An OpenAI-backed fund now owns 50 accounting firms. It just raised $2B to buy more.

Thrive Holdings raised $2 billion from SoftBank, D1 Capital Partners and Altimeter Capital, at a $12 billion valuation. The firm is a spinout of Thrive Capital, the Josh Kushner fund that was one of OpenAI's earliest large backers, and it runs what amounts to a private equity playbook for AI: acquire traditional service businesses, then wire agents into the work they sell.

OpenAI is not a bystander here. It took an ownership stake in Thrive Holdings in December 2025 and has been sending its own employees into the portfolio companies to speed the rebuild.

The portfolio is bigger than most of the coverage let on. Current, the accounting arm, holds more than 50 firms and over 2,000 professionals. Shield, the IT arm, holds around 20 managed service companies. Across both platforms Thrive counts more than 70 businesses, and part of the new $2 billion goes to a third platform for regulatory work on physical assets: getting buildings approved, certified and kept in operation.

And for once, an AI announcement came with checkable operating numbers instead of adjectives. They are worth reading slowly.

The model is simple enough to describe in one breath. A traditional accounting or IT services firm sells labor by the hour, so its margins sit where labor margins sit. Buy that firm, hand its repetitive compliance work to agents, and the cost base drops while the client list and the licenses stay put. The cash that throws off funds the next acquisition. Repeat until you own a meaningful slice of the market.

Private equity worked out the first half of this years before the AI part arrived. The International Federation of Accountants counts 177 direct PE investments in accounting firms between 2015 and 2025, which triggered 875 roll-up acquisitions and about 900 follow-on deals in 2025 alone. By early 2026, roughly half of the 30 largest US firms had PE money or an alternative ownership structure behind them. The consolidation happened while nobody outside the profession was looking.

Thrive's accounting platform shows the speed. It started as Crete Professionals Alliance in 2023, partnered with Thrive Capital in May 2024, and by mid 2025 had passed $300 million in annual revenue and announced a plan to spend over $500 million acquiring more US firms and fitting them with OpenAI-powered tooling. It rebranded as Current in 2026 and now holds the 50-plus firms in Tuesday's announcement.

So the flywheel is not a pitch deck. It has been spinning for two years, and the $2 billion is fuel for a wheel already in motion.

If Thrive were alone, this would be a curiosity. It is not alone.

General Catalyst has allocated $1.5 billion of its latest fund to the same creation strategy: funding AI-native teams that buy call centers, property managers, MSPs and accounting firms, then automate them. Anthropic joined Goldman Sachs and Blackstone on a $1.5 billion venture to push Claude into PE-owned companies. OpenAI spun up a consulting arm, DeployCo, with $4 billion of initial capital at a $14 billion valuation. Microsoft committed $2.5 billion and 6,000 employees to a dedicated AI implementation unit. The day after Thrive's raise, IBM announced an OpenAI partnership that certifies thousands of its consultants on OpenAI models. Two days after that, OpenAI's CFO told investors enterprise revenue had passed consumer revenue on roughly $40 billion of annualized sales.

Read those together and the shape is hard to miss. The labs and their backers stopped waiting for service firms to buy AI. They are buying the service firms. We covered the talent half of this land grab when 70% of companies started chasing the same 2,000 forward-deployed engineers; the ownership half is the same strategy executed with a checkbook instead of a job posting.

Because the labor math was already broken before the agents showed up.

The Bureau of Labor Statistics projects about 124,200 openings for accountants and auditors every year through 2034, against a median salary of $81,680. Meanwhile 61% of finance and accounting leaders say skilled people are harder to find than a year ago, 75% say the shortage has delayed critical projects, and 62% have canceled projects outright, per Robert Half data. Accounting enrollment finally rose 8.9% this spring, the third straight year of growth, but a freshman declaring the major this year signs your returns in 2031 at the earliest. The gap is now.

The software market had already priced this in. Basis, an agent platform for accountants, raised $100 million at a $1.15 billion valuation in February with 30% of the top 25 US firms using it. Intuit says its agents hit 85% repeat usage, with invoices paid in full five days faster and manual work down about 30%. In accounts payable, 72% of finance leaders call it the obvious starting point for agentic AI, with an average reported ROI of 80%.

High volume, rule-bound, deadline-driven, chronically understaffed. If you were choosing one profession to point agents at, you would choose this one too. Thrive just chose to own it rather than sell to it.

Now the number everyone repeated and nobody worked.

7,000 returns at 98% accuracy means roughly 140 returns with something wrong in them. That is not a gotcha; human preparers make errors too, and any firm processing 7,000 returns has a stack of amended filings somewhere. The difference is that a human firm's error rate is nobody's press release, while this one is the headline proof that a $12 billion valuation rests on.

Three things about the claim deserve your skepticism before your applause. Nobody published what accuracy means here: measured against an IRS notice rate, an internal review, or a spot check, over what period, by whom. The figure is company-reported in the middle of a fundraise, not audited by anyone with something to lose. And accuracy on tax prep says nothing about the judgment calls, the entity structure question or the deduction gray zone, where getting it defensibly right is the actual product. Research on AI-prepared returns consistently finds the same shape: calculation errors go down, interpretation errors do not.

Same lens for Shield's 36x help desk number. Resolution time is the easiest metric in IT services to compress by closing tickets faster than you should. Maybe Shield's agents genuinely resolve 36 times faster. The point is you cannot know from a funding announcement, and neither can the businesses whose accountant gets acquired next quarter.

And liability has not moved an inch. A credentialed preparer still signs the return, and the taxpayer still owns the penalties unless the engagement letter explicitly says otherwise. When the 2% lands on you, the rollup's press release will not answer the IRS notice.

Suppose the numbers hold. Your compliance work gets cheaper and faster, whether or not your own firm is owned by a rollup, because a market where one player closes books 30% faster reprices everyone eventually. The 124,200 openings a year you have quietly been competing in every time you tried to hire a bookkeeper stop being your problem. Month-end stops eating the first ten days of the month.

And you inherit a benchmark. The moment a vendor across the table claims agents on the work, you now have public numbers to hold against theirs: 7,000 returns, 98%, 30% faster, and an 80% average ROI in AP. A provider who will not put their own figures next to those is telling you something. We have written before about what AI actually returns when someone counts honestly, and the honest counts are exactly what these claims should be judged against.

Now suppose you are on the wrong side of it.

Your accounting firm sells to a rollup mid-engagement, and the service changes under you without your signature on anything. The partner who knew your business takes the earnout and leaves. The work moves onto agents tuned for the portfolio's median client, which you are not.

Pricing is the quieter risk. The rollup's cost base drops 30%; your invoice does not, because you have no visibility into their costs and switching accountants in October is misery. Margin re-rating is the entire thesis, and the re-rate comes out of the gap between what the work costs them and what you keep paying.

Concentration compounds it. Follow the ownership in Tuesday's announcement and your accountant, your IT help desk and the models both run on can now share one cap table. When half the largest US firms already have PE behind them, the number of genuinely independent providers with capacity to take you is shrinking while you read this.

And when something goes wrong, accountability diffuses. The firm blames the platform, the platform points to the review process, and the IRS notice still has your name on it.

You cannot control whether your providers sell. You can control three things this quarter.

Ask your accountant, in writing, what runs on agents today, who reviews agent output, and whether your fee changes as their costs fall. A good firm answers in a paragraph. A firm that has not thought about it answers in a meeting request. Both answers are information, and we walked through the same style of vendor questioning in the red flags that should stop an automation proposal.

Benchmark before you renew. The claimed numbers above are now the public bar for agent-run compliance work. Hold your provider's turnaround, error handling and price against them, in the same spreadsheet where you track any other vendor.

And take the slice you can own. The repetitive layer of your own back office, the invoice coding, the reconciliation, the report that gets rebuilt every Monday, does not need to wait for anyone's acquisition strategy. An AI agent for accounting chores at that size is the most proven build in the category, and it is the work we put agents on at agentclaw: a starter build runs $1,500 to $2,500 fixed, a two-week production sprint is $5,000, and the point is that your team owns the output instead of renting it from whoever owns your vendor next. If your books are one person and ten clean hours a month, skip it; agents on that workload are a solution shopping for a problem. The threshold question is the same one in our comparison of hiring in-house versus buying the build: whether the repetitive hours you would automate actually exist on your payroll today.