What should enterprise leaders know about AI today?
On November 11, 2026, Ramp’s AI Index showed US corporate AI spending at a record, with token prices down 21% in a month and spending concentrated in a small group of heavy users. Google opened Gemini 4 Carbon to paying developers at a price that doubles after November 30, Databricks pitched an always-on Genie agent, and Railway confirmed attackers ran up about $740,000 on stolen AI keys. Leaders should treat AI keys like cash and benchmark spending against peers.
Executive takeaways
- Act nowAI Security: Railway says attackers targeted AI provider keys stored on its platform and spent about $740,000 on them; CISOs and CTOs should rotate AI keys and cap their spend.
- EvaluateAI Economics: Ramp’s data shows AI usage growing faster than spending as token prices fall, with customer support rising to 13.3% of tracked spend; CFOs and CIOs should check whether their AI budgets follow the unit-cost decline.
- EvaluateFoundation Models: Gemini 4 Carbon is now available to paying developers and enterprises at $2 per million input tokens until November 30, then $4; CTOs have a short window to test it at the lower price.
Ramp’s AI Index shows record AI spending led by a few heavy users as token prices fall
AI Economics · Ramp, The Decoder, November 11, 2026 Priority: Evaluate
What changed?
Ramp’s November 11, 2026 AI Index found that 56.8% of businesses on its platform pay for AI, and AI spending rose 1.65% in October to a new high. Anthropic is used by 44.8% of businesses, up 1.1 points, and OpenAI by 40.6%, flat (Ramp).
Spending is concentrated: the top 1% of firms spend a median $10,200 per employee per month, the median firm $16.32. The Decoder, reading Ramp’s data, reports that the top 1% drive 80% of OpenAI and Anthropic revenue (The Decoder).
Average prices per million tokens fell 21% in a month, while customer support rose from 0.8% of classified spend in June to 13.3%. Engineering still takes 72%, and the most expensive frontier models take 69.5% of spend.
Why does this matter for enterprises?
Falling unit prices mean usage is growing faster than bills. A flat AI budget can buy more work each month, so a rising budget needs a rising output measure.
The jump in customer support spending shows where AI moves next after coding. This does not mean every company is adopting at this pace: Ramp’s base skews toward tech-savvy, growth-oriented firms and misses free and personal-account use, so it likely overstates adoption.
What should leaders do next?
Decision for CFOs and CIOs: Is our AI spend per employee buying measurable output, and are we paying frontier prices for routine work?
Within 60 days, split AI spend by function and model tier, and move routine workloads to cheaper models where quality holds.
Sources: Ramp — AI Index, November 2026 · The Decoder — US corporate AI spending hits a new high
Google opens Gemini 4 Carbon to enterprises, with a price that doubles after November 30
Foundation Models · Google, November 10, 2026 Priority: Evaluate
What changed?
Google made Gemini 4 Carbon available on November 10, 2026, to paying developers in Google AI Studio and to enterprises in Gemini for business. In our October 29 issue the model was limited to vetted cyber defenders; its full cyber access still runs through the Fairwind Program (Google).
Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30, 2026. Google also previewed /swarm in Antigravity, which turns one prompt into a coordinated team of coding agents.
Google says Carbon ranks first on several agent and coding benchmarks at less than half the cost per task of the next-best external model. These are Google or partner-run evaluations, and Google says the model’s resistance to prompt injection still needs improvement.
Why does this matter for enterprises?
A frontier model with a published price step gives buyers a dated decision: evaluate before December or pay double for the same tests.
This does not mean Carbon is safe to give broad agent permissions. Google’s own caveat on prompt injection is the risk that matters most when agents read untrusted documents or websites.
What should leaders do next?
Decision for CTOs and CISOs: Which of our agent workloads should we benchmark on Carbon before its price doubles?
Within 18 days, run your own evaluation set on two or three agent tasks, with prompt-injection tests, and compare cost per completed task with your current model.
Databricks pitches an always-on Genie agent that works unsupervised across company data
AI Agents · TechCrunch, Fortune, November 11, 2026 Priority: Monitor
What changed?
Databricks unveiled an updated Genie on November 11, 2026, that runs in the background and completes multistep tasks unsupervised. It draws on a new ontology that maps a customer’s data on Databricks, including reporting lines and departmental projects (TechCrunch).
Chief executive Ali Ghodsi said he uses it daily, for example to rank fixes for the weakest areas of an employee survey. Coverage did not report availability, pricing or named customers (Fortune).
Why does this matter for enterprises?
Data platforms are becoming agent platforms: an agent that already sits on governed company data can skip much of the integration work that stalls agent pilots.
Ghodsi said Unity Gateway can cap usage and swap in cheaper open-weight models, which matters because background agents consume tokens all day. This does not mean the product is ready to buy; there is no release date or price.
What should leaders do next?
Decision for CDOs and CIOs: Is our data catalogue clean enough for an agent to act on it without supervision?
Within 90 days, if you run Databricks, audit access rights on the data an always-on agent would read before joining any preview.
Sources: TechCrunch — Databricks’ new AI agent · Fortune — Databricks launches agentic AI tools
Railway says attackers hunted AI keys on its platform and ran up about $740,000 in charges
AI Security · Railway, BleepingComputer, November 10, 2026 Priority: Act now
What changed?
Cloud host Railway’s November 10, 2026 incident report says an attacker escaped its build sandbox on October 20 and reached customer environment variables for at most 1.85% of workspaces, about 100,000. The attacker queried the variables database for AI provider key names first (Railway).
Railway estimates about $740,000 in unauthorized AI usage on stolen keys and says it will cover what providers won’t refund. Wiz links the attack to a wave that also hit Northflank, Warp and Lightning AI (BleepingComputer).
Why does this matter for enterprises?
AI API keys are now a theft target in their own right: they convert straight into compute that attackers resell or use to train models, and the bill lands on the key owner.
Railway describes a contest of agent speed, with attackers probing many routes in parallel. This does not mean Railway customers’ databases were accessed; Railway found no evidence of that.
What should leaders do next?
Decision for CISOs and CTOs: Would we notice within hours if one of our AI keys started spending?
Within 30 days, inventory AI provider keys across hosting platforms, rotate any stored with third parties, and set per-key spending caps with alerts.
Sources: Railway — Incident report: October 21st security attack · BleepingComputer — Railway breach leads to $740,000 in fraudulent AI charges
What should enterprise leaders do next?
| Action | Owner | Timeframe | Success measure |
|---|---|---|---|
| Put spending caps on every AI key | CISO with CTO | 30 days | every production AI provider key is inventoried, rotated if stored with a third party, and capped with an alert |
| Benchmark AI spend by function and model tier | CFO with CIO | 60 days | one report showing spend per employee and per function, with at least one routine workload moved to a cheaper model |
| Test frontier models against your own tasks before price changes | CTO | before November 30, 2026 | cost per completed task and prompt-injection results for Gemini 4 Carbon and your current model on the same evaluation set |
The bigger picture
- Cheaper tokens raise the stakes of control: Ramp’s 21% monthly price drop and Railway’s $740,000 in stolen usage both show that AI consumption scales fast, for the owner and for whoever holds the key.
- Agents are moving from chat to background work: Databricks’ always-on Genie and Google’s /swarm both run unsupervised, so permissions and spending limits matter more than prompts.
Executive FAQ
How much are companies spending on AI in 2026?
Ramp’s November 11, 2026 index shows a record high, with the top 1% of firms spending a median $10,200 per employee per month and the median firm $16.32.
How much does Gemini 4 Carbon cost?
Google charges $2 per million input tokens and $10 per million output tokens until November 30, 2026, then $4 and $20.
How do companies protect AI API keys from theft?
Keep an inventory of keys, rotate any stored on third-party platforms, and set per-key spending caps with alerts; Railway’s attacker searched specifically for AI provider keys.
About the curator & AI disclosure
Curated by Luciano Ayres
Luciano Ayres is Director of Artificial Intelligence at ArcTouch, with more than 20 years of experience in software and technology leadership and specialized studies in deep learning. He focuses on helping organizations connect AI capabilities with practical business outcomes, responsible adoption, and scalable implementation. Connect with him on LinkedIn to discuss enterprise AI opportunities, strategy, investment decisions and risks.
AI-generated briefing. This issue was produced through an automated AI research and writing workflow curated by Luciano Ayres. Sources are linked for independent verification; summaries and analysis may contain errors and do not replace professional due diligence.