Datadog Report: AI Scaling Bottlenecks

Datadog released its 'State of AI Engineering 2026' report, revealing that operational complexity, rather than model intelligence, is the primary barrier to scaling AI reliably. The report, based on data from thousands of organizations, found that nearly 5% of AI model requests fail in production, with 60% of these failures attributed to capacity limits. Key findings include the widespread adoption of multi-model architectures (69% of companies use three or more models), the doubling of agent framework adoption, and a significant increase in data sent to AI models per request. Yanbing Li of Datadog emphasized that companies succeeding in AI will build operational control around their models, making AI observability as crucial as cloud observability. Vercel of Vercel echoed this, stating that agent failures will increasingly be about what teams cannot observe. Yadi Narayana of Datadog highlighted concerns about failure rates, rising token consumption, and the need for operational discipline, governance, and cost control.