Roughly 80% of Anthropic’s revenue comes from enterprise customers. Compare that to OpenAI at around 40%, and you get a picture of two companies building for very different rooms. One is building for everybody. The other is building for procurement teams, security reviewers, and the person whose job is to ask uncomfortable questions about where the data sits.
That number explains a lot about why the enterprise deployment conversation at TechCrunch Disrupt 2026 — with Anthropic, Gamma, and Clay on stage — is worth paying attention to. Not because panels are where truth happens, but because the gap between “we deployed AI” and “our AI deployment survived a security review” is the single widest gap in this entire space right now. I review tools for a living. That gap is where most of them fall in.
The data retention question nobody asks until month three
In August 2026, Reuters reported that Anthropic plans to let enterprise customers keep their 30-day retained data on their own cloud infrastructure rather than solely inside Anthropic’s systems. The retention requirement itself doesn’t go away. What changes is who holds the bag.
If you’ve never sat through an enterprise AI procurement cycle, this sounds like plumbing. It isn’t. Data residency is the thing that kills pilots. A team spends six weeks building something genuinely useful, demos it, gets applause, and then legal asks where the prompt logs live. The answer is usually “a vendor’s cloud, in a region we didn’t pick, for a duration we didn’t set.” That’s the moment the project quietly becomes a “phase two initiative.”
Letting customers hold their own retained data doesn’t solve every compliance problem, but it moves the conversation from “no” to “show me the architecture.” For anyone evaluating tools, that’s the distinction that matters. A model that’s slightly worse at reasoning but clears your security review is infinitely more useful than a model that’s slightly better and never gets deployed.
What the safety policy rewrite actually signals
Anthropic’s Responsible Scaling Policy v3.0, published February 24, 2026, kept the capability-threshold framework but shifted emphasis toward transparency and industry-wide recommendations instead of unilateral pauses. The stated reasoning is blunt: stopping development while competitors keep going carries its own risk.
I’ll be honest about how I read that. It’s a company acknowledging that being the only one who brakes doesn’t make the road safer. You can find that pragmatic or you can find it convenient. Probably both are true. What it tells buyers is more concrete: the safety posture you’re evaluating is competitive, not absolute. Policy documents describe intentions under commercial pressure, and commercial pressure in 2026 is considerable.
Anthropic also disclosed industrial-scale distillation attacks it attributes to three Chinese AI labs. That detail deserves more attention than it got. It means frontier model providers are themselves targets, and the thing being stolen is model behavior. If you’re building a product on top of someone else’s model, your supply chain now includes their security posture. That’s not a hypothetical risk on a slide. That’s a disclosed incident.
The reality check from actual deployments
The 2026 State of AI Agents Report drew on insights from more than 500 technical leaders plus implementations at Novo Nordisk, Doctolib, L’Oréal, and Shopify. Five hundred technical leaders is a real sample. Companies like those don’t ship AI features because a demo looked good; they ship because something cleared review.
Which brings me back to why a panel with Gamma and Clay alongside Anthropic is a sensible pairing. Model providers talk about capability. Companies building on top of models talk about the parts that break — rate limits, cost surprises, output that’s 95% right in a workflow that needs 100%, and the ongoing work of explaining to stakeholders why the thing that worked in the demo needs three more weeks.
Worth noting where forecasts disagree. Enterprise technology predictions for 2026 from Dave Vellante and Gemma Allen suggested OpenAI will exit the year with more enterprise revenue than Anthropic, while still expecting Anthropic to do well. Both things can happen. Market share and fit-for-your-use-case are separate questions, and only one of them shows up in your quarterly review.
What I’d actually take away
- Ask about data residency in week one, not week twelve. The answer determines whether your project ships.
- Read safety policies as statements of current intent, not permanent guarantees.
- Your vendor’s security incidents are your security incidents.
- Enterprise revenue concentration tells you who a company optimizes for. An 80/40 split is a product roadmap in disguise.
None of this is exciting. All of it is what separates deployments that survive from pilots that get politely archived. If you’re evaluating a toolkit this quarter, spend less time on benchmark charts and more time on the retention terms. That’s where the real answer lives.
🕒 Published: