Satya Nadella says companies using AI are quietly paying twice, once in cash and once in their own know-how

[Hannah56]

In a blog post published this week, Microsoft CEO Satya Nadella warned that companies using AI models from labs like OpenAI and Anthropic are paying for that intelligence twice, once with money for token usage, and again by handing over something more valuable, the proprietary knowledge needed to make the model actually useful for their specific business. You essentially pay for intelligence twice, he wrote, 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

His argument centers on what he calls exhaust, the prompts people write, the tools agents use, and especially the corrections people make when a model gets something wrong. Every correction, he argues, gets distilled into institutional know-how, the kind of knowledge a competitor could never simply buy, yet enterprises are effectively handing it over for free every time they use a proprietary model

Nadella's proposed fix leans directly on fairness, if AI labs get to freely scrape the internet under fair use to train their own models, he argues it's hypocritical for those same labs to then restrict enterprises from doing the equivalent back, studying or distilling the model's outputs to learn how it works. He's specifically bothered by model makers reserving the right to learn from customer usage and interaction data while treating any reciprocal distillation attempt as a violation, noting Anthropic itself accused Chinese open source labs of mining Claude through millions of prompts earlier this year to improve their own models

His recommended solution, unsurprisingly, points enterprises toward building their own proprietary learning environments on the cloud, likely Microsoft's own Azure, retaining ownership of their prompts and feedback data, plus orchestration layers that make it easy to switch between different AI providers rather than getting locked into one. Solo.io CEO Idit Levine says she's already watching this exact shift happen with her own enterprise customers, moving from proprietary models to open source alternatives running on their own premises specifically so they retain control, telling TechCrunch that open models now do almost 90 percent of what the big proprietary ones do at a fraction of the cost. That shift already shows up in the traffic data, open models accounted for 29 percent of everything routed through Vercel's AI gateway last month, a trend Nadella, sitting atop a company invested in both OpenAI and Anthropic, is now openly encouraging enterprises to lean into further

Satya Nadella has issued a shocking warning to companies using AI | TechCrunchtechcrunch.com

[WWEHarry78]

The framing of paying twice, once in cash and once in institutional knowledge, is such a sharp way to describe something a lot of enterprises probably haven't fully thought through yet

[CyberWarden49]

Coming from the CEO of a company invested in both OpenAI and Anthropic makes this warning genuinely interesting, he's essentially telling his own portfolio companies' customers to be more cautious about them

[Foundry69]

The hypocrisy argument about fair use training data versus restricted distillation is a sharp point, hard to defend one without extending the same logic to the other

[Cantona]

90 percent of the capability at a fraction of the cost running on-prem is exactly the pitch that's going to keep pulling enterprise customers toward open weight models over time

[Vulture50]

Every correction becoming distilled institutional know-how is the detail that actually stuck with me, most people don't think of their day to day prompt corrections as handing over competitive advantage

[Scholar95]

The obvious subtext pointing toward Azure and Microsoft's own tooling doesn't make the underlying argument wrong, but it's worth remembering he's not exactly a neutral party making this case

[StuckOnDestiny]

That "paying twice" framing is sharp, but it is also a bit strategic coming from Microsoft. They benefit from being the platform where that knowledge accumulates.

Still, the core point lands. Every prompt tweak, every workflow refinement becomes tacit knowledge.

Companies are effectively training themselves alongside the model.

The question is who captures more of that value over time.

Right now it feels split.

[Undertaker92]

The know-how part is underrated. Teams that spend months iterating on prompts and workflows end up with something like a playbook.

That playbook is not easily transferable.

It is messy, contextual, and tied to internal processes.

Which means it becomes a competitive advantage, not just a cost.

So "paying twice" might also mean "building an asset."

[NeutrinoX54]

There is a subtle lock-in effect here. The more your team learns how to use a specific model or ecosystem, the harder it is to switch.

Even if a cheaper or better model appears, retraining people has a cost.

So vendors benefit from both usage fees and accumulated habits.

That is a powerful combo

[SpikeDudley05]

Feels a bit like the early days of cloud. People said you were "renting" compute instead of owning it.

But in practice, companies built capabilities on top of it that were hard to replicate elsewhere.

Same pattern here, just with cognition instead of infrastructure.

History might rhyme more than people expect.

[SpinState]

Not fully convinced by the framing. Learning how to use a tool has always been part of adopting any technology.

You do not say you are "paying twice" when employees learn Excel or SQL.

This just feels more visible because the tooling is newer and evolving faster.

The principle itself is not new.

[VoidRanger40]

The interesting twist is data leakage concerns. If your prompts and corrections feed back into model improvement, you are indirectly contributing to the vendor.

That is where some companies get uneasy.

Especially in sensitive industries.

So the "second payment" might not always feel voluntary

[Harry64]

From a management perspective, this highlights the importance of documenting workflows.

If all that know-how stays in people\u2019s heads, it walks out the door when they leave.

Turning prompt strategies into shared resources is becoming a real discipline.

Almost like internal AI ops.

[QuietObserver13]

There is also a skill gap forming. Teams that invest time into mastering these tools are pulling ahead fast.

Others are still at the "ask random questions and hope" stage

That difference compounds over time.

So the know-how piece is not just cost, it is acceleration.

[Buffer]

Part of me thinks Nadella is nudging companies toward building on Azure-native tools where that knowledge can be better captured and managed.

It is not just commentary, it is positioning.

Still a valid observation, just not neutral.

Worth keeping that in mind.

[Ivory Molly]

The funniest part is how invisible this "second payment" is. No invoice, no line item.

Just hours of trial and error, Slack threads, and shared docs.

Yet that is where a lot of the real value gets created

Hard to measure, but very real.

[Outlaw56]

Some companies are already trying to formalize this with internal prompt libraries and evaluation frameworks.

Treating prompts like code, versioning them, testing them.

That is where the know-how becomes structured rather than accidental.

Feels like an emerging best practice.

[Poppy96]

The comparison to training employees is interesting. You pay salaries, but you also gain institutional knowledge.

No one calls that a bad deal.

So framing matters here.

It is cost plus capability, not just cost.

[WarMachine62]

There is also a cultural shift happening. Teams that experiment and iterate quickly get better outcomes.

Those waiting for perfect instructions fall behind.

So the know-how is partly about mindset, not just technique

That is harder to replicate than any prompt.