In 2026, as generative AI applications rapidly penetrate, enterprise large-model call volumes have surged, making compute supply constraints increasingly apparent. The Financial Times has reported exclusively that Google has imposed capacity limits on Meta’s use of its Gemini model, a move that directly reflects the severe supply-demand imbalance in current AI infrastructure.

According to the report, Google Cloud, which provides underlying compute for some of Meta’s Gemini models, has taken restrictive measures to ensure service stability amid overall AI demand far exceeding expectations. This incident is not isolated but exemplifies the compute bottlenecks cloud giants face in the AI boom, and is expected to have profound effects on model development, deployment costs and competitive dynamics across the AI ecosystem.

Background: explosive AI demand outstrips supply

In 2025–2026, generative AI expanded from consumer chat tools into enterprise agents, code generation, multimodal analysis and other high-intensity use cases, driving exponential demand for high-performance GPUs/TPUs. Meta, as a heavy AI player, relies on multiple cloud providers including Google Cloud for training and inference of Gemini.

The Financial Times noted that Google’s capacity limits primarily targeted Meta’s high-intensity call requests. Industry sources say this reflects Google’s own data-center capacity nearing saturation while it must prioritize internal work (such as Google DeepMind model training) and core enterprise customers. Similar situations have emerged at other cloud giants, indicating AI compute has become a scarce strategic resource.

Core causes: supply-demand mismatch and infrastructure bottlenecks

1. Demand-side surge

Model scale increases: Frontier models such as Gemini continue to grow in parameter counts and context windows, so single inferences consume far more compute than in the early GPT-4 era.
Deepening application scenarios: Enterprises are shifting from simple chat to complex agent workflows, multimodal generation and real-time analysis, increasing call frequency and duration.
Industry-wide sprint: OpenAI, Anthropic, Google, Meta and others are simultaneously advancing next-generation models, further elevating compute demand.

2. Supply-side constraints

Lagging data-center construction: Although cloud providers are accelerating expansions, tight GPU supply chains (notably H100/H200/B200 series), power limits and construction timelines prevent new capacity from rapidly meeting demand.
Priority allocation: Cloud providers must balance internal R&D, large customer contracts and public-market supply, so heavy external users such as Meta are among the first to face limits.
Cost and energy pressures: The substantial electricity and cooling costs of AI training and inference compel cloud vendors to allocate resources more cautiously.

Direct impacts: short-term pain and cost transmission

Impact on Meta:

Service stability may be challenged, causing Gemini response delays or reduced availability, which would affect user experience for AI products such as smart assistants and content-generation tools.
Meta may need to accelerate building its own compute capacity or migrate to other cloud providers (such as AWS or Azure), increasing deployment complexity and short-term costs.

Impact across the AI value chain:

Slower model development: Limited compute could delay new model iterations and large-scale experiments.
Rising costs: Cloud compute prices may face upward pressure, ultimately passing through to enterprise customers and end users.
Reshaped competition: Companies with in-house data centers or a