The Investor

Why the unwind hits leverage before it hits the growth story

Pull-forward volume and A2L pre-buy revenue inflated the 2023 base, and OEM price pass-through inflated margin with it. Post-pandemic OEM price increases flow through a distributor's P&L on far better economics than volume growth does. The A2L refrigerant transition pulled orders forward while supply chains were snarled, which is the condition under which a distributor's pricing power looks strongest (and the condition that does not repeat). Strip roughly 15% of phantom EBITDA out of an entry model and the leverage multiple the deal was approved at rises by about the same proportion. No covenant gets tested.

So the existing book gets normalized before anyone looks at a new deal, and the allocation cost is real: diligence hours go to re-underwriting marks instead of sourcing. That base year sits in the mark and the covenant headroom, and in the exit assumption on any platform being taken to market.

Watsco's cost of capital wins the banked processes

The consolidation math is real: 2,000+ independent regional distributors against a leader holding only about 20% of residential systems. That leader carries very little debt and clears 15%+ ROIC, so it wins any banked process on cost of capital alone. The edge for a levered buyer is proprietary sourcing in states the leader has not entered, since 50-state coverage remains incomplete, from owners who sell into platform breadth and parts availability they cannot build themselves.

The same file names the terminal risk. E-commerce accounts for a third of the leader's revenue across a 120,000+ contractor network, which is the best public evidence anywhere of how deep workflow penetration runs in a trades channel and also the precise revenue base that OEM direct-to-contractor selling threatens. The digital front end is the moat and the disintermediation vector at once.

Shein prints the same lesson with a different distortion

The second mechanism is policy rather than pull-forward, and it clears at a harsher number. Shein's disclosures tie a shrinking U.S. business in the first quarter to the repeal of the sub-$800 de minimis exemption, the rule that let a cross-border parcel enter duty-free. Duty-free treatment was doing the work the merchandising got credit for. Cross-border commerce inherits that discount rate whether anyone volunteers it or not, and so do DTC-import and import-logistics positions. The listing price is the comparable now, rather than a 2022 round.

Revenue produced by a rule change is worth less than revenue produced by a customer, and a growth rate does not distinguish between them.

Where the reads agree, and what to discount

Both reads converge from opposite ends of the market: public comps have absorbed conditions that private marks still capitalize, and the index gives no cover. The S&P 500 rose 3.6% over a trailing one-month window with the Fear & Greed Index at Neutral. Idiosyncratic de-ratings inside a calm index are where stale private comps do the most damage, because nothing in the LP conversation forces the question.

The Watsco figures are relayed second-hand from charting services and investor-relations decks, so treat the drawdown and the ROIC as directional until pulled first-hand. And the bull case is an assumption stack: a $10B revenue target times an assumed 8% net margin times an assumed 25x forward multiple gives $20B and 56% upside, with no stated time horizon. The horizon decides whether 56% reads as attractive or embarrassing, and it is the one input the memo leaves out. That verdict may be too harsh, since a near exit year makes the arithmetic fine, but where the assumed exit multiple supplies most of the modeled IRR the same treatment applies.

The official rate now undercuts the ramps by four points

Until April 2025 an Argentine ramp sold access to a scarce dollar, which is a lovely business to be in and requires almost no product work. Since that change the legal official market sells the same dollar roughly 4% cheaper as of August 28, 2026, so the main competitor is the state, at par. What is left to sell is round-the-clock availability, payout speed, programmability and self-custody. Users still pay about 4% for that bundle, which is genuine willingness-to-pay and not the same thing as a moat, because global exchanges and neobanks can price a convenience spread down the moment they decide these users are worth acquiring.

The larger prize, or rather the more interesting one, is not crypto-native at all. Argentine households have hoarded physical dollar cash since the 2002 pesification, when the peg broke from 1 to about 4 pesos per dollar and roughly 75% of deposit value was destroyed. Converting mattress dollars into on-chain dollars is the biggest remaining growth vector in the country, and it is won with trust plus cash conversion (agent networks, offline on-ramps), which is precisely the capability crypto-native competitors are not building. Every quarter spent on order-book features is a quarter not spent on agents.

The payroll series and the wallet series disagree

The payroll data is macro-levered. The share of Argentina-based contractors paid in USDC rose with inflation to an April 2024 peak, when prices were running 289% year over year, then retraced to roughly a fifth of that peak by July 2026 as monthly inflation fell to 2.1%. That series was a bet on inflation, and inflation went to 2.1% a month.

The wallet data is product-levered. Lemon, one of the country's largest wallets, grew downloads every single quarter through the entire disinflation, which is the only clean evidence of durable demand in the dataset and argues for underwriting retained users rather than transaction flow. Context for both: 94% of peso crypto trading volume goes to stablecoins, the highest share of any major currency Artemis tracks, with roughly 9 million users on a 46 million population and top-15 app downloads up 93% in 2024.

Discount the source, then use it anyway

This is high-quality thesis marketing from an investor with direct exposure to the conclusion. The payroll dataset comes from Deel, a disclosed a16z portfolio company. Both series are indexed to January 2024 rather than absolute, which obscures actual adoption levels, and adoption itself is proxied by app downloads, which conflate new users, reinstalls and device churn. The author is an editorial intern, and a16z explicitly disclaims having independently verified the third-party data. None of that makes the argument wrong. It makes it a hypothesis with a marketing budget.

The falsifiable version of the habit claim is absolute stablecoin balances and 90-day retained actives, which is both the data request to send and the first question to ask any founder pitching this narrative. a16z has framed Argentina as the opener of a country-by-country series, so entry prices in Turkey, Nigeria and Egypt will drift up over the following two to three quarters as installments land. Argentina has also reversed FX policy at least three times since 2002, and stablecoins are valued there precisely because they sit outside the controls, which invites targeted regulation in either direction. The view breaks if the wallets are retaining balances rather than downloads, in which case the 4% spread is a floor and not a decay curve. It breaks differently if the state widens the official gap again and the scarce-dollar business simply returns. Underwriting the retained user survives both.

The demand floor sits 80% below the crisis peak, and a floor is still a floor. A business model built on a 100-point spread does not survive a 4-point one.

A 0.07-point spread is not a ranking

In a blind evaluation of 203 tasks scored by 163 expert judges, Tencent's Hy4 came in at 2.99 out of 4, Moonshot's Kimi K3 at 2.94, Zhipu's GLM-5.3 at 2.92. A spread that narrow across three labs sits inside human-judge noise, which leaves the license plus the price as the only things a buyer still holds once the press release scrolls off. Hy4 also arrived 53 days after its predecessor at 2.6x the size and 4x the context window, a cadence that depreciates any per-model differentiation claim on roughly a two-month schedule.

The same compression showed up in the least glamorous category

Document parsing is where this gets concrete for unit economics. Datalab's Marker v2 scores 76.0% on Allen AI's olmOCR-bench, a public harness of 1,403 PDFs, against MinerU's 72.7% and IBM-originated Docling's 50.3%, while running over 5x more pages per second. The engine is Surya 2, a single 650M-parameter model covering OCR, layout, reading order and table recognition across 90+ languages. At 2.9 pages per second in accuracy mode on one NVIDIA B200, that is roughly 10,400 pages an hour, or well under a tenth of a cent per page in raw compute. Commercial per-page ingestion APIs charge 10-50x that against a free download with a public benchmark receipt.

Allen AI owns the benchmark and also ships a competing parser, so reproduction on the buyer's own corpus is the price of quoting the number. And the 23.7 pages-per-second headline requires OCR disabled, which means model cost of goods comes off the accuracy mode rather than the marketing mode.

Where "free" stops

Free weights still bill for inference. A 770B model needs serious GPU capacity even at roughly 49B active parameters and FP8, which pushes most users back to hosted APIs: bullish for gateways and routers, bearish for the self-hosting addressable market. Ollama's v0.33 release makes the point with unusual precision, or rather the more interesting version of it, by inserting itself as a third-party gateway inside Claude Desktop, where one reversible toggle surfaces Qwen, DeepSeek, Kimi and GLM in Anthropic's own model picker, running on the user's hardware. Anthropic keeps the interface relationship and loses the inference revenue on every toggled session. The switch is where pricing power moved.

The demand side moved too. Uber published a full agentic cost decomposition, users times sessions times turns times requests times tokens times price, and took cost per session down 52% from peak. That is probably a floor on what a motivated procurement team can do with the same arithmetic. Every enterprise buyer now has a public playbook for halving agentic spend, which makes net revenue retention built on unmanaged token growth an optimistic input to any 2027 renewal model.

What survives, and where the reads diverge

Independent reads converge on model quality no longer paying rent, and diverge on speed. One argues custom-silicon inference deflation is a scenario rather than a base case, because the performance-per-watt numbers are vendor-supplied and long-context agentic workloads are untested. Another notes hosted price floors have drifted back up before, once free tooling needed support contracts. Both cautions are fair, and neither rescues a per-unit pricing model.

What holds value is the layer that certifies output and controls spend: confidence routing and human verification, domain schemas for legal, clinical and scientific ingestion, immutable audit trails, batching and scheduling architecture that scales throughput off one shared inference server rather than a model copy per worker, and provenance attestation over weights that originate almost entirely from Chinese labs. Verification is the billable part, which is why Marker's own documented failure mode has fast mode reading equations from the text layer while math accuracy "falls off sharply." Engineering time spent rebuilding the parser is time not spent on the schema and audit work that outlives it.

When the weights are free and the benchmark is public, the only defensible thing left to sell is proof that the output is right.