# ByteDance Unleashes AI “Wildcards” in Early 2026: Seedance 2.0 Goes Global, 1,000-Person Chip Team Takes Shape As 2026 began, ByteDance dropped one blockbuster announcement after another. In the large model space, the company recently launched **Seedance 2.0**, its AI video generation model. During closed beta, this model — capable of creating cinematic videos from text or images — swept the internet with its stunning performance. On the hardware front, sources say ByteDance has built an **in-house chip team of over 1,000 people**, with four self-developed product lines: AI chips, CPUs, VPUs, and DPUs. Amid the surging global AI boom, top tech giants worldwide are doubling down on investment, sparking an all-out race for AI dominance. As a major player in this bloodless battle, ByteDance has spared no expense: - It poured **80 billion yuan** into AI in 2024; - Its 2025 capital expenditure is expected to hit **150–160 billion yuan**; - For 2026, the firm plans another massive **160 billion yuan** investment in AI. Rough calculations show ByteDance will plow an **astounding 400 billion yuan** into AI in just three years — a sum nearly matching the annual GDP of a medium-sized prefecture-level city in China. This aggressive AI spending has made ByteDance a favorite in capital markets. Last November, **Capital Today**, led by “Venture Capital Queen” Kathy Xu, bought $300 million worth of ByteDance shares, pushing its valuation to **$480 billion** (roughly 3.4 trillion yuan). Tianyancha data shows Zhang Yiming founded ByteDance in 2012. As the firm’s valuation hit new highs, Zhang’s net worth soared. On the **2025 Hurun China Rich List** released last October, Zhang ranked second with 470 billion yuan in wealth — a 34% year-on-year jump. ## Catch Up, Then Surpass: Seedance 2.0 Takes the World by Storm Many netizens’ first taste of generative AI video came in February 2024, when OpenAI unveiled Sora. With short or detailed prompts — or a single static image — Sora could generate cinematic, lifelike scenes with multiple characters, varied actions, and detailed backgrounds, producing up to one minute of 1080P HD video. Sora rewrote expectations at the time, with many calling 2024 the “first year of AI video generation”. ByteDance’s rival models arrived much later. In September 2024, it released **PixelDance** and **Seaweed**, developed by its AI Lab and Seed teams respectively — three months behind Kling AI, built by competitor Kuaishou. The two models only launched publicly on Jimeng, ByteDance’s video generation app, in November 2024. In 2025, the AI video race intensified as OpenAI’s Sora 2 and Google’s Veo 3 arrived. ByteDance fought back. In April 2025, it merged the AI Lab into the Seed team. The next month, it combined PixelDance and Seaweed to launch **Seedance 1.0 lite**, accelerating its push into video generation large models. In June, it released Seedance 1.0 and 1.0 Pro, supporting text and image inputs to generate seamless multi-shot 1080P videos with drastically shorter creation times. Last December, Seedance 1.5 Pro arrived with audio-visual sync, multi-language voice acting for multiple characters, and other advanced features. Just two months later, **Seedance 2.0** exploded across global internet platforms with its groundbreaking AI video capabilities. Compared with previous versions, Seedance 2.0 delivers vastly improved output quality, higher usability in complex interaction and motion scenes, and significantly better physical accuracy, realism, and controllability — perfectly fitting industrial-grade creative workflows. One user commented: “The lighting, camerawork, sound design… Sora looks like a relic from a bygone era next to this.” The global tech tsunami triggered by Seedance 2.0 even stunned Elon Musk, the world’s richest person. He reposted and commented on Seedance 2.0 on social media, marveling at the model’s rapid evolution. Acclaimed director **Jia Zhangke** was equally impressed: “Seedance 2.0 is truly amazing — I plan to make a short film with it.” Ironically, at a “Youth Green Plan” film masterclass less than a year earlier, Jia had said AI could not fully replace traditional filmmaking. Several US directors also raved about Seedance 2.0 after using it for short films. Director Charles Curran wrote that he used Seedance 2.0 to make a trailer for an unreleased live-action game film in just 20 minutes, costing only $60. Director Andrew Oleck released a 30-second short of a boy in a wheelchair sliding into water in a forest — tight, thrilling, and cinematic. “It’s unbelievable,” he said. “One prompt, and Seedance 2.0 does it all.” “The childhood era of AIGC is over,” gushed Feng Ji, producer of *Black Myth: Wukong*. “I’m grateful — at least today’s Seedance 2.0 comes from China.” He described it as **leading, all-around, and low-threshold**. But Feng also warned: - Production costs for ordinary video will no longer follow traditional industry logic, moving toward the marginal cost of computing power. - Content will face unprecedented inflation, and traditional organizational and production structures will be completely restructured. His biggest fear: **rampant fake videos and a crisis of trust**. “That’s why I posted this — realistic fake videos will have zero barriers to entry, and existing IP and censorship systems will face unprecedented shock.” Hollywood pushed back. The **Motion Picture Association (MPA)** — representing Netflix, Paramount Pictures, Amazon MGM Studios, Sony Pictures, Universal Pictures, Walt Disney Pictures, Warner Bros. Discovery, and other major studios — issued a formal protest. The MPA claimed Seedance 2.0 saw mass unauthorized use of US copyrighted works within a day of launch, demanding ByteDance immediately suspend the service. MPA Chairman Charles Rivkin said ByteDance ignored copyright laws protecting creators, threatening tens of thousands of US film and TV jobs. ByteDance acted quickly: it suspended users’ ability to upload real-person images and pledged to respect IP and copyright — but the move did not fully ease Hollywood’s concerns. According to an Axios-obtained letter, **Walt Disney Company** sent a cease-and-desist notice to ByteDance on Friday (local time), accusing it of using Disney works without permission to train and develop Seedance 2.0. ## 1,000-Person In-House Chip Team, Four Product Lines Launched The success of ByteDance’s AI products — from Seedance 2.0 to Doubao — relies on powerful computing hardware. As reported by *Securities Times*, ByteDance previously bought most computing power from third parties, but has recently decided to build and operate its own AI data centers, controlling everything from infrastructure to hardware procurement. The company has spent heavily: - 2024 AI capital expenditure hit **80 billion yuan** — nearly matching the combined 100 billion yuan spent by BAT that year. - 2025 AI spending is projected at **160 billion yuan**:  - ~90 billion yuan for AI computing chips (40 billion in China, 50 billion in Southeast Asia);  - 70 billion yuan for IDC infrastructure and network gear (50 billion in China, 20 billion abroad). According to 36Kr’s *Smart Emergence*, ByteDance plans **160 billion yuan** for AI in 2026, including **85 billion yuan** for AI processor R&D and procurement. By the end of last year, ByteDance had stockpiled: - ~120,000 NVIDIA GPUs (A100, A800, H800); - Nearly 1 million NVIDIA H20, L20, L40 chips; - Tens of thousands of domestic AI chips: Cambricon 5/6/9, Ascend 910B/C, Hygon, Kunlunxin, etc. ByteDance also wants control of core chip technology. Its chip team has quietly grown to **over 1,000 people**, per *Smart Emergence*. Insiders say the AI chip division has over 500 staff (more than half), with ~200 in the CPU team. On February 14, *Lan Jing News* reported ByteDance’s chip team is scaling up hiring across Beijing, Shanghai, and Shenzhen, for roles including chip architecture and SoC design. The chip team also underwent a restructuring: - Former chip lead Wang Jian (CTO of Data Systems) no longer oversees the unit; - Shi Yunfeng (AI Chip & DPU) and Yu Hongbin (CPU) now report directly to Yang Zhenyuan, Vice President and head of Volcano Engine. ByteDance’s chip initiative began in 2020, with four lines: 1. **AI chips**: for Doubao large model inference; 2. **CPU**: server chips for general data center computing; 3. **VPU (Video Processing Unit)**: video decoding and content moderation; 4. **DPU (Data Processing Unit)**: network & storage optimization for data centers (e.g., RDMA). Reuters reported that after six years of secret development, ByteDance’s AI inference chip (**SeedChip**) is set to tape out by the end of March 2026. It plans to produce at least 100,000 chips this year, ramping up to 350,000, with Samsung as the likely foundry. *Smart Emergence* added that ByteDance’s AI chips have already gone through 3–4 secret iterations, with talks held with other foundries besides Samsung. Analysts say successful in-house chips will: - Cut computing costs and reduce external reliance; - Enable deep software-hardware co-optimization; - Better fit and iterate ByteDance’s algorithms; - Strengthen its technical moat. ## Global Tech Giants “Burn Cash” Fiercely for AI As ByteDance bets 400 billion yuan on AI, tech giants worldwide are also spending heavily. At last September’s Apsara Conference, Alibaba CEO Wu Yongming announced a **380 billion yuan, three-year AI infrastructure plan** with more spending to come. Overseas, Google unveiled 2026 capital expenditure of **$175–185 billion** — nearly double 2025’s figure. Reuters reported that Google, Microsoft, Amazon, Meta, and others will spend **over $630 billion** on AI in 2026 — even as return on investment lags spending growth. Tencent said in its latest quarterly report that strategic AI investment boosted ad targeting, gaming engagement, and efficiency in programming, gaming, and video production. Alibaba reported in its H1 2026 earnings that cloud revenue (excluding consolidated Alibaba revenue) rose 28% year-on-year, driven by public cloud — including AI product adoption. Yet Tencent and Alibaba face the same headwinds as Western giants: - Tencent’s H1 2025 net profit rose 16.7% year-on-year, but growth slowed by ~45 percentage points; - Alibaba’s H1 2026 net profit fell 5.92% year-on-year, with growth down nearly 16 points. For ByteDance, growth and profit opportunities extend beyond consumer apps like Doubao and Jimeng. Its cloud platform **Volcano Engine** is poised to benefit directly from the AI boom. Omdia’s *2025 Global Enterprise MaaS Market Analysis* shows that as of October 2025, Volcano Engine processed **over 30 trillion tokens daily** — ranking third globally with 15% market share, behind only OpenAI and Google Cloud. **MaaS (Model-as-a-Service)** has become the fastest-growing, highest-margin AI cloud product. By December 2025, daily Doubao model calls topped **50 trillion tokens** — up 66.7% from October and over 10x from 2024. Thanks to model architecture and software-hardware optimizations, Doubao delivers high cost-performance and strong margins. Volcano Engine President Tan Dai said Doubao MaaS margins are far superior to traditional IaaS cloud products. After its blockbuster start to 2026, can ByteDance maintain its lead in the global AI race? **Leda Finance** will continue reporting.