Mark’s Musings

CFOs already understand concentration risk. We just usually apply it to banks, customers, and supply chains — not to the intelligence layer now writing board packs, cash forecasts, and diligence notes.

If half your finance workflow depends on a model you do not host, do not control, and cannot audit end to end, you have built a single point of failure into the operating system of the business. That is not an IT preference. It is a governance decision. And boards should treat it as one.

The Dependency Problem

Cloud AI is extraordinary. It is also leased. You rent capability by the token, subject to vendor policy, pricing power, outages, and political weather.

In June 2026 the US Commerce Department put export controls on Anthropic’s frontier models, and access was switched off globally overnight while the company worked out how to comply. The controls were later lifted after new safeguards. The point is not the politics. The point is the switch. One policy decision, and a production capability disappeared.

A fortnight later, OpenAI previewed GPT-5.6 to a limited set of “trusted partners” first, explicitly at government request under a voluntary pre-release review framework. OpenAI said it did not want that model of access to become the long-term default. Fair enough. But enterprises now have a live example of frontier intelligence arriving through a gate, not a pipe.

Then the market did what markets do: it routed around the constraint. Orchestration layers and open-weight alternatives appeared fast — Sakana’s Fugu among them — because capability that can be withheld will always attract substitutes.

None of this requires a conspiracy theory. It requires a CFO’s instinct: if a critical input can be gated, censored, repriced, or reversed by someone outside your control, you should not build the entire house on it.

Cost, Latency, Confidentiality

Run the ledger properly.

Cost first. Public API pricing looks cheap until usage compounds. Board packs, monthly closes, covenant models, contract review, management accounts commentary, buyer Q&A — token volume scales with ambition. Cloud spend is opex with a vendor’s hand on the dial. Local inference has capex and energy cost, but the marginal cost of the next confidential memo is near zero once the box is paid for. That changes the unit economics of “use AI everywhere.”

Latency second. Interactive finance work hates round trips. Cash models, scenario trees, and live diligence chats feel different when the model sits on your network rather than three jurisdictions away. Speed is not vanity. It is whether people actually use the tool under pressure.

Confidentiality third — and this is the one boards understand immediately. Sending draft SPAs, customer concentration analyses, working capital bridges, or management presentations to someone else’s model is a data-processing decision. Even with enterprise contracts, retention policies, and “we don’t train on your data” promises, you have expanded the attack surface and the counterparty list. For sensitive M&A work and PE portfolio reporting, that is not a technical footnote. It is risk acceptance.

Encryption, access control, and data residency are not lifestyle choices for finance teams. They are controls. Local or private deployment puts those controls back under your policy, not a vendor’s product roadmap.

What “Local” Actually Means

Local does not mean a dusty server under the FD’s desk and a vow of technological poverty.

It means a spectrum:

• On-prem or colo hardware running open-weight models for high-sensitivity workloads.

• Private cloud tenancy where you control the network boundary and keys.

• Hybrid: public frontier models for low-sensitivity drafting; local models for cash, people, contracts, and deal data.

• Open weights where the model parameters are yours to run, pin, and version — not a black-box endpoint that can change behaviour between Mondays.

The mature pattern is the same one we used for banking systems and ERP: classify the data, then choose the environment. Public web search and generic writing can stay in the cloud. Anything that would hurt if it leaked — or freeze the close if it vanished — should have a home you control.

You do not need the absolute best model for every task. You need a good enough model that is available, private, and accountable when the board pack is due at 7am.

The PE Angle

Private equity should care more than most.

Portfolio companies are already wiring AI into forecast packs, pricing tools, collections, and customer support. That creates three diligence questions buyers will eventually ask:

Where does the intelligence run? What happens if the vendor changes terms, price, or access? How much of the “AI-enabled” value creation is transferable at exit?

Vendor lock-in used to mean ERP and CRM. It now includes model dependency. If a company’s operating edge is a prompt library glued to one closed API, the exit story is thinner than it looks. If the same company can run core finance workflows on owned infrastructure with portable open weights, the capability survives a change of control.

There is also a portfolio resilience angle. One policy shock or regional outage should not simultaneously degrade reporting quality across twelve companies because they all rented the same brain. Concentration risk is concentration risk, whether the asset is a bank facility or an inference endpoint.

For GPs and operating partners, local AI is not a gadget budget line. It is part of exit readiness, cyber diligence, and operational independence.

Practical Starting Steps for a Finance Leader

Skip the manifesto. Do this:

1. Map the workflows, not the hype.

List where AI already touches finance: board packs, commentary, covenant testing, invoice capture, contract review, data-room Q&A, FP&A scenarios. Rank each by confidentiality and operational criticality.

2. Draw a hard line.

Anything involving deal data, payroll, customer-level margin, unpublished results, or lender packs defaults to private or local processing unless there is a documented exception.

3. Pilot one high-value, high-sensitivity use case.

Local review of contracts. Private drafting of management accounts narrative from internal numbers. On-network Q&A over a diligence folder. Prove cycle-time and control benefits on something the board cares about.

4. Measure like a CFO.

Track cost per pack, hours saved, rework rate, incidents, and whether the process still works when the public API is slow or unavailable. If it only works on a perfect internet day, it is not production.

5. Separate “assistant” from “system of record.”

Models draft. Ledgers, workpapers, and approvals remain controlled systems with audit trails. Do not confuse fluency with authority.

6. Demand architecture options from vendors and internal IT.

“We use ChatGPT” is not a strategy. Ask for data flow diagrams, retention, key custody, fallback models, and an exit plan. If nobody can answer, you already know the risk posture.

Control the Controllables

You cannot control Washington’s export calendar, a lab’s safety incident, a vendor’s price sheet, or the next voluntary “trusted partner” gate. You can control where your sensitive numbers go, which systems are load-bearing, and how badly a third-party outage damages the close.

Cloud AI will remain useful. Frontier models will remain impressive. Use them where the data is dull and the upside is speed. For the work that defines enterprise value — forecasts, cash, contracts, diligence, board decisions — ownership of the intelligence layer is becoming as important as ownership of the general ledger.

The finance function’s job has always been to keep the business solvent, informed, and free to act. Depending entirely on rented cognition works against all three.

Local AI is not nostalgia for servers. It is the boring, adult version of resilience: keep the critical path under your own keys.

Four days ago I wrote that the Fable 5 reversal wasn’t the story — the switch was. One signature darkened the world’s best model for nineteen days; another switched it back on, but only after the market had already routed around the damage. I flagged a second data point in that piece almost in passing: OpenAI had released GPT-5.6 as a government-gated preview, available to “trusted partners” only, at the explicit request of Washington. I called it the new shape of the pipeline. This week, that loop closed too.

The Gate Came Down on Schedule

OpenAI has confirmed that GPT-5.6 — Sol, Terra and Luna — launches publicly this Thursday, 9 July, with global preview access expanding now. The announcement lands roughly two weeks after the preview was first locked to a hand-picked list of partners under the government’s pre-release “voluntary framework.” The most capable model OpenAI has shipped goes from Washington-approved keyholders to the open market in the space of a fortnight.

If that arc feels familiar, it should. It is the Fable 5 sequence again, run a second time in a single month: access restricted for policy reasons, then quietly widened once the restriction stopped being tenable. Two frontier labs, two gates, two reopenings — inside four weeks.

Two Reversals Are Not Reassurance

The comfortable reading is that the system is self-correcting. Models get gated, concerns get aired, models get released — no harm done. But a CFO doesn’t underwrite continuity on the assumption that the gate always reopens. He asks who controls the gate, and on what timeline.

Here’s what both episodes actually establish: the frontier now routinely passes through a government checkpoint before it reaches you, and the duration of that checkpoint is entirely discretionary. Fable 5 spent nineteen days behind it. GPT-5.6 spent about fourteen. The next model might spend a day, or a quarter, or arrive gated the week you’re mid-integration and depending on it. You don’t get to know in advance, because the people setting the timer owe you nothing.

Two fast reversals don’t prove the gate is harmless. They prove the gate is now standard — and a mechanism that reopens quickly today can hold shut tomorrow for exactly the same reasons it opened.

The Pen-Stroke Test, Reapplied

I ended the last piece with a question for any board treating AI as core infrastructure: which parts of your operation survive a pen-stroke? GPT-5.6 sharpens it. This isn’t a rogue export ban from one agency anymore — it’s the release process itself, at the largest AI company in the world, running through a state framework as a matter of course. The checkpoint isn’t the exception. It’s the pipeline.

Which changes the risk calculus. When gating was an incident, you could treat it as tail risk. Now that it’s the default posture for frontier releases, it’s a structural feature of every model you rent through an API in that jurisdiction. Your access is conditional by design, not by accident.

The Answer Hasn’t Changed — It’s Just More Obvious

Nothing about the response shifts. If anything, the second reversal makes the case duller and more certain. Orchestration tools like Sakana’s Fugu that let you fail over between providers aren’t a performance play — they’re continuity planning for exactly this environment. And local open weights on your own hardware remain the only layer no framework, no executive order and no terms-of-service update can reach into and switch off. The lifeboat doesn’t need to be the fastest boat. It needs to be the one that’s still yours when the gate closes.

Use GPT-5.6 on Thursday. It’ll be excellent, and I’ll be using it too. But watch the mechanism, not the model. We’ve now seen the frontier gated and reopened twice in a month. The switch works in both directions, and you don’t hold it.

Own your intelligence, or accept that you’re renting it on someone else’s terms.

In January, the government announced it had “abandoned” plans for mandatory digital ID. Headlines celebrated. Campaigners claimed victory. And the machine kept building.

Here’s what actually happened: the compulsion didn’t disappear. It moved. You won’t be required to carry a digital ID — but by the end of this Parliament, every employer in the country will be legally required to run digital right-to-work checks. You’re free to refuse the ID. You’re just not free to earn a living without it touching you.

That’s not a U-turn. That’s a redesign.

Voluntary Is How Mandatory Arrives

No government in modern Britain will ever pass a law saying “citizens must carry identity papers.” We fought that battle over the 2006 ID card scheme and won it. The lesson Whitehall learned wasn’t “don’t do this” — it was “don’t do it like that.”

So the new model is voluntary. Free to download. Stored on your phone, “secure as a banking app.” The official explainer is a masterpiece of reassurance: no central database, you control your data, alternatives for people without smartphones.

And then, quietly, the perimeter closes:

Want a job? Your employer must verify you digitally. Want a pint? From autumn 2026, digital ID becomes valid age verification for alcohol in England and Wales — optional today, default tomorrow, as retailers standardise on the cheapest compliance path. Benefits, childcare, banking, age-gated websites: each sector gets its own “convenient” integration via the Digital Access to Services Bill now moving through Parliament.

Nobody mandates the frog into the pot. You just make the water comfortable and let network effects do the rest. When every checkout, landlord, employer and bank asks for the same credential, “voluntary” is a word that describes the law, not your life.

The Threat Isn’t the Card. It’s the Chokepoint.

Be precise about the danger, because it isn’t the technology. Cryptographic identity done properly — keys you hold, selective disclosure, no phone-home — is genuinely useful. That’s not what’s being built.

What’s being built is a chokepoint: a single credential that mediates your access to work, money, services and age-restricted life, operated under government-defined rules that can change with a statutory instrument. The threat model isn’t today’s minister. It’s the permanent capability handed to every future one.

We’ve just watched this movie in another theatre. In June, a US export order switched off two frontier AI models globally overnight — and switched them back on eighteen days later. One signature each way. The lesson wasn’t about AI. It was that centralised access is a switch, and someone else’s hand is on it.

A digital ID chokepoint is the same switch, wired to your identity. Once every employer verification, every purchase check, every service login routes through one credential, the infrastructure for conditional citizenship exists — regardless of whether anyone currently intends to use it. “We would never” is not an architecture. It’s a mood.

And the mood changes. Ask anyone whose bank account was closed for their politics, or who watched Canadian trucker-protest donors get frozen out of their own money in 2022. The tools get used because they’re there.

What “No Central Database” Actually Means

The government’s flagship privacy promise deserves scrutiny. “No centralised database of personal information” sounds decisive. But the ID scheme sits alongside GOV.UK One Login — a single sign-on across government services — and a planned UK Wallet for official documents. You don’t need one big database to build a surveillance capability. You need linkable identifiers and logging at the verification layer. Every time the credential is checked, somewhere a record can exist: who, where, when, for what.

Distributed storage with centralised observability is not privacy. It’s a database with better PR.

How to Resist — Practically, Legally, Now

Resistance here isn’t dramatic. It’s a set of unglamorous habits that keep the analogue paths alive and the pressure on. The paths only stay open while people use them.

1. Use cash, deliberately. Every cash transaction is a vote for an economy that doesn’t require identity to function. Cash is legal, private by design, and the single most effective everyday act against transactional surveillance.

2. Keep and use physical documents. Passport, driving licence, paper records. From autumn, when a checkout offers digital age verification, hand over the physical card instead. Friction is the point — acceptance rates are the metric that decides whether alternatives survive.

3. Respond to the consultations and the Bill. The Commons Library briefing is the best neutral summary of where the legislation stands. Write to your MP about the Digital Access to Services Bill — specifically demanding statutory guarantees: no verification logging, true offline alternatives with equal legal standing, and a prohibition on private-sector demands for the ID where physical documents suffice. The January climbdown proved pressure works.

4. Support the organisations doing the heavy lifting. Big Brother Watch and the Open Rights Group have fought this fight since the 2006 scheme. They killed mandatory ID once. Fund them.

5. Master the tools of self-sovereign identity. Encryption, keys you control, money you custody. The skills compound. A population fluent in cryptographic self-custody is structurally harder to herd through a single government credential — and if the state ever offers genuine self-sovereign ID (keys on your device, zero-knowledge age proofs, no logging), the people who understand the difference will be the ones who can tell.

6. As an employer or director, choose maximum-privacy compliance. Those of us who run companies will be conscripted as enforcement points. Comply with the law — and implement it with minimum data retention, no gratuitous identity harvesting, and documented pushback through trade bodies. Conscripts can still drag their feet.

The Line Worth Holding

I’m not against digital identity. I’m against this shape of it: state-defined, employer-enforced, scope-creeping, observable at the point of verification, and sold as voluntary while the exits are bricked up one by one.

Rights that depend on infrastructure are only as durable as the infrastructure’s owner is benevolent. The British instinct — the one that killed ID cards — was never anti-technology. It was the older, sounder instinct that the state serves the citizen, and a citizen who must be verified to work, buy and exist has quietly become the servant.

The water is warming slowly, and comfortably, exactly as designed. Get out of the pot.

On 13 June 2026, the United States government switched off the most capable AI model on the planet with a single signature. On 1 July, another signature switched it back on. Nineteen days. That’s how long the world’s best publicly available intelligence spent in the dark — and if you think the story here is that it came back, you’ve missed the point entirely.

I wrote about the shutdown the weekend it happened. The US Department of Commerce issued an export order barring foreign nationals from accessing Anthropic’s Claude Fable 5 and Mythos 5 — the first export control ever applied to AI models rather than the chips they run on. Anthropic, unable to reliably geofence a global product, pulled both models worldwide overnight. One day the frontier existed; the next it didn’t.

The Reversal Nobody Should Celebrate

This week the arc completed. Anthropic confirmed that Commerce has lifted the export controls, and Fable 5 returned online globally on 1 July — now fitted with a new safety classifier to block jailbreak techniques, a condition of its resurrection. Access restored, service resumed, everyone back to work.

The temptation is to read this as the system working. Concerns raised, concerns addressed, model restored. But look at the sequence with a CFO’s eye for causality. The controls weren’t lifted into a vacuum. They were lifted after Sakana AI shipped Fugu, a commercial-grade answer to the Fable ban, benchmarking above the models Washington was still gatekeeping (the code is on GitHub, which is rather the point). They were lifted after roughly $2.87 billion flowed into decentralised AI — Bittensor, Venice, Morpheus — in the weeks following the shutdown. The state didn’t reconsider. The market routed around the damage, and then the state walked it back.

That’s not oversight functioning. That’s a control mechanism discovering its own limits.

Being the Best Is Not the Same as Being Unstoppable

Fable 5 was, by most measures, the strongest model money could rent. It didn’t matter. Its existence depended on a policy posture, and policy postures change with administrations, with headlines, with whichever adviser had the last meeting. The model you build your workflows on is only as durable as the political consensus that permits it.

And this isn’t a one-off anymore. The same week Fable came back, we learned that OpenAI released GPT-5.6 as a limited preview to “trusted partners” only — at the explicit request of the US government, under a new executive order creating a voluntary framework to review frontier models before public release. OpenAI said publicly that government-gated access “is not their preferred long-term model.” It doesn’t need to be preferred. It’s the new shape of the pipeline: the frontier now passes through Washington before it reaches you. A voluntary standards framework is expected to be formalised within days.

One export ban is an incident. A pre-release government gate plus a ban plus a conditional reinstatement is an architecture.

The Cypherpunks Called This Decades Ago

Phil Zimmermann was criminally investigated in the 1990s for publishing PGP — strong encryption classified, then, as a munition. His response wasn’t to lobby. He published the source code as a printed book, protected by the First Amendment, and the control regime collapsed under its own absurdity. The lesson wasn’t that governments are evil. It’s that anything you can only access by permission can be revoked by the same permission — and the only durable answer is possession.

The same logic now applies to intelligence itself. The 19-day Fable blackout was a live demonstration: your API key is a licence, not a right. It can be suspended by someone you’ve never met, for reasons you can’t appeal, on a timeline you don’t control.

What a Rational Operator Does Now

I run the finance function thesis on this, because that’s my lens. You wouldn’t build a treasury function on a single bank account in a jurisdiction with capital controls. So why build your firm’s intelligence layer on a single frontier API in a jurisdiction that has now demonstrated — twice in three weeks — that it will gate access to models for policy reasons?

The posture that survives this environment has two layers:

Orchestration as failover. Tools like Fugu exist precisely because the market demanded a way to keep working when a frontier model vanishes. Multi-model routing isn’t an optimisation anymore; it’s continuity planning. If your workflows die when one provider goes dark, you don’t have an AI strategy — you have a dependency.

Local open weights as the lifeboat. A model whose weights sit on your own hardware cannot be switched off by the Department of Commerce, an executive order, or a terms-of-service update. It may not be the best model. It doesn’t need to be. It needs to be yours. Lifeboats aren’t judged on cruising speed.

The Pen-Stroke Test

Here’s the question I’d put to any board now treating AI as core infrastructure: which parts of your operation survive a pen-stroke? Because we now know, empirically, what a pen-stroke can do. It darkened the best mind on earth for nineteen days. The next one might last longer, or target a different model, or arrive the week you close a deal.

Fable 5 is back. Be glad, use it — I do. But don’t mistake reinstatement for reliability. The switch still exists. It has now been flipped in both directions, and both flips were made by people who owe you nothing.

Own your intelligence, or accept that you’re borrowing it.