AI Business & Industry

Today's Stories

1. 1

   Big Tech AI spending surges past analyst forecasts

   Major technology companies are investing billions of dollars into artificial intelligence infrastructure and development at levels that exceed the estimates Wall Street analysts have made. The gap between actual spending and forecasted spending has left investors concerned and forced analysts to reassess their models; the scale of the buildout suggests companies see AI as strategically essential and are willing to outpace consensus expectations.

   The sustainability of this spending pace and whether it translates into revenue that justifies the capital outlays—the body does not specify timing or forward guidance on this point.

2. 2

   Snowflake builds AI-ready semantic layer to unify millions of internal data sources

   Snowflake created an internal semantic layer — a standardized data interface that sits between raw tables and AI agents — to resolve conflicting definitions across petabytes of business data. In July 2025 alone, over 400 internal users ran more than 5,400 queries through the product data science agent, while over 5,600 employees across all teams ran more than 320,000 queries using the broader semantic layer in sales, HR, support and other areas. Without a shared semantic layer, the same question (e.g. "What is an active customer?") returned different answers across different tables and metric definitions, forcing data science teams to manually validate every query. AI agents amplified the problem because new metrics and data sources were built at lightning speed, creating information gaps. A semantic layer standardizes meaning once and reuses it across all dashboards and AI agents, reducing token usage and improving text-to-SQL accuracy from 20% to more than 90% in AtScale's benchmark testing on 40 business questions.

   Snowflake's approach uses semantic view materializations to preaggregate metrics and manage performance at scale. The semantic layer is version-controlled with dbt integration, tested against popular dashboard tiles (which represent real user questions), and routed intelligently by Cortex Agents across all available semantic views. Semantic views are also interoperable with Apache Ossie™ (incubating), a vendor-neutral open standard.

3. 3

   Snowflake expands CoCo AI governance with access and behavior controls

   Snowflake is rolling out per-user AI credit quotas (now generally available) and three new governance layers (generally available soon)—organization-wide policy management, team-level agent profiles, and restricted session scope—that let administrators set spending limits, control which external tools agents can access, and constrain SQL permissions before a session starts. Platform and security teams can now approve broader access to CoCo across their organizations instead of confining it to sandboxes or manual approval workflows. The controls map to specific security review questions: spend bounds, policy consistency across surfaces, explicit data reach, and logged external tool access. Developers inherit these guardrails automatically without extra configuration work.

   Tools by Cortex AI Gateway (preview soon) centralizes MCP server governance and adds four controls that did not previously exist—server-level allowlisting, tool-level policy, per-server rate limits, and comprehensive audit trails of tool calls. Per-user quotas reset daily and monthly independently, preventing unexpected end-of-month charges and enforcing spending ceilings across CoCo in Snowsight, CoCo CLI, and CoCo Desktop.

4. 4

   Snowflake's Observe AI SRE Cuts Incident Investigation Time 4x on Average

   Snowflake announced Observe by Snowflake, an AI site reliability engineering (SRE) tool built on three integrated layers—unified telemetry storage, a context graph modeling semantic relationships, and an agent-optimized AI layer. Analysis of 3,163 AI SRE conversation spans from October to November 2025 found productivity gains consistently clustered in the 3x–10x range, with 30% of interactions showing more than 5x improvement and 5% exceeding 10x improvement. Today's incident investigations are slow despite abundant telemetry because most AI SRE tools are bolted onto architectures not designed to support them. The baseline for a complex incident spans 10 minutes to detect, 120 minutes to investigate, 15 minutes to remediate, and 370 minutes for root cause analysis—with only 30% completion. Snowflake's unified design means investigators spend less time stitching data from multiple tools and writing custom queries, especially for investigations requiring synthesis across large data volumes.

   Three automotive SaaS provider reduced incident investigation time from over three hours to minutes, and a location intelligence company reported that Observe's AI SRE and MCP Server could transform how teams investigate incidents. The largest gains came from investigations requiring rapid synthesis of data across multiple sources—a direct result of unified telemetry and context graph integration.

5. 5

   Snowflake Adds Smart Model Routing to Cut AI Costs Without Losing Quality

   Snowflake introduced dynamic model routing in Cortex AI Gateway, which automatically selects the lowest-cost model capable of completing each task, alongside new open models DeepSeek-V4-Flash 0731 (in private preview) and GLM-5.3 (private preview coming soon). Internal tests showed the routing approach achieved up to three times greater token efficiency on data pipeline workloads compared to always using the most powerful model, while maintaining comparable quality. Organizations pay for more AI capability than most tasks require—a weekly summary does not need the same reasoning power as portfolio risk analysis. By matching task complexity to model cost, enterprises can reduce inference spending without rebuilding applications or managing routing logic themselves. Snowflake's approach also ensures open models run inside the company's data governance boundary, keeping sensitive data within existing compliance controls.

   DeepSeek-V4-Flash 0731 is available today in private preview; GLM-5.3 comes soon and will be self-hostable. On Snowflake's internal ADE-bench test, DeepSeek-V4-Flash scored 74.4%, outperforming the proprietary model tested, while GLM-5.2 (the predecessor) achieved 66% accuracy with the lowest token footprint in the benchmark, making it suited for high-volume, cost-sensitive workloads.