The Frontier Is Leaking

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.

…the American one, at least. This week Commerce let Anthropic turn Mythos 5 back on, Fable 5 is expected back within days once the Pentagon and NSA sign off, and OpenAI agreed to hold the public launch of its new GPT-5.6 family so the government can vet who gets it first. The most capable models on the market now ship on a government clock.

Meanwhile, the market voted with its tokens. The same benchmark costs about 20x more on a frontier model than an open one, US models' share of OpenRouter traffic is sliding, and UBS says 60% of companies are routing work to cheaper Chinese open models to cut their bills.

Elsewhere: the BIS warned of an AI investment bust, global M&A hit a record first half, Permira and Warburg took Clearwater private for $8.4bn, and French firms found adopting AI and getting anything from it are very different things.

But first, a few posts from the people actually building this, all circling the same question: what a frontier model is worth now that a cheap open one does most of the same job.

In today's Acquisition Intelligence:

The Frontier Is Leaking

A few posts from the frontier worth your time this week. X is still where the AI debate actually happens, ahead of the press releases. You'll spot themes we've covered these past few weeks, this time from different angles.

@trevornoren (Trevor Noren, Sage Road Research), citing JPMorgan (June 26)

The post: he argues most future tokens will come from smaller open models rather than the frontier, and that the efficient frontier is now dominated by Chinese labs. The figure underneath it: running the same Artificial Analysis Intelligence Index reportedly costs about $3,700 on Claude Opus 4.8 and around $186 on DeepSeek V4 Pro. Same benchmark, roughly 20x the bill.

If you are pricing an AI rollout off frontier rates, you are pricing off the most expensive option on the board. For most deal tasks you should be asking what the cheapest model that clears your bar is, and for a lot of them that stopped being the frontier months ago.

@DeryaTR_ (Derya Unutmaz, immunologist and prolific AI commentator) (June 26)

The post: a one-shot reproduction of a research paper cost $6.21 on GLM 5.2 against a much heavier bill on Opus 4.8. His sharper claim is that Dario Amodei is lobbying to restrict Chinese open source because it threatens Anthropic's pricing power.

I don't know if he's right about the motive, but the logic holds. When your moat is price, cheap competition is the whole problem, and the natural reflex is to reach for policy rather than fix the product. So keep an eye on the regulation. The "US frontier model as default" a lot of firms are building around may be held up as much by export controls as by a real performance lead, and export controls are a shaky thing to bet your stack on.

@zerohedge, citing Bloomberg (June 26)

The post: a Bloomberg chart showing the share of OpenRouter tokens going to US models falling off a cliff as developers shift to (largely Chinese) open models.

This is the argument turning into data. The developers closest to it have already moved a chunk of their usage across, which makes sitting tight on the frontier the more expensive default rather than the safe one. The switching costs you assumed were permanent are getting paid down by other people while you wait.

@rohanpaul_ai (Rohan Paul, AI writer), citing UBS (June 26)

The post: UBS finds 60% of companies are watching their AI budgets and shifting spend to cheaper open-source models, Qwen, DeepSeek, MiniMax, GLM, Kimi. They are routing rather than cutting, easy tasks to cheap models and the premium frontier kept for genuinely hard reasoning.

This is the most boring and most important of the six. Model routing has gone from a clever engineering trick to standard cost discipline in about a quarter. Your portfolio companies are probably already doing it, and if they aren't, "are we routing or sending everything to the most expensive model out of habit" is a fair question to put to any management team waving an AI line item at you.

@brian_armstrong (Brian Armstrong, Coinbase CEO) (June 28)

The post: the how-to. Coinbase cut its AI spend nearly in half while token usage kept climbing, without capping anyone. Five levers: default to cheaper open models, route each prompt to the right model instead of letting engineers pick by hand, cache aggressively (one tool went from a 5% to a 60% cache-hit rate), and keep context lean.

A public-company CEO laying out the operating playbook for AI cost control, in detail, for free. If you are running a rollout at your firm or inside a portco, those are the levers to copy, and halving your spend while usage goes up is a fair benchmark to hold yourself to.

@AravSrinivas (Aravind Srinivas, Perplexity CEO) (June 28)

The post: he argues every enterprise will end up running its own model-harness-sandbox-eval flywheel, and that the part worth defending is the tacit knowledge each company holds about its own domain and workflows, which the model can't supply.

Srinivas makes the point that matters most for a deal firm. If the model is the commodity, your edge is whatever it can't get anywhere else, which is the structured record of how you actually source, judge and close deals. You rent the model and swap it when something cheaper shows up, and the record of how your firm works stays with you.

Where do we go from here?

Interesting to see how this plays out from here. My sense is we get a wave of lobbying from the frontier labs to restrict open-source models on the grounds they're too dangerous to leave lying around, which is the same national-security card Washington just played on Fable and GPT-5.6. The three-way pull between the frontier labs, their open-source rivals and the US government is the one I'd watch for the rest of the year.

News Digest

Washington Becomes The On/Off Switch For Frontier AI

Two weeks ago I wrote about Washington pulling Anthropic's two best models offline overnight. This week we got the other half of the story, and it now stretches well beyond Anthropic.

On Friday, Commerce let Anthropic switch Mythos 5, its strongest cybersecurity model, back on, the first real sign of de-escalation. Fable 5, dark since June 12, is expected to follow within days, though the Pentagon and the NSA still have to give their approval. And on the OpenAI side, Sam Altman confirmed the company is holding back the broad public launch of its new GPT-5.6 family at the administration's request, releasing it first to a small set of government-approved enterprise partners while officials build pre-release testing for cyber and national-security risk.

The details:

Why it matters: Access to the most capable models on the market is now gated by Washington, granted and revoked on a government timetable rather than a commercial one. For anyone running production workflows on a single frontier model, that is a new and very real category of operational risk.

My take: I made the lock-in argument two weeks ago and I'll spare you the repeat, but the GPT-5.6 piece goes a step past pulling a model after the fact. The government now wants to vet frontier models before the public ever sees them, which turns "when can I actually use this" into a regulatory question as much as a release date. And note the irony with the section above. The harder Washington works to keep its best models inside a controlled perimeter, the more it nudges everyone else towards the cheap open (and largely Chinese) models it has no switch for. Guarding the frontier and slowing your own side down can end up being the same move.

Read more

The BIS Warns The AI Boom Could End In A "Protracted Investment Bust"

The Bank for International Settlements, the central bank for central banks, used its annual economic report to warn that the AI spending boom risks ending in a "protracted investment bust," with knock-on damage to the wider economy.

The five biggest hyperscalers are on track to spend more than $1tn on AI capex across 2025 and 2026, ahead of their own earnings and free cash flow, with the gap increasingly filled by debt. The BIS reached for three historical rhymes: the canal mania of the 1830s, Britain's railway mania in the 1840s, and the dotcom boom. Each was a genuine breakthrough that pulled in more capital than the returns could ultimately justify, and each ended in a bust.

The details:

Why it matters: This is the macro version of what the builders are showing at micro scale. If the frontier can't command premium pricing because open models are 20x cheaper, the revenue meant to justify a trillion dollars of capex gets a lot harder to underwrite. And as I wrote two weeks ago, much of that financing now sits in private credit and alternative-manager balance sheets, which is to say, with your LPs.

My take: Two of the most credible institutions on the planet read the same numbers the same week and landed on opposite calls, which is a fair measure of how little anyone really knows here. My own read is that it isn't binary, the buildout is genuinely productive and genuinely overfinanced, and both can run together for a good while. The piece I'd watch is the debt. The maths that justifies a trillion in capex assumes the frontier can keep charging a premium, and the builders above are the first real sign it can't, so if that premium keeps eroding the strain turns up in the credit behind the chips and the data centres long before it reaches a share price. That's the exposure worth understanding now, while it's still an abstract debate.

Read more

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

Global dealmaking just had a record first half (Axios, June 25) - First-half 2026 global M&A hit roughly $2.7tn, with PE deal value up around 54% year on year to about $583bn, almost all of it in very large transactions. Turns out the cure for valuation paralysis was simply writing much bigger cheques.

Permira and Warburg Pincus complete $8.4bn take-private of Clearwater Analytics (June 26) - The investment-management software firm is going private specifically to accelerate its "Gen AI agentic platform." Nothing says conviction in the public markets like leaving them to build the AI roadmap in peace.

Qualcomm to buy Modular for ~$4bn (June 26) - An all-stock deal for an AI-native compiler company whose whole pitch is running models across different chips, bought to loosen Nvidia's grip on the stack. The entire industry is now shopping for ways to not depend on one expensive supplier. Sound familiar?

French firms have adopted AI, but the gains are missing (Bpifrance survey, June 23) - 77% of mid-sized company heads say they use generative AI; just 17% report any time savings from it. Adoption is a vibe, ROI is a number, and so far only one of them shows up in the data.

Law firms are still hunting for a clear payback on AI (FT, June 25) - 83% of lawyers now have access to AI tools, only 22% trust the output, and clients armed with the same software are starting to ask what they're paying for. "I have Harvey too, so what am I using you for" is a sentence every professional-services roll-up should read twice.

The Economist says the AI backlash is only getting started (June 25) - Polls souring, protests disrupting data-centre projects, and politics finally catching up with the buildout. The models got cheaper and the public got angrier, and only one of those is on anyone's roadmap.

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. The builders spent this week proving the model is the cheap, swappable part. We build the other part: the structured deal knowledge and workflows a model plugs into, so you can route to whatever is best and cheapest this month without rebuilding anything. If you want to see what that looks like inside a deal firm, reply here or have a look at dealsage.io.