/Tech12h ago

Satya Nadella argues every company must build its own custom AI model to embed proprietary knowledge in model weights

Naveen Rao notes this reverses the centralized industry consensus.

1721.1K18796209.6K

#379

Original post

Karl Mehta@karlmehta

Satya Nadella reveals why every company may need its own AI model: the model becomes the new company database.

"To me, a model is like the database market."

"A firm should be able to take the tacit knowledge it has and embed it inside weights in a model that they control."

"When somebody asks me how many models should there be, I'll say as many models as firms in the world."

The contrarian part: the value may not sit in one universal frontier model. It sits in each company turning its private operating knowledge into a controlled model.

1:00 AM · Jun 23, 2026 · 202.4K Views

Sentiment

Positive users agree companies should build custom AI models around proprietary data as the real enterprise opportunity, while negative users call the idea too expensive, impractical due to lack of expertise, or just an Azure sales pitch.

Pos

60.7%

Neg

39.3%

15 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS10KBOOKMARKS18LIKES41RETWEETS3REPLIES3

Naveen Rao@NaveenGRao

Lol I said this 4 years ago. The prevailing wisdom was that there will be only one or two model providers

Karl Mehta@karlmehta

Satya Nadella reveals why every company may need its own AI model: the model becomes the new company database.

"To me, a model is like the database market."

"A firm should be able to take the tacit knowledge it has and embed it inside weights in a model that they control."

"When somebody asks me how many models should there be, I'll say as many models as firms in the world."

The contrarian part: the value may not sit in one universal frontier model. It sits in each company turning its private operating knowledge into a controlled model.

3h10K4118

Matt Furnari@matthewfurnari

No. Corporate data should not be stored in the opaque weights of a neural network. That's stupid.

Every company should have their own agent for sure. They should be able to switch between models at leisure and keep their data. Data should be stored in relational databases, ontologys, and documents.

10h1.2K122

dotSQUARE@dt_sqr

@karlmehta He has no idea what he is talking about. The LLM uses the database. The LLM is not a database.

9h4895

RG@rgvrmdya

@karlmehta @satyanadella describing Orquestra without realizing it -

4h20210

Joe B. Lewis@vettijoe

@karlmehta Surprised to see such naive takes coming from a deep thinker like Sathya. Models are incapable of forgetting information and adjusting to new information. And org knowledge evolves every day.

7h30911

Matt Finch@matt__finch

@karlmehta This is a really really dumb take, would be way too expensive and 99.99% firms don't have the expertise to fine-tune a model based on their data, data prep alone would be a nightmare

4h3884

Aieconomics_shailey@d_shailey

@karlmehta Satyas idea is a good one. But there has to be a large mothership model which is then finetuned with corporate domain knowledge. But to be useful the model needs not just a company's data, it needs to know about competition, related products, the supply chain, etc. it's complex.

11h84721

Ken Granville@Ken_Granville

Satya is asking the right question. As many models as firms in the world.

But fine-tuned weights are an opaque ledger. You can’t audit them. You can’t govern them. You can’t switch models and keep what you built.

Essence® separates the knowledge layer from the inference layer. Firms encode their operating intent as Meaning Coordinates, not model parameters. Meaning Coordinates (a periodic table for computing) are mapped to natural languages under the hood. Any compliant model can reason over them. Data and systems built this way are governed at the foundations. Intent-native Computing.

The model is not the database. The intent substrate is.

@Ken_Granville | MindAptiv

8h2301

Deus Machina@deus_machinea

@karlmehta No and Never. This is a costly affair just to push their own objective. Spec driven developement is already handles everything for a huge organisation, Why depend on training AI models when models are themselves changing every month?

5h4843

Tre Smith@TRESMITH322

this is exactly what microsoft needs to happen to be successful

but it's not going to happen.

he's trying so hard to save microsoft from their own mistakes

but it won't change the facts that this is the wrong direction technology needs to go

we need to move away from the microsoft suite and to an easier faster cheaper ecosystem

not just jam AI into the current ecosystem by forcing us to use microsoft to make our own models

that's just old school thinking and would be bad for everyone

2h1993

Christian Soschner@Soschner

@satyanadella @Jason I generally agree with your point. Companies really do need to think outside the box, and for very large corporations, training their own LLM often makes a lot of sense.

It really depends on the use case though. SMEs usually don't have enough proprietary data to train a model effectively on their own, so public offerings are usually the better option for them. Larger models can also pick up some advantages in reasoning by pulling in more external data beyond what's inside the company.

To me, it all comes back to basic software implementation. It's an ongoing iterative process, and the key is to start by figuring out the ROI for each specific use case.

That whole "tokenmaxxing" narrative mostly just helps the providers make more money and doesn't do much for the companies actually using the tools. That's why consulting firms end up as the big winners on the implementation side. They do well when they focus on projects that deliver real, lasting ROI for the clients who hired them.

12h189

Stephen@stephen_novesai

Corporate data should not, but the context is where the value lies. All that refined ontology, all that rich data sitting on databases spread across ERP, CRM, document stores... providing your own LLM (or SLM) with context and fusing when appropriate with the frontier models... that's the gold that your agent can gild its actions with.

5h271

Turner NextGen AI@TurnerNextGenAI

What exactly are we embedding?

If we’re embedding:

* procedures * documents * policies * code * emails * manuals

then we’re still largely embedding knowledge artifacts.

What organizational architecture produced that knowledge?

That’s different.

Environment

↓

Load

↓

Organization

↓

Load Negotiation

↓

Resource Allocation

↓

Waste

↓

Recovery

↓

Adaptation

↓

Function

↓

Independence

A company’s true value isn’t just what’s written down.

It’s how it organizes itself to solve problems.

⸻

Think about two aerospace companies.

They may have:

* identical CAD files * identical manufacturing procedures * identical AI models

Yet one consistently outperforms the other.

Why?

Traditional AI might say:

Better data.

We would ask:

How does each organization negotiate load?

* Where are resources trapped? * Where is waste accumulating? * How do transitions occur? * Where does organizational integrity drift?

Those aren’t documents.

They’re organizational dynamics.

7h711

ProudIndependent@CEOTech2020@Bhasker36731383

@karlmehta This coming from a company that had a monopoly with Microsoft Windows Operating System lmao 🤣

10h4271

The Singularity Project@01Singularity01

@karlmehta The light 💡is finally starting to come on, at least for Satya. But he still does not get that it is the SYSTEM which is the Intelligence, not the MODEL. Although he alludes to that with his comments about model orchestration.

1h101

jagatjiva@jagatinfinity

@karlmehta "33 koti Devata" #AI #servers

9h397

Peter Rex@PeterRex

@karlmehta He's right. The general models are the commodity. What you build narrow on your own data is the part that matters

6h4982

Dr Raju@drraju2

@karlmehta Why may every company need its own AI model?

Because AI is becoming the new company database. Instead of searching files, employees can simply ask questions and get answers instantly. Example: Amazon's AI knows its inventory. Tesla's AI knows its vehicles. Apollo's AI knows

7h41

Ant A. 🇺🇸@AntDX316

@karlmehta @Grok I would prefer if @OpenAI is the main firm.

5h27

Dan Harms@harms_daniel

@karlmehta Is this supposed to be a novel, transformative idea? How does this play out any other way?

6h7871