Inside General Catalyst’s $1.5B AI Roll-Up Machine

Summary

• General Catalyst is deliberately restructuring from a venture firm into “a company,” pairing traditional funds (Ignition, Endurance, a customer value fund, Creation) with permanently-held “transformation companies” — an Ohio hospital system via HatCo, a founder wealth-management arm, and AI consultancy Percepta. Marc Bhargava frames GC as “kind of one of the three largest venture players,” with every piece answering one question: “what are the best ways to work with the most founders and help them the most?”
• The Creation strategy put about $1.5 billion to work last fund cycle incubating applied-AI companies, with roughly half converting into AI-enabled roll-ups that “buy their distribution, buy their client list, and the data that comes with it.” GC invests $100–150M per project over three to four rounds, leads at least the first two, and has won eight of nine founder offers in the space — flagship names: Long Lake, Udia, Titan MSP, Beacon, Crescendo.
• The screen was systematic: 70 services industries mapped against four AI-automatable buckets — customer support/service, data entry, content/copy, and “basic logic and reasoning,” which has emerged in the last nine months — yielding just 10 industries where 20–30% of tasks (not people) can be automated. Proof points: Crescendo automates “50 to 70%” of what a call center does and is well on its way to turning 10% EBITDA margins into 40%; Titan MSP showed 38% ticket automation across six pilots before buying New York MSP RFA.
• The core arbitrage claim: services is a $16 trillion global market versus $1 trillion for software, and AI can let these businesses earn software-like margins. “Once you take out thirty, forty, fifty percent of what people do and free them up to do the hard stuff, you can double the revenue.” Case studies include roughly 10–15% EBITDA businesses reaching about 30% within a year while keeping costs flat, so “the pool of one trillion that VC has been fishing in has actually expanded… to multiple trillion.”
• This is explicitly not private equity: GC screens hard for whether a company wants to change and implement AI, holds seven to ten years toward an IPO, and adds costs (engineers, LLM subscriptions) rather than simply adding debt and cutting costs. The end-state vision is “AI-native compounders” — next-generation TransDigm, Danaher, and Constellation Software — with plenty of $100B public companies a decade out, founders owning 10–30% at IPO and GC ideally 20–25%.
• The discipline that makes it tradeable: target at least 30% automation but “we actually don’t want more than 70%” — at 80–100%, a software or agentic solution is more likely, potentially pushed by an incumbent like Google or Microsoft through its existing distribution. Add fragmentation (HOA management is “extremely hard to sell into”) and stickiness (two-year contracts, low-churn accounting/insurance/MSPs) and you have the roll-up filter.
• Bhargava calls AI roll-ups underrated and dates the moment to “2016 in crypto” — GC, Elad, Thrive, and 8VC believe; most of the market is politely skeptical. Some companies in the roll-up “will hit a hundred million of EBITDA, and they’re, like, less than two years old” yet stay stealth because “if you’re generating hundreds of millions of free cash flow, you don’t need to be out fundraising every three to six months… and promising people AGI is around the corner.” Over the next year, GC plans to make the thesis much more public and begin preparing companies with crossover firms to go public.

Deep dive

1. GC is now “a company,” not just a fund complex

• Bhargava’s opening architecture: GC builds in-house “transformation companies” it may hold “maybe forever” — HatCo, which incubated and then bought a hospital system in Ohio; a wealth-management business for founders; and Percepta, an AI consultancy that “changes Fortune 100 companies” using everything GC has learned from AI investing. Alongside sit the funds: Ignition (early stage), Endurance (later), a customer value fund for low-dilution scaling, and Creation, “all about manufacturing outliers.”
• The connective logic is founder gravity: a portfolio founder can use GC’s wealth arm, hire Percepta to build “their own mini-model or inference,” or sell into the Ohio hospital system to test healthcare tech. “There are more ways to partner with GC than any other firm out there” — and Bhargava claims that flywheel, not any single fund, is what’s lifting the brand.
• Creation’s distinction from the transformation bucket: incubated companies GC wants to take public in seven to ten years, typically businesses that don’t exist because “there’s not enough capital for it, or they’re multidisciplinary… or it’s a bit more traditional or outside the box.”

2. The roll-up thesis is an all-of-the-above bet

• Molly’s setup: what she described as an existential moment for SaaS in early 2023 “or something like that,” when “everyone thought venture was dead, SaaS was dead,” and GC leaned in hard. Bhargava’s answer is that three approaches can all win: the model layer (GC is a large Anthropic investor across the last and current rounds), SaaS itself (“it’s been overestimated. SaaS is not dead”), and a third bucket — fragmented industries where you build the AI-native platform “and then go buy our distribution.”
• Creation deploys about $1.5B, and incubations fork two ways: scale organically with growth capital, or — “in about half the cases” — fund acquisitions. Crescendo is the canonical example: GC teamed Andy Lee, who ran call-center chain Alorica for over 30 years, with two CTOs (someone from GC joined full-time), proved 50–70% task automation with about 10 pilot clients over a year, then funded buying a call center — now “well on their way” to taking 10% EBITDA margins to 40%.
• The target industries “don’t really buy technology products. When they do buy tech products, they call it IT, and they cap it at 2 or 3% of revenue” — one reason Bhargava sees buying distribution and the underlying businesses as useful alongside selling software.

3. From monkey JPEGs to a 70-industry screen

• Bhargava’s origin story, as told: crypto since 2016, built a business with Jen and Greg Tusar and sold it to Coinbase (“today, it’s the heart of Coinbase Prime”), then angel investing in Harvey, Together, and Windsurf — “if I was giving money to people putting monkeys on the internet, probably I should support these PhDs in AI from Stanford.” Second-time founders kept telling him the hard part was go-to-market in unsexy, low-trust industries — “it’s not the same as a Ramp or Brex selling into all the hot startups” — and the roll-up concept was co-conceived with founders behind companies like Long Lake and Crescendo.
• The systematization, a year and a half to two years ago: 70 services industries scored against four things AI does well — customer success, support and service; data entry and evaluation; content/copy/marketing (FAQs, NDAs, translation); and, “which has only emerged in the last nine months,” basic logic and reasoning, e.g. insurance underwriting recommendations “almost as good a job as a human.” Sixty of seventy failed the bar; ten survived with conviction that 20–30% of tasks are automatable.
• The margin math as he runs it: keep cost basis flat, free up 20–30% of worker time for more clients, cross-sell, and harder tasks — “around ten or fifteen percent EBITDA margin businesses can become thirty percent EBITDA margin businesses… in a one-year period.” Titan MSP is the template: seed/Series A to build, six pilots proving 38% automation, second round to buy RFA, third for tuck-ins.

4. A $16T TAM with software margins — held like a compounder, not a PE asset

• The sizing: services globally are about $16 trillion versus about $1 trillion for software — “sixteen times larger” — while many services businesses have historically been near break-even or otherwise not very profitable. Bhargava’s thesis is that AI can change that: strip the repetitive tasks and “your margin profile now looks much, much more similar to software,” delivered by a mix of “AI-native software, agentic workforces, people working for agents, agents working for people.”
• The exit model is the public-market compounder — TransDigm, Danaher, Constellation Software: $100B+ names that buy, improve, and reinvest free cash flow. Bhargava’s variant: “AI-native compounders” where the driving improvement is automation rather than pricing or sourcing, and he expects “plenty of these $100 billion AI-native compounders in the public markets 10 years from now.”
• The anti-PE stance is explicit and structural: no three-to-five-year flip incentive, hard screening for businesses and people that want AI-driven change from the outset, and — the inversion worth keeping — “you can’t go in, add debt, cut costs. We’re actually going in, and we’re adding costs. We’re saying, ‘Hire engineers.’” Worth it because the added cost is expected to be much smaller than the added revenue.

5. When to roll up vs. when to just fund SaaS — and how performance is scored

• Answering Kyle Harrison’s listener question, Bhargava gives a three-part filter: percent automation (at least 30%, “we actually don’t want more than 70%” — above roughly 80%, software or agentic solutions are more likely, and an incumbent like Microsoft, Amazon, or Google may have the distribution to push them); fragmentation that makes SaaS sales “very, very slow”; and stickiness — HOA management’s two-year contracts, low-churn accounting, insurance, and MSPs — to absorb implementation hiccups.
• The stage-gated metrics: first, demonstrate 30% automation of man-hours with 5–10 pilot clients; second, a detailed plan to double EBITDA margin from 15–20% to 30–40%; third, entry-multiple discipline via tuck-ins and debt; fourth, dilution — founders should own “at least 10% of their business when they go public, ideally 20 or even 30,” while Bhargava says GC could ideally hold 20–25%. Molly’s reaction: “That’s highly concentrated.” Bhargava: “Yeah, for sure.”
• On implementation risk, he cites an MIT study and says maybe 95% of Fortune 100 companies have tried AI but found implementations ineffective — “kind of a waste” and “janky.” He argues effective implementation needs an AI-native team; he looks to people who may have led applied AI at Rippling, Scale, or Stripe, or worked at places like Figma, alongside traditional client bases — “we don’t necessarily think that the world is going straight to AGI.” Anthropic informs where models are going; portfolio bets like Rocks (vs. Salesforce), Udia (legal), and Serval (vs. ServiceNow, out of stealth last week) feed Percepta’s Fortune 100 work.

6. On the labor question: squarely the “more powerful teams” camp — with a caveat about outsourcing

• Pressed on Coatue’s Michael Barton framing — AI shrinks teams versus AI supercharges them — Bhargava is “pretty squarely in the second camp.” His specimen: Hippocratic AI, GC’s incubated AI nurse, where “one nurse could manage a team of five AIs” against a massive nursing shortage. Someone three times as efficient needing “only a slight salary increase” doesn’t necessarily get fired — “you probably are hiring three or four more of them.”
• His hedged exception: “sadly, yes, some jobs might not be there anymore,” and the first things to go are outsourced repetitive work — legal, call centers, insurance claims — so the pain lands “probably abroad in places like India, Philippines… even Mexico.” GC is committed to exploring retraining and reskilling, including whether it can invest in companies doing that. Abundance, in his telling, means legal representation, real-time accounting, and nine-to-twelve-month patient follow-up that today “just doesn’t make economic sense.”

7. Underrated, stealthy, and — to Bhargava — exactly where crypto was in 2016

• Geography as signal: of the eight or nine deals, six are in San Francisco and two or three in New York; the acquired businesses are almost never in those cities and are spread across the country. The strategy is increasingly global — Dwelli rolling up UK property management, a German accounting deal, and India under close study. On talent, the divide is “less the US versus international and more the San Francisco Bay Area versus everyone else.”
• His biggest lesson from Hemant Taneja: pay up and catch iconic companies early — a $100M check and a $1M seed check often both buy 10%, so GC has pushed hard into seed via the La Famiglia acquisition, India’s Venture Highway, and Yuri leading US seed.
• Asked to be unbiased on roll-ups: “I think they’re underrated right now… very few folks really grasp” that the VC pool has expanded to multiple trillions. Some companies in the roll-up “will hit 100 million of EBITDA, and they’re, like, less than two years old” — and stay silent precisely because they don’t need to fundraise or do “the press tour, promising people AGI is around the corner.”
• His closing analogy: he compares today’s believer set — GC, Elad, Thrive, and 8VC — with the few firms he says believed in crypto in 2016, naming USV, a16z, and Founders Fund. The rest of the market is “just trying to be polite.” Over the next six months to a year, GC plans to make the thesis much more public, work with more crossover firms, and prepare companies to go public.