Permanent Capital.
Permanent Edge.

"Capital is magnetically attracted to mediocrity when hurdle rates are lowered." Mark Leonard, 2017

Before examining what Thrive Holdings is building, it is worth understanding the model it is drawing from - not to suggest imitation, but because Constellation Software is the best available case study for what disciplined, permanent-capital compounding looks like over decades. The principles that made it work are directly applicable to what Thrive is building.

In 1995, Mark Leonard raised $25 million - mostly from the OMERS pension fund and former colleagues at Ventures West Capital - and started buying small, vertical-market software companies. The kind that run parking garages, municipal utilities, and public transit systems. Not glamorous. Nobody wanted to hear about it at dinner parties. But mission-critical, sticky, and cash-generative.

Three decades later, Constellation Software has compounded shareholder returns at roughly 30% annually since its 2006 IPO. It owns over 1,000 businesses. It employs 64,000 people. It generates $10 billion in annual revenue. It completes more than 100 acquisitions a year. The Economist calls it "Tech's Berkshire Hathaway." Leonard has rarely sold shares, maintaining one of the largest insider positions of any Canadian public company CEO.1

The numbers are extraordinary. But the reason to study Constellation is the principles Leonard codified along the way - principles that translate directly from vertical-market software to the service businesses Thrive is acquiring. Four stand out.

Define the game narrowly, play it repeatedly

Leonard didn't "do software." He did vertical-market software with sticky recurring revenue and low churn - municipalities, transit authorities, homebuilders. The narrower the box, the faster the pattern recognition. Every deal starts to look like the last one, and you get better at saying no.

Thrive has drawn the same kind of box: three verticals (accounting, IT services, HOA management) where the businesses are mission-critical, the revenue is contractual, and there is a structural oversupply of forced sellers. Narrow enough to build pattern recognition. Wide enough to deploy a billion dollars.

Buy to own forever

When you cannot sell, you must buy right. Every Constellation acquisition clears a high IRR hurdle - 25% for mid-sized deals, 30% for small ones, 20% for large. Leonard tracked returns on every single acquisition and used ROIC plus organic growth as his favorite single metric. When internal initiatives didn't meet thresholds, they were killed - quickly, and without apology.2

Thrive's permanent capital structure mirrors this exactly - there is no fund lifecycle forcing an exit. The open question is whether the same IRR tracking discipline gets built from day one, or whether it develops later after expensive lessons.

Radical decentralization

Constellation's head office is absurdly small for a company of its scale. Leonard pushed both operations and capital allocation down to operating groups and individual business units. He described this as "delegation to the point of abdication" - and it worked because the rules were simple and the incentives were right. Head office did four things: set a few clear rules, taught capital allocation, facilitated best-practice sharing, and stayed out of the way.

The three-platform structure at Thrive already reflects this: Crete runs accounting, Shield runs IT, Long Lake runs community management. Each platform operates autonomously. HoldCo sets direction, allocates capital, and shares the AI infrastructure. The question is whether the platforms get pushed to allocate their own capital over time - that is where Constellation's compounding really accelerated.

Recurring revenue is what you actually own

Leonard tracked maintenance revenue growth, customer attrition by product and vertical, and net revenue retention obsessively - because they tell you whether your intangible assets are appreciating or decaying. At Constellation, customer attrition in non-cyclical verticals runs about 4% annually. The average customer stays for 26 years. That is a business you can underwrite with confidence.

All three Thrive verticals share this trait. Accounting firms run on monthly retainers and annual engagements. MSPs charge per-seat, per-month. HOA managers bill per-unit, per-month. The recurring revenue is already there. The metrics to track are the same ones Leonard obsessed over: net retention, attrition by vertical, and maintenance revenue growth. The data infrastructure to measure them should be a first-year priority.

These ideas are not complicated. They are just hard to execute with discipline over long periods of time. The question is what happens when you apply them not to software, but to service businesses - and with one ingredient Leonard never had.

Rather than licensing AI tools to the private equity firms that already own America's service businesses, Thrive Capital decided to own and operate the businesses directly. The logic was straightforward: if you believe AI transforms service delivery, you want to be on the inside, not selling shovels to people who don't think there's gold.

Josh Kushner launched Thrive Holdings last year as a $1 billion permanent capital vehicle built to own and operate businesses in industries like accounting, IT services, and community management - and use AI to make them fundamentally better. The idea grew out of a desire for Thrive Capital to look more like the companies it invests in. Not just funding software startups, but owning and operating real businesses with real employees and real customers.

OpenAI came in not as a vendor, but as an equity partner - embedding engineers and researchers directly inside portfolio companies to build AI tools on top of the actual workflows, the actual data, the actual edge cases these businesses encounter every day.3

The core belief behind Holdings is worth stating plainly, because it is the thing that makes it different from every other PE firm buying service businesses right now. For decades, technology transformed industries from the outside in - a startup builds software, sells it to incumbents, eventually replaces them. The thesis here is that the AI paradigm will be different. The data and the domain experts who fine-tune the models already exist inside these businesses. The accountant who has reviewed 10,000 K-1s knows what the model needs to learn. The IT technician who has triaged 50,000 tickets knows what auto-resolution should look like. Innovation will occur from the inside out.

The permanent capital structure is the forcing function that makes this possible.4

Three platforms. Three industries. The same playbook.

All three of these industries share a few things: aging owners, a shrinking talent pipeline, customers who can't go without the service, and massive fragmentation that creates a steady supply of willing sellers. That's the roll-up setup. It's also the setup that's attracted every PE firm in America, so the question isn't really "are these good businesses to own" - it's what Thrive can do with AI that a plain-vanilla consolidator can't.

Accounting - Crete Professionals Alliance

The U.S. accounting industry is a $145 billion market, and 81.7% of it sits outside the Big Four. It is also in demographic free-fall. Over 300,000 accountants have left the profession since 2020 - a 17% decline in the workforce. Only 1.4% of college students choose accounting as a major, down from 4% a decade ago. CPA exam candidates have dropped 27%. There are over 200,000 unfilled positions. Seventy-five percent of the current CPA workforce are Boomers approaching retirement, and most firms have no succession plan.5

This is the demographic crisis that makes the roll-up possible - partners must sell because they cannot replace themselves. Private equity has noticed: tens of billions deployed across 147 deals since 2020, per CPA Trendlines. But most buyers just consolidate back-office and wait for margins to improve. The Crete thesis is different: AI doesn't just cut costs, it creates new capabilities.

Full-population audit testing. Today, auditors test a statistical sample of transactions. AI can test all of them. PKF O'Connor Davies built a digital worker that logs into 100+ investment accounts, pulls consolidated data, and reconciles every line to the general ledger. The audit becomes more defensible, not just faster - and a more defensible audit is worth more to the client.6

K-1 processing at scale. K1x uses AI to digitize and process K-1s, K-3s, and 1099s - 90% reduction in manual data entry, 66% faster processing. For firms that handle hundreds of partnerships, this alone transforms the economics of tax season.

New revenue lines that didn't exist two years ago. Continuous close replaces the month-end scramble with real-time financial monitoring. Firms can now sell always-on cash flow analytics and fractional CFO services to SMB clients. These aren't cost savings - they're new products at new price points, and AI is what makes them economically viable to deliver.

The talent multiplier. A junior accountant with AI tools handles data entry, transaction coding, and reconciliation prep - freeing the human for oversight, judgment, and client relationships. Firms report saving 7 weeks per employee per year. That doesn't mean fewer employees. It means each employee serves 2-3x the client load, which is exactly what you need when 300,000 accountants have left the building.

IT Services - Shield Technology Partners

If you've never heard the term MSP, you've probably used one. Managed service providers are the companies that keep your office internet running, your laptops patched, and your data backed up. There are somewhere between 150,000 and 200,000 of them in the U.S., though per MSPAlliance only 5,000 to 10,000 meet any kind of verifiable maturity standard. There's no licensing body. You want to be an MSP, get a website. The market is roughly $64 billion, headed toward $100B+ by 2030.7

The interesting thing about Shield is that the AI thesis here is more concrete than in accounting. Most L1 IT work - password resets, license renewals, cache clearing, patching - is genuinely automatable today. Platforms like Mizo, ConnectWise Sidekick, and Atera claim 50-80% auto-resolution rates on these tickets. Even the conservative end of that range changes the math.

But the bigger play is vertical specialization. Generic IT support is a commodity - there's almost no switching cost. A healthcare MSP that offers continuous HIPAA compliance monitoring, though, commands 15-30% higher rates and creates a structural reason to stay. Same logic applies to construction, energy, and legal, each with specific compliance regimes. The question is whether one platform can credibly go deep in all of those simultaneously, or whether focus wins.

The valuation implications are worth spelling out. Small traditional MSPs trade at roughly 5x EBITDA (the market median across 120 disclosed deals is closer to 9x). If you can demonstrate that revenue grows without proportional headcount growth - which some AI-forward MSPs are beginning to show, though the data is still thin - multiples push into the 10-15x range. That's the rerating Shield is betting on.8

Community Management - Long Lake

Seventy-seven million Americans live in communities with an HOA. There are 373,000 associations. FirstService Residential and Associa - the two largest national players - hold just 11% combined market share. Everyone else is local.9

HOA management is dues collection, maintenance coordination, vendor management, and fielding complaints about the color of a neighbor's fence. It's contractual, recurring, and essential. It also might be the hardest of the three verticals to transform, because the moat is a single property manager's relationship with a single board president - and that person can walk out the door and take their clients with them tomorrow.

The AI opportunities here are real but more incremental than in accounting or IT. A resident asks about paint policies at 2 AM - AI searches the CC&Rs and returns the exact rule. Platforms like Stan AI and HOAi already do this. Budget prep that used to take hours compresses to minutes. Predictive dues collection - flagging likely delinquencies before they hit 60 days past due - shows 15-20% reductions in delinquency rates in early deployments, with maybe 70% of past-due accounts resolved through automated outreach before a human ever gets involved.

The common thread: one property manager handles three communities instead of one or two. That's not nothing. But the question I keep coming back to is whether the property manager even wants AI tools, or whether they see technology as a threat to the relationships that keep them employed. HOA boards are not known for their enthusiasm about change.

The point of the structure is not diversification. It is compounding.

Acquire a business with disciplined underwriting and high hurdle rates. Integrate it into the platform's standardized data backbone. AI-enable its operations - not by bolting on a chatbot, but by embedding models trained on the actual workflows, documents, and edge cases the business encounters daily. Watch margins expand and new service lines emerge. Redeploy the compounding cash flow into the next acquisition. Each turn of the flywheel makes the next one faster, because every new firm that plugs in makes the AI engine smarter and the integration playbook sharper.

The obvious risk is that AI commoditizes everything. If building features becomes trivial, what stops a competitor from replicating 80% of your tools in six months? If you're a software company, this is an existential problem. Your moat erodes as complexity gets cheap.

But these aren't software companies. They're accounting firms, IT providers, and HOA managers. The things that make them hard to leave - regulatory workflows, fiduciary obligations, the fact that someone has to sign off on a $50 million audit and take responsibility when the numbers are wrong - don't get cheaper to replicate when AI improves. If anything, the liability layer gets more valuable as the commodity layer gets cheaper. An AI can generate an audit workpaper. It cannot go to prison for fraud.

That's the argument, anyway. I find it mostly persuasive but not fully. The question I'd push on: does this logic apply equally to all three verticals, or is it really an accounting argument being stretched to cover IT support (where switching costs are genuinely low) and HOA management (where the moat is a personal relationship, not a regulatory obligation)?

This is where the "inside out" thesis either works or doesn't. When AI gets trained on the actual data inside these businesses - 10,000 K-1s, 50,000 IT tickets, years of HOA governing documents - the resulting models should be better than generic ones. Each acquisition adds more training data. Each integration refines the playbooks. That's the theory, and it's a good theory. The part I'm less sure about is how long it takes for the data advantage to actually compound into something a competitor can't replicate by just buying similar businesses and doing the same thing six months later.10

The piece that makes all of this work is tripwires. Leonard started tracking Initiative IRRs in 2004 and found returns were "nowhere near as good as originally hoped." He killed the poor ones. The discipline improved returns on the survivors. That same rigor matters for every LLM lab investment across the Thrive portfolio: IRR model, kill date, post-mortems that actually get read. AI projects have a way of surviving on vibes because they feel innovative. Tripwires keep capital honest.11

If the thesis is "hold forever," then the question isn't what these businesses look like today - it's what they can become with sustained LLM lab investment and disciplined capital allocation over the next decade. Here is a practical view of where value gets created, and the harder questions that permanent ownership forces you to answer.

Where LLMs create value - by business

The "Type" column matters. Cost savings are good but temporary - competitors will eventually match them. Revenue opportunities create new lines that justify the LLM lab investment. Moat opportunities produce a fundamentally different product or experience that raises switching costs. Scale opportunities decouple growth from headcount. A portfolio that leans too heavily on cost savings will compress margins industry-wide. A portfolio that builds moats and new revenue will compound.

The big questions for permanent owners

Holding forever means you cannot exit your way out of a mistake. These are the questions that matter most for each business if you are never selling.

I don't have answers to most of these. I'm not sure anyone does yet - it's early. But the fact that the questions are hard is itself a good sign. The businesses that compound over decades are the ones that confront ugly questions early instead of papering over them with narrative. Leonard figured that out around 2004, when he started tracking initiative returns and didn't like what he found.

Constellation compounded at roughly 30% a year for three decades. The model was never secret - Leonard published it in shareholder letters anyone could read. The hard part was doing it, every day, for 30 years. Thrive has the structure and the capital. Given its track record, I have no doubt it has the patience.