AI Compute Crunch Intensifies: Anthropic and Alibaba Face Supply Shortages

The AI industry is confronting a significant compute crunch, a bottleneck that threatens to impede the rapid adoption and scaling of advanced AI models. Initial predictions of a "coming" shortage have proven understated, with evidence now indicating an active struggle among major providers to meet escalating demand. Anthropic, a prominent frontier AI lab, recently experienced severe uptime issues, at one point dropping to "one 9" reliability. This degradation was attributed by staff to "unprecedented growth" that was "genuinely hard to forecast." In response, Anthropic has reportedly taken corrective actions, including reducing the default effort for Opus 4.6, temporarily removing access to older Opus models (4, 4.1, Sonnet 4.5) from Claude Code, and disabling prompt suggestions. Such measures, particularly impacting a high-profile product like Claude Code, underscore the severity of the compute scarcity.

This issue is not isolated to Anthropic. The broader industry, as evidenced by reliability challenges on platforms like OpenRouter, suggests systemic compute contention. Alibaba Cloud's CEO, as early as November, acknowledged an inability to keep pace with customer demand for new server deployments, a situation that remains dire months later. The observed median output token/s of 6 for Alibaba Cloud's Qwen3.5 397B A17B model on OpenRouter further illustrates extreme inference resource contention. This data directly contradicts any "AI bubble" narrative suggesting an abundance of idle compute resources.

The surge in demand is significantly fueled by the "agentic inflection point," where models like Opus 4.6 and GPT 5.4 demonstrate advanced capabilities in complex tasks, particularly in software engineering and other professional services. These agentic processes inherently consume a substantially higher volume of tokens compared to traditional LLM uses, creating a compounding effect on compute requirements. Despite Anthropic's Claude Code generating an impressive $2.5 billion in annual run rate revenue, the underlying infrastructure is visibly strained. The current compute environment suggests that while AI capabilities are rapidly advancing, the physical infrastructure required to support widespread deployment is lagging, posing a critical challenge for the industry's near-term growth trajectory.