Dave Vellante, in a 2026 enterprise tech predictions discussion, said he expects OpenAI to exit 2026 with more enterprise revenue than Anthropic, while adding that he thinks Anthropic will do very well anyway. Two things that sound contradictory, sitting comfortably next to each other. That’s the most honest sentence I’ve read about enterprise AI all year, and it tells you more about what’s actually happening inside companies than any panel title ever will.
So when Anthropic, Gamma, and Clay sit down at Disrupt to talk about what happens when enterprises actually deploy AI, I’m less interested in the stage energy and more interested in whether anyone says the boring part out loud. Because the boring part is the whole job.
The unglamorous thing that actually decides deals
On August 20, 2026, Reuters reported that Anthropic plans to let enterprise customers exercise greater control over their data by keeping their 30-day retained data on their own cloud infrastructure rather than solely inside Anthropic’s systems. The 30-day retention requirement stays. What changes is where that data physically lives.
Read that again and notice how little it resembles a product launch. There’s no new model, no benchmark chart, no demo. It’s a storage location. And yet I’d argue it moves more enterprise budget than the last three capability announcements combined, because it’s the thing that unblocks the person who says no.
Anyone who has sat through a procurement review knows the pattern. The pilot works. Engineering loves it. Then legal asks where the data sits, security asks who can subpoena it, and the whole thing stalls for two quarters. Deployment at enterprise scale isn’t a modeling problem. It’s a “whose cloud account is this in” problem. Anthropic appears to have figured that out, and the revenue mix reflects it: roughly 80% of Anthropic’s revenue reportedly comes from enterprise customers, against roughly 40% for OpenAI. One of those companies is built around people who read contracts. One is built around people who don’t.
Why the revenue mix matters more than the leaderboard
Reports have also put Anthropic at 73% of new enterprise spending, and the company has floated a $30 trillion total addressable market claim that reportedly exceeds SpaceX’s. I’d treat the TAM number the way I treat all TAM numbers, which is as a fundraising artifact rather than a fact about the world. Nobody has ever been held accountable for a TAM slide.
The revenue composition is different. That’s a measurement of who is actually paying, and it suggests two companies running genuinely different businesses under the same category label. Vellante’s split prediction makes sense in that frame. OpenAI can win on total enterprise revenue while Anthropic wins on enterprise concentration, and both can be true because they’re selling to different buyers inside the same building.
The safety posture got more honest, which is uncomfortable
Anthropic’s Responsible Scaling Policy v3.0, published February 24, 2026, keeps the capability-threshold framework but shifts emphasis toward transparency and industry-wide recommendations rather than unilateral pauses. The stated reasoning is that stopping development while competitors keep going carries its own risk.
I’m not going to pretend that isn’t a retreat from the earlier posture. It is. But it’s a retreat stated plainly rather than buried, and the underlying logic is hard to dismiss. A company that pauses alone doesn’t stop the frontier from advancing. It just stops being at it.
The company also disclosed industrial-scale distillation attacks it attributes to three Chinese AI labs. For enterprise buyers, that’s a useful signal about what your vendor is defending against. Your model provider is a target. That threat surface is now part of your threat surface, whether or not it shows up in your vendor questionnaire.
What I’d actually ask on that stage
If I had the mic at Disrupt, I’d skip the vision questions entirely and ask the three that decide whether a deployment survives contact with a real company:
- Where does our data physically live, who can compel access to it, and can we prove that to an auditor without a call with your team?
- What breaks when the model updates, and how much notice do we get before it does?
- When your safety policy changes, as v3.0 shows it can, what contractual protection do we have that isn’t a blog post?
The 2026 State of AI Agents Report drew on over 500 technical leaders and implementations at companies including Novo Nordisk, Doctolib, L’Oréal, and Shopify. Those are not companies that move on demo quality. They move on data residency, change management, and audit trails.
Which is my whole point. The interesting era of enterprise AI is the one where the interesting parts stop being interesting. Where the conversation is about storage locations and retention windows and who gets paged when a model changes behavior. A panel that spends its time there is worth watching. A panel that spends it on the future of work is worth skipping.
🕒 Published: