Your Data Is Your AI Advantage, Keep it

Satya Nadella's latest essay is getting a lot of buzz. For the middle market, it oversells the problem to pitch the fix.

Nadella's piece flips Kenneth Arrow's information paradox. Arrow said a seller can't sell knowledge without giving it away. His theory is that AI reverses it: now the buyer gives knowledge away just by using what they bought. The buyer pays twice, once in money and again in the proprietary know-how you feed the model to make it useful. Learning flows one way; what is learned is not unlearned, so value converges to whoever owns the learning infrastructure. His recommendation, framed to Microsoft's benefit, is a trust boundary where every firm's human capital and token capital compound behind a wall nothing crosses without consent. He says you should own your learning loop, and he'd like to sell that to you.

This was aimed at Microsoft's enterprise clients. A sales pitch shaped as thought leadership. That's okay and we can still get great takeaways from it for middle market and lower middle market companies. He defined the problem in the shape that sells his solution. But the middle market is not one market, and it will not have one answer. This really breaks down to two questions, who do you use as your AI provider and do you need to be custom training.

You have 3 general choice categories for AI providers:

• Big labs. These offer end to end services, with Anthropic, Google, and OpenAI as the premier examples. If you are working with these you are relying on your contract with them to manage the access to your data. Do you trust and do you believe the terms will stay the same over the lifecycle of your build. Example: Anthropic recently carved Fable out of zero-data-retention (ZDR). Prompts and outputs are now retained 30 days, and prior ZDR agreements don't cover it.
• Inference providers. These are companies that just provide the inference compute for your AI use. They don't build their own models. They typically run open weight models, and they are not in the business of gathering your data. Probabilistic compute rented by the token. Examples are Together AI, Fireworks AI, Baseten, and Replicate.
• Self hosted. You spin up your own inference system and gain access to probabilistic compute entirely in an environment under your control. There are various levels to this. You can rent bare metal, put in your own server rack, or stack Mac Studios in a closet. Here you take on the uptime and management overhead.

Which you choose really depends on how you balance the risk vs how much effort you are willing to take on. Moving away from the bundled tools like Claude Cowork, Antigravity and Codex can be a heavy lift. They create a great end to end experience that replicating fully is challenging.

Now training
Do you need to train your own LLM, for almost everyone the answer is no.
The reason to train is to improve the accuracy and results of your LLM usage. You can gain a lot of benefit in most more accessible and lower cost efforts.

Organize your data. Giving your llm easy, safe, access to your company's data is the best way to improve performance of your AI operations. Engage in RAG or via API or MCP you can give your AI the tools it needs to cleanly rationalize the answers you need.

For workflow operations build your own harness. This is a high value and low lift one time task that can really improve your output. This is also where your evals live. Define good from bad for your use case, and the harness measures every run against it. Best of all it comes with no token cost.

The last and heaviest option is building your own adapter weights. You can do this with tools like Tinker from Thinking Machines. This helps AI respond in the way that best fits your use case but it is unnecessary and even counter productive for most use cases.

So what is the right answer for a middle market company. The honest answer is it depends, and that is the point, not a dodge.

A compliance SaaS company and an HVAC rollup can read the same article and share almost none of the same problem. The compliance company lives or dies on how it handles data. For them the provider question is the whole game. The rollup just needs AI to run the business better, and the retention fine print barely moves the needle.

Nadella raised real points. His solution is not for everyone. A company is successful because it is unique. The right tools flow from that.

Let me know if you want to talk about what is right for you, your business, or your portfolio.