Datadog bets DIY AI will mean it dodges the SaaSpocalypse
The theory is that its domain-specific model will beat generalist LLMs on results and economics
Datadog is close to releasing an updated AI model that it thinks will help it avoid the so-called SaaSpocalypse – customers using AI to build their own tools.
The observability tools vendor already created a model called Toto-Open-Base that the company's explanatory paper says it built with 151 million parameters, trained on more than two trillion time-series data points – apparently the largest pretraining dataset for any open-weights time-series foundation model. All the data used to train the model came from Datadog itself, gathered in the course of operating its SaaSy observability services.
In conversation with The Register, Datadog chief product officer Yanbing Li said the company is reviewing its next model but sees that effort as the means to an end.
"What is the SaaS company's role?" she asked, before answering: "To innovate in their domain."
For Datadog, that means creating a model specific to its domain – observability – rather than relying on a generic LLM.
Li thinks developing models brings two things to Datadog.
One is that AI becomes part of its platform, rather than requiring customers to set a token budget on another service. The other is better agents that detect and predict anomalies more effectively.
She claimed Datadog's site reliability agent can already investigate incidents, provide root cause analysis, and suggest remediation actions.
- AIOps is so powerful, vendors are building tools to clean up after agents break your infrastructure
- Cisco looses Splunk to probe and tame its growing agentic menagerie
- Snowflake buys Observe to make 'Days Since Last Outage' counters obsolete
- ServiceNow's new AI agents will happily volunteer for your dullest tasks
AI remains a flaky field and agents make mistakes. The Register therefore put it to Li that operators of mission-critical IT must be wary before letting agents suggest changes to their systems, let alone enact those changes without supervision.
She agreed and said for AI systems to win trust, their output must be both explainable and verifiable. Using its own models makes that easier for Datadog, she said. They have also helped the company to create a tool that watches AI platforms while they work and can detect signs they are producing hallucinated output.
"I do not worry about a race to develop models, but applying them," she said, adding that she thinks users will apply Datadog's models because they allow constant monitoring of health – a bit like wearable devices.
"Today, when we see a doctor, it is an expensive hassle, so we only visit when we are ill," she said. Smartwatches packed full of sensors, plus AI to analyze those signals, mean it's now possible to detect and predict illness.
Li thinks Datadog offers a similar change from occasional to constant diagnosis and can dodge the SaaSpocalypse.
"What is vulnerable in this transition is point tools, when customers do not act in your tool," she said. "Those things are more easily disrupted."
She reckons AI has seen Datadog transcend SaaS to become a platform.
Every vendor aspires to that status because it makes it harder for customers to leave. Maybe AI can solve that one day. ®
