Paying Twice for AI: What Nadella's Warning Means for Your LLM Costs
Satya Nadella argues that companies buying proprietary AI pay twice, once in cash for tokens and again in the proprietary knowledge their usage teaches the model, and that second payment should change how you count AI costs.
Key takeaways
- In a blog post this week, Microsoft's CEO wrote that AI buyers "essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful."
- Models learn from usage "exhaust": the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong.
- His prescription: retain ownership of your data and add an orchestration layer so you can switch models freely.
- The shift is already visible. Open models were 29% of traffic through Vercel's AI gateway last month, and enterprises report on-prem open models doing about 90% of the job at far lower cost.
- The second payment never appears on an invoice, but you can price the insurance against it: portability.
What did Nadella actually say?
The fear he is joining has circulated in Silicon Valley for months: that the big labs selling proprietary models act like Trojan horses, absorbing sensitive business context from the companies using them. In a blog post published Sunday, Nadella made it concrete: "You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!"
Models, he argues, learn from exhaust: the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong. Every correction distills institutional know-how, "the kind of knowledge a competitor could never buy." And enterprises are handing it over as a byproduct of usage.
He also points at an asymmetry. Labs claim fair use to train on the world's public data, then impose restrictive terms on customers who want to distill their models in return. Earlier this year, Anthropic accused Chinese open source labs of mining Claude through millions of prompts. Wherever you land on that dispute, it shows where the leverage currently sits.
One caveat worth keeping: Nadella runs a giant cloud provider, and his solution, building proprietary learning environments on the cloud, rhymes conveniently with Azure. The incentive does not make the argument wrong, but it tells you how to read it.
Why is this a cost story and not just a privacy story?
The first payment, tokens, is visible and negotiable. The second payment compounds quietly: the more a vendor's model internalizes how your business works, the higher your switching cost, and the weaker your position at every renewal. Lock-in is not a legal abstraction. It is a price you pay later, at terms you did not set.
That is why the market response is infrastructure, not outrage. Orchestration layers and AI gateways that swap providers per request have become mainstream, with open models at 29% of traffic through Vercel's gateway last month. And the on-prem argument keeps landing on the same shape: Solo.io's CEO says her enterprise customers keep concluding that an open model run on their own hardware does about 90% of what the big proprietary model does, at far lower cost, under their control.
How do you keep the second payment small?
- Keep your workload portable. A gateway that can route to a second provider turns "we should renegotiate" into an actual option instead of a wish.
- Keep your data yours. Prompts, corrections and workflow context are the training set of your business. Treat contract clauses that let a vendor learn from them as a price term, not a legal footnote.
- Keep a live number on the alternative. Whether 90% quality at lower cost clears your bar is an empirical, per-task question, not a slogan. Benchmark it.
Here is the operator translation: your walk-away price is only real if it is quantified. The teams that negotiate well with model vendors are the ones holding a current, per-task cost comparison of the alternatives. That comparison is the cheapest insurance you can buy against paying twice, and you can keep it live across providers in Calcaas.
Whoever owns the exhaust owns the upside. Make sure some of it is you.
Frequently asked questions
What does paying twice for AI mean?
Nadella's argument is that companies pay cash for token usage and simultaneously hand over proprietary knowledge through prompts, tool use and corrections. That second payment teaches the vendor's model how their business works.
What is an AI orchestration layer or gateway?
It is a routing layer between your product and model providers that lets you switch or mix models without rewriting your application. Nadella recommends it as the way to avoid lock-in, and open models already account for 29% of traffic routed through Vercel's gateway.
Are open models good enough to replace frontier APIs?
For many workloads, enterprises report that on-prem open models handle roughly 90% of the job at much lower cost, per the Solo.io CEO quoted by TechCrunch. The right answer is task-specific: benchmark quality per task and compare per-task cost before switching.
How should founders account for AI lock-in?
Treat switching capability as a priced asset. Maintain a current per-task cost comparison of alternative providers and models, because optionality strengthens every renewal negotiation and the comparison itself is cheap to keep updated. Note: place this JSON-LD in a <script type="application/ld+json"> tag in the page head.
More from the blog
LLM EconomicsRenting vs Owning AI: When Do Open Models Beat Frontier APIs on Cost?
Companies typically start on frontier APIs and shift work to open models as usage scales, so rent versus own is a break-even calculation driven by volume and task mix, not an ideological choice.
LLM EconomicsToken Pricing Economics: Will LLM Providers Keep Their Pricing Power?
Benedict Evans argues that today's token prices reflect a temporary supply crunch and that every visible market dynamic points toward frontier models becoming commodity infrastructure, which means founders should plan for falling LLM costs rather than assume today's rate cards.
LLM EconomicsTokenizer Inflation: Why $/1M Token Prices Are Not Comparable Across LLMs
The same file can become up to 73% more tokens on one frontier model than another, so a $/1M token price is only comparable after you adjust for each model's tokenizer.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.