ByteDance Is Weighing $70 Billion in AI Spending This Year — Funded by $50B in Profit, Backed by a Qualcomm Chip Deal

What Bloomberg Actually Reported

Bloomberg reported on May 27, 2026, citing people familiar with the matter, that ByteDance is discussing capital expenditures of as much as $70 billion this year as it builds out data centers and AI infrastructure. The company would fund most of that spending from roughly $50 billion of profit earned in 2025. ByteDance has not publicly disclosed these plans, and the figures are preliminary — described as subject to adjustment at least every quarter. The exact range given was 400 billion yuan to 500 billion yuan, which at current exchange rates translates to approximately $55–70 billion.

Bloomberg also reported a forward figure: ByteDance has discussed boosting its capex to roughly $100 billion in 2027 if economic and business conditions are favorable. That would put the company firmly in the same tier as the largest US hyperscalers — Amazon is projecting approximately $200 billion in 2026 capex, Alphabet $175–185 billion, Meta $115–135 billion. ByteDance at $70 billion sits just below this group in absolute terms, though analysts note that data center construction costs in China are significantly lower, meaning ByteDance can likely build equivalent compute capacity for lower nominal spend than a US peer.

Why ByteDance Is Spending at This Scale

The primary driver is Doubao, ByteDance's AI assistant. Doubao is China's most popular AI chatbot with over 300 million monthly active users — a figure that rivals ChatGPT's global user base. The product has dominated China's AI market for most of the past year, competing directly against Alibaba's Tongyi and Baidu's Ernie. ByteDance is now preparing to charge subscription fees for Doubao, a significant shift in a market where Chinese users have historically resisted paying for online services. That monetization push requires the kind of compute infrastructure that can support inference at hundreds of millions of users at low latency.

Beyond Doubao, ByteDance's core recommendation engines — the algorithms powering TikTok, Douyin, and its short video and e-commerce products — are increasingly AI-driven. Generative AI features for content creation, video synthesis, and search across TikTok and Douyin require on-premises inference capacity at a scale that Bytedance cannot outsource. The company's international expansion ambitions, particularly challenging US AI leaders in markets outside China, add a further compute demand layer.

The chip access picture is also relevant. US export controls have restricted ByteDance's ability to purchase NVIDIA H100 and H200 GPUs at scale. The company has responded by aggressively diversifying its chip supply — purchasing domestic Chinese chips to satisfy Beijing's call to reduce dependence on foreign semiconductors, while also pursuing deals with suppliers not covered by US export restrictions. The Qualcomm deal is an example of the latter.

The Qualcomm Deal — What Was Agreed and Why It Matters

Bloomberg reported earlier in the same week that ByteDance struck a deal to purchase millions of Qualcomm custom ASIC chips for AI infrastructure. The chips are application-specific integrated circuits (ASICs) designed for AI inference workloads — not Qualcomm's Snapdragon smartphone chips. Qualcomm CEO Cristiano Amon had referenced "engagement" with several prospective data center clients during an April 2026 earnings call; the ByteDance deal is the materialization of that engagement. Qualcomm shares jumped nearly 5% on May 26 when the deal was confirmed.

For Qualcomm, the deal is strategically important. The company has been building a data center AI chip business to diversify beyond its saturating smartphone market. The ByteDance deal is one of the first major public proof points for its AI200 and AI250 ASIC products, slated for 2026 and 2027 releases respectively. Qualcomm is targeting the inference market — where AI models are run on production traffic — rather than the training market dominated by NVIDIA. At $250 per share and a $4.97% single-day gain, the market read the ByteDance deal as a significant validation of Qualcomm's data center strategy.

For ByteDance, Qualcomm ASICs offer two advantages over NVIDIA GPUs for inference workloads: energy efficiency and supply availability. NVIDIA chips remain constrained by export controls and high global demand. Qualcomm, as a US company, is not subject to the same export restrictions to China as NVIDIA, though any future policy changes could affect this. Analysts believe ByteDance will use the Qualcomm chips primarily for inference — running Doubao at scale — while using NVIDIA and domestic chips for model training.

How ByteDance Compares to the Global AI Capex Landscape

The gap between ByteDance and its Chinese peers is stark. Alibaba and Tencent are spending at $11–17 billion. ByteDance at $70 billion would be spending 4–6x more than either. Ke Yan, a Singapore-based tech analyst at DZT Research, told Bloomberg: "ByteDance, Tencent, and Alibaba are all converging on the view that AI infrastructure is now a strategic asset rather than a discretionary line item — and the spending gap between China's tech champions and US peers is narrower than headline figures suggest once adjusted for local cost structures."

What the Chip Strategy Tells Us

ByteDance has allocated a proportionally larger budget to domestic Chinese AI chips compared to prior years — a deliberate response to both geopolitical risk and Beijing's explicit directive to use more domestically produced semiconductors. Domestic Chinese AI chips from companies like Huawei (Ascend series) and Cambricon have improved substantially but still trail NVIDIA on raw training performance. ByteDance is reportedly spending around $14 billion on NVIDIA chips this year — the maximum it can source within export control constraints — and supplementing with domestic chips and now Qualcomm ASICs.

The Qualcomm deal is notable precisely because it represents a third option. Qualcomm is a US company but not an entity covered by the same export control restrictions as NVIDIA for data center AI chips. If ByteDance can build out a meaningful inference capacity on Qualcomm ASICs, it reduces its dependence on both NVIDIA (export-constrained) and domestic chips (performance-constrained), giving it more flexibility in its infrastructure planning.

Frequently Asked Questions