Technical prowess, a commercialization closed loop, and real-world scenario implementation are all indispensable.

Written by | Zhao Weiwei

"First, enable the most advanced enterprises to utilize the safest AI. Second, transform AI's hallucinations and risks into manageable engineering problems. Only by excelling in these two areas can AI deeply penetrate the semiconductor industry through the cloud, helping everyone scale the peaks of AI," said Tan Dai, President of Volcano Engine.

On the afternoon of the Volcano Engine FORCE Conference, 15 parallel industry and product sub-forums covered sectors such as automotive, film and television, and finance. Tan Dai appeared at the AI chip semiconductor sub-forum.

The importance is self-evident. The "most advanced enterprises" in chip semiconductors have become the cornerstone of AI industry development. Over the past decades, the chip industry has never been as deeply integrated with upper-layer applications as it is today.

Previously, Tan Dai demonstrated a real-world scenario of Doubao 2.1 Pro. In a chip design RTL test, Doubao ran continuously for nearly 18 hours, underwent 9 rounds of iteration, and completed a full engineering workflow including simulation, testing, and comprehensive checks, validating the model's production-grade code delivery capabilities in real-world engineering scenarios.

At the chip semiconductor exhibition area of the Volcano Engine Conference, five representative companies in China's GPU ecosystem—Iluvatar CoreX, Moore Threads, Huaqin Technology, MetaX, and Arm China—made a collective appearance, covering the entire domestic GPU industry chain.

Today, they are all partners of Volcano Engine, relying on its elastic computing power or using AI code tools like TRAE for chip development, verification testing, and accelerating R&D processes and project deliveries.

The Volcano Engine FORCE Conference marks the annual renewal of the Doubao Large Model. The latest Doubao 2.1 Pro and Seedance 2.5 models showcase technical prowess; chip semiconductor implementation cases validate the model's deep adaptability to hardcore industrial scenarios; and Doubao's paid model aims to prove the feasibility of a C-end closed loop for AI commercialization in China.

Technical prowess, a commercialization closed loop, and real-world scenario implementation are all indispensable. Will this snowball keep growing? Volcano Engine is already in the first tier of domestic AI vendors.

1. Is the Professional Doubao Worth It?

Based on the latest Doubao 2.1 series large models, Doubao officially launched a professional paid subscription with three monthly plans: Standard at 68 RMB/month, Enhanced at 200 RMB/month, and Premium at 500 RMB/month.

This is a clearly segmented offering, catering to the diverse needs of 200 million daily active users: The Standard plan includes all benefits of the free version, with access to the 2.1 Pro model, offering over 5 times the usage quotas for office tasks, expert modes, and other functions compared to the free version; the Enhanced plan provides 4 times the quotas of the Standard plan; and the Premium plan offers 10 times the quotas.

Compared to overseas products like ChatGPT and Claude, Doubao's pricing strategy is more localized: with moderate entry-level pricing and significantly lower prices for high-tier plans than overseas competitors, catering to both general and professional users.

ChatGPT and Claude charge 54 RMB and 135 RMB respectively for their entry-level plans. Doubao's Standard plan is priced lower than Claude but higher than ChatGPT. Meanwhile, the top-tier subscription plans for ChatGPT and Claude cost up to 1,355 RMB per month, whereas Doubao's corresponding tier is significantly cheaper, attracting budget-conscious professional users and SMEs.

Of course, Doubao Professional emphasizes that the free version's daily functions and quotas remain unaffected. For most users' daily life scenarios, Doubao's existing functions and quotas already suffice.

However, price is just the surface; the key lies in the match between price and task, i.e., the actual output per unit quota.

For budget-conscious professional users and SMEs, if Doubao Professional can only generate basic usable code for complex coding needs while Claude outputs directly runnable finished products, even with a lower price, longer debugging times would offset Doubao's price advantage.

On another level, Doubao emphasizes that in office tasks, users can better accomplish professional work such as Office suites, application development, data analysis, professional design, process automation, and financial analysis. In these professional scenarios, where the model can precisely match business needs, price becomes less of a core consideration, and output value is key.

Notably, Tan Dai believes that the current Doubao Large Model 2.1 Pro has reached a usable standard, "comparable to Claude Opus 4.6, entering the usable threshold for Agents."

Doubao Professional represents a significant step for China's large model products: boosting C-end user willingness to pay is challenging, but domestic large model vendors must continue exploring paid models.

Because post-C-end charging, user retention, renewal rates, and word-of-mouth spread (spread) are the true indicators of product strength. Even ChatGPT faces issues where subscription user penetration falls far short of expectations. Relevant media has also revealed that ByteDance will not prioritize paid penetration as a core KPI in the short term; the primary task is to validate whether the functional value of the Professional version gains market recognition.

First, broaden functional coverage scenarios, then deepen professional capabilities in each field. How to make users not just try but truly form usage habits is a challenge Doubao will face in the future.

Ultimately, from early multi-modal capability integration to deep embedding in various professional toolchains and laying out AI agent-based office solutions, this development path holds long-term value.

2. Chip Development: How Can AI Accelerate?

Undeniably, the fusion of artificial intelligence and semiconductor technology is the core driver of industrial development in the AI era.

Gao Siyuan, Head of Semiconductor Industry Solutions at Volcano Engine, believes that "AI For IC" is gradually permeating the entire semiconductor industry chain, from hardware design to software development and operations management. AI is progressively unlocking various scenarios, with implementation steadily advancing through a three-tier architectural system: model layer, business layer, and interaction layer.

In particular, ByteDance's AI-native integrated development environment (IDE) tool, TRAE, is playing an increasingly critical role in the chip semiconductor industry.

Iluvatar CoreX, a provider of general-purpose GPU chips and high-performance computing systems, shared two scenarios where AI accelerates chip development:

In accelerating library porting, code interpretation is time-consuming and labor-intensive. After agent intervention, engineers still need to assist initially in correcting the agent's outputs. However, as processes accumulate, various rules and constraints gradually solidify into reusable skill modules. Later, the agent automatically generates test code (test bench) and completes functional and performance verification on hardware.

In chip verification, writing assembly code requires at least two engineers to collaborate closely: one understands chip verification, and the other understands compilers. Their knowledge systems rarely overlap, leading to high communication costs, while the supply of versatile talent (interdisciplinary talent) is extremely scarce.

With the help of agents, engineers' implicit professional knowledge is precipitate ( precipitate ) into reusable assets on the platform, forming a complete compiler skill library, ultimately creating a complete compiler Skill for subsequent verification and direct reuse.

"An AI coding agent can save 90% of manpower, with costs about one-tenth of a senior engineer's," said Du Ning, Head of Enterprise Agent Product Solutions at Volcano Engine.

Chip R&D often takes years, with tape-out costs reaching tens of millions of RMB. An AI product that compresses R&D cycles and breaks bottlenecks means a chip design company can theoretically advance more chip design projects simultaneously without Large scale expansion (massive expansion) of R&D personnel, shortening tape-out cycles and accelerating product iteration.

In fact, AI programming has become a trend for internal enterprise applications.

The forum revealed that ByteDance internally uses AI programming tools to accelerate product development like Douyin. 97% of employees use AI programming tools, with over 90% of new business code generated by AI, and even for rewriting existing legacy code, AI participation exceeds 20%.

However, for chip companies, data security and IP protection are more critical.

For many semiconductor companies, uploading core chip code to public cloud AI services is an untouchable security red line, meaning privately deployed AI coding tools are the truly viable solution.

This is why Tan Dai emphasized the safest AI from the start. Volcano Engine launched a full-stack Security for IC solution for the semiconductor industry, covering the entire workflow of computing power, data storage, model invocation, and agent operation, while synchronously optimizing overall enterprise security protection levels.

3. Seedance 2.5: Short Films Are Not the Goal

Throughout the Volcano Engine FORCE Conference, besides the Doubao Large Model 2.1 Pro, another highlight is the Doubao video generation model, Seedance 2.5, set to launch in July.

Since its release in February this year, Seedance 2.0 quickly gained popularity in the film and television short film and advertising industries due to its rich character performances and professional shot compositions, directly stimulating the short film industry's shift from live-action to AI-generated content.

Chengdu Jinrui Hebang is a traditional film and television company transformed by AI. At the Seedance Film and Television Forum, the person in charge (person in charge) introduced that the company now cuts into AI skill training, gradually expanding to AI audiovisual content production and IP development, constructing three major business segments. Leveraging Seedance's production capabilities, the company reduced the price of wedding microfilms, which previously started at five figures, to 999 RMB.

The popularity of AI videos is now a reality. There have been rumors that Seedance 2.0 generates over 1 billion RMB in monthly revenue, but Tan Dai debunked this as inaccurate in a media interview, stating that short films are just a small scenario, with building a world model as the ultimate goal.

Short films are only part of the implementation in physical industries. The value of Seedance's extension into industrial data infrastructure is gradually emerging. Currently, Seedance has landed in fields such as embodied intelligence, industrial manufacturing, and autonomous driving, providing new tool capabilities for business needs like data synthesis, scenario simulation, and process demonstration.

Notably, in embodied intelligence, many companies currently rely on manually collecting real-world data for training, which is relatively costly. Using Seedance to synthesize training data essentially uses the virtual world of the video model to compensate for the scarcity of real-world data. The industry's core need is not high-quality finished videos but highly realistic and diverse simulation training data.

Seedance 2.5, set to launch in July, further enhances capabilities, enabling direct output of 30-second single-segment native videos, supporting joint generation of up to 50 full-modal assets while maintaining visual consistency in local edits. During the demonstration, Seedance collaborated with Bingo Group, owned by Stephen Chow, obtaining AI creation authorization for three Stephen Chow films.

Another point worth noting is the productivity transformation brought by Seedance.

Tan Dai provided a methodology: look at the ratio of API calls between weekdays and weekends. Before Seedance 2.0's launch, platform call peaks concentrated on weekends, mostly for user leisure and entertainment trials, making the product more of an entertainment tool. After Seedance 2.0's release, the data reversed, with weekday calls far exceeding weekend calls.

This indicator relies not on subjective evaluations or user surveys but directly reflects real usage behavior. Tan Dai refers to this node as the "productivity tipping point," defined by the boundary: "Examine the existing business processes in each industry and what each process requires in terms of model capabilities. Only by meeting these requirements do you cross the boundary."

Productivity transformation is the beginning of snowballing in the AI era. Every manufacturing product video, every frame of embodied intelligence training material, continuously rolls up the industrial value snowball, carrying more information about the real world for the model to keep moving forward.

Reviewed by | Chen Qiulin

END