JD.com's retail division just publicly revealed its internal AI architecture for e-commerce for the first time. They're calling it Oxygen. This is not a product announcement — it's a platform showing the structural logic of how it routes products, merchant tools, and fulfillment through a unified AI layer.

The timing is not accidental. This lands the same week 36Kr breaks that Alibaba is rebuilding its commerce stack around the concept of Token. Two platforms. One week. The same architectural insight: commerce's new organizing principle is AI infrastructure, not product catalog. This is not a coincidence — it's a race to see whose infrastructure becomes the merchant dependency layer for the next decade.

What Oxygen means strategically: JD is making its AI architecture legible to the outside world. That's either a recruiting play, a signal that JD wants merchants building against this infrastructure, or both. Compare this to Amazon's architecture — a black box wrapped in a pricing structure. JD is publishing the blueprint. That's a different competitive posture, and the Western platforms should be paying attention.

This is the most important item in today's feed. 36Kr's exclusive: as of April 1 — the start of Alibaba's new fiscal year — the core OKR for the China e-commerce group is merchant-side AI tool retention rate and AI-driven GMV growth. The structural thesis: rebuild commerce around Token — the computational unit of AI models — rather than around SKUs or search queries.

When a platform's primary metric shifts from GMV to AI tool retention, the platform is measuring whether merchants are becoming dependent on its AI stack. Dependency is a stickier moat than price. This is lock-in by design — not through marketplace network effects but through software attachment. Once your pricing AI, your content AI, and your demand-forecasting AI all run on the same platform's infrastructure, you don't leave that platform. That's the game.

We covered Alibaba's ¥10 billion profit drop as the investment thesis in financial terms. This 36Kr piece shows what that thesis looks like inside the org chart. It connects directly to Meituan's MTGR framework treating products as tokens — the same architectural metaphor now organizes China's largest commerce platform. Two of the three biggest players in Chinese e-commerce are running the same conceptual playbook. The third is watching.

Alibaba's employee count fell 34% year-over-year as it accelerates AI transformation. A Korean tech outlet frames this not as a layoff story but as a strategic metric — the AI stack doing more work with fewer people.

The arithmetic is stark: Alibaba absorbed a ¥10 billion profit hit on the AI investment thesis. The headcount data shows what that thesis buys: a platform that needs a third fewer people to operate at the same or larger scale. Fashion brands modeling their own AI transformation programs should be looking at this ratio, not the press release version. The press release version says AI augments your workforce. The Alibaba version says AI replaces about a third of it.

极睿科技 (JiRui Technology) closed a 100M+ RMB Series A led by Sequoia China — its third funding round in two years. The founder is a 1990s-born Tsinghua graduate building AI specifically for fashion retail supply chains.

What JiRui actually does: AI-generated content that drives product discovery for e-commerce merchants, with reported results in the tens of millions RMB in monthly GMV gains. This is not generic image AI. It's vertically specialized for fashion, with supply chain data baked into the model — knowing which styles are available at what margin informs what gets surfaced and to whom. The content and the inventory are not separate problems. JiRui is treating them as one.

Sequoia China leading is the signal. This is not speculative infrastructure. JiRui has merchant traction and a defensible thesis: AI-generated product content as the supply chain's last mile. Compare this to the catalog taxonomy bottleneck we've written about — JiRui is building the layer that fills the gaps hand-curated catalogs leave empty.

China's tech press is covering 数字替身 (digital doubles) entering fashion at commercial scale — AI-generated virtual models now deployed in product imagery, with no regulatory framework catching up. Technology advancing fast, regulation not yet in place: that's the headline, verbatim.

This is the ground-level backstory to today's AI Mannequin Arrives With a French Accent piece. The Chinese market has been running AI models in product photography longer and at greater scale than any Western market — lower costs, faster iteration, and a regulatory vacuum that the EU AI Act will eventually close, but hasn't yet.

The practical risk isn't philosophical. Consumers are increasingly aware of AI-generated imagery, and the first cross-border brand caught in an AI avatar dispute without a compliance framework will write the precedent. The technology moved faster than any compliance layer. That's not unique to fashion. It's just most visible here, because fashion is where the avatar is the product.

AI e-commerce startup Spangle raised $15M Series A at a $100M valuation. Six people. The product: product pages that reconfigure themselves for each visitor, AI-shaped rather than static.

The skeptic's take: expensive A/B testing with a better narrative. The optimist's take: Shopify is a platform, Spangle is a specification — and sometimes the spec wins because it bets on one thing faster than the platform can ship. Today's Cold-Start Recommendation explainability piece covers what happens when you add transparency to the recommendation layer. Spangle is the extreme version of that thesis: not just recommendations that explain themselves, but pages that reconfigure themselves.

The funding multiple (six headcount, $100M valuation) tells you what the market is pricing: the thesis, not the traction. That's fine for now. The question is whether Spangle has enough merchant data to differentiate from Shopify's personalization layer by the time the next round comes. The answer to that question will determine whether this is a company or a feature.

Canva acquired Simtheory (agentic AI) and Ortto (marketing automation and customer engagement) simultaneously. The thesis: Canva is building the design-to-distribution stack for mid-market brands.

For fashion specifically, the combination means a brand could theoretically design a campaign, automate its distribution, and personalize delivery without leaving one platform. That's a direct threat to the Adobe → Salesforce → Klaviyo chain most mid-market fashion brands assemble themselves — at significant cost and integration overhead that only the largest teams can sustain.

The timing connects to ByteDance Seedance dropping production costs to fractions of a dollar per second. Cheap creative is only powerful if you have the distribution infrastructure to deploy it at scale. Canva just acquired that infrastructure. VC already repriced fashion on software terms — Canva's dual acquisition confirms the creative tools market is running the same logic. The integration is the product.

Modern Retail's AI-in-stores roundup lands the same day our analysis of Google's Aritzia AI ads result goes live. The roundup's central observation — no single AI use case fits all physical retail — is accurate and also slightly evasive. The harder point: different store formats require fundamentally different AI architectures, and most retailers are deploying the same chatbot regardless of format.

The store is becoming an AI operating system — but that's not a uniform transformation. Puma's Las Vegas concierge is novelty. Home Depot's Magic Apron is operational. The difference is whether AI reduces friction for the customer or generates data for the retailer. At most retailers right now, it's attempting both and optimizing for neither.

WantedLab's 2026 AX Insight Report (Korea): 92% of workers use AI for their jobs. Only 5% of companies have completed enterprise-wide AI deployment. 97% believe AI transformation will impact their business environment.

This gap — between personal AI adoption and enterprise AI deployment — is the central tension in every 'AI is transforming retail' conference deck that goes unspoken. Workers use AI because the tools are immediately useful. Companies don't deploy at enterprise scale because integration is hard, compliance is unclear, and the ROI frameworks that boards require don't exist yet.

This isn't a Korean number. It's universal. At Shoptalk Spring, retailers were asked to prove results rather than announce plans. The 92%/5% split is precisely why so few can. Meanwhile, today's piece on Brazilian Gen Z purchasing behavior shows that consumers are running ahead of the enterprise gap entirely — young shoppers are already using AI to decide purchases while the companies selling to them are still in governance committee. 92% of your employees are running experiments. 5% of your companies have the data infrastructure to know what they found.

Korean local commerce platform Daangun launched '말로 쓰기' (Write by Voice): AI asks questions about your visit experience, you answer by voice, AI writes the review. Results: 43% lift in review reuse rate, 40% improvement in quality scores.

The interface barrier for reviews is e-commerce's dirty secret. Most reviews are either stars-only or three words of nothing. Conversational collection removes that friction — and the AI's ability to ask follow-up questions captures specifics (fit, texture, durability, occasion) that unprompted reviewers consistently omit. You can't recommend what you don't know. And you don't know it because nobody wrote it down.

Fashion is where this matters most. Poshmark centered its entire redesign on curation. Conversational review collection is the input layer that makes curation possible at scale. Better reviews → better recommendations → better discovery. This is a flywheel that starts with a voice prompt, and the 40% quality gain suggests the AI is asking better questions than the average reviewer is volunteering.

Foot Locker joins DoorDash for on-demand delivery. The logistics story is obvious. The category signal matters more: footwear is now a convenience purchase for a segment of buyers large enough that Foot Locker is optimizing for it.

Myntra's M-Now captured 10% of orders in active Indian markets within a year by treating apparel as a convenience category. DoorDash reports over a third of its monthly active users already shop retail and grocery. Foot Locker isn't converting sneaker consideration into impulse. It's capturing convenience intent that was never going to enter a sneaker store — people who need a specific pair today, not a shopping experience.

The inventory constraint is the story that will surface next quarter. On-demand delivery only works for SKUs maintainable within a reasonable fulfillment radius. That curated subset will tell us which sneakers the market has decided are convenience products and which ones still require the store. That's actually useful competitive intelligence about where the category is going.

Henkel is acquiring Olaplex for $1.4 billion. Standard acquisition language from the CEO: accelerate product innovation, expand reach. The interesting layer is what Henkel actually brings to the deal.

Henkel has the R&D infrastructure — including AI-assisted formulation tools — to run bond-repair chemistry at industrial scale. Olaplex has the cult following and the patent estate. The combination is a template for what premium beauty AI looks like when you have both the data infrastructure and the consumer trust. One without the other produces either expensive science nobody buys or a brand with nothing defensible under it.

The pattern bears watching: European beauty conglomerates (Henkel, L'Oréal, Beiersdorf) are acquiring category-defining brands and applying AI R&D infrastructure to them. This is slower and more capital-intensive than the American CPG roll-up playbook, but the margin structure of premium hair care gives it room to compound. The question for Olaplex loyalists: does Henkel's scale accelerate the innovation or sand down the edges that made the brand interesting?

Tubi is now a native app inside ChatGPT. Media industry headline: OpenAI is becoming an entertainment hub. Retail industry headline: the discovery interface for content and commerce are converging in the same container.

Sephora's ChatGPT app launched the same day OpenAI retreated from Instant Checkout. The pattern: the commerce that works inside ChatGPT isn't transactional — it's discovery. Tubi's content library is curated by taste and occasion. Fashion is curated by taste and occasion. The next Tubi inside ChatGPT will be a fashion discovery experience, not a product feed.

There's a structural read here that connects to Southeast Asian video commerce. The West keeps trying to embed checkout into entertainment formats and keeps failing. ChatGPT is doing the reverse — embedding entertainment into the interface where purchase intent already forms. That's the architecture that actually works, and Tubi just demonstrated it is technically feasible at consumer scale.