WAIC 2026 Wrap-Up: 4,486 Exhibits, ¥40.9B in Signings, and ¥20.36B in Procurement
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
Recap WAIC 2026's official post-event figures: 4,486 exhibits, 127 world premieres, ¥40.9 billion in signed agreements (investment intent), and ¥20.36 billion in confirmed procurement deals.
Decision in 20 seconds
Recap WAIC 2026's official post-event figures: 4,486 exhibits, 127 world premieres, ¥40.9 billion in signed agreements (investment intent), and ¥20.36 billion i…
Who this is for
Product managers, Developers, and Researchers who want a repeatable, low-noise way to track AI updates and turn them into decisions.
Key takeaways
- WAIC 2026 Closing Data Summary
- After 4,486 Exhibits, It’s Time to Assess Product Maturity
- 351 Global Firsts and 198 Debut Products—Time to Re-Categorize by Maturity
- The ¥40.9 billion in signings refers to total project investment
Last checked: 2026-07-22
The World Artificial Intelligence Conference (WAIC) 2026 concluded on July 20. While the opening days were packed with headline-grabbing product launches, the post-event data offers a clearer picture of where real industry momentum is taking hold:
- Exhibition space exceeded 100,000 m²
- 4,486 exhibits showcased — including 351 global debuts and 198 world premieres
- 32 Shanghai-based AI projects signed collectively, with total investment exceeding ¥40.9 billion
- 212 procurement needs published at the conference, representing an estimated ¥20.36 billion in procurement intent — up 25% year-on-year
- Additional ¥1.18 trillion in intended credit support for innovative enterprises
These figures are substantial — but they must not be conflated into a single “WAIC transaction volume.” The ¥40.9 billion reflects committed project investment; the ¥20.36 billion represents procurement intent, not finalized orders; and the ¥1.18 trillion is indicative lending capacity — not actual disbursement. Each figure differs in scope, timeline, execution conditions, and reporting unit. Adding them together would misrepresent financing capability, infrastructure development, and purchasing demand as if they were equivalent commercial outcomes.
This page is tailored for product, procurement, and engineering teams. It does not rehash the mid-conference trend summaries released on July 20. Instead, it focuses on questions that only become clear after the event ends:
- Which outcomes are verifiable and actionable?
- Which remain limited to debut or demonstration status?
- How can teams distill 4,486 exhibits into a realistic, two-week validation pipeline?
WAIC 2026 Closing Data Summary
| Metric | Reported Value at Closing | What It Indicates | What It Does Not Indicate |
|---|---|---|---|
| Event Duration | July 17–20 | A four-day conference window | Cannot be used to infer the project’s subsequent execution timeline |
| Exhibition Area | >100,000 m² | Physical scale of the offline exhibition | Does not reflect effective product display area or actual procurement volume |
| Exhibited Items | 4,486 | Number of projects showcased on-site | Does not mean 4,486 commercially deployed products |
| Global First Releases | 351 | Density of “world premiere” announcements at the event | “First release” ≠ formal delivery or existing customers |
| Public Debut Items | 198 | Number of items shown publicly for the first time | “Debut” may still refer to prototypes or controlled demos |
| On-site Visitors | >400,000 person-visits | Public and industry interest level | “Person-visits” ≠ unique individuals—and certainly not buyers |
| Embodied AI Products | 208 | Embodied AI emerged as a major exhibition theme | Quantity does not indicate stable operational hours or real-world reliability |
| Centralized Signings | 32 projects, total investment >¥40.9 billion | Scale of Shanghai’s priority infrastructure initiatives | Does not equal on-site transactions or companies’ immediate revenue |
| Procurement Requests | 212 items, from 177 key procurement groups | Real demand aggregated and publicly announced | Does not mean purchase orders have already been issued |
| Intended Procurements | ~¥20.36 billion, up 25% YoY | Scale of preliminary buyer–seller matchmaking | Does not equal signed contracts, payments received, or recognized revenue |
| Valid Matchmaking Sessions | 3,874 | Number of supply–demand connections facilitated by the event | Follow-up conversion rate is not disclosed |
Core data in the table comes from the Shanghai Securities News’s closing summary published on July 21 (link). Event dates, venue size, participating enterprises, and visitor figures are cross-checked against CCTV’s closing report dated July 20 (link). The two sources differ slightly in scope: CCTV reported “over 1,100 enterprises and more than 4,000 exhibited items,” while Shanghai Securities News provided the more precise post-event count of 4,486 items. This article uses the finalized figure, while retaining the original publication dates of both sources.
After 4,486 Exhibits, It’s Time to Assess Product Maturity
Compared to last year, the number of exhibitors and exhibition area both grew by over 30%, while the number of exhibits surged nearly 50%. On the surface, this signals expansion—but the real structural shift is “from model demos to system-level delivery.” Shanghai Securities News listed 108 chips and more than 200 AI computing products—including a 100,000-GPU AI supercluster, ultra-nodes, domestically developed interconnects, and compute-in-memory solutions. Buyers no longer evaluate just chip specs; they now assess cluster interconnects, scheduling, power efficiency, operations & maintenance, and software compatibility.
Embodied intelligence featured 208 products—more than triple last year’s count. This surge confirms that robotics and physical AI are gaining momentum at the show—but it still doesn’t answer the question most critical to factories: How long can a robot operate continuously outside a demo environment? How often does it require human intervention? How much new data is needed when switching scenarios? And how does it recover from falls, jams, or misrecognition? A single successful grasp on stage reflects “visible capability”; uninterrupted task execution in a real customer setting reflects “deliverable capability.”
Agents have evolved from forum topics into core product themes. Over 10% of conference sessions focused on agents—and 4 of the 10 “flagship exhibits” were agent-based products. Alibaba and Moonshot’s Qwen3.8 and Kimi K3, plus office agents and agent-native smartphones unveiled at the event, all signal a clear shift: generating a paragraph of text is no longer the endpoint of product storytelling. The new unit of competition is task completion: Can the agent understand context, invoke tools, execute multi-step workflows, and produce auditable outcomes?
Yet agent count and session share don’t equal commercial readiness. An agent that performs flawlessly in the conference’s controlled network will face identity permissions, legacy system APIs, exception handling, data retention policies, and blurred lines of human accountability once deployed in an enterprise. Procurement evaluation must shift—from “Did the demo succeed?” to “Where does it fail? Who can roll back? Are logs complete?”
351 Global Firsts and 198 Debut Products—Time to Re-Categorize by Maturity
“Global first” and “debut” are marketing labels—not indicators of technical maturity. To make the data actionable, we can classify all conference projects into five maturity tiers:
| Level | Minimum Evidence Required | Next Action Permitted | Handling When Insufficient |
|---|---|---|---|
| L0 — Lead | Official product name, publishing entity, on-site photos | Add to observation list | Do not schedule procurement meetings |
| L1 — Watchable Demo | Full demo video (including inputs, outputs, and failure cases) | Schedule technical Q&A session | Reject projects offering only highlight reels |
| L2 — Repeatable Trial | Public access point, documentation, sample data, clearly stated limitations | Conduct internal sandbox testing | Remain at observation level if no documentation exists |
| L3 — Customer-Validated in Production | Defined tasks, acceptance criteria, logs, and records of human intervention | Launch a 2- or 4-week pilot | Refuse to scale beyond pilot without clear customer-defined boundaries |
| L4 — Commercial Delivery | Contract scope, SLA, pricing, delivery & maintenance responsibilities | Enter formal procurement review | “Already signed” does not substitute for actual验收 (acceptance) |
A product can be labeled “world premiere” yet still sit at L0. Conversely, an older, long-released product may already have reached L4. After the conference ends, systematically request the specific evidence needed to advance each product’s maturity level. Projects with only a product name and booth poster qualify solely as leads; only those with public documentation and repeatable tasks warrant engineering time.
The ¥40.9 billion in signings refers to total project investment
The 32 projects signed collectively during the closing ceremony span AI infrastructure, agent-based applications, embodied AI, and scientific AI—representing over ¥40.9 billion in total investment. This signals strong industry commitment: funding is flowing not just into foundational models, but also into compute infrastructure, real-world applications, and physical systems.
However, “total investment” typically covers capital expenditures across the construction timeline—and includes contributions from multiple stakeholders. It does not mean ¥40.9 billion in confirmed revenue for the 32 AI vendors on signing day. To assess actual progress, readers should track: the project’s implementing entity, funding sources, land and data center readiness, equipment procurement status, groundbreaking date, expected commissioning timeline, and operational responsibility. Without these details, the signing reflects intent and procedural advancement—not execution.
The same closing report also cites ¥1.18 trillion in intended credit support for innovative enterprises, plus Shanghai Unicom’s planned investment of over ¥25 billion in its “UniAI·Smart Connect Shanghai” initiative. Credit lines represent maximum lending capacity—subject to due diligence, credit approval, and drawdown conditions. Unicom’s plan has its own defined scope. None of these figures should be mechanically added to the ¥40.9 billion.
The ¥20.36 billion in intended procurement is a closer proxy for demand temperature
The conference organized 177 key procurement delegations, published 212 procurement requirements, and projected a total intended procurement value of approximately ¥20.36 billion—up 25% year-on-year. These figures reflect market demand more closely than the number of exhibits, as they at least include identifiable buyers and specific requirement items. Yet “intended” procurement remains far from finalized orders—it still must pass through requirement clarification, budget approval, tendering, testing, contracting, and acceptance.
If we simply divide ¥20.36 billion by the 212 requirements, the average comes to roughly ¥96.04 million per item. But this average holds little practical procurement value: funding for large-scale infrastructure projects and small-scale application purchases is likely extremely skewed. The right approach is to segment by procurement category—compute & data centers, model & platform subscriptions, industry-specific software, robotics & hardware, consulting, and systems integration. Each category differs sharply in contract duration, gross margin structure, and delivery risk.
The 3,874 valid supply-demand matchmaking sessions show strong on-site facilitation—but the conference did not disclose how many of these progressed to RFPs, pilot programs, signed contracts, or actual payments. Going forward, conversion rates should be tracked across 30-, 90-, and 180-day horizons. Exchanging contact details onsite does not equal commercial execution.
Post-Conference Procurement Team: Compressing 4,486 Exhibits into a Validation Pipeline in 14 Days
Imagine a manufacturing enterprise with three factories sends a five-person team to WAIC—representing production, IT, security, procurement, and finance. They identify 120 promising solutions, of which 35 relate to visual quality inspection, predictive equipment maintenance, knowledge assistants, and material handling. Back at headquarters, they can’t schedule 35 follow-up demos. Instead, they must narrow those down to no more than five executable pilots within 14 days.
Day 1: Scope Cleansing
Each candidate solution must clearly define: the end user, current workflow, input data sources, required output actions, and the budget owner. Vague statements like “use AI to improve manufacturing efficiency” don’t qualify. A valid task looks like: “Ingest alarm and maintenance logs from 20 stamping presses over the past 90 days; predict high-risk unplanned downtime at least 24 hours in advance; and route alerts to equipment engineers for confirmation.”
Days 2–4: Evidence Collection
Teams request full demo recordings, system architecture diagrams, data flow maps, permission models, pricing units, customer references—and crucially, documented failure boundaries. For robotics: add metrics on continuous cycle time, manual takeover frequency, maintenance intervals, and safety incident history. For AI agents: require tool whitelists, rollback mechanisms, audit logs, and exception-handling logic. Projects offering only marketing decks—and refusing to clarify technical boundaries—are immediately downgraded to “observation status.”
Days 5–7: IT and security teams vet non-accessible items. Test whether anonymized data can be used, whether the solution integrates with existing identity systems, whether data stays within the designated region, and whether the vendor provides clear deletion and retention policies. Model performance is not the focus here—this phase is about confirming whether the pilot can legally and securely begin.
Days 8–10: Define minimal tasks. Each candidate is allowed only one core use case and one set of quantifiable metrics:
- Visual QA: Track missed defects, false positives, and manual review time.
- Knowledge Assistant: Track citation accuracy, refusal rate, and outdated answers.
- Robotics: Run 200 cycles and record task completion rate, handover frequency, and unplanned stoppages.
- Equipment Maintenance: Measure prediction lead time and false alarm rate.
Days 11–14: Finalize pricing, staffing, and exit criteria. Projects missing any of the following do not proceed to pilot: a business owner, test data, deliverables promised within two weeks, or willingness to document failure samples. The final five candidates must each have a single-page task card—not a stack of presentation decks.
| 14-Day Milestone | Must Deliver | Pass Criteria | Pause Signal |
|---|---|---|---|
| Day 1 | Owners assigned for task, user, data, action, and budget | All five fields fully specified | Still only describing “industry enablement” vaguely |
| Day 4 | Full demo, architecture diagram, pricing, constraints, and failure samples | Evidence verifiable by technical staff | Only edited videos and sales narratives provided |
| Day 7 | Security & access conclusion | Pilot can start in an anonymized or isolated environment | Unclear data flow paths or retention rules |
| Day 10 | Minimal pilot scope and success metrics | Comparative data can be generated within two weeks | Requires long-term platform overhaul first |
| Day 14 | Staffing plan, quote, acceptance criteria, and explicit exit conditions | Signed off jointly by business, IT, and procurement | No budget owner assigned—or refusal to define exit conditions |
Why Physical Intelligence Has Become the Post-Conference Priority
A consensus emerged during the conference’s technical sessions: training solely on static internet data is hitting diminishing returns. The next frontier lies in learning from observation, interaction, and feedback drawn directly from the physical world. For embodied AI, this means co-optimizing models, embodiments, sensors, control systems, hardware components, and real-world scenario data. The 208 embodied products showcased reflect how industry competition is shifting—from isolated motions to full-stack systems. Yet this volume does not translate directly into a procurement list.
This also explains why humanoid robots, AI-powered smartphones, and edge devices are all heating up simultaneously. They all deploy models into constrained real-world environments—where compute and battery resources are limited, networks drop out, sensors introduce noise, and errors can lead to physical harm or privacy breaches. A larger model doesn’t automatically solve these constraints. Real product moats come from system-level trade-offs, closed-loop data pipelines, and long-term operational records.
🔗 Conclusions That Cannot Be Drawn Yet
- You cannot infer procurement conversion rates from 400,000 visitor registrations.
- You cannot conclude that Chinese companies have captured corresponding markets based on 351 global first launches.
- You cannot claim that robots are already running stably at scale in factories just because 208 embodied products were showcased.
- You cannot treat the ¥40.9 billion in signed projects as confirmed revenue for vendors.
- You cannot equate the ¥20.36 billion in intended procurement with actual transaction value.
Each number must be interpreted within its proper context. WAIC brings R&D, products, capital, procurement, and policy together under one roof—but post-event tracking is essential: pilot completion rates, procurement conversion, stable operation, and cash collection.
🔗 Five Data Sets Worth Tracking Over the Next 90 Days
- Of the 32 signed projects, how many have publicly announced commencement dates and named responsible entities?
- Of the 212 procurement needs listed, how many have progressed to formal RFPs or pilot phases?
- Of the 351 globally launched products, how many provide official documentation, pricing, and delivery timelines?
- How many embodied intelligence projects publish continuous operation logs—including human intervention frequency and maintenance records?
- Of the ¥1.18 trillion in intended credit lines, how many have been converted into actual credit approvals and drawdowns?
If the conference or related institutions release follow-up conversion reports, those updates should appear incrementally on the same page—not as a new, synonym-rich “WAIC Outcome Review.” This page currently captures a verified snapshot as of July 22, 2026. All future data should retain its original date and definition.
❓ Frequently Asked Questions
How many exhibits were shown at WAIC 2026?
According to a post-event report by Shanghai Securities News, the exact count was 4,486. CCTV’s closing report used the phrase “over 4,000.” This article adopts 4,486, while noting both sources’ publication dates and precision differences.
Is ¥40.9 billion the event’s “transaction volume”?
No. It represents the total investment value across 32 key Shanghai AI-related signing projects—spanning infrastructure, agent systems, embodied intelligence, and scientific AI. It does not reflect vendor revenue or on-site transaction value.
Has the ¥20.36 billion in intended procurement been completed?
No. The figure reflects anticipated procurement value. Actual purchases still require formal procurement processes, contract execution, and acceptance procedures.
🔗 Sources
- Shanghai Securities News: WAIC 2026 Post-Event Report
- CCTV: WAIC 2026 Closing Coverage
- WAIC Official Website: Project Signing List
Can the ¥1.18 trillion in credit support be added to the signed contract value?
No. This is an intended credit facility for innovative enterprises—not actual funding disbursed. It differs fundamentally from committed investments or procurement agreements, and does not reflect funds already drawn down.
What’s the most valuable takeaway for procurement teams from WAIC?
Procurement teams should walk away with a few concrete, verifiable next steps:
- A clearly assigned business owner,
- Defined input data requirements,
- Clear technical boundaries,
- A two-week pilot plan,
- Quantifiable success criteria—and
- Explicit “pause” signals if things go off track.
Official & Trusted Sources
- World Artificial Intelligence Conference (WAIC) official website, accessed on 2026-07-22 — for official event access and identity verification.
- CCTV’s Focus Talk: Scale, internationalization, and outcomes all reach new highs, published on 2026-07-20.
- Shanghai Securities News: New technologies, new achievements, new collaborations—WAIC 2026 delivers results, published on 2026-07-21.
- China News Service: From “able to converse” to “capable of doing work”, published on 2026-07-20.
Continue Reading
- WAIC 2026 Highlights: Qwen3.8, Kimi K3, Agents, Robots, and Industrial AI
- 2026 Humanoid Robot Fighting Tournament Recap: URKL, WAIC Mech Arena, and the Control Tech Behind Them
Related reading
FAQ
How much time does this take? 20–25 minutes per week is enough if you use one signal source and keep a strict timebox.
What if I miss something important? If it truly matters, it will resurface across multiple sources. A consistent weekly routine beats daily scanning without decisions.
What should I do after I shortlist items? Pick one concrete follow-up: prototype, benchmark, add to a watchlist, or validate with users—then write down the source link.