AI spending isn’t growing anymore. That’s the picture from new data out of Ramp, the corporate card and expense management platform, which shows that while AI usage continues to climb, actual spending has flatlined.
The timing is awkward. Anthropic is reportedly close to its IPO, and OpenAI just raised another $11 billion at a $1.25 trillion valuation. The narrative has been consistent: AI is growing fast, companies are spending more every quarter, and the infrastructure buildout justifies itself through explosive demand.
But Ramp’s data suggests something different is happening. Companies are using more AI, but they’re not paying proportionally more for it. The gap comes from two things: aggressive price cuts from providers trying to maintain market share, and smarter cost controls from the biggest spenders who’ve figured out how to optimize their usage.
If you’re building on AI APIs, this is mostly good news. Prices keep dropping. Claude, GPT-4, and competing models cost a fraction of what they did a year ago. You can do more with the same budget.
But if you’re building an AI company or trying to justify infrastructure spend, the math gets harder. The assumption has been that usage growth translates directly into revenue growth. If that breaks down, the unit economics of AI infrastructure start to look worse.
GitHub published a blog post this week positioning Copilot as more than autocomplete. The pitch: AI can help you plan work, review changes, and investigate production issues, not just write code faster.
It’s a reasonable expansion. Writing code did get faster. The work around it didn’t. If you’re spending half your day in pull request reviews, incident response, or trying to understand what broke in production, there’s room for tooling to help.
The question is whether GitHub’s implementation is actually useful or just more surface area for the product. Copilot works well for code completion because the task is narrow and the feedback loop is immediate. Planning and investigation are messier. It remains to be seen whether the same approach scales.
Underdog released a demo of a 32MB speech model that can generate both speech and music. Thirty-two megabytes. For context, that’s smaller than most product photos on an e-commerce site.
The demo apparently works, though details on quality and limitations are thin. Small models matter if you’re deploying on-device or at the edge, where bandwidth and storage are real constraints. A 32MB model you can run locally beats a cloud API if latency or privacy matter.
Worth watching if you’re building voice interfaces or need audio generation without the overhead of a full-scale model.
Anthropic released new research on how agent memory works. The paper breaks down the components that let AI agents remember context across interactions and use that memory to inform future actions.
This is foundational work, not a product release. If you’re building agents or trying to understand why they behave the way they do, it’s worth reading. If you’re just using AI tools, you can skip it.
The research does suggest that memory architecture is becoming a bigger focus as companies try to make agents more useful for long-running tasks. Right now, most AI interactions are stateless or rely on brittle workarounds. Better memory could change that.
The Ramp data is the story that matters most. AI usage is real. The tools work well enough that developers keep using them. But if spending growth has stalled while usage keeps climbing, someone’s margins are getting squeezed.
For now, that’s mostly a problem for the companies raising billions at trillion-dollar valuations. For developers, it means continued price competition and better deals. But if the economics don’t work long-term, the current pace of improvement could slow down.
One email at dawn. The five stories that mattered, with the bits removed and the meaning kept. Free, for now.