Anthropic’s IPO shows the growth. Its risks show the work ahead.

The filing says Anthropic had more than 6,000 customers generating at least $100,000 of revenue over the trailing twelve months as of June 30, up from more than 600 a year earlier. Those customers accounted for about 74% of first-half 2026 revenue. That is what caught my attention. Claude is moving beyond individual experiments and into budgets, teams, and the processes they return to every day. Once a tool is there, the question changes from “Should we try it?” to “How do we use it well and keep our company protected?”

Then I got to the risk section. It starts on page 26 of the prospectus and runs to page 120. The Associated Press counted 93 pages of risk factors, from ordinary public-company concerns to AI misuse and the possibility of far more serious harm. Axios focused on the scale of Anthropic's compute commitments and the tension between commercial pressure and safety work. Anthropic also describes risks involving customer information, model errors, and reliance on a small group of partners. These are disclosures about what could go wrong, not proof that every scenario has happened. They are a useful reminder that rapid adoption and careful use have to grow together.

The financial side deserves the same care. Compute and infrastructure expense reached $8.945 billion in the first half of 2026, equal to 55% of revenue, down from 201% in the comparable period a year earlier. Anthropic reported a $693 million operating loss for those six months. Some headlines focus on its much larger $80.7 billion net loss, but the filing says $72.4 billion reflects a fair-value change in financing instruments, rather than cash spent running the business. The growth is extraordinary, and the cost of supporting it is still substantial.

For a company with 10 to 100 employees, this comes down to everyday decisions. An employee may want Claude to summarize a client conversation, prepare a proposal, or work through a spreadsheet. Which information can go into the tool? Should the employee use a company account or a personal one? What can Claude draft, and what must a person check before it reaches a client? If those answers live only in one person's head, the company is leaving too much to chance.

This is where Ruby's work starts. We sit with the people doing the work, map the recurring tasks, and learn what information each process needs. From there, we help a company create an AI policy, practical guidelines, and governance that employees can actually use: approved tools and accounts, rules for sensitive data, permission boundaries, a human review point, and a clear owner when something goes wrong. We train employees on those rules inside the workflows they already do, then revisit them as the tools change. In our work with IT providers, that also means connecting the company's security and access controls to the way its team uses AI day to day.