The Board Room

The reversal is not the detail worth tracking. Security review, used as a release gate, is survivable on its own. The part worth tracking is that a well-connected competitor can now trigger that gate against a rival's roadmap, and by these accounts it has already happened twice.

The reported pattern runs like this: Andy Jassy, Amazon's CEO, is said to have raised concerns about vulnerabilities in Anthropic's models directly with senior officials. Newsletter accounts credit that intervention with triggering a crackdown that became the industry's template. The same template was reportedly then run on OpenAI, which shipped GPT-5.6 in a staggered release said to come at the administration's request rather than on its own schedule. Competitor lobbying now functions as a regulatory attack vector, timed to a rival's launch window, alongside the older competition over product and talent.

Two forces are converging on the same outcome from different directions. Federally, an administration that promised in January 2025 to dismantle Biden-era rules was, eighteen months later, reportedly gating flagship models, with a proposed FINRA-style body positioned to make that gating permanent. Locally, 300-plus moratoriums from cities, counties and states already govern where compute gets sited, driven by energy prices and jobs anxiety that no federal preemption can override. The result is a compute siting map that federal policy alone cannot simplify.

There is a counter-position inside this threat. The same fear driving regulators, that capability itself is a cybersecurity liability, is the same fear driving buyers in regulated industries. A secure-by-design narrative now sells where a capability-only pitch invites scrutiny. The caveat: an administration that reversed once can reverse again. The sound response is optionality, not a bet on political constancy.

The sharpest evidence was a teardown, not a launch. A hands-on rebuild showed Claude Code's edge over a naive agent loop isn't the model — it's the harness: planning, memory, sandboxing, orchestration. The same rebuild reconstructed most of that harness on CrewAI, an open-source framework running identically across Anthropic, OpenAI and Google backends, and took a broken test suite to fully passing.

That tells you where value is — and isn't — pooling. Raw model capability has commoditized and the harness plumbing is now free. So the value shed by both layers lands on the two places no framework automates: the orchestration layer, which scores weakest on enterprise-readiness, and the irreducible engineering of prompts, tool selection and execution environment.

Three independent sources converge on the same map. The commoditization thesis, the harness teardown, and the reliability field all point to the layer above the model as the contested ground. The proof is a persistent use-to-ship gap that widens to 57% versus 73% at enterprise scale. That gap is the difficulty — and the margin.

The sleeper risk sits in the same layer. Agents act on operational metadata at machine speed, so a stale service catalog flips from a productivity tax into a production risk. With the agent-CVE count climbing far faster than mature governance spreads, the agentic push carries an unpriced liability. Sandboxing and approval gates are baseline controls, not features — prompt restrictions are not a control.

One caveat: parts of today's harness are transient. Anthropic dropped context resets once its model was capable enough — scaffolding is a depreciating asset, so build it modular and expect to retire it each model generation.

Start with durability, because it decides what's worth buying. Where feature velocity is automatable, the appreciating moats are physical, regulatory and data-network — Copart owning its salvage yards behind zoning and permit barriers is the benchmark, not any software feature set. Everything an LLM can replicate is depreciating; everything it can't is the acquisition target.

Against that backdrop, the timing signal is loud. Adyen — a company that made building everything in-house part of its identity — completed its first-ever acquisitions, buying Orb (billing) and Talon.One (loyalty) to become a full commerce platform. A disciplined organic-growth culture doesn't break its own religion for marginal gain; it does so when the window to own an integrated stack is closing faster than any roadmap can build into it.

The supply side is a buyer's market. Roughly 160 enterprise software startups are reportedly for sale, OpenRouter is fielding multibillion-dollar interest, and an AI-disruption discount is depressing software valuations even where the threat hasn't materialized. Paychex's $4.1B Paycor deal shows the intent: not a feature patch, but a move up-market to raise switching costs and pricing power. M&A as moat-deepening.

The two forces meet in one conclusion: capability is available at a discount precisely because the market fears AI, while demand and infrastructure underneath remain intact. Disciplined acquirers with a ready target list convert that fear into position. The discipline that separates winners is corp-dev readiness, not deal appetite — watch whether Adyen absorbs its first acquisitions cleanly, because integration is where value leaks.

Caveat: much of the M&A signal comes from adjacent headlines; treat specific valuations as directional, not confirmed.