The run of briefs this week tracked agent authorization as it descended: from UI, to policy layer, to token, to a line item procurement teams price. Today it keeps falling. NVIDIA's OpenShell Reference Architecture puts security enforcement inside the inference and networking stack, with Sentry running on BlueField-4 DPUs. The yesterday's argument said agent identity is a procurement line item. Today's says the control can be hardware-enforced, and banks including BBVA are already specifying it.
That is the move since the 2026-11-12 brief. For the Identity Control Surface, non-human identity governance stops being a software policy that a clever prompt or a stolen token can route around. When the DPU enforces, the assurance argument for regulated industries changes shape. This is why the IBM and Microsoft Imagine Enterprise launch lands the same week: 6,000 prebuilt agents with governance and observability built in from the outset. Two of the largest infrastructure coalitions in the market are converging on the same thesis, that regulated buyers will not source agents until the control plane ships with them.
Hold that against the capability signals, because the contrast is the point. Amplitude's Wave compressed the product decision loop to one to two days, with the human moved to intent-setting. Google's Gemini Decisions API exposes calibrated decisioning as a formal interface. The capability to let an agent decide and act end-to-end is in production now. The enforcement to make that safe in a bank is arriving at the hardware layer this same week. Enterprises that bought the capability story in 2026 are now being handed the enforcement story, and the two were engineered by different vendors with different incentives.
The gap that remains is credential lifecycle. Agentic payments infrastructure across Google, Mastercard and Visa now lets an agent initiate a financial transaction, but neither Google's protocol nor Mastercard's Agent Pay has published how the delegation credential binding a human principal to an agent action is issued, rotated or revoked. Hardware enforcement is only as good as the identity it enforces against. A DPU that trusts a credential nobody can revoke is a faster theater.
The practical move for principals this morning: when you evaluate agentic infrastructure, treat it as a reference architecture decision, and make credential lifecycle an explicit purchasing criterion, not a feature you discover after signing. Ask where the delegation credential lives and who can revoke it in under a minute.
Watch item: Nscale's regulatory filing says it must raise additional financing to fulfill $45B in Anthropic contracts, casting doubt on its $35B IPO. If that financing does not close, Anthropic loses inference capacity at scale. Watch Anthropic's response and whether rival GPU clouds move to absorb the gap.
Index Reference · Applied AI Index 2026-W43, the latest published issue (Sunday, October 25, 2026)
WatchNscale's regulatory filing states it must obtain additional financing to fulfill $45B in Anthropic contracts after committing to Nvidia chip orders, casting doubt on its planned $35B IPO. If Nscale cannot close that financing, Anthropic loses a significant inference capacity commitment at scale. Watch for Anthropic's response and for whether alternative GPU cloud providers move to absorb the capacity gap.