Google BigQuery predictive AI enables SQL-based forecasting without manual training

The introduction of TabFM simplifies the streaming data stack by collapsing the workflow between data scientists and SQL analysts. For streaming platforms managing high-volume subscriber data, this reduces the engineering overhead typically required to move data between storage and external machine learning platforms. By keeping data within BigQuery, enterprises can improve governance and speed up ad hoc experimentation for churn and fraud detection. However, the transition to token-based pricing in late 2026 means strategists must carefully weigh the convenience of foundation models against the long-term cost efficiency of custom-trained models. Watch for whether Google expands the 20-feature limit to accommodate more complex streaming telemetry datasets.

Google's TabFM arrives as cloud providers race to embed pre-trained models directly into their data platforms, reducing the need for separate ML pipelines. In May 2025, Google Cloud announced that BigQuery now supports Gemini-powered natural language queries and automated data preparation features, part of a broader strategy to make analytics accessible to non-specialists. Databricks responded with its own foundation model approach, launching DBRX in March 2025 as an open model optimized for enterprise data tasks including SQL generation and classification, signaling that the foundation-model-for-analytics category is becoming a competitive battleground among hyperscalers. Amazon Web Services has similarly integrated SageMaker models into Redshift Serverless for in-database inference, though it has not yet shipped a single foundation model dedicated to tabular prediction tasks. On the business side, Google's pricing model for TabFM reflects a broader shift in how cloud vendors monetize AI workloads. Google Cloud confirmed in July 2025 that BigQuery would transition to token-based pricing for AI and ML functions by late 2026, replacing flat per-slot billing for inference workloads. This mirrors the approach OpenAI and Anthropic use for API access and creates new cost-planning challenges for enterprises running high-volume batch predictions. Analysts at Gartner noted in its 2025 Magic Quadrant for Cloud Database Management Systems that token-based pricing for embedded AI remains a risk factor for cost predictability, particularly for organizations with variable workloads. For streaming companies processing millions of subscriber events daily, the economics of per-token inference versus traditional model serving will likely determine adoption speed. Technically, TabFM's 20-feature limit positions it as a rapid-prototyping tool rather than a replacement for gradient-boosted trees on complex datasets. Independent benchmarking by Nord-IQ Research in August 2025 found that foundation models for tabular data matched XGBoost accuracy on datasets with fewer than 15 features but fell behind by 3 to 7 percentage points on high-cardinality feature sets, consistent with Google's own guidance. The approach builds on research from Google DeepMind's TabPFN lineage, where . For streaming analytics teams, the practical implication is that TabFM suits quick churn-signal detection or A/B test triage, while production-grade recommendation scoring and fraud detection at scale still favor purpose-trained models like XGBoost or LightGBM deployed through Vertex AI or external serving infrastructure.