Summary

Sosin believes Carvana now offers its most straightforward risk-adjusted setup yet: an enormous addressable market and margins well in excess of 2.5-3 times the industry’s on one side, a better-capitalized business able to trade price for volume on the other. His illustrative long-range math is deliberately huge—$3,000-$3,500 of net income per car across 20 million cars equals $60 billion in today’s dollars, perhaps 10-20 years out. More immediately, he argues that shocks might reduce profit or growth, but “they won’t lose money.”
Carvana’s nearly $2,000 of financing profit per vehicle is not simply an anomalous subprime gain: roughly two-thirds of its loans are prime, dealers ordinarily earn for originating loans, and vertical integration captures economics that otherwise accrue to finance companies. The genuinely debatable advantage is closer to a subset of roughly $800 per car. Sosin attributes that remainder to somewhat higher prime rates, lower vehicle prices and LTVs, higher subprime down payments, better cars, and possibly superior online underwriting—while stressing, “I’m not 100% sure that that’s the case.”
Alternative data helped Sosin understand Carvana’s business but did not make the stock, which Walker framed as traveling from roughly $300 to $3 and back toward $300, easy to trade. Daily sales contain weather, seasonality, and logistics noise, while knowing an earnings result does not reveal the market’s reaction. Its best use is slower validation—whether Carvana consistently offers the cheaper make-model-trim-year-mileage combination—not forecasting every weekly move: late 2022 instead became “water torture.”
Carvana’s 2022 collapse showed how a correct structural thesis can coexist with worsening operating facts that are difficult to interpret in real time. Omicron stretched delivery times from roughly three or three-and-a-half days to seven and cut displayed inventory from about 90% to half, initially explaining weak conversion; when logistics recovered, underlying demand still did not. By early 2023, falling headcount, stabilizing units, a visible delivery line, and the banking crisis’s support for a strong supposition that lending markets were no longer firming rapidly and would normalize finally suggested that the setup had turned.
Walker’s hardest downside cases—a 30% used-car contraction, a compelling $20,000-$25,000 EV, and autonomous fleets replacing ownership—do not strike Sosin as existential, though sustained vehicle-price deflation would shrink Carvana’s market. His cyclical stress case uses estimated price elasticity of negative 7 to negative 8: a 4% price reduction could offset a 30% demand decline, with perhaps six points of total price concession against an 11%-plus margin. For robotaxis, 30%-50% empty miles, wait times, storage, personalization, and heterogeneous consumer economics undermine the vision of “the same golf carts.”
Sosin treats concentration as a continuing sell decision, not a mechanical portfolio rule: trimming every winner above 30% guarantees that an investor never owns a truly transformative compounder at full weight. His warning is Stanford’s sale of Google around its IPO, but Walker’s pushback matters—Google, Amazon, Facebook, and Apple developed profit engines early buyers could not foresee. Walker also raises Walmart, Costco, and Home Depot; Sosin responds that robotics, self-driving trucks, and scale could keep strengthening Carvana’s core system.
His Herbalife exit illustrates how updating the reference period and monitoring a named tail risk can matter more than being right about the original controversy. Sosin still considers Bill Ackman’s pyramid-scheme claim “entirely incorrect,” but years of sideways performance displaced decades of teens growth as the relevant base rate; then effective GLP-1 weight-loss drugs triggered the exact risk he had identified. His process remains intentionally open-ended: let experts talk, “shut the fuck up,” and learn one brick at a time. Walker’s trite Cliff saying is to wait for the stocks to go up.

Deep dive

1. Public visibility made Carvana harder to own, but worth memorializing

Sosin describes investment publicity as “sort of a one-way ratchet”: followers can impede buying and selling, create an expectation that every move be explained, and form “a big crowd” ready to celebrate failure. He had therefore decided to stop appearing publicly.
Carvana’s near-death and recovery felt too unusual to let “disappear gradually into the mists of time.” After reaching out to Patrick, Sosin returned to Walker because of their five-year friendship and the chance to address finer points—“the details of retail margin”—for a more specialized audience.

2. Prime lending distributes capital; non-prime lending manufactures information

Sosin divides consumer lending into two different economic activities. Prime lenders identify borrowers with established repayment reputations, then compete largely through distribution; thinner margins and apparently safer loans encourage greater leverage, leaving the lender effectively “selling disaster insurance” against synchronized unemployment or other shocks.
Non-prime lenders search for “the good borrowers amongst the ones that were thrown out by the prime market.” They add value through underwriting information and behavior modification—engaging customers in ways that increase repayment—rather than merely performing distribution.
The higher rate is compensation for genuine work and risk, not necessarily predation. Many borrowers live with income roughly matching expenses, so an interruption or unexpected bill makes credit important; a lender that identifies the repayable customer can become “a valuable sort of partner” and improve that borrower’s life.
That combination of human behavior, finance, and banking created Sosin’s circle of competence. Complexity and historical blowups keep generalists away, while repeated study can reveal businesses whose economics are better than the market’s broad treatment of “subprime” implies.

3. Subprime blowups begin when repayment signals become self-reinforcing illusions

Sosin’s core warning is that lending is an information business, so “false or misleading signals can really screw you up.” In the late 1990s, non-prime auto and unsecured borrowers refinanced from one lender to another; each lender saw an apparently performing loan, while the industry collectively passed borrowers around “like a hot potato.”
Reported performance attracted more capital, which funded the refinancing that created the reported performance. The system “works great until it doesn’t”: once capital stopped expanding, lenders discovered that borrowers had been rolling debts rather than repaying them from sustainable cash flow.
Housing from roughly 2004-2007 repeated the mechanism through collateral. Rising prices appeared to validate borrowers, encouraged more lending, financed additional purchases, and drove prices still higher. Sosin’s general lesson is to identify the feedback loop that may be manufacturing the very evidence used to justify underwriting.

4. Carvana’s financing profit decomposes into ordinary and differentiated pieces

Carvana generates nearly $2,000 of financing profit per retail unit, but Sosin resists treating all of it as evidence of exotic subprime economics. Ordinary dealerships are paid to originate loans, while CarMax retains loans and earns the economics over time; selling those loans would produce a recognizable gain on sale.
Carvana does over-index to non-prime, yet roughly two-thirds of its originations are prime, depending on the date and definition. In prime lending it earns more than CarMax largely because it charges somewhat higher interest rates, although Sosin says lower vehicle prices can still make the consumer’s total transaction better.
The non-prime loans in CarMax’s business still get made—by a finance company such as Westlake, Sosin says. Carvana instead integrates that finance-company layer and captures economics that ordinarily sit outside the dealership.
ABS data provide a reality check: interest income less expected charge-offs, servicing expense, and funding costs leaves excess profit. With an average pool life of about two years, Sosin says Carvana’s roughly 9% non-prime gain on sale is supportable when those cash flows are capitalized on a discounted basis.

5. Better cars and richer digital signals might explain the remaining edge

Walker narrows the dispute: if an ordinary dealer might earn $1,200-$1,300 against Carvana’s nearly $2,000, the roughly $800 gap is material to a thesis built on superior scale and execution. Sosin narrows it further—the portion attributable specifically to better non-prime performance is only a subset of that difference.
Sosin’s operational explanation starts with vehicle quality. He says the primary reason a subprime borrower defaults is that the car breaks down; Carvana’s cars should fail less often, while lower prices, lower LTVs, and a higher down-payment mix reduce loss given default.
Online origination may also permit more verification with less friction. Every lender balances confidence against adverse selection: onerous requests drive good borrowers elsewhere, but Carvana can embed verification “in a click” inside an already streamlined purchase, financing, and delivery journey.
He offers illustrative—not specifically Carvana—digital signals: applicants with little phone battery have historically underperformed those with more, while Internet Explorer users once proved worse credits than people who downloaded Chrome. Such correlations might enrich underwriting, but Sosin preserves the hedge: “I’m not 100% sure” online data explain Carvana’s advantage.

6. Alternative data validates the flywheel better than it predicts the stock

Carvana lends itself to unusually detailed measurement, and Sosin hired a consulting firm to build more of it. Yet “this week’s sales are light” says almost nothing definitive about next week; weather, subtle seasonality, inventory availability, and delivery constraints introduce meaningful high-frequency variation.
His preferred analysis asks whether a customer seeking a precise make, model, trim, year, and mileage can find it at Carvana or CarMax, then compares total price in a randomly selected market after shipping fees and delivery time. Carvana wins “the vast majority of the time,” though the result moves as both companies change prices.
That evidence tests the structural proposition—selection, price, and convenience—not whether a quarterly estimate will move shares. Sosin recalls a blog about someone who hacked a law firm and stole earnings releases yet got only about 65% of the trades right: knowing the result is not the same as predicting the reaction.
Carvana demonstrated the distinction in 2024. Sosin felt prepared for several earnings beats, but similarly strong quarters produced opposite stock moves: Q3 rose and Q4 fell. Alternative data can make volatility easier to endure by keeping the investor connected to the business; it cannot make “life easy.”

7. Omicron obscured a demand collapse before operational data revealed recovery

Early in 2022, Omicron disrupted Carvana’s logistics. Average delivery lead times expanded from roughly three or three-and-a-half days to seven, and customers saw about half the inventory rather than the roughly 90% available in a healthy system; weak conversion therefore had an obvious temporary explanation.
The difficulty was that Omicron also masked weakening underlying demand. Logistics improved, but sales did not rebound; competitors charged interest rates Sosin believed were uneconomic, yet he could not know when that behavior would end. Selling because an unsustainable distortion persists and buying because it must eventually disappear can both sound rational.
By late 2022, sales weakened almost every week. The information advantage became “water torture”: instead of suffering one bad quarterly release, Sosin received another daily drip showing that the company genuinely was not performing well.
The 2023 turn appeared through several imperfect signals together. Headcount was falling, units finally stabilized, and a delivery line had emerged; meanwhile, the regional-banking crisis supported a strong supposition—before the data arrived—that lending markets were no longer firming rapidly and would normalize. Sosin saw the setup and told Walker, but a poison pill prevented him from buying more.

8. Carvana’s long runway supports an extreme—but explicitly distant—upside case

Some 40-some-odd million cars are sold annually in the United States, and Sosin considers most of that market addressable, with new cars offering another potential avenue. Scale improves the experience, and “there’s nothing like it,” so he expects Carvana eventually to sell “many, many millions” of vehicles.
Margins are already well in excess of 2.5-3 times those of the industry despite rapid growth. Fixed-cost leverage remains available, as do granular improvements such as deciding which parts to replace during reconditioning by balancing immediate cost against warranty and vehicle-service-contract expense.
His illustrative destination is $3,000-$3,500 of net income per vehicle multiplied by 20 million units, or $60 billion in today’s dollars. Sosin places that outcome perhaps 10, 15, or 20 years away and emphasizes the cash generated along the path; it is a framework for magnitude, not a near-term forecast.
Walker’s challenge is that famous “never sell” winners—Google, Amazon, Facebook, and Apple—developed major businesses their original investors could not anticipate. He also raises Walmart, Costco, and Home Depot. Sosin responds that technology, including robotics and self-driving trucks, is working in Carvana’s direction and could make its advantage over dealerships bigger.

9. Margin advantage now gives Carvana room to absorb an industry shock

Sosin says Carvana’s resilience is “far greater than it’s ever been.” Competitors generally cannot sustain losses indefinitely, while Carvana’s higher margins let it choose how to trade off price and volume; an evenly apportioned industry shock should hurt results without recreating the old solvency risk.
Used-car demand is more stable than many assume: 2022’s roughly 20% contraction was the worst decline on record, while the Great Recession produced a decline in the teens. Walker nevertheless asks Sosin to stress a 30% fall combined with reverse fixed-cost leverage.
Sosin estimates Carvana’s price elasticity at perhaps negative 7 to negative 8, while admitting it is uncertain. In that framework, a 4% price cut could recover roughly 30% lost demand; allowing another couple of percentage points for industry price compression produces about six points of concession.
Against an expected margin of “11 and change” percent for the year, that scenario leaves Carvana near the average competitor’s present margin. It would be unpleasant, but competitors would simultaneously be “losing lots of money and disappearing at breakneck speed,” limiting how long the adverse pricing environment could persist.

10. Cheap EVs threaten long-run market size more than today’s inventory

Walker imagines a compelling $20,000-$25,000 EV making Carvana’s gasoline inventory obsolete and leaving too little value for used-car intermediation. Sosin separates an immediate inventory shock from sustained real-price deflation and considers the overnight stranding scenario implausible.
The United States has roughly 300 million vehicles, and even a spectacular manufacturer cannot replace a meaningful share quickly. A superior low-cost EV might lower expectations and used prices at the margin, but production constraints would make the transition unfold over many years.
Carvana holds only about $4,000 of inventory for every annual vehicle sold because it turns inventory roughly six times. A 10% price shock therefore costs about $400 per annual unit, versus roughly $4,500 of incremental margin and prospective per-unit EBITDA in the mid-$3,000s, eventually approaching $4,000.
Sosin concedes the limiting case: if new cars fell to a few thousand dollars and became disposable, “that would be bad for Carvana.” A future with $30,000 new cars would merely reduce industry size at the margin, partly offset by households owning more vehicles or replacing them more frequently.

11. Autonomous driving may strengthen personal ownership instead of destroying it

The standard robotaxi case says idle privately owned cars incur depreciation and capital costs, so shared autonomous fleets should reduce cost per mile. Sosin’s rebuttal is deadhead mileage: Uber, taxis, and even long-haul trucking can run 30%-50% of miles empty because vehicles must reach passengers and reverse directional commuter flows.
Empty driving is more expensive than stationary ownership. Once deadhead miles, cleaning, payments, and fleet overhead enter the calculation, pooled autonomy may be slightly cheaper or slightly more expensive, but the supposedly overwhelming cost advantage “largely disappear[s].”
Average economics also hide customer heterogeneity. A price-sensitive driver can buy an eight- or ten-year-old Toyota with minimal depreciation and capital cost; a new BMW 7 Series buyer knowingly pays more for quality. A shared fleet cannot simultaneously beat the Toyota on cost and satisfy the BMW buyer’s preferences.
Waiting several minutes for a 15-20 minute journey carries meaningful time and planning costs, while cars store child seats, toys, tools, gym bags, and shopping. Self-driving may turn the car into “a personal room,” strengthening ownership; it could even reduce Uber usage when one’s own car can handle drinking, parking, or an airport drop-off and then drive itself home.

12. Selling, interviewing, and idea generation all require resisting premature closure

Sosin uses Stanford’s sale of Google near its IPO to argue that sell decisions deserve the same work as buys. Mechanical trimming may reduce risk, but anyone who automatically diversifies above 30% can never experience a Berkshire-scale winner; Carvana’s concentration therefore remains an underwriting judgment, not a rule.
Herbalife shows the opposite decision. Sosin bought after Bill Ackman’s pyramid-scheme attack, which he calls “wildly” and “entirely incorrect,” but by 2021 the sideways 2014-2021 record deserved more weight than decades of teens growth. His unproven explanation is that gig work weakened MLM recruitment, where losing 2% of distributors compounds through the network.
Effective GLP-1 weight-loss results then activated a tail risk Sosin had explicitly identified years earlier, so he sold Herbalife—while ruefully failing to buy Novo Nordisk or Eli Lilly. Experiences with Celanese and Ashland similarly increased his respect for forecasting difficulty and the need for valuation’s margin of safety.
His expert-call method follows Robert Caro: ask open-ended questions, write down follow-ups, and write “STFU” in his notepad while the source keeps talking. The same openness governs research—health insurance, life sciences made more approachable by AI, or even a mistakenly downloaded 10-K. “One brick at a time,” the objective is simply to learn something every day while “waiting around for my stocks to go up.”

Episode 140 - May 2, 2025 Apple and the Power of Platforms, AI Comes for Google Search, ChatGPT Comes for Higher Education