What are Volcano Engine's latest strategies and insights on commercial implementation?

At this year's Volcano Engine FORCE Conference, my most direct impression was two words: heavier focus.

Last winter's conference featured crowded booths for video generation and AI short dramas, attracting many developers and creators. This year, 'AI + Industries' dominated the exhibition hall, with each sector showcasing major enterprise cases—top energy, manufacturing, automotive, consumer companies, and renowned universities.

I attended a financial sub-forum where speakers and panelists included general managers of bank intelligent operation centers, securities firm digital offices, and insurance company technology departments...

Volcano Engine is becoming more B2B, or more precisely, more focused on large enterprises (B2BigB).

The conference theme was straightforward: AI has advanced rapidly in the past six months, with token consumption rising exponentially. This stems from models crossing a tipping point—they can now genuinely perform tasks and impact productivity, enabling broader AI adoption across industries and enterprises.

01 Models Surpass Production 'Tipping Point'

'By 2026, tokens will absolutely be the hottest tech term,' said Shan Zhiguang, Director of the Information and Industrial Development Department at the State Information Center. From the second half of 2025 to now, token calls on OpenRouter have surged 'super-exponentially.'

Data from the National Data Administration is even more striking: token calls in China have grown over 1,000-fold in two years. Tokens are no longer just a technical concept but now carry computational and monetary value, giving rise to 'Token Economics.'

The view of Zheng Weimin, an academician at the Chinese Academy of Engineering, is gaining wide recognition:

Tokens now serve three roles: the basic unit for models to process information, the measurement unit for computational consumption, and the standard billing unit for industries. Industrial data, once tokenized, becomes a 'priceable production factor.'

The latest data announced by Tan Dai, CEO of Volcano Engine, is astonishing: as of June, the daily average token calls for the Doubao large model exceeded 180 trillion. In December last year, 100 companies on Volcano Engine consumed over 1 trillion tokens annually, forming the 'Trillion Token Club.' This number has now doubled to 200.

Why such growth? Tan Dai gave a keyword: 'tipping point.' In coding and agent domains, Anthropic's Opus 4.6 was the first global model to cross this threshold, immediately integrating into personal and enterprise production systems. Reports suggest Anthropic's Q2 revenue could reach $10.9 billion. At this growth rate, it may become one of the world's highest-grossing tech companies next year.

This demonstrates that once a model crosses the tipping point, paired with a hit product, its adoption and monetization can far exceed expectations—even without promotion, users spread it voluntarily, achieving 10x, 100x, or even 1,000x growth in a short time.

In video generation, Tan Dai highlighted Seedance 2.0, the first global model to cross the tipping point. 'Many clients told me video models felt like toys before Seedance 2.0.' But once it crossed the threshold, with improved consistency, cinematography, and realism, it unlocked advertising, film, television, and science communication markets.

However, Tan Dai believes this surge extends far beyond short dramas—'this may just be a small scene long-term.' He is more interested in Seedance's penetration into various industries.

For example, manufacturing and retail companies use it for product explanations, embodied AI companies for synthetic data generation, and autonomous driving companies for simulating extreme weather scenarios. Looking ahead, this year and next will be peak periods for world model training. Video generation, capable of extensive unsupervised learning on existing data, offers an effective path for training world models.

At the conference, Doubao Large Model 2.1 Pro was released. Tan Dai stated it outperforms Opus 4.6 in multiple coding and agent evaluations, also crossing the production-grade tipping point. It is now available via API on Volcano Engine and integrated into products like Doubao, TRAE, and Kouzi.

How to define the 'productivity tipping point'? When asked, Tan Dai replied that the boundary is determined by whether a model meets industry and process requirements—data doesn't lie. For instance, before Seedance 2.0, weekend calls far exceeded weekdays; now, it's the opposite, indicating real work usage during business hours.

At the conference, Tan Dai outlined four core capabilities for LLMs crossing the tipping point: writing 'production-grade' deliverable code for integration into coding and R&D workflows; excelling in instruction following, hallucination control, and adapting to various harnesses for complex, long-term agent tasks; leading multi-modal understanding to better address real-world problems and enable GUI interaction; and operating stably and efficiently at scale in enterprise environments.

02 Business Above the Tipping Point

With the tipping point arrived, Volcano Engine aims to dive deeper into various industries.

The exhibition hall showcased diverse 'AI + Industry' applications, categorizable into four dimensions:

- Smart terminals (e.g., automotive, mobile devices)

- Traditional pillar industries (e.g., manufacturing, finance)

- Content creation and creative expression

- Cutting-edge research (e.g., chips, embodied AI)

Visually, creative expression seems closest and most compatible with Volcano Engine, while traditional pillars appear distant and heavily fortified. The question is: how can it expand from its doorstep to the farthest regions?

Let's start with the closest: creative expression. Seedance 2.0's launch before the Spring Festival triggered explosive growth in short dramas, with AI-generated series reaching 1 billion views. A 200–300-minute film could be produced by AI in a week, compared to at least a month and 10–20 times higher costs for human production. By May, e-commerce advertising surged, with advertising's share of token consumption visibly rising.

How is it used specifically? In Anmuhi's 'Whole Blueberry Yogurt' ad, a blueberry flies past landmarks, transforms into a football for a goal, and reverts to a blueberry falling into yogurt—maintaining consistent product packaging details across large-scale camera movements is its core strength.

Cost calculations are more direct: compared to traditional filming and editing, Seedance generates a 30-second ad for 30–45 RMB per piece, stably producing hundreds of materials daily.

Ruichi Advertising goes further, reusing hit product (hit product) structures by deconstructing a hit and applying its content logic to new SKUs, achieving viral production of hits.

Volcano Engine is largely driven by demand here—once models launch, markets explode independently, representing the lightest approach.

Next is smart terminals. Recently, a new metric emerged for automakers: 'AI concentration,' with 2026 dubbed the 'first year of intelligent agents in vehicles.'

Yang Liwei, VP of Volcano Engine, shared data: Doubao has entered 50 automotive brands, 150 models, and over 7 million vehicles, with daily calls exceeding 30 million. Vehicle control accounts for 53% of scenarios, navigation 29%, and media 10%. The paradigm is shifting: past cockpits relied on a 'classification hub' for task allocation, but this year, stronger models began end-to-end reasoning for dialogue.

What's the in-car experience like? In Mercedes' all-electric GLC, Audi E7X, and Buick Zhijing E7, the system understands and executes complex, simultaneous requests from multiple passengers.

A heavier collaboration is AIVA, a new brand formed by industrial capital like Seres and CATL. Volcano Engine, as a key technical partner, joined from the definition stage ('AI first, then car'). For example, when set to 22°C, the system considers season, clothing, and user state to understand true comfort, adapting more closely to user preferences over time.

At the automotive sub-forum, Chery's global expansion case impressed me most. Chery has been China's benchmark for automotive exports for 23 years, shipping to over 130 countries and 6 million+ vehicles. Now, with Volcano Engine, it's expanding into ASEAN—where 60% of the population is under 35, making smart cockpits a core differentiator for brand premiumization. The hardest part is true localization, not just translating Chinese to foreign languages but 'ecosystem-level' work. Chery and Volcano Engine integrate local ride-hailing and entertainment ecosystems while ensuring data compliance.

At this stage, the approach becomes heavier: from selling cockpit capabilities to AIVA's involvement from the definition stage and Chery's joint ecosystem localization, Volcano Engine starts using manpower and deep co-creation to build benchmarks.

Further out are the most fortified pillar industries. In finance, securities lead, with six or seven of the top ten brokerages developing AI investment advisory products with Volcano Engine.

Huatai Securities launched China's first AI-native trading app, 'AI Zhangle,' where users ask, 'What are today's hot stocks?' AI scans market news, sentiment, and announcements to capture trends and excavate (dig out) related sectors, even executing voice-ordered trades. More critical is their division of labor: Huatai's self-developed financial large model handles professional judgments, Doubao summarizes financials, sentiment, and policies, while Volcano Engine provides the base with web-connected question-answering agents, privatized computing power, and large model firewalls to meet financial-grade security and compliance.

CICC Wealth goes 'heavier,' embedding research from 300+ analysts and experience from thousands of advisors into a jointly customized financial large model, creating a new AI app set for beta testing.

In the most fortified sectors, clients' professional expertise increasingly takes center stage.

In manufacturing, Feihe Milk is a prime example of 'full adoption.' In 2025, Feihe and Volcano Engine formed a strategic partnership, first integrating data across R&D, production, marketing, and animal husbandry. They deployed Volcano's search and recommendation system on the company's private e-commerce platform, using native GenAI technology for chat-like intelligent shopping guidance, achieving over 90% accuracy in recommendations. In 2026, based on Volcano's HiAgent platform, they built an intelligent decision-making hub within Feihe, offering optimization suggestions at every stage. Employees have created over 4,000 self-built agents, with factory wastewater specialists using it for monitoring and alerts, finance for invoice audits, and audit for gift reconciliations.

Finally, in cutting-edge research, chip industry cases shine. In chip RTL design testing, Doubao 2.1 Pro ran for nearly 18 hours, completing six core modules, 1,303 lines of code, and nine iterative rounds, fully executing simulation, testing, and comprehensive checks—tasks that previously required three to five engineers over weeks.

From demand-driven advertising to deploying hundreds for automotive benchmarks, then to client-led traditional pillars, AI implementation follows an increasingly 'heavy' path. Currently, Volcano Engine aims to influence more industries through benchmarks and deeper AI integration.

03 Where Do Volcano Engine's Levers Come From?

What are Volcano Engine's levers for entering various industries? I summarize five key points from the conference:

First, models. Tan Dai repeatedly emphasized that broader industry coverage relies on model advancements: 'Stronger models make barriers easier to overcome.' Seedance's nearly 50% overseas user base exemplifies this.

Many attribute ByteDance's strength in video generation to its models, but building a strong video generation model requires robust video understanding, VLM, large language, and even coding capabilities. This comprehensive ability (comprehensive capability) allows it to break through more industry barriers, such as wastewater monitoring at Feihe and pipeline inspections at State Grid—'if AI can understand, it knows what to do.'

I also noticed a trend in ByteDance's large model training team: increasing end-to-end training. For example, Doubao Real-Time Voice 3.0 is no longer trained in three separate stages ('sound-to-text → large language processing → synthesis') but end-to-end, enhancing emotional understanding, naturalness, and human-likeness. Future models and agents may also emphasize end-to-end training with less manual orchestration.

Second, Harness. Tan Dai stated, 'Before models reach the tipping point, improving model capabilities is crucial. After crossing it, Harness becomes vital,' with priorities shifting rapidly: 'This month, model upgrades matter; next month, Harness does.'

Priorities alternate between the two: 'The goal isn't just a good model or API but how this model and Harness suite land in enterprise environments to solve specific business problems.' This involves system integration, data connectivity, security, and agent authentication and compliance.

Third, diversified entry points. This year, I've seen many firms emulate Anthropic, launching Coding tools like Claude Code and general-purpose personal assistants like Claude Cowork. Volcano Engine's approach differs with its '1+N+X' system.

Among them, '1' represents the AgentSphere Digital Employee Dispatch Station, which integrates all digital employees into a unified system for operation and measurement. 'N' represents multiple out-of-the-box intelligent applications that address general enterprise needs. 'X' represents an infinite number of continuously evolving business intelligence applications, allowing enterprises to continuously create, run, observe, and optimize their exclusive digital employees based on TRAE, Kouzi, and HiAgent.

Tan Dai told Shuzhi Qianxian that solving all problems with a single product at present is 'quite challenging,' so a diversification (diversified) approach is necessary. 'There may be more in the future, or it may converge; we'll have to see as AI develops.'

The fourth initiative is the new FDE team. This year, Volcano Engine has specially established an FDE (Frontline Deployment Engineer) team to deeply collaborate with industry benchmark clients.

Tan Dai said that the essence of collaboration is to 'enable clients to better understand what AI can do, while we also gain a deeper understanding of what AI can do for them.' FDEs are not sales or pre-sales personnel; they must possess 'the ability to implement AI code' and have backgrounds in various industries. 'Some work in bioengineering, so they naturally have strong industry-specific know-how.' Tan Dai also mentioned a crucial strategy: identifying super individuals within client organizations who are the first to adopt AI and working with them to achieve success.

This year, Volcano Engine also plans to co-host 500 'Volcano Cup' events with enterprises to further amplify the impact. Currently, the FDE team covers industries such as automotive, healthcare, education, finance, semiconductors, and other representative sectors.

The fifth initiative is the ecological boundary. In any sub-forum, clients, partners, and Volcano Engine cannot avoid this topic. At the automotive sub-forum, Yang Liwei admitted that on the one hand, platform companies must create real value in scenarios, and on the other hand, 'how can we do less ourselves and combine partner capabilities to improve experiences and products while achieving lower-cost implementation?'

Zhang Yongwei, Chairman of the Automotive 100 Association Research Institute, bluntly stated, 'No one wants to lose the C-end market, but how can we capture it?' This is a core issue that AI poses for the entire automotive manufacturing industry. He believes that in the AI era, no single enterprise will dominate everything; instead, there will be a greater emphasis on industrial division of labor and collaboration, as requirements for technical quality continue to rise across all aspects.

An on-site attendee summarized that currently, AI has penetrated various industries, moving from single points to multiple points, but it has not yet truly permeated entire enterprise organizational systems. At present, 'it is mainly tool-based, efficiency-enhancing, or auxiliary.'

Tan Dai believes that we are still in the early stages of implementation. 'Last year, we had run 500 meters; this year, we've run a little over a kilometer. But this kilometer has already crossed the qualitative change point in production.' In 2025, Volcano Engine ranked first in the Chinese public cloud market for large model invocation volume, with a 49.5% market share. Tan Dai believes that this market still has 10x or even 1000x growth potential, and short-term wins or losses are not as important. 'What matters more is whether we can use better AI capabilities to serve better enterprises.'

In a video message at the conference, ByteDance CEO Liang Rubo stated that climbing the AI peak is ByteDance's most important task at present, and Volcano Engine's MaaS business is becoming a foundational business for ByteDance. 'Our commitment will be long-term and resolute.'