The Rise of Open-Weight Models

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.

…and it is bigger than anyone budgeted for. Uber blew through its entire 2026 AI budget by April and now caps employees at $1,500 a month. Even Sam Altman concedes cost has gone from a non-issue last year to a "huge issue" for customers this one.

The pressure valve is open weights. A Chinese model, GLM-5.2, landed within touching distance of Claude Opus for roughly a sixth of the price, and even Microsoft is now testing a Chinese open model inside Copilot to cut its own bill. In the markets, SpaceX's record IPO became a $60bn war chest it spent on Cursor within days, and in Berlin private equity bosses admitted cheap AI now threatens their law and accountancy roll-ups.

But first, my take on open models. What they are, why they cost a fraction of the frontier, the catch nobody mentions, and why they are becoming increasingly relevant for anyone sitting on confidential deal data.

In today's Acquisition Intelligence:

From The Trenches

The Rise of Open-Weight Models

"We created a monster." That was the CIO of a software firm called Workato, talking to the FT this week. When Anthropic moved them onto token-based pricing in May, their spend went up sevenfold the first day. His internal sessions used to be about innovation. Now they are about "AI financial responsibility."

He is not alone. Amazon, Walmart, Cisco, Uber and Meta have all started capping AI usage or pushing staff onto cheaper models. The thing everyone was promised would be cheap or free is turning out to be neither, and the bill arrives one token at a time.

I flagged this a while back, and I think it only gets louder from here. The cost of running AI was always going to become the story, because the economics of an agent are nothing like the economics of a chatbot.

A chatbot answers when you ask. An agent runs on its own, often in a loop, often spawning more agents to handle sub-tasks. Cisco's product chief put it plainly: for every human you might have 10, 100, or on the aggressive end 1,000 agents, and they just keep working. Each one burns tokens the entire time.

So when a firm moves from "summarise this" to "go do this," consumption does not tick up. It explodes. And that is exactly why the model you run those agents on is about to matter far more than the model you chat with.

You can already see who is reacting. Microsoft, of all companies, which has every reason to push its own partners' models, is testing a fine-tuned, Azure-hosted DeepSeek (a cheap, open Chinese model) inside Copilot specifically to control costs. When the company selling you OpenAI access starts routing you to a Chinese open model, the cost pressure is not theoretical.

All of this is pushing firms towards a different kind of model. Not a better frontier one, a cheaper, open one they can run themselves.

Think of it as rent versus own. A closed model, Claude or GPT, is rented: it lives on the lab's servers, you send your text and pay per word, and you never see the thing itself. An open model is owned: the lab publishes the actual file that makes it work, so you can download it, run it on your own machines, and tune it to your needs.

Two distinctions people muddle. "Open weight" describes what you get: the finished model, downloadable, though not the full recipe used to build it. "Open source" is really about the licence, how freely you can use, change and redistribute it. And "Chinese" is just who built it. GLM is all of these: a Chinese open-weight model from a lab called Z.ai, released under the permissive MIT licence. DeepSeek, Meta's Llama and Alibaba's Qwen sit in the same camp.

Three reasons they cost less. No one is taking a frontier margin, because once you can download a model, nobody can charge $25 a million tokens to serve it. They run anywhere, on your own servers or any budget host. And Chinese labs in particular have got good at squeezing more out of each token.

Now the catch, because the per-token price is only half the picture. A cheaper model that needs more tokens to finish a job can wipe out its own advantage, and "run it yourself" means buying or renting the GPUs, which is a real bill. The headline 90x gap narrows once you account for both. But the direction of travel is not subtle.

There is a second reason this matters more in our world than most. A closed model means your data leaves the building: every prompt, every document, every line of a CIM goes to someone else's servers. For a fund sitting on confidential deal data, that is not a small thing.

Open weights change it. Because you hold the model, you run it on your own infrastructure, with nothing going back to the vendor. That is the entire pitch behind Prem, the Swiss startup raising a $100m Series A this week at a $500m-plus valuation, backed by Jim Breyer, Index and Sequoia China, to let hedge funds and law firms own their AI rather than rent it. And after Washington pulled Anthropic's two best models for most of the world last month, "what happens when the model you depend on gets switched off" stopped being hypothetical.

This used to be an easy trade-off: cheap models were cheap because they were worse. That is no longer true.

GLM-5.2, released by Z.ai on June 13, scores 81 on Terminal-Bench against Opus 4.8's 85. On the harder SWE-bench Pro it trails, 62 to 69. But it is open-weight, MIT-licensed, and costs $1.40 / $4.40 per million tokens against Opus at $5 / $25.

List prices, June 2026. Per-token only, before infrastructure and token efficiency.

The people building these models think the gap closes entirely, and soon. Asked on X this week when China would reach "Fable class," meaning Anthropic's frontier, Elon Musk guessed the first quarter of next year. Jie Tang, the academic behind Z.ai's GLM models, replied: "won't take that long."

Let me be clear about what I am not saying. I am not saying everyone rips out Claude and self-hosts GLM next quarter. They won't, and for good reason: running models yourself means buying the infrastructure, standing up the inference, and keeping it running. For plenty of firms the frontier API is still the right call.

But the default is going to move. You are probably reaching for the most expensive model in the room for everything, because that is what the labs trained us to do. Most of the work in a deal does not need it: pulling figures from a CIM, tagging documents, first-pass summaries, reconciling a data room. That is high-volume, repetitive work an open model handles for a fraction of the price, and increasingly on hardware you control.

Satya Nadella, who I quoted here last week, made the macro version of the point: nobody wins in "a world where every company across every sector is ceding value to a few models that eat everything they see." Open weights are the most direct answer to that.

So this is the trend I would watch over the coming months. Not a wholesale switch, but a steady pull, as the cost questions get louder and firms weigh the price of a token against the cost of running the inference themselves. You are going to hear a lot more about which model runs where, and why.

News Digest

SpaceX's Record IPO Becomes a $60bn War Chest

SpaceX priced the biggest IPO in history on June 11, raising around $75bn at a $1.75tn valuation. By Monday the stock had run to $192, a market cap of $2.5tn, roughly $740bn added in under four trading days. Days after listing, it agreed to buy Cursor, the AI coding startup, for $60bn entirely in stock.

Here is the part worth sitting with. That $60bn is about 3.4% dilution at the IPO price, and the stock appreciated by the entire cost of Cursor in a few hours of first-day trading. As Fortune put it, Musk has minted a "super-currency." The deal cost him almost nothing.

This is xAI's play. Musk folded xAI into SpaceX earlier this year, and the war chest now feeds his AI ambitions: compute, capital, Starlink distribution, X, and with Cursor, the software layer on top.

The details:

Why it matters: When a newly public stock compounds on AI optimism faster than you can spend it, a $60bn acquisition becomes rounding-error dilution. That is an M&A weapon almost nobody else has, and it tilts the whole field.

My take: This is convergence, and it is the part people keep underpricing. The biggest players are no longer just building better models, they are assembling the entire stack, compute, capital, distribution and now applications, and paying for the missing pieces with equity that compounds faster than they can spend it. A traditional sponsor cannot compete on currency with a company that adds a Cursor's worth of market value before lunch. And note the tension with the From the Trenches above: the same week open models push the cost of intelligence down and out to everyone, the capital and influence to control AI are concentrating into a few hands at a pace the market has nowhere near digested. The landgrab is just getting started, and almost everyone is underestimating its scale.

Read more

Private Equity Bosses Warn AI Threatens Their Law and Accountancy Bets

At the SuperReturn conference in Berlin this week, the private equity industry stopped pretending. Some of the businesses they have poured billions into, law firms, accountancy practices, consultancies, are among the most exposed to AI, and the people who bought them know it.

"Apologies to the lawyers, accountants, consultants in the room," Apollo's Scott Kleinman told delegates. "You're going to see a lot of pressure." Man Group's Kevin Marchetti was more specific: claims auditing, billing automation, proxy voting, legal services, all of it now looks automatable. Accenture, the world's largest listed consultancy, has lost almost half its market value in a year.

The worry is structural. These are asset-light businesses that bill by the hour, exactly the model that breaks when the hours collapse.

The details:

Why it matters: Software was the first sector AI re-priced. Professional services is shaping up to be the second, and PE is heavily long it, often with debt stacked on top.

My take: This is the same story as the From the Trenches above, seen from the other end. The cost of intelligence is collapsing, and when intelligence gets that cheap, the businesses that sold it by the hour have the most to lose. The honest tell is the flight to the halo trade. When the smartest buyers start preferring factories to law firms, they are telling you which revenue they think a cheap model can copy and which it cannot. The diligence question has moved from "how durable are these earnings" to "how much of this work survives a capable model that costs pennies a task." Not everything is exposed equally: regulated audit, and anything where a human has to sign and carry the liability, holds up better. But "we bill by the hour" has stopped being a moat. It is a countdown.

Read more

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

Salesforce buys AI customer service platform Fin for $3.6bn (June 15) - Salesforce keeps buying its way to an agentic future, bolting Fin onto Agentforce to do customer service without the customer service team. The agents Salesforce sells you will, eventually, talk to the agents it bought.

Montagu takes a majority of BMC Helix in a carve-out from KKR (June 18) - Montagu carves the agentic ServiceOps platform out of KKR-owned BMC Software, with KKR keeping a minority stake. One sponsor's "non-core asset" is another's "AI-native platform play," depending on the quarter.

BCG: The AI-First Finance Function (June 18) - BCG reckons leading finance teams go "AI-first" within two to five years, with agents running core workflows and routine headcount roughly halved. File under "things your portfolio company CFO should read before next budget season."

Warren wants answers on PE's data-centre deals (June 15) - Senator Warren wrote to BlackRock, Blackstone, Brookfield and KKR asking about overlaps between their data-centre bets and the utilities that power them. The AI buildout was always going to attract Washington. It just got a name attached.

Half of AI job cuts will be reversed by 2027 (June 19) - Gartner predicts firms that cut headcount without redesigning the work end up rehiring by 2027. Turns out "fire everyone and add AI" was a strategy, just not a good one.

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 built the platform to be model-agnostic from the start, so each job runs on the model that does it best and cheapest. You're never paying frontier prices to summarise a CIM, and never stuck on yesterday's model when a better, cheaper one lands next month.