Large companies are paying a premium for AI that their smallest competitors aren't, and it has nothing to do with capability. Several of the strongest and cheapest open models on the market right now come from Chinese labs. Startups put them into production without a second thought. A Fortune 500 company, a bank, or a public agency almost never does, for reasons that are reasonable on their face: regulatory exposure, customer perception, procurement risk, and the fact that nobody wants to explain the model's origin to a board or an auditor. That caution is a posture, and the posture has a cost. The difference between what your risk policy lets you deploy and what your least constrained competitor deploys is a price you pay for where a model came from. We call it the origin premium, and most enterprises are paying it on every workload, including the ones that never needed the protection.
What is the origin premium?
It's the cost gap between the model your policy permits and the cheapest model that would do the job, where the gap exists because of the model's provenance rather than its performance.
Every premium in AI has a reason, and the reason decides whether it's worth paying. The frontier tax is the price of running the best model on work that didn't need it, and the fix is matching the model to the job. The origin premium is different. The matching is already right. The cheaper model clears the job. You're paying more because of who trained it, and the premium shows up in two forms. Sometimes it's a direct price gap, because the model you're allowed to run costs multiples per token of the one you aren't. Sometimes it's a capability gap, because the permitted model is a generation behind on the task. Either way, the bill is for provenance, and it should be itemized the way any other premium is.
Is the caution wrong?
No. The caution is often right. The mistake is applying it uniformly.
We're not going to argue a regulated enterprise should put a model of uncertain provenance behind its customer data, and we don't build systems that do. The exposure is real, and the regulatory picture around model origin is a reported fact that may tighten. What's wrong is the blanket. Most companies have one rule for model origin, and it's the rule written for the most sensitive workload they have. That rule then governs the internal document summarizer, the test-data generator, the code-comment drafter, and the overnight classification job, none of which carry the exposure the rule was built for. A posture designed for the top of the consequence line gets applied all the way to the bottom, and the premium gets paid on every call.
Why is this a bigger deal than it sounds?
Because the companies not paying it are the ones coming for your customers, and the premium compounds with volume.
The startup building in your category runs whatever model is best per dollar today, swaps it next quarter, and never writes a policy memo. The largest enterprises and public agencies are operating at a cost disadvantage against companies willing to take the origin risk, and the disadvantage is widest exactly where AI volume is highest. At pilot scale, nobody notices a few cents per task. At production scale, where an agent fleet runs millions of calls a month, a two-to-five-times price gap on the cheap tier is the difference between an AI budget that grows with the business and one that gets capped. We've written about the frontier freeze, where growth lands on cheap capacity under a flat premium bill. The origin premium is what happens when your policy forbids the cheap capacity.
Does the premium ever go to zero?
It's shrinking, and the reason is a market the enterprise buyer should be watching closely.
A year ago the strong open models were almost entirely Chinese, and the only alternatives were smaller releases from Western labs that trailed by a generation. That's changing. Western labs are now shipping open-weight models aimed squarely at the enterprise and public-sector buyer who can't use the Chinese ones, and several of them are priced and licensed to compete. The gap hasn't closed, and the best open model on most leaderboards still isn't Western. But the premium is now a moving number, and a company that priced it in January is working from a stale figure. This is also where the weight line and the license line earn their keep. A Western open model under a permissive license, held on your own infrastructure, answers the provenance question and the control question in one move, which is a combination the closed frontier can't offer at any price.
How do you find out what you're paying?
Price two stacks for the same workload, one under your current policy and one under the policy a startup would use, and the difference is your premium.
Pick a high-volume, low-consequence workload. Classification, summarization, extraction, anything where a wrong answer is cheap. Cost it on the model your policy allows. Cost it on the best open model per dollar regardless of origin. The gap is the origin premium for that workload, and multiplied across your volume it's a number a CFO can read. Most companies that run this exercise find the number is large on exactly the workloads where the risk justification is weakest, and small or zero on the sensitive workloads where the caution belongs. That's the finding that opens the fix, because it separates the premium you're paying on purpose from the premium you're paying by default.
What should you do about the origin premium?
The question you're really asking is whether you can keep your risk posture and stop overpaying for it. You can. The move is a tiered origin policy instead of a single rule.
Write the policy in three tiers against the consequence line. The top tier, customer data, regulated workloads, anything you'd have to report, keeps the strict rule, and the premium there is a deliberate cost of doing business. The middle tier, internal work with no sensitive data, gets a documented review path, where a model of any origin can be approved for a named workload after the security and license checks that any component would get. The bottom tier, synthetic data, test environments, disposable experiments, runs the best model per dollar with origin logged and nothing else. Then put a Western open model behind your socket as the default for the middle tier, so the premium there shrinks toward zero without the policy argument ever reaching the top tier. Revisit the tiers quarterly, because the market closing the gap moves faster than policy does. The caution stays. The bill for it stops landing on work that never needed protecting.
If you want the origin premium measured on your workloads and a tiered policy that keeps the posture without the bill, start with an AI Blueprint or reach us at contact@theyor.com