Welcome to the Enterprise AI Hangover

It feels like we're having an Enterprise AI hangover — just as Anthropic and OpenAI  climb toward IPO territory.

A few months ago, the cultural mood was tokenmaxxing: engineers competing on how many tokens they could burn, with productivity measured in AI consumption. That phase is ending. The next one doesn't have a name yet — call it tokenmindfulness — but everyone is suddenly worried about AI spending. Consider the recent record:

Uber's CTO reportedly blew through the company's entire 2026 AI budget in the first few months, with per-engineer API costs running between $500 and $2,000 a month.

Microsoft cancelled thousands of internal Claude Code licenses last month after costs spiralled past expectations, six months into the pilot. 

Meta engineers built an internal leaderboard called Claudeonomics on the company intranet — tracking AI token consumption across 85,000+ employees, with the top 250 power users ranked. Over 30 days, total usage exceeded 60 trillion tokens — somewhere between $100M and $900M depending on actual pricing terms. The leaderboard gamified usage with titles like "Token Legend," "Cache Wizard," and "Session Immortal." Meta killed the dashboard two days after The Information broke the story.

An AI consultant tells Axios one of their clients recently spent half a billion dollars in a single month after failing to put usage limits on Claude licenses for employees.

Everyone is worried about AI spending now. And worries about AI transformation efficiency haven't gone away either. It's become obvious that Enterprise AI adoption takes more than buying seats at Claude and ChatGPT.

The market response: cheaper models, smaller models, and routing

On OpenRouter, DeepSeek's share of tokens consumed has now overtaken Anthropic's. 

The big labs have noticed. Google's recent price cuts on Gemini are aggressive enough to make OpenAI and Anthropic visibly nervous, and OpenAI is now reportedly mulling its own price slash ahead of Anthropic's next round of competition. The pricing war that the model layer has been dancing around for two years is finally starting.

On the SaaS side, the SaaSpocalypse might be softening

Private equity mogul Orlando Bravo declared on CNBC this week that "the SaaSpocalypse is over," arguing that AI will benefit software firms rather than undercut them.

The market hasn't entirely agreed yet. The BVP Nasdaq Emerging Cloud Index sits at 1,488 — up around 29% from the April low of 1,154, but still 16% below its 52-week high and well off where SaaS multiples traded historically.

But specific names are starting to validate Bravo's thesis. Snowflake jumped more than 35% to its highest point since December after delivering one of the strongest quarters in the company's history:

- Product revenue: $1.334 billion

- Q1 growth: 34% YoY — accelerating from 30% last quarter and 26% a year ago

- Consensus beat: 5.3% — well above the typical 2–3% range

- Strongest sequential dollar growth ever

On the analyst call, CEO Sridhar Ramaswamy pointed to two products driving the acceleration: Snowflake's AI coding agent and its data-search tool, which pulls from Snowflake databases and from apps like Microsoft, Salesforce, and SAP.

Well, customers are buying AI that's already wired into their data and workflows. That's the part of the stack the model layer cannot commoditise.

Where this takes us

I'd bet on three things:

- AI spend observability becomes standard — adopted both from the FinOps layer (Ramp, Finout) and from the aggregators (Databricks, Snowflake, Salesforce, the hyperscalers).

- Open-weight and smaller models keep gaining share — alongside edge computing for sensitive or high-volume workloads.

- Model-routing tech becomes a core infrastructure layer — lowering near-term AI spend by offloading queries to cheaper models.

Ramp is the clearest beneficiary of the AI-spending fears narrative. Now worth $44 billion after raising $750 million, CEO Eric Glyman argued in a blog post that the world is moving away from the traditional two-pillar accounting setup (people and vendors) toward a three-pillar world — one where AI spending needs to be accounted for and managed as its own category.

Ramp's annualised revenue is now over $1 billion (Bloomberg puts the run-rate above $1.5 billion). It has reached positive free cash flow and serves over 70,000 customers — Visa, Uber, Shopify, Anduril, Figma — up from 50,000 last November.

Glyman's pitch: "When token spend shows up in the budget lines for every enterprise, the ability to track and optimise it only gets more valuable."

Stripe acquired Metronome, which specialises in usage-based billing. Fal AI and OpenRouter help companies buy tokens from lower-cost providers (DeepSeek is roughly 1/30th the cost of frontier US models).

The platforms positioning around this:

Databricks — Mosaic AI Gateway provides routing, observability, rate limiting, and cost tracking. System tables give query-level cost attribution by user, workspace, and job. Foundation Model APIs offer their own hosted open-source models (Llama, DBRX) at lower price points — but only inside the Databricks environment.

Snowflake — query-level cost visibility and built-in cost management interfaces, with Cortex AI delivering LLM functions natively in the warehouse.

Salesforce Agentforce — Digital Wallet for real-time consumption tracking and a Command Center (introduced in Agentforce 3) with health metrics and conversation replay. Customers must choose one consumption model per organisation — Flex Credits or per-conversation billing.

Microsoft Azure AI Foundry — cost observability built in, though developers have reported being charged for observability features that were enabled by default.

AWS Bedrock — native IAM/CloudWatch integration, model customisation, granular billing.

Google Vertex AI — integrated cost management within GCP.

One last thing. The most interesting financial data in tech right now sits in two private companies. We'd love to see Anthropic and OpenAI's revenue mix by model, how usage shifts between generations, what gross margin on inference actually looks like, and how much of the spend comes from a handful of power users.

Anthropic, OpenAI — please go public already. We could use the data transparency.