ByteDance’s $14 Billion Gambit: Betting Big on NVIDIA While Building a Singapore Safety Net
Ddos 
According to sources cited by the South China Morning Post, ByteDance plans to allocate roughly RMB 100 billion (about USD 14 billion) in 2026 specifically for the purchase of NVIDIA AI chips. This figure represents a notable increase from the RMB 85 billion budgeted in 2025, underscoring that the company’s appetite for computing power shows no sign of abating.
Insiders say the USD 14 billion allocation is predicated on one key assumption: that NVIDIA will be permitted to sell its H200 GPUs to China. The H200 is among NVIDIA’s most powerful AI chips currently available, and if the U.S. government grants approval—whether in full form or as a downgraded variant—ByteDance is expected to place orders without hesitation.
This, however, is only one component of ByteDance’s vast AI spending plans for 2026. Total AI-related capital expenditure is rumored to reach as high as RMB 160 billion, encompassing not only chip purchases but also infrastructure expansion and talent acquisition. At the same time, the company is careful not to put all its eggs in one basket. While pouring billions into NVIDIA hardware, ByteDance also maintains an internal chip design team of roughly 1,000 engineers, quietly advancing a “Plan B.”
Reports indicate that ByteDance’s in-house semiconductor unit has already achieved a major milestone by successfully taping out a self-developed processor. According to sources, its performance is comparable to NVIDIA’s H20—a version tailored for the Chinese market—but at a lower cost. This suggests that, at least for entry-level and inference workloads, ByteDance has begun to establish a degree of self-sufficiency.
Beyond compute, the company is also aggressively positioning itself in the field of high-bandwidth memory (HBM) through a combination of internal R&D and investments in startups, seeking control over one of the most critical components of AI accelerators. Why does ByteDance need such enormous computing capacity? The answer becomes clear when looking at its AI chatbot, Doubao.
Data shows that Doubao’s daily token throughput has surpassed 50 trillion this month—a staggering leap from 4 trillion in December last year. Meanwhile, ByteDance’s cloud platform, Volcano Engine, has secured an exclusive AI cloud partnership with China Central Television’s Spring Festival Gala, an event that will subject its infrastructure to extreme traffic demands.
Notably, in response to intensifying U.S.–China tech tensions, ByteDance appears to be reshaping its organizational structure. The report notes that its internal chip division was transferred in September to a Singapore-based subsidiary called Picoheart. Some Chinese employees were reportedly asked whether they would be willing to relocate to Singapore, a move widely interpreted as an effort to shift sensitive technologies and operations away from geopolitical flashpoints.
In my view, ByteDance’s strategy is both pragmatic and assertively ambitious.
On one hand, as long as NVIDIA hardware remains accessible—even in downgraded form and even at a premium—ByteDance is prepared to buy. This remains the fastest route to training large-scale models, supported by the most mature ecosystem available. A USD 14 billion budget sends a clear signal: ByteDance has no intention of falling behind in the AI arms race.
On the other hand, achieving “H20-level” performance with an in-house chip may not sound as formidable as H100 or H200-class hardware, but it represents a critical safety net. Should U.S. restrictions eventually extend to products at the H20 level, ByteDance would still have a domestic solution capable of sustaining its massive inference workloads and keeping platforms like TikTok and Doubao running.
This hybrid approach—buying at the high end, building at the mid and lower tiers—combined with a gradual shift of R&D toward Singapore, suggests that ByteDance is deliberately constructing the strongest possible moat around its AI empire amid an increasingly volatile global landscape.
Donate