Insource the Model

Canada just disclosed $66 million in Palantir contracts it hadn’t been telling anyone about. The headlines are about transparency. That’s the wrong thing to be upset about.

The real question isn’t whether the contracts were disclosed. It’s whether, five years from now, Canada will own anything that matters — or whether it will have spent $66 million building a dependency.

This isn’t a Palantir problem. It isn’t even a government problem. It’s the beginning of a transition every serious institution is about to go through, and the ones that understand it early will be in a fundamentally different position than the ones that don’t.

We’ve Seen This Curve Before

In the late 1990s, technology was widely treated as a cost center. You outsourced it to vendors and consultants. You signed enterprise contracts. You bought platforms. Technology was something you bought, not something you built.

Then companies started noticing that the businesses winning their markets weren’t outsourcing technology — they were insourcing it. Amazon didn’t outsource its logistics platform. Google didn’t buy its search algorithm from a vendor. The realization spread: if technology is integral to how your business runs, it’s a core competency. You own it or you lose the competition to someone who does.

The insourcing wave reshaped entire industries through the 2000s. Companies that made the shift built durable advantages. Companies that stayed in the enterprise contract model found themselves perpetually behind, dependent on vendors whose incentives didn’t align with theirs.

AI is going to run the same curve. Faster.

What an LLM Actually Is

Here’s the frame that makes everything else clear: a large language model is a regression of knowledge.

Train it on the public internet and you get a regression of public internet knowledge — broad, generic, useful for many things, optimized for none of them. That’s what every commercial API gives you. It reasons the way the average of the internet reasons.

Now consider what it means to train a model on your data. Your decisions. Your institutional history. Your domain expertise, your customer interactions, your operational logic, the accumulated judgment of your organization across years of activity.

That model doesn’t reason the way the internet reasons. It reasons the way you reason. It reflects your standards, your risk tolerance, your institutional knowledge. It can make decisions the way a senior person at your organization would make them — because it was trained on exactly that.

This is not a vendor product. This is institutional DNA.

If an AI is making important decisions for your company — underwriting risk, triaging cases, advising customers, flagging anomalies — do you want it trained on what everyone knows, or do you want it trained to your spec?

The Coming Frontier

Every serious company will eventually have its own models. This is not a prediction that requires any exotic assumptions. It follows directly from the same logic that drove technology insourcing in the 90s. When AI becomes integral to how a business operates — and it is already integral, and it will become more so — the question of who owns the model becomes a strategic question, not a procurement question.

The competitive moat in five years won’t be which company is using GPT-6. It’ll be which company has a model that knows its business, trained on its proprietary data, reflecting its institutional judgment. That model can’t be replicated by a competitor who signs the same API contract. It’s not for sale. It’s not available in a vendor catalog.

Custom models will be deeply proprietary — more protected than source code, because the training data and the human judgment embedded in that data are harder to reconstruct than any codebase.

The Palantir contract is a signpost on the wrong road. Buying a platform that reasons generically about your data is the outsourced-IT equivalent. It looks like a solution. It’s actually a deferral — of the capability-building that will eventually have to happen anyway, at a higher cost and from a weaker position.

Doing It Right Is Another Thing

None of this is easy.

Training custom models requires data infrastructure, ML engineering, evaluation frameworks, and institutional discipline that most organizations don’t have. Getting the training data right — clean, representative, annotated correctly — is expensive and slow. Getting the evaluation right is harder. Knowing when the model is wrong is harder still.

The 90s insourcing wave was also painful. Companies that moved too fast built systems that didn’t work. Companies that hired wrong built technical debt. The capability had to be learned, not just acquired.

The same will be true here. Most organizations will do this badly before they do it well. The ones that start building the internal capability now — even imperfectly — will be ahead of the ones that wait for the perfect solution and keep signing vendor contracts in the meantime.

Canada’s $66 million doesn’t have to be a mistake. It can be a phase. The question is whether it’s followed by the harder work: building the institutional capability to own what you depend on.

The institutions that understand that AI is a core competency — that an LLM trained on your decisions is an asset, not a subscription — will not be buying it from a vendor in 2030.

— J.P. Howlett

Related: The Ungoverned AI Looks Like This — when there’s no protocol between institutions, vendor capture is how the gap gets filled.

Related: The Trials Exist for a Reason — autonomous AI doing drug development raises the same accountability question: who answers when something goes wrong?