ByteDance bets big on AI independence, shuns US model shortcuts

The gist
ByteDance is betting billions on building world-class AI from scratchdrefusing to copy U.S. rivals and doubling down on homegrown innovation despite regulatory and hardware hurdles.
What to know
- Since 2023, ByteDance has banned distilling U.S. AI models and reaffirmed this stance in 2026, choosing slower, independent progress over shortcuts.
- Led by ex-DeepMind VP Wu Yonghui, ByteDance is training a 10 trillion-parameter Mixture of Experts model, pouring over $30 billion into AI in 2026 alone with plans to hit $100 billion by 2027.
- Facing U.S. GPU export controls, ByteDance is mixing older Nvidia chips, Chinese hardware, and custom TSMC GPUs to keep its AI ambitions on track.
ByteDance's No-Distill Doctrine
ByteDance’s internal ban on using competitor AI models as training shortcuts is a calculated sacrifice of speed for proprietary innovation and regulatory safety.
Since 2023, ByteDance has enforced a strategic internal ban on distilling AI models originating from U.S. companies and rivals, a policy reaffirmed publicly in 2026 amid heightened scrutiny of AI ethics and security. Founder Zhang Yiming explicitly directed the Seed AI team to avoid relying on distillation techniques that use competitor models as “teachers,” emphasizing a willingness to accept short-term competitive disadvantages in favor of long-term independent innovation. This nuanced stance permits self-distillation within ByteDance’s own models but prohibits leveraging external or open-weight models, underscoring a deliberate choice to prioritize proprietary development over rapid capability gains.
While early reports linked the distillation ban to mitigating U.S. regulatory backlash—especially given TikTok’s fraught national security scrutiny—the decision was primarily driven by ByteDance’s strategic vision rather than immediate geopolitical pressures. Internal analysis clarifies that TikTok’s regulatory challenges played almost no role in the 2023 policy’s inception, although by 2026, concerns about attracting Washington’s attention reinforced the company’s cautious approach. This reflects ByteDance’s heightened sensitivity to geopolitical risks, informed by its prior experience with U.S. government scrutiny, and a conscious trade-off to avoid shortcuts that might invite further regulatory complications.
ByteDance’s ban on AI model distillation illustrates a broader strategic dilemma in China’s fiercely competitive AI landscape, where the company willingly sacrifices rapid advancement opportunities to maintain compliance and safeguard its long-term goals. As rivals like Moonshot and Kimi K3 accelerate their development partly through distillation, ByteDance’s leadership, including Zhang Yiming, underscores the importance of regulatory compliance and sustainable innovation over short-term gains. This approach signals a calculated bet that enduring success depends on navigating geopolitical constraints as much as technological prowess.
It is important to note that the distillation ban applies specifically to the Seed AI research team and does not constitute a companywide prohibition on all forms of model distillation. This targeted policy suggests a calibrated internal governance approach, balancing innovation flexibility with risk management within ByteDance’s AI divisions. By limiting the directive to a core research unit, ByteDance maintains room for experimentation elsewhere while clearly signaling its strategic priorities in sensitive areas of AI development.
Building AI From Scratch
Led by a former DeepMind VP, ByteDance is pouring billions into training a 10 trillion-parameter model entirely on its own data and research, refusing to shortcut with outside models.
ByteDance’s strategic commitment to long-term independent AI innovation is exemplified by its explicit ban on distilling rival AI models, a policy established as early as 2023 under founder Zhang Yiming’s directive. This approach prioritizes original large-scale pre-training over shortcuts like leveraging competitors’ outputs, with Zhang emphasizing the willingness to accept temporary setbacks rather than compromising on foundational research integrity. While self-distillation using ByteDance’s own models remains permissible, the company firmly prohibits using external proprietary or open-weight models as “teachers,” underscoring a deliberate focus on building proprietary frontier AI capabilities.
By early 2026, ByteDance had mobilized massive investments and top-tier talent to realize this vision, spearheaded by Wu Yonghui, a former Google DeepMind VP and Google Fellow who joined ByteDance in 2025 to lead the Seed AI research division. The company is pre-training a groundbreaking 10 trillion-parameter AI model employing Mixture of Experts (MoE) architecture, designed to be both physically trainable and economically deployable. This ambitious project involves around 2,000 employees and is backed by capital expenditures exceeding $30 billion USD in 2026 alone, with total AI infrastructure spending potentially reaching $100 billion the following year, reflecting a deep commitment to original AI research rather than reliance on existing models.
Centralizing Data Power
A newly unified AI Data & Security unit now governs ByteDance’s proprietary data pipeline, signaling data quality and control as the company’s decisive AI edge.
In August 2026, ByteDance solidified its commitment to proprietary data as a strategic cornerstone by creating a centralized AI Data & Security unit led by Adam Wang, elevating data governance to the same level as its flagship business units like Seed and Douyin. This move reflects founder Zhang Yiming's 'no distillation' philosophy, emphasizing long-term proprietary R&D fueled by exclusive, high-quality data rather than short-term gains, a stance reinforced by CEO Liang Rubo's pledge to accept temporary setbacks in pursuit of sustained innovation.
The consolidation effort, initiated in June 2026, merged disparate teams such as Global Data, DMC, and AIDP into a massive, integrated organization overseeing the entire data lifecycle—from sourcing and synthetic data generation to cleansing and quality evaluation—thereby streamlining operations and optimizing personnel deployment. This comprehensive approach supports ByteDance’s expansive AI ambitions, including training large language models and the Seedance video generation model, where data evaluation alone involves over 1,000 specialists, highlighting the company’s multimillion-dollar investments and its view of proprietary data as a decisive competitive advantage.
Hardware Hurdles and Hacks
Blocked from top-tier NVIDIA chips, ByteDance is forced to juggle older GPUs, Chinese hardware, and custom TSMC chips, making technical workarounds a necessity for AI advancement.
By mid-2026, ByteDance’s AI ambitions have been significantly shaped by stringent U.S. export controls that bar direct access to NVIDIA’s most advanced GPUs, such as the H100 and B200 models. This blockade forces ByteDance to adopt a hybrid hardware strategy, allocating roughly half of its massive 200 billion yuan AI budget to procure a mix of older Nvidia GPUs and domestically produced Chinese chips, hedging against further geopolitical restrictions. While competitors like OpenAI and Google enjoy unfettered access to cutting-edge hardware, ByteDance must innovate within these constraints to sustain its AI model training scale.
To overcome the computational bottlenecks imposed by limited hardware capabilities, ByteDance employs a Mixture of Experts (MoE) architecture that activates only a fraction of its staggering 10 trillion parameters per query. This architectural choice is critical, as it enables the company to train and deploy massive models economically and physically despite relying on less efficient, older GPUs. The MoE approach thus becomes a technical necessity rather than a mere optimization, allowing ByteDance to stretch constrained resources without resorting to copying rival models.
In parallel, ByteDance is investing heavily in developing custom GPUs manufactured by TSMC, with plans to scale production to 350,000 units. However, these bespoke chips face a steep uphill battle to match the performance and ecosystem maturity of NVIDIA’s hardware, a process expected to take years. This long-term bet reflects ByteDance’s strategic response to hardware scarcity but also underscores the immediate limitations: training on inferior or nascent hardware not only prolongs model training times but can fundamentally degrade the AI’s capabilities due to less efficient training steps.
Risk Management Over Speed
Leadership’s refusal to distill U.S. models—even under competitive pressure—reflects a governance-first strategy to avoid regulatory backlash and protect ByteDance’s long-term interests.
Since 2023, ByteDance has enforced a strict internal ban on distilling AI models originating from U.S. companies, including OpenAI’s GPT, as part of a strategic choice to prioritize long-term independent innovation over short-term competitive gains. Founder Zhang Yiming emphasized that the Seed AI lab should accept temporary setbacks rather than rely on outputs from stronger competitors, underscoring a governance approach that balances proprietary innovation with caution amid increasing geopolitical tensions and regulatory scrutiny.
By 2025 and into 2026, despite internal debates triggered by competitors like DeepSeek releasing strong models, ByteDance maintained its stance against AI distillation, reflecting a deliberate risk management strategy to avoid attracting U.S. government scrutiny and potential sanctions. Founder Jiang Yimin explicitly instructed employees to forgo shortcuts even at the expense of falling behind domestic rivals, highlighting the company’s prioritization of national security and reputational risk over rapid AI enhancement.
In response to escalating U.S. export controls restricting access to advanced NVIDIA GPUs critical for AI training, ByteDance has strategically diversified its AI hardware investments by allocating roughly half of its $30 billion 2026 AI capital expenditure to both Nvidia hardware and domestic Chinese chips. This hedging strategy, coupled with significant reliance on Microsoft’s cloud services—where ByteDance reportedly spends over $1 billion annually—and development of custom GPUs manufactured by TSMC, exemplifies the company’s corporate governance approach to mitigating geopolitical risks while sustaining ambitious AI innovation.
Chasing the AI Titans
With a 10 trillion-parameter model and $100 billion at stake, ByteDance is betting on scale and original research to compete with Western AI giants on the world stage.
ByteDance is aggressively positioning itself as a formidable contender in the global AI arena by developing a massive 10 trillion parameter model designed to rival Anthropic’s Claude Mythos, which itself is estimated to have 8 to 10 trillion parameters. This ambitious project, led by former DeepMind executive Wu Yonghui, underscores ByteDance’s strategic emphasis on building large-scale, long-term AI capabilities rather than relying on short-term shortcuts like model distillation, as directed by founder Zhang Yiming. By focusing on scale and sustained innovation, ByteDance aims to close the gap with leading Western AI labs and assert itself at the forefront of advanced AI technology.
The scale of ByteDance’s AI initiative is staggering, involving around 2,000 employees and an investment trajectory that could reach $100 billion in 2027, up from $70 billion in 2026. This level of commitment not only reflects a desire to surpass domestic rivals but also signals a clear intent to establish global leadership in AI development. Such massive resource allocation highlights ByteDance’s strategic prioritization of AI infrastructure and talent, positioning the company to compete head-to-head with Western giants and reshape the competitive dynamics of the international AI landscape.


