Golden Data

Issue #10 · July 12, 2026

The AI budget got a meter

Paul BrownWrites Golden Data from Ogden, Utah6 min read

Every Sunday I look at the line item the CFO signed but never asked about. This Sunday it is the AI budget.

For two years the CFO signed a lump sum against a per-seat subscription and hoped the vendor would say something before usage got out of hand. The vendor did not. Uber burned four months of its 2026 AI budget through Claude Code by April. Tesla capped engineer AI spend at $200 a week after the invoice arrived. The finance team got surprised. The vendor got the money.

Then OpenAI shipped spend caps on June 18. Microsoft, Anthropic, and Databricks matched inside eighteen days. Spend caps. Credit-level attribution. Threshold alerts. The AI budget stopped being a black box. The AI budget got a meter.

What’s actually shipping

1. OpenAI shipped credit-level spend caps for ChatGPT Enterprise on June 18, 2026. The Global Admin Console breaks credit consumption down by user, product, and model, with a unified Cost API behind it. Workspace default cap, group-level caps, individual overrides. The frontier lab shipped the same governance layer the FinOps team already ran for cloud. Everyone else had eighteen days.

2. Microsoft made metered billing mandatory for Copilot Cowork on July 1, 2026. Copilot Credits at $0.01 each. Cost per task depends on the model, the context retrieved, the tools called, and the run length. Tenants with a Frontier user active between March 30 and June 16 had until July 1 to configure usage-based billing. After that, Cowork access was suspended. Overage enforcement triggers at 125% of prepaid capacity. The productivity vendor’s grace period ran out. The customer either configured the meter or the agent stopped working.

3. Anthropic added spend alerts and model-level entitlements to Claude Enterprise on July 2, 2026. The admin dashboard shows usage and cost by group and by user, with output — artifacts created, files edited, skills invoked — displayed next to the cost that produced it. Spend-threshold alerts fire at 75% and 90% of the org cap; users get in-app notifications at 75% and 95% and can request an increase without leaving Claude. The Admin API scripts the same cost-control workflows. The model vendor stopped waiting for the CFO to notice.

4. Databricks turned Genie into pay-as-you-go on July 6, 2026. No seat fees. No per-user licenses. Each identified user gets 150 free DBUs a month — about $10.50 at US East rates — and every DBU above that is billed on underlying LLM consumption. Budgets and alerts run at the account, workspace, group, or user level. Automation against a service principal gets zero free allowance. The BI vendor stopped selling the tool. The BI vendor started selling the tokens.

That last one is the tell. A meter is a budget instrument and a billing instrument. It is the same object. Databricks did not hand finance a cost control — it moved off a capped seat fee onto an uncapped consumption bill and shipped the alerts that make an uncapped bill survivable. Read the fortnight from the other side and the trade is legible. The CFO got visibility. The CFO gave up the ceiling. Every vendor on this list now reports the number that sets its own revenue.

What I’d ship in your app this week

The vendors report their own meters. Your customer’s finance lead now has four dashboards and no way to check any of them. Two tools you can ship in two weeks, so your product’s AI cost is a row they can audit instead of a number they take on faith.

Feature one: the per-tenant AI-cost telemetry endpoint. For each AI feature you expose, publish a per-tenant endpoint the customer’s FinOps team can pipe into Cloudability, Apptio, or a BigQuery view — real compute cost by user, model, and task category. Your product lands on the same page as the rest of the AI bill, and lands there auditable.

  • Shape. One REST endpoint per tenant plus a nightly aggregation. FOCUS-compatible JSON schema so it drops into the customer’s pipeline.

  • Data shape. Per (tenant_id, user_id, feature_id, model_id, task_category, input_tokens, output_tokens, compute_cost_cents, timestamp).

  • System shape. Aggregation query on the model-call log. No new LLM call. Under 200ms per tenant read, under $20 per month per 5,000-seat tenant.

  • Eval. Every pilot’s FinOps lead reconciles the endpoint total against the invoice. Under 2% variance before general availability.

  • Instrumentation. Track pull rate. Hourly scrapes mean production FinOps is running against your product; silence means the value isn’t visible yet.

  • Two weeks in. Three customers pipe it into their governance stack. If none have, the schema is wrong — sit with a FinOps engineer for an afternoon.

Feature two: the graceful-degradation cap on the customer admin surface. When a customer’s meter crosses their configured threshold, degrade instead of blocking. Route to a cheaper model. Batch instead of stream. Cache-first. Surface it on the admin surface so the finance lead sees what changed and can lift the cap. Microsoft’s alternative is a hard cutoff at the worst possible moment.

  • Shape. One per-tenant config (soft_cap, hard_cap, degradation_policy). One switch in the model-selection code path. One row on the admin surface: current spend, cap, mode.

  • Data shape. Per (tenant_id, current_spend_cents, soft_cap_cents, hard_cap_cents, mode = full | degraded | blocked, changed_at).

  • System shape. Model-selection reads mode from a cached lookup. Degradation swaps to a cheaper model and batches the call path. Degraded-mode latency is documented so the tradeoff is visible.

  • Eval. Review every degradation event for a month on one question: did the user notice before the finance lead did.

  • Instrumentation. Track soft-cap crossings, cap-raise requests, and downstream conversion. Customers who raise the cap validate the pricing; customers who stay degraded tell you the feature isn’t worth it.

  • Two weeks in. Three customers cross the soft cap and either raise it or live with degraded mode. If none do, the defaults are too tight.

Both ship in two weeks with the team you have. The vendors who shipped the meter made the AI budget a row on the general ledger. They also made themselves the only source for the number in that row. The vendor who publishes a number the customer can check is the one still on the invoice at renewal.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown