The Board Room

The inventory is the hard part

Microsoft owns the migration. The inventory belongs to whoever ships the software, and it runs wider than the release train: installers, kernel drivers, endpoint agents, auto-updaters, build-time signing steps, and the signed third-party components a vendor redistributes without having written. Auto-updaters go first, because a client that cannot validate a new signature cannot deliver its own fix. After that point remediation takes a human touch per install, which makes it a support-cost problem rather than an engineering one.

Risky Business's reporting leaves no room on the developer warning. Microsoft has said apps will break, and has put completion at mid-October. That converts a cryptography roadmap slide into a release-calendar dependency with no negotiating counterparty. There is no extension to request, and by default nobody on the software vendor's side owns the date.

Why finding it late costs roughly ten times more

The same fix carries two prices. Found now, it is an inventory exercise and a test matrix. Found in a customer escalation, it becomes an emergency release, a support surge, a trust conversation with the largest accounts, and the opportunity cost of engineers who were supposed to be shipping Q4 features, all of it landing during quarter close. An internal cutover of October 1 buys two weeks of buffer against a date held by someone else.

Turn the fire drill into a capability

This is one in a series of forced cryptographic migrations, not a one-time event. Treated as an emergency each time, it carries the emergency premium annually. Funded as a named platform capability, crypto-agility gets paid for once: centralized signing, pluggable primitives, no hardcoded algorithm assumptions, and an enterprise security questionnaire that answers cleanly. That is the part with revenue attached: within three years, crypto-agility is a procurement question in regulated and enterprise deals, not platform hygiene competing for sprint capacity.

The concentration quantified

The same body of reporting covers a 10.0-severity remote code execution flaw in Entra ID (CVE-2026-69836), a one-click Copilot data-exfiltration bug (CoSnitch, CVE-2026-24301), an advisory published with an exploitation flag enabled by mistake, and this breaking change to the distribution trust those products depend on. All patched, and individually routine. Collectively they put a number on something most technology companies have never actually modeled: identity, agentic data access, and software distribution trust in one supplier.

A reasonable skeptic says Microsoft remains the safest place to hold those dependencies, and the skeptic is probably right. The question was never whether Microsoft is secure. It is the size of the blast radius, and whether detection survives a vendor advisory being wrong, which already happened. The detail worth sitting with is that the Copilot bug was surfaced by Copilot itself during normal use. AI assistants create self-revealing data paths existing monitoring was never designed to observe, which means the vendor's telemetry and a customer's own are not interchangeable.

Post-quantum cryptography stopped being a 2030 roadmap item the moment Microsoft put a mid-October date on it.

The board-reportable version is two lines: the date signed artifacts are verified against the new chain, and the detection held in-house that does not depend on a Microsoft advisory being accurate. Neither line requires a strategy change. Both require an owner.

The sequence carries the argument

Commercial ties were strengthened first. The equity talks came second. A financial investor does not order it that way; a supplier converting a commercial relationship into ownership of demand for its own hardware does. That ordering is what makes this a template rather than an event. If compute supply agreements convert into application-layer equity once, they convert again, and every AI company's cap table becomes a disclosure about who holds preferential access to capacity and price.

Qualify it properly before anyone builds a slide. The Information describes a round under discussion, not closed, sourced to people with knowledge of the talks, and the roughly 40x figure is measured against annualized revenue rather than an audited annual result.

What the multiple is actually paying for

Perplexity's annualized revenue went from under $250M in January 2026 to over $750M by August, credited in part to Perplexity Computer, an agent professionals use to complete work on their own machines. That is an architecture change wearing a product name. Retrieval and synthesis produce an answer engine and a consumer subscription. Task completion produces a professional tool with budget authority behind it. The market paid roughly 40x for the second thing while repricing the first downward. A 2027 roadmap that still sells assistance rather than completion is carrying the wrong multiple into its next financing or its next renewal cycle.

The channel gets conflicted on the same timeline

Anthropic's enterprise arm bought a consultancy while the company prepares supervoting founder shares ahead of a mega-IPO, per The Information. A model provider with delivery capability competes with the systems integrators it depends on, and eventually with any differentiation that amounts to knowing how to deploy the model. Neutrality is a wasting asset. The window to sign integrator alliances as the credible non-competing partner is measured in quarters, and it is the only item here that moves without another party's cooperation.

The accounting correction nobody wants to run

Both reports flag the same soft input: Nvidia's mixed messaging on chip longevity. Useful life drives depreciation, refresh cadence, and the AI gross margin already sitting in the board deck. Ambiguity from the supplier means that number is softer than the model implies. A reasonable skeptic would say this is a footnote adjustment, not a strategy question, and for one quarter that is fair. Run the sensitivity at a 30–40% shorter competitive life anyway; if it moves the margin story materially, the board should hear it from management rather than from an analyst. It also hands transparent-TCO competitors an opening.

One timing note sets the working assumption. Hyperscalers respond to accelerator price increases by accelerating custom silicon, and that takes 18–36 months. Most planning windows sit inside that gap.

When the chip supplier starts buying equity in the application layer, compute neutrality stops being a procurement detail and becomes a budget line against 2027 inference demand.

Three absences, not one algorithm

The ruling did not penalize automation. It penalized automation arriving with three absences: no human in the loop, no reason given, and no route to contest. Uber's automated driver-account deactivations supplied all three at once. Written as a checklist, that reads less like a legal finding than a product specification: a named human reviewer, a rationale stored where it can be retrieved, and an appeal path the subject can actually reach. The fine is the price of shipping without them.

The scope inside a given product is wider than legal's list. Suspensions, bans, denials, payout holds, deranking, fraud declines, risk scores, moderation actions, and any AI-driven enforcement still on the roadmap all qualify. The test survives a review meeting: a model decides against a person, and the question is whether that person can see why and whether a human can overturn it. Where the answer is no both times, the exposure is already priced.

The liability travels with the asset

Two readings of the same $400M privacy settlement, and both are worth keeping. Risky Business notes that TikTok's payment partly covered infractions inherited in a 2017 acquisition. The Information treats the figure as the working enforcement benchmark for how AI assistants reach personal data. The diligence conclusion is the same either way: legacy privacy and automated-decisioning liability belong on the M&A scorecard in the same pass as security debt, because the acquirer buys the enforcement exposure along with the code.

This is the cheapest control on the page. Two more questions on a diligence template cost nothing. Finding a decade-old automated-decisioning practice after close costs whatever the regulator decides it costs.

The surface is growing, not shrinking

AI assistants are heading toward operating-system-level access to personal data, per The Information. That means more automated decisions, on more sensitive data, with less explainability per decision. Every assistant feature that acts on a user's own files and accounts widens the set of consequential actions nobody can reconstruct afterward. Where the roadmap includes agentic actions on customer data, the audit is a prerequisite rather than a follow-on.

Why this shows up in win rates

Finance prices this as compliance cost, and finance is usually right about compliance spending. The competitive read is the more useful one here. Enterprise buyers probe human-in-the-loop design and permission scoping during security and procurement review, and they ask where the decision rationale is stored; almost nobody answers with artifacts rather than assurances. A capability sheet built before the market demands one turns a legal expense into a differentiator in the buying committee. Privacy posture works as a go-to-market wedge precisely because most competitors will wait for an incident.

The order of operations is unremarkable. The audit of what already ships comes before the gate on what ships next, and both come before the diligence template stops importing other companies' exposure. None of it is a strategy change. All of it needs someone senior owning the inventory of automated adverse actions, which almost certainly does not exist in one place today.

An automated decision your customer cannot see the reason for has a published price, and it was set in euros.