> ## Content Index
> Fetch the complete content index at: https://research.porbital.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# When Efficiency Runs Out: What PDD Got Right
- URL: https://research.porbital.io/when-efficiency-runs-out-what-pdd-got-right/
- Published: 2026-09-28T12:48:22.000Z
- Updated: 2026-10-17T07:43:30.000Z
- Description: PDD mastered algorithmic retail efficiency, yet its margin is narrowing. Comparing PDD with JD, Alibaba and Amazon reveals why a stronger business position may matter more than faster execution.
- Author: Roverzon
- Tags: AI and Productivity, #i18n-pdd, Platform Economics, PDD Holdings, E-commerce, China Consumption

PDD’s competitive advantage is its ability to bring buyers and merchants together at low cost. Recommendations, social sharing and low prices concentrate demand, while merchants and logistics partners carry much of the inventory and fulfilment burden. That makes PDD efficient at selling more goods without building a retail infrastructure of its own.

But a cost advantage does not guarantee lasting profit growth. Competition can force the savings back to shoppers and merchants. Comparing PDD with JD, Alibaba and Amazon shows why the business model matters as much as operating efficiency: each company bears different costs, and some have customers and sources of profit beyond retail.

On March 17, 2021, Colin Huang used Pinduoduo's [annual shareholder letter](https://www.globenewswire.com/news-release/2021/03/17/2194340/0/en/Pinduoduo-2021-Shareholder-Letter.html?ref=research.porbital.io) to announce that he was resigning as chairman, less than a year after giving up the CEO job. He surrendered his super-voting rights, locked up his shares for another three years, and said he was going off to do research in food and life sciences.

What makes the letter worth rereading is the reason he gave. Competition, Huang wrote, had become intense and in places distorted, and it had made him realize that competing through scale and efficiency "has its inevitable limitations." Later in the same letter he made the point about his own company's work in agriculture: PDD had raised farmers' incomes and saved consumers money by matching supply and demand more efficiently, but better distribution did not fundamentally add value to the product itself.

This was the founder of a company built on efficiency, three years after its IPO, saying that efficiency runs out. Five years later, PDD's numbers are making his argument for him.

The model Huang was stepping away from is the one that same letter describes as a combination of Costco and Disney, savings plus fun. The Disney side makes shopping a game people want to keep playing, and the Costco side makes the goods behind that game as affordable as possible through supply-chain efficiency.

The two sides feed each other. Engaging discovery, social sharing, and rewards aggregate demand; concentrated demand lets merchants produce and sell at lower cost; lower prices bring buyers back. Put simply, efficient engagement plus efficient supply equals efficient GMV: more merchandise sold with less friction and fewer resources. That explains PDD's scale. What it doesn't explain is how much profit the machine gets to keep.

That question is pressing because China's e-commerce market has spent the last few years attacking the model from every direction: [Douyin came in through content](https://research.porbital.io/bytedances-ai-bet-needs-more-than-users/), Alibaba renewed its focus on price, and JD subsidized food delivery. Better merchandising was no longer enough; platforms had to decide how long they could keep funding the same low prices.

Consider what PDD is. It runs a marketplace with an estimated $780 billion of annual merchandise volume, and it had 25,474 employees at year-end 2025\. Property and equipment come to less than 1% of assets, because merchants and logistics partners carry the inventory and the delivery burden (those costs still exist; they just sit on someone else's balance sheet). It has no interest-bearing debt, and cash worth about 60% of its market capitalization. And it has stayed within commerce: no cloud, no model business, no food delivery, no content platform, no physical stores. Temu took the marketplace abroad and the first-party brand initiative moved upstream in the supply chain, but both extended the existing business.

This is about as far as algorithmic operating efficiency has been pushed in retail, which makes PDD a useful test of the standard investment case for AI: higher output per worker leads to wider margins, and wider margins lead to higher shareholder returns. (Recommendation and matching systems are applied machine learning, not generative AI, but the logic is the same.) PDD has taken the first step further than anyone, and it has the lowest valuation of the four companies I compare below. AI can make a defined task faster and cheaper; whether more value stays in that task is a separate question.

## What Efficiency Bought

Two qualifications first. Chinese consumption is weak: retail sales grew 1.1% nominally in the first eight months of this year, roughly zero in real terms. And PDD's governance is genuinely unfriendly: no dividend ever, no stated plan for one, and RMB9.4 billion of "other losses" in the first half that the company has not explained. Both surely explain part of the valuation discount; I can't tell you how much, and while a capital-return announcement would change the picture, it would not necessarily remove the whole governance discount. Both can also change. What I want to focus on is the thing that is harder to change, which is the limit on matching and monetization.

PDD's history has three phases, and separating them shows the difference between creating new efficiency and changing who gets the proceeds.

Through 2021, matching drove the gains. Recommendations replaced search, and algorithms allocated traffic to individual items instead of relying on people to manage categories, storefronts, and campaigns. Merchandise volume per employee roughly doubled in three years. Buyers got lower prices, merchants got access to demand, and PDD won share.

From 2021 to 2024, operating profit per employee rose several-fold while volume per employee flattened. The gain this time came from charging merchants more for similar volume: buyers kept getting cheaper goods, and PDD kept its position and widened its margin, but it did so by changing the division of the proceeds, not by matching any better.

Then came the limit. In the second quarter of 2024, PDD earned RMB32.6 billion of operating profit on RMB97.1 billion of revenue, a margin near 34%. Two years later it earned RMB27.8 billion on RMB112.4 billion, under 25%, after committing RMB100 billion to merchant support; in its Q2 2026 results the company said it had stepped up ecosystem investment and was prioritizing merchant support and sustainable development. At current revenue, that margin gap is worth roughly RMB40 billion a year.

The parcel data show the same thing from the other end of the chain. In 2025 Chinese express volumes rose 13.6% while courier revenue rose 6.5%: many more boxes, each worth about 6% less. In 2026 revenue per parcel stopped falling, and volume growth immediately dropped to about 5%. Aggregate parcel data can't isolate PDD, and product mix matters, but this is the pattern you would expect if price concessions were what was holding volume up.

And it isn't just PDD. Its domestic advertising revenue now grows about 3%, against 4.3% growth in Chinese online retail of goods; Alibaba's customer management revenue fell 7% (or rose 1% excluding new-business contra-revenue), and JD Retail fell 4.7%. The definitions differ, but when every major operator is weak, the problem is bigger than one company's execution.

## Three Forms of Defence

PDD's restraint isn't only about its business mix; it shows up in how the company invests and holds capital. It stays within commerce: advertising and commissions on the main app, and transaction services on Temu. It keeps assets off its own balance sheet: capital expenditure runs around a tenth of operating cash flow, merchants carry inventory, third-party couriers handle delivery, and fixed assets remain below 1% of the balance sheet. (JD chose the integrated model instead, with more than 900,000 people in its ecosystem, RMB171.7 billion in twelve-month costs, and a 1.3% operating margin.) And it retains capital: no debt, no dividend, no buyback, and roughly RMB553 billion in current assets, enough to fund four years or more of marketing at the current rate.

Asset-light is not the same thing as low-cost, though. Specialist couriers do bring scale efficiencies, but moving an expense to a merchant or a logistics partner doesn't eliminate it; the real test is whether the network’s unit cost falls, and whether the counterparties can live on their share. I read the RMB100 billion support programme as a sign that this boundary was being tested.

I also read the restraint as a choice, not a constraint. PDD had capital, substantial engineering capacity (in 2022 roughly half its workforce was in product development), and a period of strong profitability while its peers were pursuing cloud and model strategies. It stayed with commerce anyway.

That choice may now be hard to reverse. PDD has no model or compute business, and entering today could be substantially more expensive than entering in 2021\. Its first-party brand initiative is running behind schedule, by management's own account. Restraint in the past doesn't buy a cheap option to expand later.

## The Other Side of the Comparison: PDD vs JD, Alibaba and Amazon

- **PDD:** a focused marketplace, earning advertising and transaction-service revenue while merchants and delivery partners carry much of the physical burden. Its constraint is how much those partners can afford to pay.
- **JD:** a more integrated retail and logistics operation. Control over fulfilment comes with a larger operating infrastructure; efficiency must cover those costs before it reaches shareholders.
- **Alibaba:** commerce sits alongside cloud and AI investment. It can pursue enterprise demand beyond shopping, but funding those businesses can reduce the cash generated by commerce.
- **Amazon:** retail and its marketplace sit alongside AWS. Cloud brings a different customer base and source of operating profit, with substantial capital requirements of its own.

This is why PDD can be an efficient commerce operator without having the strongest position across the whole group. The distinction also matters for [Alibaba’s attempt to connect AI shopping with commerce and cloud](https://research.porbital.io/double-eleven-remade-alibaba-can-ai-do-it-again/): a new shopping interface may change who controls demand, rather than simply make the existing retailer more efficient.

Set each company's retail profit at 100\. The next question is what its other businesses add or subtract.

__Operating profit above retail · Q2 2026 (retail = 100)__
| Segment         | PDD | JD  | Alibaba | Amazon   |
| --------------- | --- | --- | ------- | -------- |
| Cloud / compute | 0   | n/d | +14     | **+153** |
| Everything else | 0   | −56 | −43     | 0        |
| Net             | 0   | −56 | −29     | +153     |
| A year earlier  | 0   | −92 | 0       | +113     |

Swipe right to compare JD, Alibaba and Amazon →

Index = segment operating profit ÷ retail baseline profit × 100\. Baselines: PDD consolidated, JD Retail, Alibaba E-commerce Group adjusted EBITA, Amazon North America + International.

JD cloud is not separately disclosed and is included in New Businesses; n/d does not mean zero. Alibaba AI Labs and Applications is counted in “Everything else”. PDD’s base includes Temu; Alibaba’s includes instant retail and international commerce. Profit definitions differ, and second-quarter seasonality limits precision.

Sources: [PDD Q2 2026 results (PDF)](https://investor.pddholdings.com/node/9631/pdf?ref=research.porbital.io), [JD Q2 2026 results](https://ir.jd.com/news-releases/news-release-details/jdcom-announces-second-quarter-and-interim-2026-results?ref=research.porbital.io), [Alibaba June-quarter 2026 results](https://www.alibabagroup.com/en-US/document-2026456290057781248?ref=research.porbital.io), [Amazon Q2 2026 results](https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/default.aspx?ref=research.porbital.io).

JD's diversification is still commerce: logistics earned RMB2.3 billion last quarter while new businesses, mostly food delivery, lost RMB9.9 billion. **Doing more things is not the same as changing customers**, and JD has paid the cost of expansion without buying the benefit. Alibaba has genuinely left the format, but is paying for it now — its AI applications segment lost RMB13.9 billion in the quarter, and quarterly free cash flow was negative RMB44.7 billion. **I'd argue Alibaba occupies the most uncomfortable position of the four**: the admission fee is paid, the rent has not arrived, and the market's patience is finite. Amazon is through: AWS is 21% of revenue and 60.5% of operating income, with a $496 billion backlog.

Amazon still faces tension between buyers and sellers inside its marketplace, and takes considerably more from merchants than PDD does. What AWS gives it is a different customer: businesses allocating capital to computing, instead of households allocating spending to retail. That changes its exposure to demand.

The sequence and the cost both matter. AWS launched in 2006; Amazon held that position long before the current AI buildout, and AI demand arrived in a business it already owned. It is still paying for it: about $220 billion of capex this year, negative $7.6 billion in trailing free cash flow, and long-term debt that rose from $69 billion to $133 billion in six months. Even a good position has to be funded.

A second customer base isn't the only route to a premium, either. Walmart and Costco have no cloud business and trade at roughly twice their market multiple, while their demand pool continues to grow. Diversification matters most when the original pool stops growing.

My interpretation is that country risk and governance explain part of the valuation gaps, and that strategic position is less fully priced. One reason I think so: Amazon gets no premium over Walmart, despite AWS. That is a hypothesis about pricing, not a claim that I have found a mispriced stock, and further AWS and Alibaba cloud results should help test it.

## The Question That Isn't About Execution

I think PDD's judgment has largely been right.

While competitors spent heavily to win share, PDD stayed focused. It pushed matching efficiency, kept its headcount and balance sheet small relative to merchandise volume, and gave some margin back to merchants when the pressure on them increased. The company ran its chosen business with discipline.

The problem is that discipline has run into the edges of the business. Further gains in matching can't be assumed. Take rates, which I estimate at around 4% for both PDD and Alibaba (neither discloses GMV), are bounded by what merchants can afford. And weak household spending limits the demand side.

The variables that matter most now sit outside the company. Chinese consumption may recover, in which case growth hides all of this and a debt-free PDD with the longest runway in the industry benefits more than anyone. Or the entry point to shopping may move: Alibaba reports 250 million users making a first AI-driven purchase through Qwen, JD has its own models, Amazon has its own assistant, and PDD has neither model nor compute. An agent that shops across platforms treats a platform as an interchangeable supplier — which is precisely how PDD taught buyers to treat merchants.

(I take that question further in [AI Had Its iPhone Moment. It Won’t Get the iPhone Economy](https://research.porbital.io/ai-had-its-iphone-moment-it-wont-get-the-iphone-economy/), which asks whether model makers, agents, or the businesses around them get to hold onto the customer.)

For an investor, then, execution is no longer the question; PDD has shown it can operate. The question is whether staying concentrated in commerce keeps paying when demand slows and someone else may control where shopping starts, and of the four companies here, PDD is the one most exposed to getting that answer right.

That is also the lesson for anyone evaluating an AI efficiency claim. Faster, cheaper work improves throughput; it says nothing about whether you picked the most valuable work to do. PDD's algorithms made matching buyers and sellers remarkably efficient. Whether matching remains the part of commerce that captures the value is a different question, and it is the one PDD now has to answer.

---

## Sources and calculations

- [PDD’s 2025 annual report (PDF)](https://investor.pddholdings.com/static-files/e15edea4-5e6d-4740-808a-63a3dd1b2ce4?ref=research.porbital.io) supplies the year-end employee count and balance-sheet figures. The [PDD Q2 2024 results](https://investor.pddholdings.com/news-releases/news-release-details/pdd-holdings-announces-second-quarter-2024-unaudited-financial?ref=research.porbital.io) and [PDD Q2 2026 results (PDF)](https://investor.pddholdings.com/node/9631/pdf?ref=research.porbital.io) releases supply the margin comparison.
- China’s [National Bureau of Statistics’ January–August 2026 retail release](https://www.stats.gov.cn/english/PressRelease/202609/t20260916%5F1965341.html?ref=research.porbital.io) supplies retail and online-sales growth; the [State Post Bureau’s 2025 annual release](https://gs.spb.gov.cn/gssyzglj/c103928/c103936/c103939/202602/1aaf04aa945e46098ec64f7002be2557.shtml?ref=research.porbital.io) and [January–July 2026 postal-industry release](https://www.spb.gov.cn/gjyzj/c100015/c100016/202608/b38ba608fef440bb89db1e7c3a9fddbf.shtml?ref=research.porbital.io) releases supply parcel volume and revenue.
- The comparison table links to the four companies’ quarterly releases. Amazon’s [Q2 2026 Form 10-Q](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm?ref=research.porbital.io) gives the $496 billion in remaining performance obligations, primarily AWS.

GMV, take rates, operating profit per employee and the annualized margin impact are my estimates from reported figures. The table compares different segment-profit definitions; the indices show scale and direction, not decimal-level precision.

**Related reading:** [How Industrial AI Can Raise Productivity Without Raising Producer Profit](https://research.porbital.io/how-industrial-ai-can-raise-productivity-without-raising-producer-profit/) examines the same gap between faster output and durable profit in manufacturing.