Google CEO Sundar Pichai admits that cloud business is forgoing short-term revenues

Google is giving up potential near-term revenue from customers that want to buy additional computing capacity. The sacrifice is not a stated dollar amount. It is the cloud income Google could earn by selling scarce chips, servers, and data-center space today.

The reason is strategic. Sundar Pichai said Google Cloud is making compute trade-offs to support frontier AI development. Google considers access for Google DeepMind non-negotiable because advanced models can reveal new abilities and help prepare future products. That makes internal research a priority even when paying customers are waiting.

The trade-off is becoming more visible as demand rises. Google Cloud produced $24.8 billion in second-quarter revenue, up 82% year over year, while its contracted backlog reached about $514 billion. Yet Pichai said serving a new customer seeking capacity in 2027 would be incredibly difficult, showing why Google is choosing long-term AI progress over some immediate cloud revenue.

Frontier AI means developing models that aim to reach the highest level of current capabilities. It matters because these systems can give Google an early view of emerging AI abilities. That insight may help Google prepare products for future uses and improve its position in a highly competitive field.

The computing burden is large at two stages. Models need substantial capacity during training, when they learn from vast amounts of data. They also need capacity while running, when users send requests and receive responses. The article says frontier systems require massive computing power for both purposes, competing directly with Google’s internal research needs.

Google’s own demand illustrates the pressure. Lumen Argon processed almost 25 trillion tokens during its first 72 hours, about three times Lumen’s volume over the same period. Google was already capacity-constrained while serving it, even before satisfying every cloud customer’s request.

Google Cloud is expanding rapidly, but demand is exceeding immediately available capacity. Second-quarter revenue reached $24.8 billion, an 82% increase from a year earlier. That figure shows the scale of current business and the speed at which customers are buying cloud services.

Its future commitments are even larger. Google Cloud’s contracted backlog was about $514 billion. Alphabet’s filing gave a more precise figure of $513.9 billion at the end of June. Slightly more than half of that backlog is expected to become revenue within two years, linking today’s capacity decisions to substantial future sales.

Alphabet expects to spend $195 billion to $205 billion on capital projects in 2026. Much of that investment will support AI and cloud infrastructure. Still, new capacity takes years to plan and build. Pichai said the computing available today reflects decisions made in 2024, so spending cannot instantly solve shortages.

Customers face delays or limits when they ask Google Cloud for more capacity. Google is reserving some of its scarce computing resources for internal frontier-AI research. That means cloud demand cannot always be met immediately, even though customers are willing to pay and Google Cloud’s business is growing quickly.

The shortage involves more than chips. It also includes servers and data-center space needed to run AI services. Pichai said fitting in a new customer seeking capacity in 2027 would be “incredibly difficult to do.” He also said everyone wished they had twice what they had planned, describing an industry-wide constraint.

The effects can last because capacity decisions happen years ahead. Pichai said today’s available compute reflects decisions made in 2024. Alphabet plans large infrastructure spending, but new facilities and equipment will not appear instantly. Until then, customers may face constrained access while Google balances contracted business against its own AI demand.

Google sees computing for Google DeepMind as essential to its long-term AI strategy. Pichai said developing advanced models gives researchers an early view of new abilities. That knowledge can help Google prepare products for future uses, rather than reacting only after competitors or customers discover those capabilities.

The company also expects frontier models to contribute to recursive self-improvement. In the article’s explanation, an AI system could assist with the research, coding, or testing needed to create a more capable successor. That makes access to frontier-scale computing useful not only for one model, but potentially for the next generation of models.

For that reason, Pichai rejected making smaller systems such as Lumen Google’s only focus. He said Google has been thinking about the AI frontier for well over a decade and would not compete any other way. Internal research therefore receives protected capacity, even when cloud customers want it.

Google could reduce pressure in some ways, but the article shows no simple escape. Smaller models generally require fewer resources, yet Pichai rejected focusing only on systems such as Lumen. Google wants to compete at the frontier, where it believes advanced models provide strategic knowledge and future product benefits.

Building more capacity is the clearest path described. Alphabet expects to spend $195 billion to $205 billion on capital projects in 2026, much of it for AI and cloud infrastructure. However, capacity decisions take years. Pichai said the compute available today reflects decisions made in 2024, so investment cannot quickly remove the shortage.

The article does not describe Google outsourcing its computing. Outsourcing could shift where capacity is obtained, but it would not clearly solve the broader shortage of chips, servers, and data-center space. Google therefore faces a continuing choice: reserve resources for frontier research, serve customers, or expand infrastructure far ahead of demand.

Cloud computing provides shared access to computing resources through large data centers. Specialized AI chips supply the processing needed for model training and for answering user requests. Servers contain and connect those chips, while data-center space houses the equipment, power, and cooling systems. Together, they form the physical foundation for AI services.

Model training uses this capacity to develop an AI system. After training, the model still needs computing power to run whenever users send prompts or products process data. The article highlights this continuing demand: Lumen Argon processed almost 25 trillion tokens in its first 72 hours and already strained Google’s capacity.

Google must allocate the same limited chain between its own research and cloud customers. Frontier models need enormous resources for training and operation. Customers need capacity for their services. Because new chips, servers, and facilities take years to plan and build, Google’s choices today shape AI access and cloud revenue later.