AI spend is now big enough to need its own budget discipline
One company's cloud model platforms took 93% of its AI spend while ChatGPT and Cursor took under 4%. The fix vendors propose: watch spend daily, and score shadow AI by the data it touches, not how often it is used.
- In one instance Larridin describes, cloud model platforms took about 93% of 30-day AI spend, ChatGPT and Cursor combined took under 4%, and a single day of cloud model usage can represent up to 15% of a month's total. Self-reported m:…2320
- Larridin cites the FinOps Foundation's State of FinOps 2026 finding that 98% of practitioners now manage AI spend, and Harness's State of AI in FinOps 2026 finding that just 13% had basic visibility into it. Self-reported m:…2320
- Larridin cites IBM's 2025 Cost of a Data Breach Report: organisations with high levels of shadow AI had breach costs averaging $670,000 more, one in five reported a breach due to shadow AI, PII was compromised in 65% of shadow AI incidents, and only 37% of organisations had policies to manage or detect it. Self-reported m:…2319
- Greybeam says customers using it for read workloads see their Snowflake bills drop 86% on average, and Greybeam costs $100 a month plus hourly compute from $0.75 an hour. Self-reported m:…3083
The spike you only see daily
Usage-based AI spend can spike in a single day, and monthly or per-seat views will not show it until the invoice arrives. That is Larridin's starting point, and the example it gives makes the case: in one observed instance, a single day of cloud model usage represented up to 15 percent of the month. Usage-based pricing ties cost to tokens, calls and compute, so a batch job, a migration, an agent left running or a new model rollout can multiply spend overnight. m:…2320
The same instance shows where the money actually goes. Cloud model platforms combined, Vertex, AWS Bedrock and the Anthropic API, took about 93 percent of 30-day AI spend. ChatGPT and Cursor combined took under 4 percent. The seat licences people argue about are a rounding error; the model calls are the budget. Larridin cites two surveys that frame the gap: the FinOps Foundation's State of FinOps 2026 found 98 percent of practitioners now manage AI spend, while Harness's State of AI in FinOps 2026 found that just 13 percent had basic visibility into it. m:…2320
Larridin's proposed discipline is straightforward. Watch four dimensions, each against a normal range: source, meaning each cloud platform, API and tool as its own line; day, meaning daily cost against a trailing average; person and team, including spend marked unassigned; and model, meaning cost by model version, since a change of default model can shift spend quickly. Set a baseline per source from the last 8 to 12 weeks, and define thresholds, for example any day above twice the trailing daily average, or any person above a weekly cap. The implication is that AI spend is now a finance operation with daily controls, not a procurement line reviewed at renewal. m:…2320
Shadow AI, scored by data
The second Larridin post argues that the same discipline is missing on the risk side. Most shadow AI metrics weight tools by how much they are used, but risk depends on what data each tool touches and how many people feed it. The tools usually underweighted, Larridin says, are meeting notetakers and coding agents: even limited activity can expose sensitive conversations, files or code. A single recorded meeting could contain an entire customer conversation or pipeline review. m:…2319
The numbers Larridin cites come from IBM's 2025 Cost of a Data Breach Report. Organisations with high levels of shadow AI had breach costs averaging $670,000 more than those with little or none. One in five organisations reported a breach due to shadow AI in 2025. Shadow AI incidents compromised personally identifiable information in 65 percent of cases, and only 37 percent of organisations had policies to manage or detect it. Those are IBM's figures as reported by Larridin, not Larridin's own measurements, but they set the stakes for the scoring method that follows. m:…2319
The method is a one-to-three score for each unapproved tool on data sensitivity, number of users, vendor review status and what each event contains. The unapproved usage rate, unapproved events divided by all AI events, is still tracked, but the score is what prioritises action. Two practical warnings come with it: a desktop client of an approved tool may be misclassified as unapproved, so the inventory needs checking before it is trusted, and tools should be re-scored monthly, reporting the count of high-risk tools rather than just the usage rate. m:…2319
The dashboard tax
Greybeam's post adds a third kind of AI-adjacent spend: the analytics bill behind the dashboards that report on it. Tableau Cloud and Tableau Server already tag the Snowflake queries they send with the ID of the workbook behind them, and Snowflake's QUERY_ATTRIBUTION_HISTORY view stores that tag next to the credits each query used, so one query prices every workbook. The waste Greybeam points to is idle time: a warehouse that serves a few dashboards all day can burn more credits waiting between queries than running them, because Snowflake bills a running warehouse per second, with a 60-second minimum each time it starts, and keeps billing until auto-suspend shuts it down. m:…3083
Greybeam's answer is to route the read workloads elsewhere. The company says it is wire-compatible with Snowflake: point the Tableau data source's Snowflake connection at Greybeam and that is the whole change, with queries running on DuckDB and no schema migration. On read workloads like dashboards, the company says, customers' Snowflake bills drop 86 percent on average. Greybeam costs $100 a month plus hourly compute from $0.75 an hour. Headset, which serves Looker dashboards to its own customers, cut the cost of its analytics warehouses by 92 percent, according to Greybeam. Those savings are the vendor's own figures, and the net saving after Greybeam's own charges is the number that matters, not the headline 86 percent. m:…3083
The three posts add up to one message for whoever owns the budget. AI spend is now large, usage-based and fast-moving enough that monthly reporting misses it, and the tools employees adopt on their own carry data risk that usage counts do not capture. The discipline Larridin and Greybeam propose is the same in both cases: measure by the unit that actually drives the cost or the risk, daily and per tool, and act on the outliers. m:…2320 m:…2319
What is still unresolved
- Whether the 98% and 13% survey figures hold across industries, or reflect the FinOps practitioners surveyed.
- How much of the shadow AI breach premium reflects the tools themselves versus the absence of policies around them.
- Whether Greybeam's savings hold for workloads heavier than dashboard reads.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.