When the Initial Buzz Fades, Who Will Set the Benchmark?

Author｜Qingyun

Editor｜Xiaobai

Produced by｜Qiangdiao Next

An intriguing trend emerged at this year's WAIC: As parameters, models, and products proliferate, companies are vying for the authority to define what constitutes true 'progress' in the AI landscape.

Step.ai and ZTE Nubia are both claiming the title of 'pioneer in AI agent smartphones.' Baidu is advocating for DAA (Daily Active Agents) as a replacement for traditional metrics like tokens and DAU (Daily Active Users), counting daily active agent counts instead. Huawei is aggregating 1,024 Ascend cards into a super node, aiming to shift the competitive focus from individual chips to comprehensive systems. Robotics companies are using offline production volumes, shipments, and factory deployment numbers to demonstrate their mass production capabilities. However, the products and delivery statuses measured by these various metrics differ significantly.

Each company is selecting a metric that best showcases its strengths. The challenge, however, lies in the fact that while naming your product is straightforward, establishing industry-wide standards is not.

01. Two 'Pioneers' Compete for User Control

At WAIC 2026, two smartphones claimed to be the 'world's first AI agent smartphones.'

One is Step.ai's STEPX Neo. Although it can operate on-site, its design and hardware specifications are still unfinalized, with no price or release date announced. The other is ZTE Nubia's Navi X Ultra, featuring the Doubao phone assistant, which is touted as a 'mass-producible flagship' but was displayed in a glass case without any hands-on or live demonstrations.

The regulatory landscape adds another layer of complexity. Among the first seven generative AI services for mobile phones announced on July 15, the Doubao model embedded in Navi X Ultra was included, while Step.ai's mobile model was not. Both companies claim to be 'first,' but one resembles a development platform, and the other a commercial product. They are not competing on equal terms.

Initially, Step.ai followed a path similar to Doubao. Earlier this year at MWC, ZTE showcased the M153 with Doubao's assistant and the Nubia Z80 Ultra with Step.ai's GUI Agent. The fact that a hardware manufacturer integrated solutions from two different model companies suggests that, at this stage, models are still interchangeable modules.

At WAIC, Step.ai aimed higher. It launched its brand, Step AOS, and the personal agent Amoo, attempting to control models, systems, and terminals simultaneously.

Step.ai's expansion reflects the broader trend of model companies being forced to move up the value chain. Smartphone hardware profits are not Step.ai's primary goal. Instead, it seeks the 'default authority to act for users' and the intent and transaction data behind each task.

Doubao, on the other hand, opts for a lighter approach. By leveraging ZTE's phones, it avoids the risks associated with full hardware development while validating whether agents can become cross-brand system capabilities.

ZTE aims to be the intermediary connecting models and terminals: collaborating with Doubao, integrating Step.ai, and launching its AIOS.

Behind this tripartite division are three different bargaining chips. Doubao has traffic and service ecosystems, ZTE has hardware and system permissions, and Step.ai hopes vertical integration ensures its irreplaceability.

Yet, the first-generation Doubao phone demonstrated that system permissions do not equate to execution rights.

After its launch, the agent faced restrictions in WeChat, Taobao, Alipay, and Meituan. Doubao later tightened controls on finance, gaming, and auto-farming scenarios. The conflict is not just about security. When agents directly complete searches, price comparisons, orders, and payments, users may bypass pages, ads, and recommendation feeds, reducing super apps to mere pipelines and collapsing their business models—something major firms find unacceptable.

Thus, agent phones face two 'sovereignties': the OS decides if it can click, while app platforms decide if the click counts. User authorization for agents does not equate to authorization from WeChat, Taobao, or banks.

Competition for smartphone agents now hinges on interfaces and revenue sharing. Stealthy simulated clicks are only a transitional phase; long-term viability requires apps to open services via APIs, MCPs, or inter-agent protocols, then renegotiate traffic, commissions, liability, and data ownership. This will take longer and may not yield results, as major firms are not short of large models.

What Doubao, ZTE, and Step.ai are competing for is not just a slice of the smartphone market but the right to represent users and command all apps.

02. Baidu Proposes a New Metric, But Peers Are Yet to Adopt It

Baidu is also vying for definition rights, though its focus is on metrics rather than terminals.

In May, Robin Li proposed DAA (Daily Active Agents), arguing that tokens reflect only model calls and costs, not agent productivity. During WAIC, IDC released research predicting global active agents to grow from 28.6 million in 2025 to 2.216 billion in 2030.

Using token consumption to prove AI value is akin to using factory electricity to measure output—a fact most firms acknowledge.

Yet DAA still isn't a definitive measure of output. Agents can replicate rapidly, and one task can spawn dozens of sub-agents. In an extreme case, if Company A uses one powerful agent while Company B splits the same task into 100 sub-agents, the latter's DAA is higher but not necessarily faster, cheaper, or more accurate.

IDC and Baidu's framework must further incorporate per-agent task volume, completion quality, single-task value, and operational costs. This admits that agent count is just one multiplier. What matters commercially is completed tasks, their quality, and costs.

IDC's involvement does not equate to industry endorsement; IDC is a commercial entity. Its January forecast of 2.216 billion active agents by 2030 used 'active Agent count,' not DAA. The WAIC collaboration lent authority but has not proven that other firms accept these metrics.

As of July 18, Alibaba still emphasizes cloud and AI product revenue; ByteDance and Tencent disclose token and product usage growth; MiniMax uses revenue, subscriptions, and overseas share. Public materials show no company adopting DAA as a core metric. Baidu hasn't released its DAA, nor is it clear if it's in any executive's OKR (Objectives and Key Results).

This makes DAA more of a PR initiative by Baidu than an established industry standard. To become common ground, Baidu must disclose its data, the industry must standardize agent identity, deduplication, task success rates, and unit costs, and allow third-party audits.

Of course, Li himself left room in his speech, stating DAA's ultimate validity will be 'tested over the next 12 months.' It remains to be seen which executive's OKR will include DAA next.

Baidu raises a valid question: What should the industry measure after tokens? DAA advances the discussion, but what truly matters is 'how much work is reliably and economically completed.'

03. Super Nodes Grow, but Computing Power Metrics Remain Unstandardized

Huawei's Atlas 950, showcased at WAIC, connects 1,024 Ascend cards into a super node, offering 256 TB of unified memory addressing. Huawei also revealed that over 750 of the previous 384-card super nodes have been deployed.

Atlas 950's broader aim is to promote another comparison method: abandon single-card metrics and focus on the complete system of chips, interconnects, memory, scheduling software, and model adaptation.

This definition favors Huawei. Its single cards may lag behind advanced GPUs, but it excels in system engineering, network equipment, software stacks, and government-enterprise delivery. If clients accept 'super nodes as the computing power unit,' Huawei needn't win every chip parameter.

ZTE proposed an alternative route at the same conference. Its OEX super node emphasizes open decoupling and multi-GPU compatibility, allowing clients to mix domestic chips and avoid early binding to a single chip or software stack.

These routes reflect two sides of China's computing power market: vertical integration reduces migration and debugging costs but risks new lock-ins. Open heterogeneity preserves choice but leaves adaptation, scheduling, and stability to system integrators and clients.

Thus, 1,024-card, 10,000-card, or 100,000-card achievements aren't inherently comparable. The industry lacks data on effective token output, performance per watt, continuous uptime, fault recovery, and migration costs under the same model and task.

Without these, 'larger' only proves companies can connect more devices, not that clients get more intelligence for the same money.

Huawei seeks more than super node orders. It wants the system of Ascend, Lingqu interconnects, and CANN to become the de facto standard for domestic computing power. Whoever defines the minimum competitive unit for computing power can shape the next round of software and hardware ecosystems.

04. Robot Mass Production: First, Define 'Mass'

Robotics firms are competing over what 'mass production' means.

By 2025, China had over 140 humanoid robot manufacturers, launching over 330 products. Omdia counted 13,318 general-purpose embodied robots shipped globally that year.

Zhiyuan announced its 15,000th embodied robot offline in June 2026, with Omdia recording 5,168 shipments in 2025. Another public dataset suggests Zhiyuan shipped approximately 5,200 units in 2025, including around 1,300 full-sized humanoids.

These figures do not contradict; they count different things: embodied robots include wheeled, semi-humanoid, and full-sized models. Offline volumes do not equal shipments, which do not equal production environment entry, which does not equal continuous paid work.

'Mass production' first proves supply chains can build them but not automatically that client ROI (Return on Investment) is valid. This is the embodied AI industry's costliest confusion: treating supply-side capability as demand-side validation.

Factories truly need to compare how many effective work hours a robot completes monthly, its average failure interval, manual intervention frequency, retooling time, and payback period relative to robotic arms, AGVs (Automated Guided Vehicles), or human labor.

Without disclosing these metrics, offline volumes at launches prove production capacity (capacity), not productivity.

Chinese robotics firms' strengths are clear: rapid hardware iteration, complete component availability, and engineers who debug repeatedly in factories. Their greatest risk lies ahead: machines can be built, but stable, paid work won't emerge just because production lines exist.

05. Products Matter More Than Benchmarks

WAIC 2026 showcased more new products but, deeper down, a battle for rules.

Step.ai, Doubao, and ZTE compete to represent users in app operations; Baidu for agent evaluation metrics; Huawei for computing power benchmarks; robotics firms for mass production definitions.

These firms' chosen metrics are no accident. Model firms emphasize intelligence, terminal makers emphasize permissions, cloud providers emphasize calls, hardware firms emphasize systems, and robotics firms emphasize offline volumes. Each metric illuminates part of the truth but obscures each firm's weakest point.

True definition rights cannot be declared at press conferences; they must be granted by others.

Agent phones gain execution rights only if app platforms open interfaces. DAA becomes an industry metric only if competitors, investors, and auditors adopt it. System standards solidify only if clients procure super nodes using unified benchmarks. Mass production becomes commercialization only if factories repurchase and disclose efficiency.

After WAIC, the 'world's first' title remains unclear. The more important question is who can make developers, clients, competitors, and regulators adopt their interfaces, standards, and rulers. This requires truly industry-leading products, ideas, and standards. Throughout tech history, standards emerge not from who has better ideas but from which products lead under different standards.

Products can launch in a day; industry rules are not established so easily.

- END -