Rent The Model, Own The Stack

I round up the most relevant AI-in-finance news - the deals being done, who’s rolling out what, and what’s actually working on the front lines

… on the market this week. Commerce ordered Anthropic to bar all foreign access to Fable 5 and Mythos 5 on national-security grounds, and Anthropic disabled both models globally overnight. The most powerful models in the world went dark for most of the planet, by government letter rather than commercial choice.

The money is moving even faster than the policy. Apollo led a roughly $35bn debt deal for Anthropic's compute, part of something like $255bn Wall Street has poured into the big hyperscalers this year. Private credit, where a lot of your LPs sit, is now underwriting the buildout.

Elsewhere: Bezos' Prometheus hits $41bn, Mistral is raising at €20bn and Thoma Bravo called the end of the "SaaSpocalypse,".

But first, my take on what Friday night should teach anyone running their firm on someone else's model, and where the value is actually going to accrue instead.

In This Week’s Issue:

From The Trenches

The Model Is Not Yours

On Friday night, Anthropic disabled Fable 5 and Mythos 5, its two most capable models, for most of the world. A letter from the US government, and the best models on the market went dark overnight for anyone outside the country. Nobody got a price change or a deprecation notice. They just stopped working. (See below for the full story in the News Digest.)

The model you run on isn't yours. A regulator can pull it, a deprecation can break it, a governance meltdown can hold it hostage, and none of that is in your control.

A post from Satya Nadella on X over the weekend kicked off a fair bit of debate, and it points straight at the deeper version of all this. He'd know, since he sells a good chunk of your access to these models. The model, he argues, isn't where the value is. "The real opportunity is not in picking the best model, but instead in building a learning loop on top of models where human capital and token capital compound." And the line for deal people: "You can offload a task, or even a job, but you can never offload your learning."

Running two models instead of one is a hedge, and a hedge is about as far as most firms get. What the labs are counting on you to miss is the deeper move: two of Nadella's largest competitors have built a business around making sure you never own the loop that would let you switch at all.

Anthropic and OpenAI are selling forward-deployed engineers into the Fortune 500, building bespoke workflows on top of their own models. Look at what that actually is. The labs are paid enormous fees to sit inside your business and learn exactly how it works: where the value is, which workflows matter, what good looks like in your domain. Normally you'd pay a fortune for data like that. They're getting paid to take it.

OpenAI's $4bn Deployment Company, Anthropic's parallel vehicle with Blackstone, PwC putting 30,000 staff on Claude. Framed as distribution, they double as the cheapest way the labs will ever buy the one input that makes the next model better: how the best firms in the world actually work.

This is the dark mirror of Nadella's own line. You can never offload your learning, he says. Except that's exactly what you do when you let a frontier lab's team rebuild your workflows on their model. How your business actually operates ends up accruing to them, not you.

And it's been worth it, which is what makes it dangerous. The ROI is fast and skilled deployment is scarce, so firms pay the premium gladly. The catch is that the payoff comes now and the dependency comes later.

So how do you actually avoid putting everything in one basket? You stop building around the model in the first place. Your systems, your data, your infrastructure become the fixed part of the stack, and the model plugs in behind them like any other component. When a better, cheaper or more private one comes along, you swap it and nothing else moves.

What stops most firms doing this is that they run on prompts. A prompt is brilliant for a proof of concept and hopeless at specifying how a system should behave on real work. To get reliable behaviour you have to do the boring part: rearchitect your knowledge base so a model can read from it, write back to it, and be held to your standard of a right answer. That's the work we've put our time into at DealSage, and it's the part that doesn't expire when the model does.

Because there's always a new model. They'll keep getting better, and I've said it a thousand times this past year, they're going to commoditise in capability. The frontier labs understand this better than anyone, which is exactly why they're racing to embed their engineers and rebuild your workflows on their model. If the model can't be the moat, your dependence on it has to be. Own the system yourself and that moat is gone.

The real question underneath all of this is who captures the learning. Every deal your firm works leaves a record: the path your people take, the corrections that reveal their judgment, the outputs they accept and the ones they bin. Today most of it vanishes inside meetings, edits and a few partners' heads. Capture it, structure it with your own context, and the judgment that lived in three people becomes part of how the firm operates. That's the asset, and it accrues to whoever can best point a model at their own information. Which is why the value pools in the applied layer, between a generic model and what your firm actually knows, rather than in the model itself.

We're all still working out the exact architecture, but the direction is clear enough. The model has become the commodity, and the value has moved to applying it against what only you know. The customers Anthropic cut off on Friday night lost a generalist overnight. The ones who owned their systems, their data and their workflows lost a supplier. Everyone else lost their capability. This was never really a build-versus-buy question. It's whether the learning from your work ends up on your balance sheet, or theirs.

News Digest

Washington Pulls Fable, And Model Dependence Becomes A Kill Switch

Two weeks after an executive order that explicitly promised no "mandatory governmental licensing, preclearance, or permitting" for AI models, the government did something that rhymes with all three. Commerce Secretary Howard Lutnick sent Anthropic an export-control letter barring any foreign national, inside or outside the US, from using its two most advanced models, Fable 5 and Mythos 5. That includes Anthropic's own foreign-national staff. Anthropic complied by disabling both models globally, abruptly, for every customer.

The trigger was security. The administration cited national-security concerns after a separate company claimed it had jailbroken Mythos, and after Amazon, one of Anthropic's largest backers, flagged the models to senior officials. Amazon CEO Andy Jassy reportedly told Treasury Secretary Scott Bessent that Amazon's own researchers had used Fable 5 to pull together information useful for cyberattacks. An investor reporting its portfolio company's product to the government is its own kind of signal.

The details:

Why it matters: The two most capable models on the market were switched off for most of the world overnight, by government order rather than commercial choice. Anything built on top of them stopped working at the same moment.

My take: This is the model-agnostic argument in its hardest form. Lock-in used to mean a painful migration on your own timetable. Now a single letter from Commerce, or tomorrow from Brussels, can take your core capability offline with no notice. If you run one frontier model and employ a single foreign national, you just got a preview. The firms that barely noticed treat the model as swappable and keep the systems, data and workflow that are actually theirs.

Read more

Apollo Wires $35bn Into Anthropic's Compute, And Private Credit Owns The Buildout Now

Apollo and Blackstone are backing a roughly $35bn financing to expand Anthropic's compute capacity, built around Broadcom custom chips and data-centre sites operated by Fluidstack. The structure is large, asset-backed and routed through private credit rather than public equity. It's not an isolated cheque, either. KKR launched Helix Digital Infrastructure with over $10bn committed to AI-ready data centres, Amazon signed a $17.5bn bank loan, and Oracle is raising roughly $40bn to fund its own buildout.

Step back and the scale is hard to absorb. Axios reckons investors have funnelled around $255bn of equity and debt into the five major hyperscalers so far this year.

The details:

Why it matters: The exposure to the AI capex cycle has migrated off tech equities, where everyone is watching it, and onto private-credit and alternative-manager balance sheets, where fewer people are looking. Those are the vehicles a large share of institutional and PE capital actually sits in.

My take: Everyone's debating the AI bubble on the Nasdaq, which is slightly the wrong place to look. If this buildout disappoints, the first losses don't show up as a share price. They show up in the credit structures underwriting the chips and the buildings, held by the managers your LPs allocate to. And you're lending billions against the one input that, as above, is depreciating and commoditising. I'd check whose balance sheet holds it before the next vintage gets marked.

Read more

Other Interesting Things I’ve Read of Seen This Week:

AI Is Turning M&A Into a High-Impact Learning Machine (BCG, June) - BCG's framing maps almost exactly onto Nadella's: the real shift is turning episodic deal work into a compounding capability, from continuous sourcing through agentic diligence. The most useful institutional read of the week, and the one to forward to a sceptical partner.

AI fears trigger a 70% collapse in PE tech deal value (GoodReturns, June 11) - Bain-cited data showing tech buyout value cratering into what the piece calls "valuation paralysis," as buyers demand steep discounts on anything software-heavy. The same week Thoma Bravo declared the SaaSpocalypse over. Both are true, which is the whole story: the market is repricing software violently in two directions at once.

Thoma Bravo to take Kneat private for ~C$650M (June 9) - An all-cash take-private of a validation-software business pitched as a foundation for AI in regulated life sciences. Thoma Bravo's answer to "is software dead" is, as ever, to quietly buy more of it.

Figure to acquire Kiavi (June 11) - A blockchain-native lender bolts on an AI-powered platform for residential real-estate investors, adding loan volume and underwriting models in one go. Fintech consolidation, now with extra acronyms.

Jeff Bezos' Prometheus is now worth $41bn (Axios, June 11) - A $12bn Series B for an industrial-AI startup promising to compress product development and manufacturing cycles. Pre-revenue valuations are back, just wearing a hard hat this time.

Governor Pritzker slows Illinois data-centre development (Axios, June 9) - Illinois pauses data-centre tax breaks and adds energy, water and clean-air requirements, citing rising bills. The buildout has met local politics, and local politics is undefeated at the ballot box.

Karp: bragging about AI layoffs means you "might as well sign the Bernie Sanders manifesto" (Fortune, June 9) - Palantir's CEO warns that executives celebrating headcount cuts are inviting a political backlash they won't enjoy. Said out loud, for once, and by the last person you'd expect to say it at all.

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We're building an AI-native deal intelligence platform to help professionals turn their institutional knowledge into better decisions. If you're curious what we're up to, check out dealsage.io or just reply here