WAIC 2026 Preview: Dongfang Suanxin DF1000 Software-Defined 3D Chip, StepFun STEPX Neo Native AI Phone, Huawei Atlas 950 SuperNode — Three Paths of China's AI Ecosystem Breakthrough

WAIC 2026 Preview: Dongfang Suanxin DF1000 Software-Defined 3D Chip, StepFun STEPX Neo Native AI Phone, Huawei Atlas 950 SuperNode — Three Paths of China’s AI Ecosystem Breakthrough

1. Introduction: Countdown to an AI Feast

From July 17-20, 2026, the World AI Conference (WAIC 2026) will be held in Shanghai. The theme is “Smart Partners, Co-creating the Future,” spanning four venues across three districts, with an exhibition area exceeding 100,000 square meters for the first time. Over 1,100 enterprises will exhibit, showcasing 4,000+ products, with 300+ claiming global debuts.

But four days before the conference opens, three major announcements have already electrified the industry:

July 13 — Dongfang Suanxin released the DF1000, the world’s first software-defined near-memory computing 3D AI chip, achieving 520 TFLOPS@BF16 at 14nm node, completely independent of advanced process nodes and HBM.

On the same day, StepFun launched the STEPX Neo, the world’s first LLM-native agent smartphone, featuring its self-developed Step AOS operating system and the built-in Step Amoo agent, rebuilding the system from the ground up.

Huawei’s Atlas 950 SuperPoD SuperNode will make its physical debut at WAIC, starting from 64 NPUs per cabinet and scaling to 8,192 Ascend NPUs.

These three paths — software-defined + 3D stacking for chips, operating system-level reconstruction for terminals, and domestic supernodes for infrastructure — represent China’s three-pronged strategy for AI ecosystem independence.

2. Dongfang Suanxin DF1000: Bypassing the Advanced Process Bottleneck

2.1 Core Specifications

DF1000 Core Specifications:
┌─────────────────────────────────────────────────────────────────────┐
│  Chip Parameters                                                     │
│  ├─ Process: 14nm (domestic mature node)                           │
│  ├─ Compute: 520 TFLOPS @ BF16                                     │
│  ├─ Memory Bandwidth: 6.4 TB/s                                     │
│  ├─ Packaging: DRAM-LOGIC Wafer-level hybrid bonding 3D vertical    │
│  ├─ Tape-out: Fully verified                                       │
│  └─ Cluster: 128-card stable operation verified                    │
│                                                                     │
│  Technology Path                                                    │
│  ├─ Software-defined chip: Reconfigurable computing architecture    │
│  │   └─ Fully asynchronous dataflow, tensor tile scheduling        │
│  ├─ 3D near-memory computing: Bump-less vertical interconnect      │
│  │   └─ 6.4TB/s bandwidth, exceeding HBM3 (3-4TB/s)               │
│  └─ Full domestic supply chain: Design, fab, packaging, test       │
│                                                                     │
│  Product Matrix                                                     │
│  ├─ DF1000 Accelerator Card (OAM 2.0)                              │
│  ├─ TY64 SuperNode (64 cards, 33 PFLOPS@BF16)                      │
│  ├─ QY100 Server                                                   │
│  └─ HS128/HS512 Smart Computing Clusters                           │
└─────────────────────────────────────────────────────────────────────┘

2.2 Technology Breakthrough

"""
DF1000 vs Traditional Approach Comparison
"""
class ChipCompare:
    def __init__(self):
        self.traditional = {
            "process": "5nm/3nm", "memory": "HBM3/HBM4",
            "bandwidth": "3-4 TB/s", "supply": "TSMC/Samsung/SK Hynix"
        }
        self.df1000 = {
            "process": "14nm (domestic)", "memory": "DRAM (domestic)",
            "bandwidth": "6.4 TB/s", "supply": "Full domestic chain"
        }
    
    def compare(self):
        print("=" * 60)
        print("DF1000 vs Traditional: Key Advantage")
        print("=" * 60)
        print(f"  14nm → 520 TFLOPS: Comparable to 7nm/5nm products")
        print(f"  6.4TB/s bandwidth: Exceeds HBM3 (3-4TB/s)")
        print(f"  Full domestic supply chain: No export control risk")
        print(f"  Software-defined + 3D stacking: Bypasses both")
        print(f"  process and memory bottlenecks")

ChipCompare().compare()

3. StepFun STEPX Neo: Operating System-Level Revolution

3.1 Product Positioning

On July 13 evening, StepFun Chairman Yin Qi took out a new phone and said softly: “Hi Amoo, welcome to the world.”

This is the STEPX Neo — the world’s first LLM-native agent smartphone. StepFun’s vision is clear: model, software, and hardware as a unified, symbiotic whole.

STEPX Neo Architecture:
┌─────────────────────────────────────────────────────────────────────┐
│  LLM Layer (Brain)                                                  │
│  ├─ StepFun proprietary LLM                                        │
│  └─ Edge-cloud collaborative: local inference + cloud enhancement   │
│                                                                     │
│  System Layer (Step AOS - Agent-Native OS)                          │
│  ├─ Traditional OS: App-centric (user manually opens apps)         │
│  ├─ Step AOS: Agent-centric (agent automatically orchestrates)     │
│  ├─ Hardware + system resources → agent-callable APIs              │
│  ├─ Dual-domain memory (short-term + long-term)                    │
│  └─ Edge-cloud architecture for latency vs complexity balance      │
│                                                                     │
│  Agent Layer (Step Amoo)                                            │
│  ├─ User expresses intent → agent plans → cross-app execution      │
│  ├─ Partners: Ctrip, Alipay, DiDi, Meituan, WPS, Jianying         │
│  └─ L3 certification (highest in AI terminal grading)              │
│                                                                     │
│  Security:                                                          │
│  ├─ "Trustworthy, Visible, Controllable, Reversible" framework     │
│  └─ Joint whitepaper with Shanghai AI Lab                          │
└─────────────────────────────────────────────────────────────────────┘

3.2 Why This Is an “OS-Level” Revolution

StepFun’s analogy is precise: “Open a door for the agent in the old system, and it’s always a visitor. Build a house for the agent, and it becomes a native resident.”

Traditional AI phones add an AI assistant on top of Android/iOS, where the AI can only access apps through APIs. Step AOS breaks down communication, file, and application capabilities into modules that the built-in agent can freely call — users don’t need to open apps one by one; just speak, and the agent completes cross-app workflows automatically.

"""
Traditional AI Phone vs Agent-Native Phone Comparison
"""
class PhoneComparison:
    def compare(self):
        scenarios = [
            ("Booking trip", 
             "Traditional: Open Ctrip → Search → Calendar → Open DiDi",
             "Step AOS: 'Book a trip to Beijing next Monday' → done"),
            ("Create PPT & send",
             "Traditional: Open WPS → New PPT → Insert → Save → Email",
             "Step AOS: 'Turn yesterday's notes into a PPT and send' → done"),
            ("Find restaurant",
             "Traditional: Open Meituan → Search → Reviews → Map → Navigate",
             "Step AOS: 'Find a good Sichuan restaurant nearby and book' → done"),
        ]
        for task, t, a in scenarios:
            print(f"  {task:20s}")
            print(f"    Traditional: {t}")
            print(f"    Step AOS:    {a}")
            print()

PhoneComparison().compare()

4. Three “World’s First” AI Phone Routes

Three AI Phone Routes Comparison:
┌──────────────────┬─────────────────┬─────────────────┬─────────────────┐
│ Dimension        │ STEPX Neo       │ Nubia×Doubao    │ Honor Robot     │
│                  │ (StepFun)       │ (ByteDance)     │ Phone           │
├──────────────────┼─────────────────┼─────────────────┼─────────────────┤
│ Positioning      │ LLM-native      │ AI agent phone  │ Embodied smart  │
│ OS               │ Step AOS (self) │ Android-based   │ Android-based   │
│ Key Differentiator│ System rebuild  │ Doubao system   │ 4-DoF robotic   │
│                  │                 │ integration     │ arm              │
│ Agent            │ Step Amoo(self) │ Doubao agent    │ Magic agent     │
│ Hardware Novelty │ N/A             │ N/A             │ Retractable arm │
│ Partners         │ Ctrip/Alipay/   │ Not disclosed   │ Not disclosed   │
│                  │ DiDi/Meituan    │                 │                 │
│ Certification    │ L3 (highest)    │ Not disclosed   │ Not disclosed   │
└──────────────────┴─────────────────┴─────────────────┴─────────────────┘

5. Huawei Atlas 950 SuperNode

5.1 Core Specifications

Huawei Atlas 950 SuperPoD:
├─ Single cabinet: 64 Ascend NPUs (entry level)
├─ Max cluster: 8,192 Ascend NPUs (SuperCluster)
├─ Compute scale: 500,000+ NPUs (SuperCluster)
├─ Interconnect: HCCS high-speed
├─ Memory: HBM2e (domestic)
├─ Cooling: Liquid
└─ Competitor: NVIDIA DGX SuperPOD

WAIC 2026 Domestic Computing Exhibits:
├─ Huawei Atlas 950 SuperPoD (H2-A101)
├─ Dongfang Suanxin DF1000 full stack (debut)
├─ Muxi "Xijing" S-series SuperNode
├─ Kunlun Core P800 OAM module
├─ ZTE + Xizhi + Biren + Muxi + Suiyuan "Matrix SuperNode"
└─ Sugon scaleX

6. Go Implementation: AI Native Terminal Evaluation Framework

package main

import (
	"fmt"
	"sort"
)

type Dimension struct {
	Name   string
	Weight float64
	Score  int
}

type Terminal struct {
	Name       string
	Type       string
	OS         string
	Agent      string
	Dimensions []Dimension
}

func (t *Terminal) ComputeScore() float64 {
	total, ws := 0.0, 0.0
	for _, d := range t.Dimensions {
		total += float64(d.Score) * d.Weight
		ws += d.Weight
	}
	return total / ws
}

func (t *Terminal) PrintReport() {
	fmt.Printf("\n=== %s (%s) ===\n", t.Name, t.Type)
	score := t.ComputeScore()
	for _, d := range t.Dimensions {
		bar := ""
		for i := 0; i < d.Score/5; i++ {
			bar += "█"
		}
		fmt.Printf("  %-20s %3d %s\n", d.Name, d.Score, bar)
	}
	fmt.Printf("  Score: %.1f/100\n", score)
	grade := "C"
	if score >= 85 { grade = "S" } else if score >= 70 { grade = "A" } else if score >= 50 { grade = "B" }
	fmt.Printf("  Grade: %s\n", grade)
}

func main() {
	terminals := []Terminal{
		{"STEPX Neo", "Agent-Native", "Step AOS", "Step Amoo",
			[]Dimension{{"Autonomy", 0.25, 92}, {"Cross-app", 0.20, 88},
				{"Speed", 0.15, 85}, {"Ecosystem", 0.15, 78},
				{"Security", 0.15, 90}, {"UX", 0.10, 88}}},
		{"Nubia×Doubao", "AI Assistant", "Android+SDK", "Doubao",
			[]Dimension{{"Autonomy", 0.25, 72}, {"Cross-app", 0.20, 65},
				{"Speed", 0.15, 80}, {"Ecosystem", 0.15, 85},
				{"Security", 0.15, 75}, {"UX", 0.10, 78}}},
		{"Honor Robot", "Embodied", "Android+Magic", "Magic",
			[]Dimension{{"Autonomy", 0.25, 68}, {"Cross-app", 0.20, 60},
				{"Speed", 0.15, 75}, {"Ecosystem", 0.15, 70},
				{"Security", 0.15, 72}, {"UX", 0.10, 82}}},
	}
	
	sort.Slice(terminals, func(i, j int) bool {
		return terminals[i].ComputeScore() > terminals[j].ComputeScore()
	})
	
	for _, t := range terminals {
		t.PrintReport()
	}
	
	fmt.Println("\nConclusion:")
	fmt.Println("  Agent-native terminals (Step AOS) lead in task autonomy")
	fmt.Println("  AI assistant integration (Doubao) leads in ecosystem")
	fmt.Println("  Embodied AI (Honor) innovates in hardware but lacks software")
	fmt.Println("  2026 is the watershed from 'AI+phone' to 'AI phone'")
}

7. WAIC 2026 Other Highlights

7.1 LLMs and Agents

• MiniMax M3 (H1-A801) debut
• Baidu “Agent Family” — Wenxin 5.1 tops Arena domestic charts
• Kunlun Wanwei releases 4 models simultaneously (world model, video, music, robot)
• Zhiyuan Robot: 15,000 units produced, revenue exceeding 1 billion CNY
• 60 Expedition A3, Lingxi X2 robots deployed for public service

7.2 Academic Lineup

• First-ever WAIC Academic international conference, chaired by Yao Qizhi
• 284 submitted papers
• 9 Turing Award/Nobel laureates attending
• Richard Sutton (RL pioneer) keynote
• Yoshua Bengio on UN AI governance framework
• Kevin Kelly (“Silicon Valley’s spiritual father”) attending

7.3 Shanghai AI Industry Data

2025: 394 designated AI enterprises, industry scale exceeding 637 billion CNY, +39.5% YoY.

8. Conclusion

The three major announcements before WAIC 2026 outline three paths for China’s AI ecosystem breakthrough:

Dongfang Suanxin DF1000 represents the chip path of bypassing advanced process nodes — software-defined + 3D stacking achieves performance comparable to advanced nodes at 14nm. This is a classic case of “changing lanes to overtake.”

StepFun STEPX Neo represents the terminal path of OS-level reconstruction — self-developed Step AOS transforms agents from “visitors” to “residents,” redefining the human-computer interaction paradigm.

Huawei Atlas 950 SuperNode represents the infrastructure path of full-stack domestication — from chips to clusters to ecosystems, all independently controllable.

Three paths, one goal: making China’s AI industry no longer dependent on external supply chains.

References

1. Shanghai Securities News - “WAIC 2026 Preview: Smart Partners Accelerate from Concept to Reality”
2. ChipThings - “Dongfang Suanxin DF1000 Chip Launch”
3. Shangguan News - “StepFun Unveils LLM-Native AI Terminal Brand STEPX”
4. 36Kr - “WAIC 2026 Pre-Week Analysis: Phones, Computing, Embodied”
5. Dongfang Suanxin Official - DF1000 Technical Whitepaper (2026-07-13)
6. StepFun - STEPX Neo Product Launch Materials (2026-07-13)

• ←Previous
• Next→