Bits Sans Borders
Anthropic charges extra for US-only inference. What can we infer from that?
(Today’s post is a guest post by economist Ashish Kulkarni, who usually writes over at EconForEverybody. Here, he looks at one small detail of Anthropic’s pricing model for customers using their Claude model via the API—an extra charge for ensuring that your data stays within a country—and derives a number of interesting conclusions and questions from it. This post is not just about the various interesting geographical and political issues that get involved in AI use, but also has examples of how to get a better understanding of things happening in the world. Note, for example, the focus on “why did this price happen” and “what are the competitors doing”)
Economics is absolutely everywhere. Even, as it turns out, in API documentation exposed by the AI companies to allow programs to use their GenAI models
Here’s a specific term in that document that got my economist’s antennae up: inference_geo tag. Let’s take a look at what that means, and what makes it interesting to my tribe.
This tag is simply a way for you, as a programmer, to tell Anthropic where you would like your work to be done. When you write a piece of code instructing Anthropic to run its models on your data, you can also specify where this should be done - that is, in which geography. With this snippet, Anthropic is saying, you can instruct Anthropic to have the code be run either a) anywhere in the world or b) specifically within the United States of America.
The really interesting thing is the price. Anthropic gives you this option, but says that you must pay 10% extra for it.
“Data residency pricing varies by model generation:
Claude Opus 4.6 and newer: US-only inference (inference_geo: “us”) is priced at 1.1x the standard rate across all token pricing categories (input tokens, output tokens, cache writes, and cache reads).
Global routing (inference_geo: “global” or omitted): Standard pricing applies.”
Bottom line: companies are big, and have big-ass rules. Not complying with the rules usually comes with a world of legal and regulatory pain, and quite real security concerns to boot. So Anthropic is saying to its new potential customers that hey, don’t worry, I got your back. It is going to cost you 10% extra, though.
Never Reason from a Price Change
If you have some exposure to economics, you would be tempted to look at this pricing structure and ask, “What does this mean?” But if you are exposed to economist Scott Sumner’s writing, you know the title of this section in your bones. Here’s a short summary from a very early post on this topic by Scott:
“My suggestion is that people should never reason from a price change, but always start one step earlier—what caused the price to change.”
So here we have a nice little problem. Folks who want inference to be run exclusively in the United States can get what they want, but they must pay 10% more - but only if they decide to go with Anthropic. That’s a price change by one company, all right. But now let’s try to solve the Sumnerian task: what caused the price to change?
Maybe it genuinely costs Anthropic (and only Anthropic) more to guarantee US-only inference (dedicated capacity, less efficient routing, can’t load-balance globally)?
Maybe corporate buyers will pay whatever it takes for compliance — the demand is inelastic, so Anthropic captures surplus?
Maybe the premium signals “enterprise-grade” positioning, even if the actual cost is lower?
Or maybe it is all of these things and something else as well! That, in fact, is the central point that Scott wants us to learn as students of economics: don’t reason from a price change, but be clear about what the underlying causes are. Or try to be, at any rate.
So in this specific case, which is it?
We may not find out “the” answer by the time we come to the end of this post (does this qualify as a non-spoiler alert?), but we will end up learning more about both economics and about the AI industry. And a good way to begin this lesson is by asking the question: Does this happen in other, related industries?
Cloud, baby, cloud
The major players in cloud infrastructure have charged regional premiums for years. All of the major players (AWS, Azure, GCP) charge their customers a premium to have processing happen in specified geographies. Each of them do it a little different than each other, but given the inflexibility of compliance requirements, it is safe to assume that demand is going to be mostly inelastic. And what that means is that if legal requirements mandate it, and companies have no choice but to comply, cloud service providers have every incentive to apply (and get away with) higher pricing in such cases.
But the analogy between general cloud pricing and AI inference pricing ain’t perfect. In traditional cloud computing, US regions are often among the cheapest — AWS itself recommends deploying to “US East (N. Virginia)” by offering lower prices for that option, and third-party research says pretty much the same thing: that US regions bring up the rear of the global price index. Think of it this way: the “traditional” cloud market is driven by real-world infrastructure costs — land, electricity, taxes. That tells you about the nature of the equilibrium in both markets - or the lack of one in one of them, rather.
AI inference geo-pricing, in other words, is a different animal. When cloud providers offer geographic guarantees specifically for AI model inference, a 10% premium shows up - at least in some cases. Here’s Claude on the topic:
“Azure’s ‘Data Zone’ tier for OpenAI models costs roughly 10% more than the Global tier. On AWS Bedrock, it’s the reverse framing but the same spread: global cross-region inference offers ~10% savings over geographic profiles. And on Google Vertex AI, regional endpoints for Claude carry a 10% premium over global — while Google’s own Gemini models have no regional premium at all. This isn’t inherited from general cloud pricing. It’s a new, AI-specific pricing pattern, and the convergence on that 10% number across independent providers is, to put it mildly, worth paying attention to.”
A question worth considering: if Google can offer zero premium here, is it subsidizing adoption, or is its cost structure genuinely different?
Different Labs, Different Bets
So the three major labs have different pricing strategies, with quite a few variants each. Mistral, the French AI lab, is EU only by default, and so compliance is offered as a baseline feature, rather than something that can be potentially upsold.
Does that mean that Anthropic is in the lead here, because it is able to charge a higher price for a feature such as this? Or does it mean that different firms are making fundamentally different bets about what enterprise buyers need and will pay for? That we’re asking these questions at all is the point: when one sees this much variance in pricing strategies, one can reach the conclusion that the market is still being defined.
That is, the fact that we see these differing price regimes, even for something as niche as this, indicates that every major player seems to be unsure of what price will be supported for which feature. But personally, I take this as an indicator that Anthropic thinks it has the ability to apply a bit of differential pricing. It is also an indicator that Anthropic is getting serious about the corporate segment - even more than it already was!
What Happens Next?
… is precisely what makes this such a fascinating case study.
If Anthropic finds out that demand is inelastic, expect to see this code snippet hang around for a while. Also, expect to see other firms reach the same pricing strategy. That tells you that there isn’t much differentiation between the models being offered by the different labs, and that demand for regulatory compliance is relatively inelastic.
If, on the other hand, Anthropic finds out that demand is elastic, expect to see this code snippet disappear on short notice. Also, expect to see that other firms never charge this premium.
What if we find out that Anthropic removes this premium, but other firms introduce it? That’ll be a rich vein of analysis for us, and really bad news for Anthropic!
Or, of course, we could potentially observe that we are in a world where regulatory compliance simply doesn’t matter as much. That wouldn’t surprise me all that much, but that’s a whole other story.
So What?
Here’s the thing: pricing decisions in AI are economic signals, and decoding them tells you something about market structure that no earnings call will spell out for you. This entire post has been an exercise in reading one such signal — a 10% premium on a single API parameter — and tracing out what it implies about costs, demand, competitive positioning, and where this industry is heading.
Which brings me to a parting thought. In a recent conversation with Dwarkesh Patel, Dario Amodei (Anthropic’s CEO) has this characterisation of the current state of the AI industry: a small number of firms, differentiated products, and margins that don’t equilibrate to zero. He even used cloud computing — three, maybe four players, sustained profits — as his analogy.
Take a look at what we just walked through. Cloud providers seem to be converging on a roughly 10% premium for AI geo-inference. Labs are making different bets on whether to charge for compliance at all. This is a market where product and pricing strategies haven’t yet converged — which is exactly what you’d expect in a Cournot oligopoly that’s still figuring out its equilibrium. But it is one thing for this to be said in an interview, and quite another to uncover pieces of evidence that corroborate this hypothesis. Helping you figure out how to do that is one of the points of this post.
Dario is betting that this market stays an oligopoly with room for margins. The inference_geo premium is one of the first places where you can actually watch that bet play out in real time.
(This is just one small economic look at AI. Ashish has designed an entire course on the economics of AI, which he then compressed into a talk which became a website and which became an advanced method of using AI in your work and which became a harbinger of things to come. You should see his blog about this too.)