<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Context Design Co — Notes</title>
    <link>https://contextdesign.tech/blog/</link>
    <description>Operator notes on substrate, operational IP, and AI as balance-sheet asset.</description>
    <language>en-ZA</language>
    <lastBuildDate>Sun, 14 Jun 2026 19:35:47 GMT</lastBuildDate>
    <atom:link href="https://contextdesign.tech/feed.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>The Learning Loop Has to Be Yours</title>
      <link>https://contextdesign.tech/blog/learning-loop-has-to-be-yours</link>
      <guid isPermaLink="true">https://contextdesign.tech/blog/learning-loop-has-to-be-yours</guid>
      <pubDate>Sun, 14 Jun 2026 00:00:00 GMT</pubDate>
      <description>Satya Nadella says every firm must own a compounding loop of human and token capital. He's right — and he named the test. The unsaid part is where the loop lives, and who owns it.</description>
      <content:encoded><![CDATA[
This week Microsoft's CEO published the argument we have been making to mid-market operators for the better part of a year — almost in our own vocabulary. [Satya Nadella's case](https://www.businesstoday.in/tech-today/news/story/satya-nadella-ai-economy-microsoft-chief-says-ecosystems-matter-more-than-models-536791-2026-06-14) is that every company will now have to build two balance sheets at once: **human capital** — the judgement, relationships, and pattern recognition of its people — and what he calls **token capital**, the AI capability the firm builds and owns. The two compound together, or neither does.

When a hyperscaler describes your business back to you, read it twice. The first read is validation. The second read is for the part that is missing.

## §1 — What he got right

Nadella's core claim is that the advantage does not come from picking the best model. It comes from building *a learning loop on top of models where human capital and token capital compound*. Offload a task, even a whole job, and you keep moving. Offload your learning and there is nothing left to compound. "You can never offload your learning," he writes — and the loop that captures it "becomes the new IP of the firm." It compounds. He calls it a hill-climbing machine.

We have a plainer word for the same object: the **substrate**. A substrate is the governed, business-specific knowledge layer that sits between an organisation and a generic model — its ontology, its decision history, its runbooks, its audit log, its retrieval graph. We have argued from the first day that [this layer is the asset](/notes/operational-ip-now-capital), not the model, and that [it accrues](/blog/introducing-the-substrate): every new model generation makes the same substrate more useful, not less. Nadella has now written the enterprise edition of that memo. The mechanism is identical.

## §2 — The test he named

The most useful sentence in the piece is the one most readers will skim. A company, Nadella writes, should be able to "switch out a 'generalist' model without losing the 'company veteran' expertise built into their learning system" — and he calls this "the key test of your control and sovereignty in the era ahead."

We could not have stated the design constraint more precisely. A substrate is built so that the model stays generic and the substrate carries the expertise. It speaks to whatever model you point at it over [MCP](https://modelcontextprotocol.io), the open protocol — one vendor today, a better one next quarter — and the company-veteran knowledge stays exactly where it lives: in your layer, on your infrastructure, unmoved by the swap. Passing Nadella's test is not a feature we bolted on. It is the thing a substrate is *for*.

## §3 — The part the platform will not say

Here is the second read — the missing part.

Nadella names the test (swap the model, keep the veteran) but not the variable that decides it: **where the loop lives, and who holds the licence.** The omission is not careless. A hyperscaler's commercial gravity pulls in exactly one direction — for your compounding loop to compound on its platform, in its data estate, behind its billing. And a learning loop you do not host, cannot audit line by line, and cannot carry out of the building is not sovereignty. It is a more sophisticated rental, and it fails his own test the moment you try to leave.

> The loop only counts as yours if you can switch the model out from under it — and walk the whole thing across the road if you choose to.

This is the line we will not blur. A CDC substrate is [self-hosted, MIT-licensed, and owned by you on day one](/operational-ip). We hand over the keys, the code, and the audit trail; we do not hold your context hostage. The private evaluations Nadella wants — scored against the outcomes that matter to your business rather than an external benchmark — run on traces from inside your operation and stay inside it. That is what makes the loop a capitalisable asset on your books instead of a line on someone else's revenue.

## §4 — A frontier ecosystem, mid-market edition

Nadella's larger worry is worth taking seriously, because it is ours too. If a few models capture all of the returns, he warns, the political economy "will simply not tolerate it" — and he reaches for the right analogy: the hollowing-out of industrial economies in the first wave of globalisation, where the headline numbers held while the displacement was real. His answer is "a frontier ecosystem, not just a frontier model," so that value flows broadly across every company, industry, and country.

Agreed. But notice who he is writing for. An enterprise with its own AI research function can build its own hill-climbing machine — Nadella is telling it to, and it will. The R 50-million-to-R 500-million operator cannot staff that. The manufacturer, the distributor, the professional firm whose institutional knowledge lives in one person's head and a folder of spreadsheets is precisely the business most likely to find its expertise quietly commoditised out from underneath it. That operator is who we build for. A substrate is how a company *without* an AI lab still ends up owning its loop: as a fixed-scope [Proof](/operational-ip), then a production foundation, then ongoing stewardship as the model layer keeps moving beneath it. The same hill-climbing machine — delivered to the businesses Nadella's framing quietly assumes already have one.

## §5 — The clause we would add

So we will take the validation, gladly. The CEO of Microsoft and a substrate shop in the Western Cape now agree on the shape of the next decade: human and token capital, compounding in a loop, owned by the firm and not by the model.

We would add one clause to his best line. You can offload a task, even a job, but you can never offload your learning — *and you cannot offload it onto someone else's balance sheet either.* The loop has to be yours: hosted by you, licensed to you, switchable underneath. That is not a philosophical preference. It is the difference between an asset you can capitalise and a subscription you can cancel.

Build the loop. Then make sure your name is on the title deed.

For the artefact list and the engagement shape, see the [Operational IP page](/operational-ip). Or [book a discovery call](/#contact).
]]></content:encoded>
    </item>
    <item>
      <title>Introducing the Substrate</title>
      <link>https://contextdesign.tech/blog/introducing-the-substrate</link>
      <guid isPermaLink="true">https://contextdesign.tech/blog/introducing-the-substrate</guid>
      <pubDate>Tue, 26 May 2026 00:00:00 GMT</pubDate>
      <description>Why generic AI doesn't know your business — and how a substrate closes the gap. The first post under the Context Design Co byline.</description>
      <content:encoded><![CDATA[
Most of what makes a mid-market business work is not on its balance sheet. It lives in the founder's head, in the working knowledge of senior staff, in the spreadsheets nobody else opens. It is in the way decisions get made when the pressure is on. For most of the history of operating businesses, that was simply the reality — unauditable, untransferable, and invisible to a buyer or a successor.

There is now a structural fix. We call it a substrate.

## §1 — What we mean

A substrate is a custom-fit knowledge layer that captures what your business already knows — documents, workflows, decisions, conversations, system records — into a queryable structure that lives on hardware you control. AI connects to it via [MCP](https://modelcontextprotocol.io), the open Model Context Protocol. Once connected, the same generic LLM that struggles with your business becomes fluent in it.

## §2 — Why this matters

Two things. First: most AI consulting work today is configuring generic models against generic data. The output is correspondingly generic. What's rare is AI that knows your specific business — your project files, your contractor history, your tax records, your accumulated decisions. The substrate is the path between your business's specific knowledge and AI's general capability.

Second: the substrate, once installed, is an asset on your balance sheet. We codify it as [Operational IP](/operational-ip/) — MIT-licensed software that the buyer's auditor can recognise as an intangible asset, and that a future buyer pays a premium for because it removes the founder-dependency discount that SMEs normally absorb.

## §3 — Where the line is

We don't sell AI. We sell the substrate that makes AI useful to you specifically. The substrate is yours from minute one — code-licensed, self-hosted, audited. We don't hold your context hostage.

This is the first post. Future entries will be both founder-voice thesis pieces (under this byline) and Editorial digests on the operational-IP literature and what mid-market operators are doing with substrate in production.
]]></content:encoded>
    </item>
  </channel>
</rss>
