Datadog has released its State of AI Engineering 2026 report, warning that operational complexity — not model capability — is now the primary bottleneck to scaling artificial intelligence reliably in production environments.

The report, based on anonymised telemetry from thousands of organisations running AI systems in production, finds that nearly one in 20 AI model requests — around 5% — fail outright, with nearly 60% of those failures caused by capacity limits rather than flawed model logic. The result: slowdowns, errors, and broken user experiences in AI-powered applications at a critical moment of enterprise adoption.

Multi-Model Complexity Now the Norm

The study finds that 69% of companies now run three or more AI models simultaneously, alongside increasingly intricate agent workflows. Meanwhile, token volumes are surging: median-usage teams are sending more than double the amount of data per AI request compared to a year ago, while heavy users have seen that figure quadruple.

On the provider side, OpenAI remains the most widely used at 63% market share, but adoption of Google Gemini and Anthropic Claude grew by 20 and 23 percentage points respectively — reflecting a broader shift toward multi-vendor AI strategies.

Agent framework adoption doubled year-over-year, accelerating development cycles but also introducing more moving parts and failure points into production pipelines.

Observability Becomes Mission-Critical

“AI is starting to look a lot like the early days of cloud,” said Yanbing Li, Chief Product Officer at Datadog. “The cloud made systems programmable but much more complex to manage. AI is now doing the same thing to the application layer. The companies that win won’t just build better models — they’ll build operational control around them.”

Li added that in this environment, AI observability is becoming as essential as cloud observability was a decade ago — a theme that aligns with growing enterprise demand for real-time visibility across GPU utilisation, model behaviour, and agent workflows.

Vercel CEO Guillermo Rauch echoed the concern: “The next wave of agent failures won’t be about what agents can’t do but what teams can’t observe. Unlike traditional software, agents have control flow driven by the LLM itself, making observability not just useful, but essential.”

Implications for Asia-Pacific

While the report draws on global data, the findings carry particular weight for Asia-Pacific enterprises racing to deploy AI at scale. The region has seen aggressive AI investment across financial services, government, and manufacturing sectors — often with teams that have scaled model count faster than their observability infrastructure can support.

Datadog has indicated it is available to provide Asia-Pacific context from a regional spokesperson for media seeking localised commentary.

The full State of AI Engineering 2026 report is available on the Datadog website.