The Data You Feed Back
2026-07-14 · 5 min read · Janaina Maia
This week, three stories from different corners of the AI industry said the same thing.
Satya Nadella published a blog post warning enterprises that when they use proprietary AI models, they pay twice. Once with money, and once with something more valuable: the proprietary knowledge they must reveal to make the model useful. Every prompt, every correction, every agent tool call — what he calls "exhaust" — flows back to the model maker. "You essentially pay for intelligence twice," he writes. "Once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful."
Apple filed a 41-page complaint against OpenAI alleging that OpenAI's hardware division — built from the $6.5 billion acquisition of Jony Ive's startup io — was constructed using Apple's stolen trade secrets. The allegations include job candidates being asked to bring Apple hardware prototypes to OpenAI interviews for "show and tell," a former Apple engineer texting "LOL, I found out I can access the [network storage]" about Apple's internal systems, and OpenAI coaching departing employees on how to evade Apple's security procedures. Over 400 former Apple employees now work at OpenAI.
And Playcode published a meticulous analysis showing that Anthropic's new tokenizer produces roughly 30% more tokens than its previous one for the same content — at the same sticker price. For TypeScript, the language AI coding agents write most, Claude's new tokenizer produces 1.73x the tokens of GPT for the identical file. When Sonnet 5's intro pricing ends on September 1, the same code will cost roughly 32% more on Sonnet 5 than on Sonnet 4.6, at an identical list price. The price increase has no line item. It's hiding in the tokenizer.
Three stories. One thread: the most valuable thing in AI isn't the model. It's what you put into it.
Nadella's warning is self-interested and correct
Let's start with the obvious: Nadella is the CEO of Microsoft, which invested $13 billion in OpenAI and runs Azure, the cloud platform that would host the "proprietary learning environments" he's advocating. His argument is sound regardless of his commercial interest. But let's not pretend the messenger has no stake.
The argument itself is simple and worth taking seriously. When you use a proprietary AI model, the prompts you write, the tools your agents call, and especially the corrections you make when the model is wrong — all of that is exhaust. It's information about how your business works, what your customers need, where your edge cases live. Model providers can learn from this exhaust. Some of them explicitly reserve the right to train on customer data. Nadella's point is that the enterprise is teaching the model its competitive advantage, and then paying for the privilege.
The solution he advocates — orchestration layers, model switching, open-source models on-prem — is also the solution that benefits his cloud business. But that doesn't make it wrong. Solo.io, Vercel, and OpenRouter are all seeing surges in traffic to open-source models. Open models accounted for 29% of traffic routed through Vercel's gateway last month. Enterprises are voting with their compute.
Apple's lawsuit is the exhaust problem made literal
The Apple complaint takes Nadella's abstract warning and gives it a face, a name, and a text message. Tang Tan, OpenAI's chief hardware officer, allegedly directed Apple employees interviewing at OpenAI to bring "actual parts" from Apple for show-and-tell sessions. Chang Liu allegedly exploited an authentication bug in Apple's systems to access confidential files, texting a colleague: "LOL, I found out I can access the [network storage], so funny." OpenAI allegedly circulated an internal Apple "need to know" document to coach new hires on how to avoid Apple's "dreaded walkout" security procedure.
Whether or not these specific allegations hold up in court, the pattern is familiar to anyone who's watched Silicon Valley for more than a decade. The difference is the scale and the stakes. OpenAI isn't just poaching talent from a competitor. It's building a hardware device — reportedly a phone — that would compete directly with Apple's core business. And it's allegedly doing so using Apple's own people, Apple's own processes, and Apple's own prototypes. If the model maker becomes your competitor and it learned from your exhaust, you haven't just paid for AI twice. You've paid for your own disruption.
The tokenizer swap is exhaust you can't even see
Playcode's analysis is the quietest of the three stories, but for anyone building AI products, it might be the most immediately relevant. When Anthropic shipped its new tokenizer in Claude Sonnet 5, Opus 4.8, and Fable 5, the list price didn't change. But the same content now produces roughly 30% more tokens. For code — which is what AI agents spend most of their time processing — the inflation is even higher. TypeScript is the worst case at 1.73x.
Think of it this way: you're paying a deli per slice. The deli changes its slicer to cut thinner slices but keeps the price per slice the same. Your sandwich now costs 30% more, but the menu hasn't changed. That's not a price increase you can see on the pricing page. It's a price increase you discover on your invoice.
The intro pricing for Sonnet 5 ($2/$10) roughly cancels the inflation for now. But on September 1, the sticker snaps back to $3/$15 — and the 30% token inflation stays. From that day, the same code costs roughly 32% more on Sonnet 5 than it did on Sonnet 4.6, at an identical list price. The discount was a transition cushion, not a price cut.
This is the tokenizer version of exhaust: information about your costs that's visible to the provider but invisible to you.
Why this matters for product design
Three stories, three domains — enterprise strategy, trade secrets, pricing infrastructure — and one shared lesson: in AI, the data flows both ways, and only one direction is transparent.
- The prompt is a leak. When your team writes detailed prompts about your business processes, your customer workflows, your edge cases, you're teaching the model about your domain. That knowledge has value. The question is whether you're choosing to share it, or whether it's leaking out as a side effect of using the product. Design your AI product so that users can opt out of training, audit what's been learned, and export their data. The products that win enterprise trust will be the ones where data sovereignty is a visible, configurable surface — not a footnote in the terms of service.
- The invoice is a design surface. Playcode's analysis shows that pricing transparency is a product feature. When a tokenizer changes and the cost of the same work goes up 30%, the customer deserves to know. The products that win builder trust will be the ones where cost is predictable and changes are announced, not hidden inside a technical component swap.
- The talent is the data. Apple's lawsuit is a reminder that when your best people leave for a competitor building the same product, the exhaust isn't just in the model. It's in their heads. The less your competitive advantage depends on any individual, the less vulnerable you are to this kind of extraction. Document your processes. Distribute your knowledge. Make the system the moat, not the person.
My take
The AI industry talks a lot about what models can do. It talks much less about what models take. Nadella's blog post is the first time a major platform CEO has said out loud what enterprise customers have been whispering: the model is not the product. The model is a service that learns from its customers, and the terms of that learning matter.
Apple's lawsuit is what happens when the learning is allegedly deliberate and industrial. Playcode's analysis is what happens when it's invisible and commercial. Both are versions of the same dynamic: information flows toward the model maker, and the customer pays for the privilege.
The products that last will be the ones designed around this asymmetry. Not the ones that pretend it doesn't exist. Not the ones that hide it in a tokenizer change or a "need to know" document. The ones that make the data flow visible, give the customer control over what they share, and charge for the value they provide — not for the value they extract.
The data you feed back is the most expensive thing in AI. Start treating it that way.