valued at up to $60 billion. In other words, capital expenditure is extending from "who has the stronger model" to "who can lock in electricity, data centers, GPUs, custom chips, and cloud capacity more quickly."
Demand is Forcing Supply to Expand
Behind the surge in spending is the unified assessment by cloud vendors and model companies of the demand for AI workloads: supply is still insufficient. On Thursday, Amazon disclosed for the first time that AWS AI services generate an annualized revenue exceeding $15 billion, accounting for about a tenth of AWS's $142 billion run rate revenue. The company also stated that current growth is constrained by capacity limitations, with customer demand far outstripping what existing infrastructure can support. This narrative aligns with Meta's outsourcing of CoreWeave capacity, Google's expansion of its Texas data center, and various companies signing long-term chip lock-in agreements. Companies aren't suddenly willing to spend more aggressively; they fear that if they don't secure resources now, they won't have enough computing power in the coming years.
This also explains why deal structures increasingly resemble long-term purchase agreements in the energy or commodities industry, rather than one-time purchases seen in the traditional software era. Many of the transactions listed by Reuters, in essence, correspond to multi-year purchasing commitments, reserved capacity, or joint construction obligations, rather than immediate cash investments. For example, CoreWeave's five-year, $11.9 billion contract with OpenAI is fundamentally a commitment to consume computing power; Meta's cloud agreements with Google exceeding $10 billion and Oracle's approximately $20 billion cloud negotiations with Meta are also "secure supply now, consume periodically later" infrastructure arrangements.
Chip Companies and Cloud Vendors Benefit Together
For companies like Nvidia, AMD, Broadcom, and Oracle, which are "selling shovels," the allure of this cycle lies in securing not scattered orders, but high-visibility demand across years. AMD has secured major AI chip agreements with both OpenAI and Meta; Broadcom signed a long-term custom AI chip collaboration with Google on April 6 and provides large-scale computing power for Anthropic based on Google's chips. For its part, Oracle, in addition to being reported to have signed a $300 billion contract for computing power with OpenAI, stated in February that it expects to raise $45 billion to $50 billion by 2026 to expand its cloud infrastructure capabilities.
CoreWeave's recent actions particularly illustrate this model. On the same day it announced the expansion agreement with Meta, this Nvidia-backed cloud infrastructure company also disclosed plans to issue $1.25 billion in bonds and $3 billion in convertible bonds. In simple terms, AI clients are locking in long-term cloud capacity, and cloud service providers, in turn, are using these long-term contracts to leverage debt and capital market financing to continue buying GPUs, building facilities, and renting power. This makes the AI infrastructure cycle increasingly resemble a high-leverage, long-cycle, asset-intensive utility model, rather than just a lightweight software story.
The Risk Lies in the Large Amounts, But Varying Standards
However, simply adding these headline figures can easily exaggerate the actual capital intensity. Both Reuters and Breakingviews highlight that many of the "tens of billions or even hundreds of billions of dollars" figures appearing in the current market are not entirely comparable: some represent purchase caps, some are investment frameworks, some are reported negotiation amounts by the media, some are multi-year phased contracts, and others involve mixed terms of equity, cloud services, and equipment supply. Breakingviews even estimated on April 7 that global AI data center plans could involve investments totaling over $6.6 trillion by 2030, while actual financing, electricity usage, construction, and revenue realization capabilities may not keep pace.
From a market perspective, the question is no longer whether AI will continue to spend money, but whether this wave of investment can convert into stable revenue over a sufficiently long period. The revenue competition between OpenAI and Anthropic, AWS's first disclosure of annualized AI revenue, and Meta's increase in annual capital expenditure to the $115 billion to $135 billion range all indicate that leading companies believe demand can justify heavy capital investment. But if the commercialization of models lags behind expectations, or if the costs of power, land, equipment, and financing continue to rise, this AI infrastructure race could quickly shift from a "growth story" to a "balance sheet story."
