ByteDance Boosts AI Budget 25% to $30 Billion, Pivoting to Chinese Chips

Just a few months ago, ByteDance's 2026 AI infrastructure budget was 160 billion yuan, roughly $24 billion. That was already a serious number. Then it got revised upward by at least 25%, bringing the total to more than 200 billion yuan, approximately $30 billion.

The revision, reported by the South China Morning Post and confirmed by Bloomberg, tells you two things about where ByteDance stands in 2026. First, demand from its AI products is running well ahead of initial projections. Second, the cost of building AI infrastructure, particularly memory chips, is rising faster than anyone planned for.

Both pressures converged at the same time. ByteDance opened the wallet wider.

This is not a typical mid-year adjustment. A $6 billion upward revision in capital expenditure, made roughly five months into a fiscal year, signals that something material shifted in the business. The original $24 billion plan wasn't just undershooting on ambition. It was undershooting on demand.

Where the $30 Billion ByteDance Capex Is Going

The money is not going to a single project. ByteDance runs a portfolio of AI-dependent products, and all of them are scaling simultaneously.

Doubao, ByteDance's consumer AI chatbot, reached 227 million users by February 2026. Volcano Engine, the company's cloud division, processed over 50 trillion tokens per day by December 2025. That is up from roughly 4 trillion tokens per day a year earlier, and just 120 billion at the platform's launch in May 2024. The growth curve is steep enough that the original infrastructure budget was already under pressure by the time the year started.

The infrastructure that supports these products is not just chips. It's everything that goes around them: data center space, high-bandwidth networking, cooling systems, power infrastructure, and the software orchestration layer that ties it all together. When memory costs rise, as they have in 2025 and into 2026, the entire stack gets more expensive. ByteDance didn't just need more hardware; it needed to pay more for the hardware it was already planning to buy.

On top of Doubao and Volcano Engine, ByteDance runs AI features inside TikTok, Douyin, CapCut, and a suite of productivity tools. Each of those products generates inference workloads that need compute, memory, and network bandwidth to deliver reliably at scale. The aggregate demand from hundreds of millions of daily active users across multiple apps is what drives infrastructure planning at ByteDance's level.

ByteDance has also been pushing Doubao toward agentic capabilities. The company released Doubao 2.0 earlier this year, positioning it for the "agent era" of AI. That shift toward agents running multi-step tasks rather than answering single queries increases the compute load per session significantly. Agents that browse the web, execute code, or coordinate multiple actions are more expensive to serve than a simple chatbot response.

Volcano Engine's enterprise business is growing alongside the consumer side. By the end of 2025, over 100 enterprise customers had each accumulated more than one trillion tokens in total usage on the Volcengine platform. By March 2026, that figure had grown to roughly 140 customers. At the 2026 Beijing Auto Show, Volcano Engine revealed that Doubao is now embedded in 145 car models across more than 50 Chinese automotive brands, handling more than 30 million in-cabin interactions every single day.

When you combine Doubao's consumer growth, Volcano Engine's enterprise expansion, and the in-vehicle AI business, you get a demand picture that easily explains why $24 billion in planned capex wasn't going to be enough. The infrastructure needed to serve 227 million Doubao users, 50 trillion daily tokens on Volcano Engine, and 30 million daily in-car interactions simultaneously is substantial, and it keeps growing.

ByteDance's Domestic Chip Strategy Amid US Restrictions

The most strategically significant detail in ByteDance's 2026 spending plan is not the total number. It's where the chips are coming from.

ByteDance is allocating a proportionally larger share of its new budget to domestic Chinese semiconductors. Specifically, the company has reportedly committed $5.7 billion to Huawei Ascend processors in 2026. It's also in active talks with Samsung to manufacture a custom AI chip internally, codenamed SeedChip.

The reason for this shift is US export controls. Washington has approved Nvidia's H200 chips for export to China, but Beijing has not yet cleared Chinese companies to actually import them. That creates a bureaucratic trap: the chips are theoretically available but practically inaccessible. ByteDance, like other Chinese AI companies, can't build a 2026 infrastructure plan around hardware it may or may not be able to receive.

Huawei's Ascend chips are a genuine alternative, though not a perfect one. The Ascend 910C processors offer competitive inference performance on certain AI workloads, but they lag behind Nvidia's H100 and H800 chips on training throughput, memory bandwidth, and software ecosystem depth. The CUDA ecosystem that Nvidia has built over fifteen years doesn't transfer. Developers building on Ascend hardware have to use Huawei's CANN software stack, which is less mature and has fewer third-party integrations.

ByteDance is choosing availability and sovereign supply chain over peak benchmark performance. That's a real trade-off, but it's one the company has decided is worth making.

The SeedChip project is a longer-horizon bet. Custom silicon development takes years. Even after the design is complete, you need a foundry to manufacture it, a packaging and testing process, software to program it, and an engineering team that knows how to use it at scale. But the companies that get it right, Google with TPUs, Meta with MTIA, Amazon with Trainium and Inferentia, end up with cost and efficiency advantages that compound over years. ByteDance is making a similar bet. If SeedChip reaches production, it would give ByteDance independent control over a meaningful share of its AI compute, reducing its exposure to both US export controls and Huawei supply chain risk.

This also aligns with Beijing's stated policy priorities. China's government has pushed domestic tech companies to reduce dependence on foreign semiconductors, and ByteDance's $5.7 billion Huawei Ascend commitment directly advances that goal while also protecting the company from future export control escalations. When national policy and commercial strategy point in the same direction, the spending tends to be decisive.

How ByteDance's Spending Fits the Global AI Race

The $30 billion figure is large in absolute terms. In the context of the global AI infrastructure buildout, it's mid-table.

Meta has projected 2026 capital expenditure of $115 to $135 billion. Amazon has raised its estimates to approximately $200 billion. Microsoft and Google are each planning around $190 billion. Oracle is targeting $35 billion. According to TrendForce, the top nine cloud service providers combined are planning $830 billion in AI-related capital expenditure in 2026, with North American providers driving the majority of that figure.

ByteDance at $30 billion is spending more than Oracle and less than any of the four major US hyperscalers. But raw compute comparison is not the right frame for this story. ByteDance is not trying to out-build Amazon globally. It's trying to own AI infrastructure in China's market, which has its own dynamics. The more relevant comparison is ByteDance versus Alibaba Cloud and Tencent Cloud inside China. Alibaba committed $53 billion in AI infrastructure investment over three years in March 2026, approximately $17-18 billion per year. On that metric, ByteDance's $30 billion in a single year puts it near the top of domestic Chinese AI cloud spending.

The NVIDIA and IREN deal announced in May 2026, where NVIDIA committed up to $2.1 billion to help build 5 gigawatts of AI data center infrastructure, illustrates how broadly the global AI infrastructure race has expanded. ByteDance's $30 billion revision is another data point in that same story.

ByteDance's $5.7 billion in Huawei Ascend commitments in a single year also has direct implications for Huawei's roadmap, strengthening the revenue base that funds next-generation chip development. The constraint created by US export controls is, paradoxically, accelerating the domestic semiconductor ecosystem ByteDance will need in 2027 and beyond.

A mid-year revision of this size also tells you something specific about the business. The original $24 billion plan was built on forecasts for Doubao usage, Volcano Engine enterprise growth, and infrastructure cost assumptions. All three variables moved in the same direction: usage grew faster than expected, enterprise demand accelerated, and memory costs went up. Raising the plan by $6 billion mid-year is a statement about return on investment. ByteDance wouldn't make this adjustment unless the AI business was generating, or was clearly on track to generate, the revenue that justifies the additional spend.

For anyone building AI products or infrastructure, the ByteDance story is a reminder that AI demand forecasting is still uncertain. The companies that win in this market won't necessarily be the ones with the most accurate initial budgets. They'll be the ones that recognize when reality has outrun the plan and move fast enough to close the gap.

For a broader view of how AI infrastructure spending, chip supply chains, and cloud capacity decisions are shaping the industry in 2026, the AI Infrastructure in 2026: Chips, Cloud, and Capacity Choices guide covers the full landscape.