JD.com’s AI Explosion: How “Lobster” Products Hit 455% Growth in Just One Week

• Developer Infrastructure: “Lobster” (Longxia) product suite & JoyAI-LLM Flash.
• Profitable Applications: Digital human livestreaming via JoyStreamer.
• Future-Proofing: JoyInside initiative for embodied intelligence and mass data collection.

On March 24, 2026, JD.com (Jingdong 京东) dropped some serious AI announcements that signal a major shift in how Chinese tech is approaching artificial intelligence.

The headlines were hard to ignore: 455% week-over-week growth in token usage for their new “Lobster” (Longxia 龙虾) product suite.

But this isn’t just vanity metrics—what’s happening here reveals something deeper about the future of AI infrastructure, digital commerce, and embodied intelligence in China.

Let’s break down what JD.com just announced and why it matters.

---

The “Lobster” Ecosystem: A New Framework for AI Agents

First, let’s talk about what’s actually driving that explosive 455% growth.

JD.com (Jingdong 京东) released the Instruct version of JoyAI-LLM Flash, their foundational large language model (LLM).

Here’s what makes this different:

• It’s specifically optimized for code development
• Built for AI agent construction
• Designed for terminal applications
• Tailored directly for the “Lobster” framework

Now, here’s the problem that JD Cloud (Jingdong Yun 京东云) saw coming: developers were getting crushed by token costs when using “Lobster” agents.

So they built multiple product formats to solve that:

• One-click deployment for lightweight cloud hosts
• Integrated hardware-software machines (no middleman complexity)
• Cloud-based SaaS versions for teams who want to scale fast

The result?

Token invocation volume for the JD Cloud “Lobster” series surged 455% week-over-week.

Translation: developers stopped worrying about token economics and actually started building.

---

Digital Humans Are Going Mainstream—And They’re Profitable

While “Lobster” is grabbing headlines, JD.com (Jingdong 京东) is also doubling down on something that’s already making real money: digital human livestreaming.

They just updated their JoyStreamer platform with some serious tech:

• Dual-Teacher DMD post-training (better model quality)
• Dynamic CFG modulation strategies (smarter video generation)
• Historical frame + “pseudo-final” frame structure (solving the uncanny valley problem)

What problems are they actually solving?

• Audio-video desynchronization (the lips don’t match the words)
• Uncoordinated multi-modal control (the hands don’t match the face)
• Identity distortion in long-duration videos (the digital human starts looking weird after 30 minutes)

The real insight here is the new “Free-State Digital Human” feature.

JD made digital human livestreaming free for all merchants last December, targeting:

• Home appliance brands
• Home furnishing companies
• Fashion apparel retailers

Why the aggressive free strategy?

Cost is the killer app.

Digital human livestreams now cost as little as 1/10th of a human-led livestream.

And the scale is already here: over 70,000 merchants are currently using JD’s digital humans.

Here’s what’s fascinating though—merchants aren’t just chasing lower costs anymore.

They’re asking for quality.

Merchants want digital hosts that actually engage like humans, because better interaction = higher audience retention = more sales.

To make this happen, JD.com (Jingdong 京东) is asking merchants to do something counterintuitive: upload their after-sales knowledge databases and FAQ details to the platform.

The digital humans then use this data to respond more naturally to customer questions.

It’s a smart loop: better data → better conversations → better retention → more GMV.

Looking ahead, digital human livestreaming is projected to drive a cumulative Gross Merchandise Volume (GMV) of tens of billions of ¥ RMB (multi-billion $ USD) by 2025.

That’s not a rounding error—that’s a real business.

---

Embodied Intelligence: The Data Collection Moonshot

Now, here’s where things get wild.

JD.com (Jingdong 京东) announced an ambitious push into embodied intelligence through its JoyInside initiative.

They’ve already locked down partnerships with:

• Nearly 100 home appliance brands
• Over 40 robot (Jiqi-ren 机器人) and AI toy brands

But here’s the bottleneck that everyone in embodied AI is facing: we don’t have enough real-world scenario data.

Robots need to learn how humans actually move, interact, and problem-solve in the real world.

That data doesn’t exist at scale yet.

So JD.com (Jingdong 京东) is doing something audacious: they’re launching what they’re calling the “largest human-scale data collection action in history.”

Here’s the scale:

• Over 100,000 internal JD.com employees will participate
• Up to 500,000 external personnel across retail, logistics (Wuliu 物流), industrial sectors, food delivery, and domestic services
• In Suqian alone, over 100,000 citizens will participate in data collection

And here are the targets:

• Within the next year: 5 million hours of real-scenario human video data
• Within two years: 100+ million hours of real-scenario human video data
• Simultaneously collecting: 1 million hours of native robot (Jiqi-ren 机器人) data

To put that in perspective: 100 million hours of video is roughly 11,000 years of continuous footage.

This isn’t theoretical—this is JD.com (Jingdong 京东) literally building the infrastructure for the next generation of AI.

---

What This Means for the Broader AI Landscape

These three announcements paint a picture of where Chinese tech companies are actually focusing right now:

• Developer infrastructure (through “Lobster”)
• Profitable AI applications (through digital humans)
• Future-proofing (through embodied intelligence)

The 455% growth in “Lobster” token usage isn’t just about hype—it suggests that removing friction from AI development actually drives adoption.

When you make it cheap and easy for developers to build, they build.

The digital human strategy shows that AI isn’t winning on novelty anymore—it’s winning on ROI.

Merchants care about cost reduction and revenue generation, not flashy tech demos.

And the embodied intelligence push reveals that JD.com (Jingdong 京东) is thinking 5-10 years ahead, investing in data infrastructure that will power the next wave of AI.

For investors and founders watching this space: this is what execution looks like.

JD.com (Jingdong 京东) isn’t just announcing research breakthroughs—they’re shipping products, measuring adoption, and reinvesting in the next layer of infrastructure.

---

References

• JD.com Announces AI Progress: Token Usage for ‘Lobster’ Products Increases 455% Week-on-Week – Yicai (第一财经)
• Official Website – JD.com (Jingdong)
• Understanding China’s Digital Human Economy – Yicai Global
• AI and Cloud Computing Solutions – JD Cloud