The AI bill now has a gatekeeper: cost controls move into the gateway

The hardware price, in plain sight

Start with the unit economics. Fal's pricing guide puts the list price for a GB200 at $9.99 per GPU-hour, with an as low as price of $5.89 per GPU-hour, and notes that the B200 costs 20% less per GPU-hour at list price. A GB200 unit consists of one Grace CPU and two Blackwell GPUs, and the NVL72 rack contains 72 GPUs in one NVLink domain with 13.4 TB of HBM3e, drawing approximately 120 kW of power. Analysts estimate the purchase price of an NVL72 rack at approximately $3 million. NVIDIA reports the GB200 NVL72 is 30 times faster for real-time inference than the H100 on a 1.8-trillion-parameter MoE model, which is the vendor's own performance claim, but the pricing is what matters for budgeting: the frontier inference tier now has a visible hourly rate. m:…6759

That visibility is new in one specific way. A 2-trillion-parameter model at FP8 requires approximately 2 TB of memory for weights, while an 8-GPU B200 server holds 1.4 TB and an 8-GPU B300 server holds 2.1 TB. The memory math decides which hardware tier a model needs before any price negotiation starts, and the NVLink domain decides how far a model can scale without leaving the rack. Teams that used to budget inference as a cloud line item now have to reason about racks, memory, and power. m:…6759

Attribution first, then caps

Databricks is bringing cost attribution, access control, and monitoring to serverless notebooks, jobs, and pipelines, a move driven by usage: serverless compute usage at Databricks grew more than 3x in the past year. Serverless Compute Controls are generally available, while rate limits for automated workloads are currently in Beta and function as maximum autoscaling ceilings. Databricks provides system tables to track serverless DBU consumption, and serverless compute tags for attribution are currently in Beta. Monitoring for serverless cost is generally available on AWS, Azure, and Google Cloud, and Databricks plans to introduce shared budgets for interactive serverless sessions. The sequence is deliberate: attribute the spend to people and workloads first, then cap it. m:…1492

Snowflake's Cortex AI Gateway, released into public preview, takes the same idea one step further by putting a gate in front of the traffic. It provides a centralized entry point for AI traffic to enable cost control, and its specification allows administrators to define an explicit list of models that the gateway will serve. The AI_GATEWAY_USAGE_HISTORY view provides granular credit breakdown for requests touching multiple models, is available in both ACCOUNT_USAGE and ORGANIZATION_USAGE schemas, and retains 365 days of request history. Shared resource budgets for the AI Gateway can trigger stored procedures at specific threshold percentages, and per-user quotas can block AI requests within minutes of reaching a limit. m:…5916

The enforcement gap to watch

A Snowflake practitioner's account of the same controls adds the caveat that matters most. Snowflake splits spend into two meters, credits for warehouse compute and AI credits for AI services, and the two are distinct units that cannot be added together. Per-user quotas are the only AI cost control with built-in enforcement, and they became generally available in August 2026, but quota blocks are not real-time: they occur within minutes, allowing for potential spend overshoot. Budget thresholds can take up to eight hours to act, or two hours with latency optimization, and budgets notify users and run attached stored procedures but do not block spend. Anomaly detection monitors spend once per day, requires at least 30 days of history, and is limited to 20 monitors per account. m:…5880

The practical reading is straightforward. If your AI bill is small, attribution and budgets are enough. If it is large, the only control that actually stops spend is the per-user quota, and even that lets some overshoot through. Gateway allowlists decide which models can be called at all, which is a cost control disguised as a governance control. The teams that will keep inference bills predictable are the ones that combine all three: attribution to see it, quotas to stop it, and an allowlist to narrow it. m:…5916