Better the Devil You Know
Every "sovereign AI" startup pitches the big model companies as villains. I'm not defending them, but they don't train on API data, you can opt out on the consumer side, and there are audits and regulators. The startup promising "security" usually has a vibe. Pick the devil with the paperwork.
Every week a new AI startup lands in my inbox promising "sovereign," "private," "secure" AI, and the pitch always has a villain in it. OpenAI is reading your data. Google is training on your customers. Anthropic is a black box. Come to us instead, we built our own model, your data never leaves our hands.

I run an AI company and I'm about to say something that'll annoy half my industry: I'd rather hand my data to the devil I can audit than the one I can't.

I'm not defending the big guys
Let me get this out of the way. OpenAI, Anthropic, Google, the rest of them have real problems. There are lawsuits that are going to take years. The terms of service read like they were written to be skimmed. Their incentives are not your incentives. I'm not carrying water for any of them.

But here's what they also have, and what the "secure AI" startup pitching you almost never has: a published data policy you can actually read. A zero-data-retention option you can turn on. HIPAA-eligible configurations with a signed agreement behind them. Third-party audits with reports you can request. Opt-outs from training that are documented, not promised. Regulators who know their names and have subpoena power. Journalists who would love to catch them lying.

Those aren't just words on a sales page. OpenAI lists SOC 2 Type 2 and ISO 27001 audits for its API and business products, and signs a HIPAA business associate agreement for eligible health care customers. Anthropic lists SOC 2 Type I and Type II, ISO 27001, and ISO 42001, the international standard for how a company manages AI. Google Cloud is certified under ISO 42001 too. Anthropic will send you its compliance reports through its trust portal if you ask.

Is the legal language annoying? Yes. Is it a pain to find the right page? Also yes. But it exists, it's specific, and if they violate it there are people whose job is to make that expensive for them. That's what "the devil you know" means. Not that he's good. That you know what he does.

Here's what that looks like in real life. In 2025, in the New York Times copyright lawsuit, a federal court ordered OpenAI to save ChatGPT logs it would normally delete, even chats people had deleted themselves. That's one of the real problems I was talking about. But look at everything else you know. According to OpenAI, business customers on ChatGPT Enterprise, Edu and zero data retention were outside the order, and the broad hold ended that September. All of it played out in public, in court, where anybody could read it. If a startup got the same order, would you ever hear about it?

Two things most people don't know, and should
These are the two facts that quietly demolish most of the fear-based pitches, and almost nobody selling against the big labs brings them up.

One: they don't train on API data. There are two different doors into these companies. There's the consumer app, the chat window you type into on your phone. And there's the API, which is what a developer plugs software into. The major providers state plainly in their business and API terms that anything sent through the API is not used to train their models, not by default and not without you specifically asking for it. So when your software calls a model to answer a customer's question, that conversation isn't going into the next version of the model. It's processed, it comes back, and that's the end of it. If you want the retention window gone too, most of them will give you zero-data-retention on request, which means they don't even keep a copy for the usual abuse-monitoring period.

And it's written down, by name. OpenAI stopped training on API data by default on March 1, 2023, and without zero data retention it keeps API data for up to 30 days to check for abuse. Anthropic says the same about its API, even when you reach Claude through Amazon Bedrock or Google Cloud. And Google says that when you pay for the Gemini API, it doesn't use your prompts or responses to improve its products.

Google's terms prove my point from the other side, too. On the free tier, Google says it does use what you send to improve its products, and human reviewers may read it. The terms say, in bold, do not submit sensitive, confidential, or personal information. That's not a scandal. That's the paperwork telling you exactly which door is which. Pay for the door that doesn't train on you.

That matters because the scary version of this story, "your customer data is becoming part of a giant AI brain," is specifically the thing the terms say doesn't happen on the channel that businesses actually use. And you don't have to take my word for it or theirs. It's written down. Go read it.

Two: on the consumer side, you can opt out. The chat app is a different deal. There, your conversations may be used to improve the model, and historically that's been on by default. But there's a setting. In every major one of these products you can turn off "improve the model for everyone" or whatever they're calling it this quarter, and once it's off your chats stop feeding training. Business and enterprise plans generally have it off from the start.

Anthropic is a good example of how this plays out. Until 2025 it didn't train on consumer chats at all. Then it changed its consumer terms and asked every Free, Pro and Max user to pick. Say yes, and your chats can be used for training and kept for up to five years. Say no, and they stay on the old 30-day retention. Its business, government and education plans and its API weren't part of the change. Plenty of people were mad, and fair enough. But notice how you found out: a public announcement, a pop-up, a deadline, and a setting you can change any time.

Is "it's on by default and buried in settings" an annoying design choice? Absolutely. I'd rather it were opt-in. But an annoying toggle that exists and works is still infinitely more than a startup telling you they take security seriously and offering you nothing to switch, nothing to read, and nobody to complain to.

The devil you don't
Now look at the startup that says "we built our own model, so your data is safe." Ask them the questions you'd ask the big guys. Where does the data live? Who can access it? Is it used to train anything? Is there an audit? Who did it? Can I see it? What happens to my data when you get acquired or run out of runway?

I've asked. The answers are usually some version of "we take security very seriously." That's not a policy. That's a vibe. A twelve-person company with a model they trained on a rented cluster has fewer controls, fewer eyes on it, and less to lose from getting it wrong than a company with a hundred lawyers and a stock price. The small company might be honest. I'd bet most of them are. But honest and accountable aren't the same thing, and I can only build a business on the second one.

Most "AI companies" are application layers, and that's fine
Here's the part the sovereign-AI pitch doesn't want you to think about. Most of the companies telling you they're an AI company are calling one of the big providers under the hood. Including a lot of the ones telling you not to trust the big providers. They've got a product on top, a prompt layer, some retrieval, a nice interface, and then an API key to the same model everyone else uses. Scratch the paint and it's OpenAI or Anthropic underneath.

And I want to be clear: there is nothing wrong with that. It's the correct architecture. I don't build my own cloud, I use AWS. Nobody thinks less of a software company for running on AWS. The cabinet shops I work with don't grow their own trees or forge their own saw blades. They buy lumber and hardware and equipment from suppliers who do that at scale, and then they do the part that actually matters to the customer: the design, the fit, the finish, the install.

The dishonest move isn't using somebody else's model. The dishonest move is pretending you didn't, and selling "sovereignty" on top of an API call. If you're an application layer, say so. Then be a great one.

Frontier models are not a startup project
Training a frontier model costs more than most startups will raise in their lifetime, and the compute to run one at scale is rationed. The handful of companies doing it are fighting over chips, power and data centers. That's not a game a seed-stage company plays. It's not even a game most unicorns play.

How big is the game? Stanford estimated that the computing power alone to train GPT-4 cost about $78 million, and Google's Gemini Ultra about $191 million. Those were 2023 models, and the bills have only gone up. In 2025, OpenAI, SoftBank and Oracle announced Stargate, a plan to put up to $500 billion into AI data centers in the U.S. Anthropic signed up for as many as one million of Google's AI chips, a deal worth tens of billions of dollars, to bring well over a gigawatt of computing online in 2026. And Amazon, Microsoft, Google and Meta planned about $700 billion in capital spending for 2026, much of it for AI data centers. Your local AI startup is not in that fight.

So when a small company says "we built our own model" the honest version is usually "we fine-tuned an open-weight model and we're running it on rented GPUs." Which can be a fine technical choice for some use cases. It is not a security feature. If anything it's the opposite: now your data is on hardware nobody's audited, run by a team that doesn't have a compliance department, with a model whose training data they probably can't fully account for either.

Where startups actually win
The last mile. Every time. The model is a commodity getting cheaper by the quarter. Stanford found that the cost of running a model as good as GPT-3.5 fell more than 280 times between November 2022 and October 2024. What isn't a commodity is knowing exactly how a cabinet shop quotes a job, or what a dental office needs to hear on a missed call, or where a landscaper's leads leak out of the funnel. That knowledge, turned into a product, priced on the outcome it produces, is where the money is for anyone who isn't OpenAI.

That's what we do. We own our platform, we control the phone lines and the texts, and the models we call come from providers whose data policies I've actually read, through the API, with retention switched off where it matters. We don't pretend we trained the brain. We built the thing the brain is plugged into, for a specific kind of customer, and we charge for what it produces. A client doesn't care which model answered the phone. They care that the phone got answered and the lead got booked.

It's symbiotic, not a war
The pitch that frames big model companies as the enemy is selling you a story, and the story's job is to make you forget to ask for the audit. The real shape of this industry isn't David versus Goliath. It's a power plant and a delivery route. The frontier labs do the thing that costs billions and requires a hundred lawyers. Startups do the thing that requires actually knowing the customer. Startups price on outcomes, build the last-mile logistics, and stand on top of providers who've been forced, kicking and screaming, to write down what they do with your data.

That's not a compromise. That's the right split. Pick the devil with the paperwork. Then go build something he can't.

Want to talk about what you're building?
Get in touch


