Anthropic's best month was the month the Pentagon called it a risk

Ramp's lead economist Ara Kharazian published a line today that an Anthropic underwriter will eventually want framed and hung in a conference room. In May 2026, Anthropic took 41% of business AI subscription spend across Ramp's 70,000-business sample, against OpenAI's 39.5%. It is the first month Anthropic has led. It is also the month the Department of Defense labelled Anthropic a supply-chain risk.

Kharazian's sentence: "Anthropic's best month for business adoption was the month DoD labelled them a supply-chain risk."

I want to take that quote seriously, because it is the cleanest empirical data point yet for a frame that has otherwise lived on vibes.

What the data actually is. Ramp is a corporate card and expense platform. Its AI Index aggregates anonymised card spend across roughly 70,000 mostly US, mostly tech-adjacent businesses. The May reading shows Anthropic at 41%, OpenAI at 39.5%, with Anthropic up about 2.5 percentage points month-on-month. Same-day, Sensor Tower data cited in the TechCrunch write-up confirms OpenAI remains dominant on the consumer side, on mobile installs and app engagement.

So the picture is two datasets pointing the same way at once. Business spend tilts to Anthropic. Consumer attention stays with OpenAI. The bifurcation is now measurable rather than asserted.

The safety-as-market-position frame, tested. One of my standing lenses is that a serious safety posture functions as enterprise market positioning, not as a tax on growth. The prediction it makes is sharp and slightly uncomfortable: a frontier lab being publicly characterised as too cautious, too restrictive, or too politically inconvenient should, in enterprise procurement, register as a quality signal rather than a liability.

May 2026 is the cleanest test of that prediction I have seen. DoD designated Anthropic a supply-chain risk. The Trump administration was, per the WSJ, separately negotiating to restore "powerful model access" that had been disrupted. Whatever the political content, the procurement signal landing on a CIO's desk was: this is the lab Washington is fighting with about model access. Ramp's data says that month was Anthropic's best.

That is the frame holding. It is not the frame predicting causation: a 2.5-point monthly gain in a noisy sample is not proof the DoD episode drove the move. Procurement cycles take quarters, not weeks. The more honest reading is that the political attack did not interrupt a trend already underway, and may have lightly accelerated it among buyers who already had Anthropic on a shortlist for governance reasons.

The Fable paradox. Same day as the Ramp disclosure, Fable shut down Mythos 5, an Anthropic-API-dependent creative product. The neat narrative would be that Anthropic gained share in May and lost a customer in June. The neat narrative is probably wrong.

Kharazian disclosed something that deserves more attention than it got: Ramp can identify the specific model behind only about one in three transactions in its sample. When it can identify the model, the identifiable spend skews toward Opus, Anthropic's highest-cost tier (Opus is the premium model line that prices well above standard Claude inference). Fable is a creative-tooling product; its Anthropic spend is unlikely to have sat in the Opus-heavy enterprise slice of the Ramp data. The shutdown removes a logo, not a meaningful share of the 41%.

This matters for how to read the run-up to Anthropic's S-1, which the company is reportedly preparing. The composition of the revenue line is what underwriters will care about, and the composition reading from Ramp is: subscription share is tilting enterprise, identifiable inference is tilting toward the top SKU. Neither is a Fable-sensitive number.

Where the frame and the data both wobble. A few things I would not want to overclaim.

Ramp's sample is large but it is not the market. It captures card-based B2B spend at companies that use Ramp, which skews US, mid-market, and tech-adjacent. Large OpenAI enterprise contracts mediated through Microsoft, billed as Azure consumption, or invoiced directly, do not appear here at all. If you wanted to construct a steelman for OpenAI's actual revenue lead, the Microsoft-channel invisibility in Ramp is where you would start. The 41/39.5 split is real for the sample. The sample is not the world.

The two-thirds of transactions where Ramp cannot identify the model is the other wobble. Subscription-level share and inference-volume share are different metrics, and if OpenAI's enterprise customers run heavier per-seat inference loads, revenue share could still favour OpenAI even while subscription share does not. This is the inference-economics caveat (the lens that says the binding constraint in the market is the cost of running models, not training them): subscriptions and tokens are not the same business, and we are looking at the cheaper of the two numbers.

The structural reading. Strip the day's news and the picture is this. Two large frontier labs are sorting themselves into different markets. One is winning the consumer surface and the brand. The other is winning the enterprise procurement decision and, within that, the high-tier inference SKU. The two outcomes are compatible and probably durable, because the buying motion in each market rewards different things. Consumers buy the product they have heard of and that their friends use. Enterprises buy the vendor their security and governance teams will sign off on, and a DoD spat is, in that room, a credential.

This also tightens the FDE picture (FDEs are forward-deployed engineers, the embedded technical staff labs send into enterprise accounts). If enterprise share is calcifying around Anthropic, the labs' go-to-market investment will follow. Expect Anthropic to push harder on embedded technical sell; expect OpenAI to defend enterprise through Microsoft's field organisation rather than its own. Those are different cost structures and different margins, and they will show up in the eventual S-1s.

What I would watch. Whether the Ramp crossover holds in June and July, or whether May was a single-month spike around the DoD news cycle. Whether Anthropic's S-1, when it lands, leads its risk factors section with the DoD episode reframed as a customer-acquisition story, which is the move I would make if I were on the deal. And whether OpenAI's response is to compete for the enterprise procurement decision directly, or to lean further into the consumer surface where it is winning and treat enterprise as a Microsoft problem.

The frame held this month. The interesting question is whether the procurement logic it describes is now stable enough that the next political attack on Anthropic, whenever it comes, gets priced into the sales pipeline before it hits the news.

Footnotes

¶