Is a 10% Edge Worth
57 Times the Cost?

The decision the giants are making

Microsoft is reportedly weighing a self-hosted version of DeepSeek's V4 model for Copilot — to cut costs, as OpenAI and Anthropic push enterprise pricing up. Read that again. The largest software company on earth is openly doing the one calculation most buyers refuse to do out loud: is the premium model actually worth the premium?

I'll say the quiet part first: we made this decision before Microsoft did. We plugged the DeepSeek API into our stack weeks ago — not because a giant blessed it, but because the math was already obvious to anyone willing to look. When the Microsoft news broke, my reaction wasn't surprise. It was: of course.

And it wasn't frictionless internally. When I made the call, one of my staff pushed back immediately — her worry was data security. China might collect our data.

How much is your data worth?

A fair instinct. But my first question wasn't about China. It was: how much is your data actually worth?

Not rhetorically — literally. What is in those API calls that a foreign government wants, can use, and can't already get more easily elsewhere? For most businesses, the honest answer is: far less than we like to imagine. We treat our data like state secrets and price it like gold, then hand it to a dozen apps a day without a second thought.

Why only the Chinese one?

And here's the part that ended the discussion. Why doubt only the Chinese one? OpenAI, Anthropic, Google — they collect data too. They run on servers I can't see, under policies I can't audit, in jurisdictions with their own subpoenas and their own politics. Nobody can guarantee what any of them do on the back end. If the concern is "a provider might see our data," that concern applies to every provider. The moment it only applies to the Chinese one, it stopped being a security argument and became something else.

Cost vs performance

So strip the flag off the model and look at what's actually left: a decision between cost and performance.

So the real question — the only one that survives once the fear is set aside — is this: is a single-digit performance gain worth paying 57 times more for?

The 90% most people miss

For a narrow set of tasks, yes — pay for the best and don't blink. If a few percent of accuracy decides whether you win a case, ship safe code, or close a million-ringgit deal, the premium is cheap insurance. But be honest about how often that's actually the job.

So the premium model isn't wrong — it's wrong for 90% of what you do. The discipline is to match the model to the task: the frontier for the 10% that truly moves the needle, the cost-efficient one for everything else. Most businesses do the opposite — they pay frontier prices across 100% of their usage to protect an edge they need maybe a tenth of the time.

The giants aren't stupid

This is where I come back to a belief I keep returning to: the people running the world's giants are not stupid. Microsoft has armies of engineers, lawyers, and risk officers. They didn't just glance at the price tag — their own testing reportedly judged DeepSeek's performance good enough that it's indistinguishable from the expensive models on the everyday tasks that make up most real work. So they weighed security, performance, and cost from every angle I could, and more — and still moved. Not out of carelessness. Because the math is the math.

And here's the quiet satisfaction in it for me: a small company in Malaysia and a trillion-dollar one in Redmond ran the same calculation and reached the same answer — except we got there first. When the right call is obvious, you don't need permission from a giant to make it. You just need to be willing to do the math before everyone else does.

I've also seen this from the inside. In a meeting with ByteDance — a company with its own large model (Doubao in China, Dola internationally) — I watched their own staff reach for DeepSeek to get the work done. Sit with that for a second. The people who build a frontier model, who run it in-house at effectively no cost to themselves, still picked the outside tool for the task in front of them. For them it wasn't even about price — it was a plain verdict on what works best. When the model-makers themselves vote with their own workflow, the cost-versus-performance debate is basically settled.

And notice what the Azure detail quietly proves: when data security genuinely matters, you don't ban the model — you host it yourself. Open-weight models can run on your own infrastructure, where nothing leaves your walls. The serious answer to a security concern is engineering, not suspicion aimed at one country.

The verdict

I'm not telling anyone to abandon the frontier models. I pay for them where the performance earns its price. I'm saying the decision should be made on a spreadsheet, not a reflex. Cost on one axis, performance on the other, and the actual value of your data — measured honestly — somewhere in between.

Published by IMA AI — June 2026. Microsoft / DeepSeek report: Axios, June 2026; pricing figures as publicly reported.