Here’s what’s on our plate:

• 🧪 Why productivity, not token usage, defines AI's true value.

• 📰 First AI-built malware, Uber's superapp sprint, Windows 11 gets faster.

• 💬 Audit your AI spend against actual productivity gains.

• 🗳️ Poll: What's the right response to runaway token bills?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Flat-rate AI is dead: Anthropic and OpenAI now bill by usage because agentic workflows consume 5–30x more tokens than chatbot Q&A, turning AI into a volatile utility bill.

• Spending outpaces strategy: Business AI token spend has grown 13x since January 2025, the typical enterprise contract is projected to hit $1M in 2026, and Uber burned its annual AI budget in weeks.

• Consumption does not equal productivity: Over 80% of companies report no measurable AI productivity gains, and tracking token volume as a proxy for output risks rewarding waste over value.

• The real test is measurement: Organizations that connect token spend to outcomes will treat AI like cloud infrastructure. Those who can’t may face a sharp correction when the bill comes due.

Why productivity, not token usage, defines AI’s true value

As conversations around artificial intelligence continue to mature, enterprises looking to integrate the technology into their workflows are working to understand what it costs to use AI tools and how to get the most out of a technology that is still evolving. In this context, tokens have emerged as a key point of discussion, both within boardrooms and within technical decision-making processes.

The importance of these discussions comes down to a simple mechanic: every interaction with an AI model runs on tokens. When a user types a prompt, the model breaks that text into smaller units and processes them sequentially. When the model responds, it generates another stream of tokens that form the output. Providers charge for both sides of this exchange, input and output, at rates that vary by model.

For a time, tokens remained invisible to most buyers. Companies purchased AI like any other software product: through seat licenses, subscriptions, and predictable monthly budgets. That arrangement held up when employees asked chatbots a handful of questions a day; however, it began to break down when AI agents entered the picture. Unlike chatbots, agents do not wait for users to type every instruction and can carry out tasks independently without constant human approval. Because every step uses tokens, usage can rise quickly in ways that flat-rate subscriptions were never designed to handle.

The financial consequences of this shift have already started showing up in the cost of running AI.

Gartner projects roughly $1T in global spending on AI services and software in 2026, a figure cited in Ramp’s research on what the firm calls the AI spending blind spot. Ramp’s own data puts it more bluntly: average monthly business AI token spend has grown 13x since January 2025, outpacing every other spending category the company tracks. And yet, the cost of a single token has never been lower, with industry-wide price drops of roughly 80% over the same period.

That is the central paradox: token prices keep falling while total AI bills keep rising, and the gap between those two facts is forcing a structural rethink of how enterprises price, purchase, and manage AI.

The end of unlimited AI pricing

Anthropic initially sold Claude to enterprises through familiar seat-based subscriptions: a flat monthly fee per employee that included bundled token usage. It was predictable, easy to budget for, and familiar to procurement teams used to SaaS licensing.

The model is now beginning to shift, and in November 2025, Anthropic started moving enterprise customers onto a usage-based plan when their contracts came up for renewal. By February 2026, all new enterprise agreements would default to a $20 per-seat base fee covering platform access only, with each token billed at standard API rates. The old Premium and Standard seat types were formally retired for new customers.

The new system also includes minimum monthly spending commitments based on expected usage. For finance teams, that turns AI from a predictable software expense into something closer to a utility bill.

What’s worth noting is that Anthropic is not alone in making this shift. OpenAI has also moved coding products toward token-based billing, reflecting a broader industry shift away from unlimited plans.

In both cases, the underlying reason is mechanical. Agentic AI workflows can consume 5–30x more tokens per task than a standard chatbot conversation, according to Gartner estimates. In this scenario, when a company’s developers are running always-on coding agents that iterate through problems for hours at a stretch, the token volumes they generate bear no resemblance to the casual Q&A sessions that flat-rate plans were designed around. Once every token has a price, how much an organization spends on AI becomes a question engineering can no longer answer on its own. It becomes a finance question, and most finance teams have not prepared for it.

The bill nobody budgeted for

The numbers confirm what the pricing shifts suggest. Ramp says the median company on its platform now spends nearly 15% of its software budget on AI tools, while the average enterprise AI contract is projected to reach $1M in 2026, up sharply from two years ago.

At the same time, adoption is accelerating among businesses that are spending on AI. Overall adoption reportedly crossed 50% in March 2026, with nearly one in four now paying for Anthropic’s Claude, up from one in 25 a year prior.

What averages obscure is how unevenly the costs land. Uber’s CTO said the company exhausted its annual AI budget in weeks, while Anthropic deployment data suggests active Claude Code users can cost around $13 per developer per day.

The widening gap between what finance teams expected to spend and what invoices now reveal has created an entirely new software category focused on controlling AI expenses. Ramp launched its AI Spend Intelligence product in April 2026, allowing companies to pull token-level usage data directly from providers such as OpenAI and Anthropic, then break spending down by model, team, and project. Finout, CloudZero, and others are building similar tools.

That shift has also entered the consulting world. In a January 2026 report, Deloitte urged business leaders to manage AI economics with the same discipline once reserved for energy bills or capital budgets, describing tokens as “the new currency” of enterprise technology.

Two years ago, almost none of these products existed, because the spending category they are designed to track barely existed either.

The sheer scale of the shift further exacerbates the problem. According to Ramp, business AI spending grew 4x between February 2025 and February 2026, yet much of it remains poorly managed.

Beyond scale, there is also the issue of visibility. One Ramp customer discovered $120k in annual AI spend that had never appeared on any provider dashboard because the charges were scattered across individual employee credit cards, invisible to anyone trying to build a company-wide picture of what AI actually costs.

Falling prices, rising doubts

Yet rising budgets are colliding with uncertain returns. A February study cited by Tom’s Hardware found more than 80% of companies using AI saw no measurable productivity gains, while Gartner says over 40% of AI agent projects could fail by 2027.

McKinsey offers one explanation: companies reporting strong AI returns were twice as likely to redesign workflows before choosing models. Process change, not model choice alone, appears to drive value.

Internal incentives may be worsening the problem. Some companies now track token usage as a proxy for productivity, and firms such as Meta and Shopify have reportedly built internal leaderboards around consumption. Databricks CEO Ali Ghodsi has criticized that logic, arguing that optimizing for token use instead of outcomes can reward waste rather than value. Spending more on AI does not necessarily mean accomplishing more with it.

Anthropic CEO Dario Amodei has described the broader risk as a “cone of uncertainty” around compute demand. Data centers take years to build, yet companies are committing billions today for capacity they assume future demand will justify. If current spending proves inflated by weak metrics rather than real productivity gains, the eventual correction could be sharp.

Just a couple of years ago, most executives had never heard the word ‘token’ outside banking, security, or linguistics. Today, tokens have become one of the most important units in enterprise technology spending, shaping budgets, procurement decisions, and debates over productivity.

That shift says something larger about the AI era: the economics are evolving faster than most organizations can adapt to them. Companies entered the market expecting software subscriptions and predictable seat licenses, only to discover a utility-style model in which every prompt, workflow, and autonomous action incurs a cost. Prices per token may keep falling, but lower unit costs have not prevented total spending from rising sharply. The next phase of enterprise AI adoption will not be defined only by better models or more capable agents, but by whether businesses can measure value with the same precision that providers measure usage.

If they cannot connect tokens to outcomes, many current budgets may prove temporary. If they can, tokens may become as familiar to finance teams as cloud spend is today.

Bite-Sized Brains

• Hackers' first AI-built malware: Security researchers say hackers used AI to develop the first known piece of fully AI-generated malware, marking a turning point in the cyber threat landscape.

• Uber's superapp sprint: Uber has always wanted to be more than rides, and now AI competition is finally giving it a reason to move fast on becoming a full-service platform.

• Windows 11 gets faster: Microsoft is rolling out a macOS-style speed boost, signaling a serious push to close the performance gap users have complained about for years.

Prompt Of The Day

Tuesday Poll

The Toolkit

• Deepgram: Speech-to-text API built for scale, handling real-time transcription and voice intelligence for production apps.

• Descript: AI-powered audio and video editor that lets you edit recordings by editing the transcript like a doc.

• Drift.ai: AI-powered conversational marketing platform that turns website visitors into qualified pipeline through automated chat. 

Rate This Edition