Tokens Are A Confession

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.

Alex Karp went on CNBC on Wednesday and gave the frontier labs both barrels. Enterprises are "livid," he said, "paying for tokens that create no value," while the labs are "stealing the weights and alpha" of their businesses. Palantir followed up with a nine-point manifesto on AI sovereignty, and the whole thing has been ricocheting around X ever since.

Meanwhile Microsoft committed $2.5bn to a new unit called Frontier Co: 6,000 engineers embedded inside customers to make AI actually work. So in the same week one side of the industry declared token pricing a scam, and the other conceded you need an army of humans to turn tokens into value.

Also this week: Blackstone walked away from the world's largest planned data centre, Zuckerberg told staff his agents are running late, OpenAI offered Washington 5% of itself, and Schneider Electric paid $3.1bn for Cognite.

But first, my take on the Karp rant.

In today's Acquisition Intelligence:

From The Trenches

Alex Karp doesn't do measured. But even by his standards, Wednesday's CNBC interview was a full demolition of the frontier labs' business model, delivered on live television while nominally there to announce a partnership with Nvidia.

What He Actually Said

The core of it: "In this country every single enterprise I deal with, these people are livid. They're like, 'I am paying for tokens that create no value.'" And then the line that stuck: "These people are stealing the weights and alpha of my business, and they're creating a wealth tax."

The best bit was the pricing challenge. "If it was so valuable and let's say I can make you a billion dollars, then tomorrow, wouldn't I say, 'I'll make you a billion dollars, and I want 30%.' Why are they charging for tokens if it's so valuable?"

Palantir then posted a nine-point manifesto on what it calls AI sovereignty: your data retention is your treasure, controlling your weights is controlling your fate, and a warning against "tokenmaxxing," the addictive feeling of progress you get from pouring money into metered models. The commercial subtext is doing a lot of work there, and we'll get to it.

Is He Right?

Now look, Karp is obviously talking his own book a bit here (who doesn't). Palantir's whole model is the sovereign stack, your data and open-weight models running inside your own perimeter, and the Nvidia deal he was there to announce is exactly that product. But he's making a good point, and anyone who has read this newsletter for a few months will know I sit close to his end of the argument.

And the argument holds up well beyond his book. Per-token pricing is how you sell a commodity. Electricity is priced per kilowatt-hour because nobody's kilowatt is special. When the labs charge per token rather than per outcome, they're conceding, in the price structure itself, that what they sell is interchangeable, which is precisely what the 20x open-model discounts we covered last week keep proving.

And an interchangeable input can't be your edge. If every firm bidding on your deal runs the same model, whatever it gives you it gives them too, and frontier rates run anywhere from 5x to 100x the open alternatives that decent engineering can match. You're paying a premium to stay level with the field.

The "stealing your alpha" charge sounds paranoid until you look at what happened in April. Anthropic's chief product officer, Mike Krieger, sat on Figma's board. He resigned on April 14, and on April 17 Anthropic launched Claude Design, a direct competitor, knocking around 7% off Figma's stock in a day. This from a lab that had been Figma's collaboration partner as recently as February.

On Tuesday the same lab launched Claude Science, an AI workbench aimed at pharma and research. Anthropic is heading for an IPO at a valuation north of its current $965bn, and the revenue to justify that number has to come from somewhere, so the labs are climbing into the application layer of every industry they serve. If your business runs on one of their models, you are simultaneously their customer, their training signal and, eventually, their target market.

"No value" sounds extreme until you look at how most firms actually buy this stuff: rent a frontier model, point it at scattered data, wait for magic. The Ramp study below shows where the returns actually live, with the minority who engineer properly around the models, and Coinbase just halved its AI bill while usage grew by doing exactly that. Which is really Karp's point. The value shows up in the system you build around the tokens, and most buyers have been sold the tokens without the system.

And still, I understand why everyone pays up. The frontier products are genuinely excellent, which is what makes them the easy button. Rent Claude or GPT, let a forward-deployed team embed it for you, and you get something that works without having to think very hard, and as of this week the easy button comes with 6,000 Microsoft engineers attached (more below). Sovereign AI, your data on models you control, takes more skill, and the difficulty is precisely what keeps it an edge rather than a checkbox.

What It Means For You

Zoom out and this week reads like a map of the next few years. The labs are climbing into applications, Microsoft is climbing into services with a $2.5bn embedded-engineering unit (more below), Palantir and Nvidia are selling the sovereign stack, and open-source keeps cutting the floor out of everyone's pricing. Every giant in the industry is converging on the same ground: the layer that sits between an enterprise's data and the model.

That tells you where they all think the value settles. Models get interchangeable, compute gets commoditised, and what's left is whoever owns the context, the workflows and the relationship with the business itself. The fight for that layer is going to be enormous, and firms like yours are the terrain it gets fought over.

So, practically. Read the data terms in your AI contracts, and your portfolio companies' contracts, because the question is what you're feeding the flywheel of a company that may end up competing with you or your portco. When you're diligencing software, the Figma episode sets the bar: a partnership and a board seat bought Figma three days' notice, so ask what a target owns that a lab can't replicate in a quarter. "We have a great UI on top of Claude" is now a risk disclosure.

To be clear, I don't think Anthropic is racing to become a private equity firm or an investment bank, or even to compete with your portcos. But the pattern of the last few months is worth taking seriously: design in April, science in June, each time straight into a customer base that had been feeding the model. The sensible planning assumption is that any profitable workflow sitting on top of a frontier model is on a product roadmap somewhere in San Francisco.

And inside your own firm, remember sovereignty is a spectrum, and you don't need Palantir's air-gapped version of it. The sovereign asset for a deal firm is the structured record of how you source, judge and close deals, the thing Palantir's manifesto calls tribal knowledge. Own that layer, rent the model on top, and Karp's wealth tax becomes a utility bill: small, predictable and swappable.

What The Builders Are Saying

A couple of interesting perspectives from the week that tie into all of the above.

@KSimback (Kevin Simback, longtime open-source AI proponent) on the context repo (July 3)

The post: coding got automated first because all the context an agent needs sits in one versioned repo with a test suite, while the rest of knowledge work is scattered across Slack, email, systems and people's heads. Fixing that takes an ontology-based context repo, and built properly it flips from a cost into an asset: owning a model of how your company operates is the final boss move.

Why this matters: This is the constructive half of the Karp argument. He tells you what to stop doing; Simback describes what you'd build instead.

My take: Translate "context repo" into deal terms and it's the structured record of every deal you've sourced, priced, passed on and closed, and why. Most firms have that history, almost none have it in a form an agent can use, and the ones who build it get compounding returns on their own judgment.

@mrp (Ron Pragides, veteran engineering exec) clipping David Sacks on the Klarna U-turn (July 4)

The post: Sacks on All-In revisiting Klarna's plan to replace its whole customer service department with AI, "peak aura farming" in his words. A year on, Klarna reversed it, and Sacks' wider point is that most CEOs have nothing original to say about AI, so they glom onto the press cycle with dramatic announcements that turn out to be fake.

Why this matters: The gap between announced AI and deployed AI is now its own category of noise, and it cuts both ways in a deal context.

My take: Discount what management announces about AI and weight what they actually spend and ship. The Klarna U-turn is the base rate for "we're replacing the department" claims, and an AI narrative in a CIM deserves the same scepticism as any other projection.

News Digest

Microsoft Puts $2.5bn Behind Making AI Actually Work

The week Karp said enterprises are paying for tokens that create no value, Microsoft launched a business whose entire premise is that he's roughly right. Frontier Co is a new $2.5bn unit staffed with around 6,000 industry and engineering experts who will embed inside enterprise customers to build AI capability, led by longtime Microsoft executive Rodrigo Kede Lima.

Satya Nadella framed it in a post that deserves a slow read: "The future of the firm is a learning loop in which human capital and token capital compound." The ambition, he says, is to help every enterprise build its own continuously improving AI capability out of its knowledge, workflows and judgment.

The details:

Why it matters: The world's biggest software company has concluded that the gap between buying AI and getting value from it is now a product category worth $2.5bn. That's Palantir's forward-deployed engineer model, vindicated at hyperscaler scale, and it should worry every consultancy whose margins live in that gap.

My take: The world's largest software company moving into services is pretty telling. Microsoft has spent four decades selling licences and leaving the messy human bit to partners, so committing $2.5bn of its own people to implementation says a lot about where it thinks the value has moved. The cynic in me questions how much of this is repurposing roles that AI made obsolete, but the signal is certainly there. Watch what Accenture and the Big 4 do next, because their pitch just got a $2.5bn competitor with root access.

Read more

The AI Spenders Are Hiring

A study from Ramp's economics lab and Revelio Labs, covered by the FT this week, is the first to join company-level AI spending data with headcount records, across 21,559 US firms. The headline cuts against the doom narrative: the heaviest AI adopters grew white-collar headcount 10.2% in the two years after adoption, with entry-level roles up 12%.

The catch is in the distribution. The gains only show up for the top third of spenders, roughly $30 per employee per month and rising. The bottom two-thirds of adopters saw no significant change against firms that never adopted at all.

The details:

Why it matters: This is firm-level evidence that AI spend and hiring rise together, published the same fortnight Oracle disclosed 21,000 job cuts and named AI as a cause. Both are true at once: the labour market effect of AI is reallocation between firms, from the ones treating it as a subscription to the ones treating it as an operating model.

My take: The threshold is the story. $30 per employee per month is a trivially low bar, two coffees, and two-thirds of adopters still sit below it, which is why they see nothing and why surveys keep finding adoption without gains. For diligence, this hands you a cheap leading indicator: ask for AI spend per employee and the trend, because below the threshold "we use AI" is a slide, and above it you should expect the effects in the numbers within a year. And note the entry-level finding, because the firms spending most are hiring more juniors, which suggests the pyramid gets rebuilt around the tools rather than abolished.

Read more

Other Interesting Things I've Read or Seen This Week:

Schneider Electric buys Cognite for $3.1bn (June 30) - All cash for the Norwegian industrial data and AI firm, which did about $170m of revenue in 2025; early backer Aker collects roughly $1.48bn. Eighteen times revenue for data plumbing, because the boring layer of the stack is where the pricing power lives.

Together AI raises $800m at an $8.3bn valuation (July 1) - The neocloud rents GPU infrastructure for exactly the open models undercutting the frontier, with bookings past $1.15bn annually, and Aramco Ventures led the round. The cheap-token trade now comes with oil money attached.

A former Goldman analyst's AI research startup raises $22m (Bloomberg, July 2) - LinqAlpha builds AI investment research for institutions and says it serves 70+ of them, including Causeway and Schonfeld. The junior analyst stack is being rebuilt one funding round at a time, mostly by people who used to be junior analysts.

SAP reins in costs to fund AI investment (WSJ, July 2) - Europe's biggest software company is squeezing hiring, external spend and travel to redirect money into AI, while pointedly avoiding a repeat of its 2024 layoffs. Cancelling the flights before the people, which in 2026 counts as restraint.

Blackstone's QTS kills the world's largest planned data centre (July 2) - The 2,100-acre Digital Gateway campus in Virginia is dead after years of resident lawsuits, with the land reverting to rural zoning. Last week the Economist said the AI backlash was just getting started; Virginia decided to prove the thesis within days.

OpenAI floats giving Washington a 5% stake (July 2, via FT) - Early talks on a structure where leading labs each cede ~5% to a public vehicle modelled on the Alaska Permanent Fund; OpenAI's slice alone would be worth about $42.6bn. Two weeks after Washington became the industry's on/off switch, the industry is offering it a seat at the cap table.

Zuckerberg tells staff AI agents are behind schedule (July 2) - In a leaked town hall, he said agentic progress "hasn't really accelerated in the way that we expected" over the past four months, while Meta stays on track to spend up to $145bn on AI infrastructure this year. The agents are late; the capex is punctual.

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. Karp's wealth tax has a simple antidote: own the layer that holds your firm's deal knowledge, decisions and workflows, and rent the models on top, swappable whenever something better or cheaper ships. That's what we build for deal firms. If you want to see what sovereignty looks like without an air gap, reply here or have a look at dealsage.io.