The YC Summer 2026 company buys straight from Chinese factories and delivers to US importers at one price, duty paid. Its trick isn't secret - it's public. It just bothered to look.
Every shipping container that clears an American port leaves a paper trail. Who sent it, roughly what was in it, which company received it - filed, logged, and by law, public. For decades that trail sat mostly unread while US importers paid trading companies to stand between them and the factories that actually made their goods. Donkey, a company in Y Combinator's Summer 2026 batch, is built on the idea that the trail was worth reading all along.
The pitch is blunt. Send Donkey one photo of a part you reorder. Within 72 hours, it names the factory that actually makes it and comes back with a price delivered to your dock in US dollars - duty paid, inspection included - that beats your current landed cost. If it can't beat your number, you keep your supplier. No switching costs, no retainer, no mystery.
Framed that way, Donkey sounds like a sourcing shortcut. It is closer to a quiet rearrangement of who holds the leverage in a container of goods. For most of the last century, that leverage sat with whoever controlled the information: the trading company that knew the factory, the agent who spoke the language, the broker with the relationship. Donkey's wager is that the information has leaked into the open, and that a company built to harvest it can hand the leverage back to the buyer.
Most American importers do not buy from factories. They buy from trading companies, which buy from factories, and the difference between those two prices is where a container's economics quietly disappear. The markup runs 15 to 30 percent, and the arrangement comes with a feature that is really a bug: the trading company never tells you which factory made your goods. That opacity is the product. If you knew the factory, you would not need the middleman.
For the importer, this means paying for coordination they cannot see and cannot audit. Prices drift. Quality varies. And every reorder runs back through the same black box, because finding the factory yourself - across language, distance, and thousands of near-identical suppliers - is harder than just paying the markup again.
Consider the small importer who has reordered the same bracket, fitting, or housing for years. They have a supplier they email, an invoice they pay, and a shipment that shows up. What they do not have is any way to know whether the price is fair, whether the plant they imagine exists, or whether the same part left the same loading dock at two-thirds the cost under a different label. The not-knowing is expensive precisely because it is invisible. It never shows up as a line item; it shows up as a number that is simply higher than it needed to be.
Here is the fact the whole company turns on: US customs records are public. Every import filing leaves a trace of who shipped what to whom. Donkey reads those filings, matches the part in your photo to the factory that has actually been shipping it, and flags exactly where the trading-company markup sits. The data was always there. What was missing was a company willing to parse it at scale and act on what it found.
The parsing is not trivial. Customs data is noisy, and a filing that names a shipper does not tell you whether that shipper is a real factory or a shell company reselling someone else's work. Donkey's software is built to tell the difference - to separate the plant with a stamping line from the trading office with a nice logo. That verification is the quiet core of the machine, and it is where the AI in "AI-native" actually earns its keep: not as a chat window, but as the thing that reads a million dull filings and finds the one factory that matters.
You send one image of a part you already reorder from China.
Donkey matches it to the real factory using public customs data.
A delivered, duty-paid US price - or you keep your current supplier.
Donkey's own inspectors clear goods before the factory is paid.
Donkey is not a broker taking a finder's fee. It takes title. It buys the goods, runs the inspection, moves the money, and sells the shipment on to the importer delivered and duty-paid. That structural choice matters. A broker is incentivized to close a match and move on; a principal that owns the container is on the hook for whether the goods are right, whether they arrive, and whether the price held. Donkey chose to carry that weight.
It also carries the financing. Qualified buyers get credit-insured net-45 terms, meaning Donkey pays the factory and waits to be paid by the importer - absorbing the gap and the risk. Bundle it together and the offer to an importer is unusually complete: factory identification, factory-direct price, inspection, freight, customs, and credit, quoted as one number.
Owning the goods also puts Donkey's own money where its quote is. When a company merely introduces a buyer to a factory, a bad inspection or a late shipment is someone else's problem. When it takes title, a defective run is inventory it is holding. That alignment is uncomfortable by design - it forces the inspection to be real, the factory vetting to be honest, and the reliability data to be more than marketing. The credit terms compound the effect: a lender that eats the loss on a bad buyer learns quickly to underwrite carefully.
The old trading houses ran this coordination with enormous headcount - tens of thousands of people managing relationships, quotes, inspections, and logistics by hand. Donkey runs the same coordination with software and a team of seven. But the more interesting difference is what happens after each deal closes. Every settled order feeds the pricing and reliability data that the next quote runs on. The system that reads the customs records gets sharper as it does more of them.
That is the part competitors can't easily copy by hiring. A traditional agent's knowledge lives in a person's head and leaves when they do. Donkey's lives in the data layer and accumulates. It is the difference between an address book and a map that redraws itself after every trip.
Donkey is led by Benjamin Martindale, its co-founder and CEO, alongside co-founder and CTO Minghao Tan. Martindale is not new to Chinese manufacturing. He previously built Prelude OS, which reached 31 paying factory customers in Guangzhou, and advised Chinese heavy-equipment makers including Sany and Sinoboom on expanding overseas. He is fluent in Mandarin and holds a finance degree from Rotman Commerce at the University of Toronto. Tan builds the data infrastructure that parses the customs filings and does the harder work of telling a legitimate factory from a shell.
The team is split across San Francisco and China, working in English and Mandarin - which for this particular business is less a nice-to-have than a requirement. You cannot quietly rewire the plumbing between American buyers and Chinese factories from only one side of the ocean.
Cross-border sourcing from China to the US is enormous and, at the buying end, still stubbornly analog. The incumbents are traditional trading companies and sourcing agents that live on the markup and the opacity, plus large marketplaces where the same shell-versus-factory ambiguity persists. Donkey's wedge is the combination almost nobody offers as one package: identify the real factory from public data, own the goods, inspect them, clear customs, and finance the buyer.
It fits into the broader wave of AI-native companies that win not by inventing new data but by finally reading the data that already existed. Customs filings, sitting in the open, are exactly that kind of overlooked resource. The bet is that transparency, once offered, is hard to take back - once an importer has seen their factory named and their markup quantified, going back to the black box is a hard sell.
There is risk in the model worth naming plainly. Taking title means holding inventory risk. Financing buyers means holding credit risk. Reading customs data means the matching has to be accurate enough that a wrong factory does not become a wrong shipment. None of that is trivial, and the whole thing lives in a corridor - China to the US - where tariffs, policy, and logistics can shift underfoot. Donkey's answer is that the same data that finds the factory also prices the risk, and that each closed order tightens both. Whether that holds at volume is the question the next year will answer.
What can an importer actually do with it? Audit a reorder they have run on autopilot for years. Learn, for the first time, the name of the plant behind their most-shipped part. Take a delivered, duty-paid number they can put straight into a spreadsheet without decoding freight, FX, and customs line by line. And if Donkey can't beat their price, walk away no worse off. For a category that has run on not-knowing, that is the whole idea.