OpenAI, Anthropic shift to enterprise

OpenAI is no longer talking about office chatbots like a novelty product. On April 8, the company said business customers now generate more than 40% of its revenue, which means the money is shifting from individual subscriptions to company budgets. (openai.com) That changes what the product race looks like. A consumer tool can be messy and still grow, but a bank, insurer, or drug company wants permissions, audit trails, and systems that connect to internal software without spraying data everywhere. (openai.com) OpenAI’s pitch is a platform called Frontier, launched in February, that lets companies run “AI coworkers” across data warehouses, customer records systems, and internal apps. The company says Frontier includes identity controls, explicit permissions, auditable actions, and built-in evaluation loops so agents can keep improving after deployment. (openai.com) Anthropic is making the same turn from selling model access to selling the plumbing around the model. This week it opened Claude Managed Agents in public beta, a service that handles the hard operational pieces like sandboxing, state management, credential handling, and tool execution for companies that want agents running in production. (anthropic.com) Anthropic is also adding the manager controls that information-technology departments usually demand before a rollout goes company-wide. On April 9, it said Claude Cowork now has role-based access controls, group spend limits, OpenTelemetry monitoring, and admin analytics for adoption tracking. (claude.com) The phrase both companies keep circling is “agents,” which means software that does multi-step work instead of answering one prompt at a time. OpenAI says customers are tired of isolated artificial intelligence tools that “don’t talk to each other,” and want one operating layer that can connect agents to company context, internal systems, and permissions. (openai.com) That is why “multi-agent workflows” matter to these vendors. Instead of one model writing one reply, a company can have one agent pull data, another check policy, another draft an action, and a final one hand the result to a human or another system. (anthropic.com, openai.com) Once the sale moves from a chatbot seat to a workflow that touches payroll, procurement, customer support, or software deployment, the bottleneck becomes infrastructure. In a memo reported by CNBC on April 9, OpenAI told investors it expects to have 30 gigawatts of compute by 2030, versus roughly 7 to 8 gigawatts for Anthropic by the end of 2027, and argued that bigger infrastructure lets it train stronger models while lowering serving costs. (cnbc.com) OpenAI is pairing that infrastructure argument with scale numbers meant to reassure big buyers. Denise Dresser said Codex has reached 3 million weekly active users, OpenAI’s application programming interfaces process more than 15 billion tokens per minute, and customers now include Goldman Sachs, State Farm, DoorDash, and Thermo Fisher. (openai.com) Anthropic is answering from the product side instead of the capacity side. Its managed-agents launch is built around the idea that most companies do not want to assemble their own orchestration stack, security wrapper, and runtime every time they try to automate a research task, coding job, or internal process. (anthropic.com) So the fight is getting less consumer and more corporate by the week. The winner may not be the company with the cleverest demo, but the one that can give a Fortune 500 technology team a working agent system with permissions, monitoring, and enough compute behind it to run every day without breaking. (openai.com, claude.com, cnbc.com)