Agents can now transact | Tuesday, August 18, 2026

Anthropic's annualized run rate reached about $65 billion, up from roughly $47 billion in May and $9 billion at the end of 2025, putting it ahead of OpenAI by that measure, with a fall IPO reportedly targeting a valuation near $2 trillion. It is a real and astonishing number. The read for a buyer is not admiration but leverage. A public vendor answers to a share price every ninety days, which makes discounting harder and list prices stickier once there is guidance to defend. Lock the terms that matter, price protection, portability, an exit that does not require a migration, while the vendor still wants your logo for the roadshow.

OpenAI launched a version of ChatGPT for users aged 13 to 17 that blocks conversations about suicide, self-harm, and romantic or sexual content, and routes younger accounts into it by default. It answers mounting scrutiny of how chatbots handle minors, and sets a template competitors will be pressed to match. For any enterprise running a customer-facing assistant, age-gating and content guardrails are moving from optional to expected, and the burden of proving them lands on you, not the model vendor. Build those controls into your own layer now, because the regulatory floor is rising toward them.

Oracle is planning another round of layoffs this month, asking managers to name affected staff ahead of its September fiscal quarter, even as it borrows tens of billions to fund an AI data-center buildout. Its workforce already shrank by about 21,000, roughly 13%, over the year to May. This is the pattern under the "AI efficiency" headline: capital redirected from payroll to compute, with AI as the story told about a partly balance-sheet decision. Trackers now flag "AI washing," companies citing AI for cuts that overhiring or weak revenue actually drove. When a vendor blames layoffs on AI, read the cash-flow statement before the press release.

The money is moving down the stack. Fireworks AI closed a $1.505 billion Series D at a $17.5 billion valuation as it crossed $1 billion in annualized revenue, and Together AI raised $800 million at an $8.3 billion valuation on roughly $1.15 billion in bookings. Both sell inference: serving and tuning models cheaply at scale, the layer now estimated at 65% to 70% of global compute demand. Training grabs the headlines; inference is where enterprise AI spend actually lands, month after month, per call. A competitive market is forming below the labs, and running workloads on a specialized inference provider rather than a frontier vendor's default endpoint is becoming a real cost lever.

Nvidia's enterprise Agent Toolkit signed 17 partners, including Adobe, Salesforce, and SAP, a push to become the default platform for building and running enterprise agents. It pairs with an expanded ServiceNow partnership and Project Arc, a long-running self-evolving desktop agent. The strategic move is familiar: having sold the picks and shovels, Nvidia wants the layer above them too. For a CIO, the caution is lock-in wearing the costume of convenience. A toolkit that spans your vendors is genuinely useful, but standardizing your agent orchestration on the company that also sells you the GPUs concentrates two dependencies into one. Adopt the convenience; keep the orchestration layer portable.