ByteDance prepares $70 billion in AI capex in 2026: China's assault on infrastructure
🔎 Why $70 billion changes everything
On May 27, 2026, Bloomberg revealed a figure that sent shockwaves through Silicon Valley: ByteDance is considering up to $70 billion in AI investment spending for 2026. This is not a hallway rumor. It is an internal projection, fueled by the $50 billion in profits generated in 2025 thanks to TikTok and Douyin.
This amount nearly triples the 2025 budget of around $25 billion. And it places TikTok's parent company at the same spending level as a Microsoft or a Google, even though the latter generate cloud revenues three to four times higher.
The meaning is clear: China is no longer content with catching up in AI. It is investing to overtake, and ByteDance is its spearhead. In a context of a trade war over semiconductors, this $70 billion plan represents the most aggressive announcement by a non-American company in the field of AI infrastructure.
The essentials
- ByteDance is evaluating a 2026 AI capex of between $30 and $70 billion, according to Bloomberg and CryptoBriefing.
- The $50 billion in 2025 profits fully fund this offensive, with a valuation reaching $550 billion.
- $14 billion would be dedicated solely to Nvidia H200 chips, subject to Beijing's approval. This chip export issue is central to the US-China chip war.
- A strategic pivot is taking place towards Chinese chips: $5.6 billion in orders for Huawei Ascend 950PR, according to Tech-Insider.
- Qualcomm joins the list of suppliers following a strategic agreement signed on May 26, 2026, according to Boursorama.
Recommended tools
| Tool | Main use | Price (June 2025, check website) | Ideal for |
|---|---|---|---|
| Hostinger | Web hosting / AI deployment | From €2.99/month | SMBs looking to deploy AI apps |
| Claude Opus 4.7 (Anthropic) | Complex agentic reasoning | Anthropic Pro/Team subscription | Infrastructure data analysis |
| GPT-5.5 (OpenAI) | General-purpose and agentic LLM | ChatGPT Plus/Team subscription | Tech monitoring and synthesis |
| DeepSeek V4 Pro (Max) | High-performance open-weight LLM | Pay-as-you-go API | On-premise deployment in China |
The figures: from 25 to 70 billion in one year
ByteDance's AI budget has experienced a dizzying trajectory. In 2024, spending on AI chips already amounted to 23 billion dollars according to OneHack, with a 70% loss in operating profits absorbed by these semiconductor purchases alone.
For 2025, the capex budget was estimated at around 25 billion dollars, including 85 billion yuan (around 11.7 billion dollars) dedicated to Nvidia chips.
2026 marks a turning point. The range mentioned by CryptoBriefing places the revised plan at 30 billion dollars, but Goldman Sachs projects that Chinese companies will collectively spend more than 70 billion dollars on data centers. The upper bound of 70 billion for ByteDance alone, reported by Bloomberg, includes all infrastructure spending.
200 billion yuan: the official plan
The plan officially announced by ByteDance amounted to 200 billion yuan (around 27.5 billion dollars), representing a +25% revision compared to the initial plan, according to AI Haberleri. This plan is already being described as the largest AI investment by a non-US company. The 70 billion range likely includes conditional options and investments spanning several quarters.
Comparison with US hyperscalers
The SCMP, citing Morgan Stanley, reports that the 5 US hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle) are expected to spend 805 billion dollars combined in 2026 capex. Alphabet alone projects 190 billion dollars.
ByteDance, with a potential 70 billion, would therefore represent approximately 8.7% of the spending of the five American giants combined. For a single company, that is considerable. For a Chinese company under sanctions, it is a political signal.
| Player | Estimated 2026 AI Capex | Source |
|---|---|---|
| Alphabet | 190 B$ | SCMP |
| 5 US hyperscalers (total) | 805 B$ | SCMP |
| ByteDance (upper bound) | 70 B$ | Bloomberg |
| ByteDance (official plan) | ~30 B$ | AI Haberleri |
| Chinese companies (GS total) | >70 B$ | CryptoBriefing |
The chip war: Nvidia, Huawei and the new third-party supplier
The most strategic aspect of this plan is not the total amount, it's the distribution among chip suppliers. ByteDance is pursuing a deliberate balancing policy between Nvidia, Huawei, and now Qualcomm.
The $14 billion bet on Nvidia H200
ByteDance plans to invest $14 billion in Nvidia H200 chips in 2026. But this plan remains conditional: it depends on a green light from Beijing, as US regulations restrict the export of high-performance chips to China.
AInvest points out that the supply chain is already expanding on a speculative basis, anticipating this approval. It's a dangerous game: if Beijing blocks it, $14 billion in budget must be quickly redirected.
This partial dependence on Nvidia illustrates a paradox. The United States restricts exports to slow down Chinese AI, but these same restrictions push companies like ByteDance to order massively whenever a window opens, in turn boosting the revenues of NVIDIA, which is itself pouring $40 billion into AI in 2026.
The Huawei pivot: $5.6 billion and 750,000 950PR chips
This is where the plan becomes truly geopolitical. Tech-Insider reports that ByteDance has ordered $5.6 billion worth of Huawei Ascend 950PR chips. Huawei plans the production of 750,000 units of this chip in 2026.
The Ascend 950PR reaches 1.56 PFLOP in AI performance. Above all, Huawei has developed a software stack compatible with CUDA, which significantly facilitates migration from Nvidia infrastructures. This is a technical detail with massive implications: it reduces the transition cost and makes the pivot toward technological autonomy credible.
AI Haberleri confirms that over $30 billion are allocated to AI infrastructure with a clear pivot toward Chinese chips for technological autonomy. ByteDance is not just buying Huawei chips: it is helping to create an alternative ecosystem to Nvidia in China.
Qualcomm joins the dance
On May 26, 2026, Boursorama revealed a strategic agreement between Qualcomm and ByteDance for chips intended for AI data centers. This diversification is not insignificant. It shows that ByteDance is deliberately building a three-pillar supply chain: Nvidia for raw performance, Huawei for autonomy, Qualcomm for diversification.
Google News sums up the situation well: ByteDance is balancing its supplies between Nvidia and Huawei, with the $14 billion plan depending on the approval of the H200s.
Data centers in Asia: building a physical empire
A $70 billion capex isn't just about chips. It's also about land, concrete, fiber cabling, cooling systems, and data centers to be built from scratch.
Infrastructure at the service of TikTok and beyond
ByteDance isn't building this infrastructure solely to train models. It's a complete ecosystem: content recommendation on TikTok/Douyin, integrated AI generation in apps, cloud services for third parties, and foundational models like those powering the Chinese AI ecosystem.
The profitability of this infrastructure relies on vertical integration. Unlike a Microsoft Azure or a Google Cloud that sell infrastructure to third parties, ByteDance first consumes its own compute for its consumer products with a billion users. The rest can be monetized through cloud services.
The Chinese model of accelerated deployment
What sets the Chinese deployment apart is the speed. Authorities facilitate construction permits, energy is subsidized in certain special zones, and companies like ByteDance benefit from priority access to land. A data center that would take 18 to 24 months to build in the United States can be operational in 12 to 15 months in China.
This physical deployment speed partially compensates for the lag on the most performant chips. The more compute you have available, even on slightly less powerful chips, the more models you can train, the more training data you collect, the more your models improve. It's a virtuous cycle that ByteDance is accelerating with these investments.
The race for Chinese technological autonomy
ByteDance's $70 billion plan can only be understood through the prism of technological autonomy. China has realized, with successive U.S. sanctions, that dependence on Nvidia chips is a strategic vulnerability.
From catch-up to substitution
Until 2024, Chinese companies bought Nvidia chips because they were better. In 2026, they buy Huawei chips because they have to, but also because the gap has narrowed. The CUDA compatibility of the 950PR is a turning point: it means that a model trained on Nvidia can be deployed on Huawei with minimal porting effort.
ByteDance and all the Chinese giants are investing heavily to make this substitution scenario the norm rather than the exception. When Moonshot AI raises $2 billion and Kimi K2.6 tops the open-weight ranking, it is this same infrastructure ecosystem that makes these performances possible.
The State's role in ByteDance's strategy
Beijing does not directly finance ByteDance's capex, but it facilitates everything else: construction licenses, access to energy, import authorizations for Nvidia chips (or their refusal, pushing towards Huawei). The conditional approval of the H200 is a negotiating lever that Beijing uses to steer investments toward local chips.
This dynamic explains why ByteDance simultaneously ordered $14 billion worth of H200s and $5.6 billion worth of 950PRs. It is not hesitation, it is strategic hedging: if the H200s arrive, so much the better for pure performance. If Beijing blocks them, the Huawei infrastructure is ready to absorb the surplus budget.
The impact on the global GPU market
$70 billion in capex, even if partially allocated to chips, has a systemic effect on the semiconductor market.
Chinese demand as a key variable
AInvest describes a supply chain that is expanding on a speculative basis. Asian Nvidia chip distributors are increasing their inventories in anticipation of orders from ByteDance and other Chinese giants. If approval does not come, it is a demand shock that propagates throughout the entire chain.
Conversely, if the H200s are approved, the additional $14 billion in demand risks creating strains on the global supply, including for Nvidia's American and European customers. It is a domino effect that analysts are beginning to model.
Huawei as a credible alternative changes the game
Until recently, US sanctions worked because there was no alternative to Nvidia. Huawei's 950PR, with its 750,000 units planned for 2026 according to Tech-Insider, changes this equation. Even if the chip is 20 to 30% less performant than the H200, the volume makes up for it.
For models like Moonshot AI's Kimi K2.6, which scores 88.1 in the agentic benchmark and 84 in generalist, the proof is there: Chinese chips are sufficient to produce leading models. This reduces the pressure on Beijing to authorize the H200s and strengthens China's negotiating position.
What this means for Chinese AI models
Infrastructure is not an end in itself. It serves to train and deploy models. And the results are starting to show up in the rankings.
A new generation of local models
The June 2025 agentic LLM ranking shows a notable Chinese presence. Moonshot AI's Kimi K2.6 ranks 7th with an 88.1 in self-host mode, ahead of models like GPT-5.4 (87.6) and Claude Opus 4.6 (84.7). As a generalist, it reaches 84, tied with DeepSeek V4 Pro (High).
Z.AI's GLM-5, another Chinese model, reaches 82 in agentic and its 5.1 version ranks at 83 as a generalist. DeepSeek V4 Pro (Max) climbs to 9th place in the generalist ranking with an 88.
These performances would not be possible without massive training infrastructure. ByteDance's capex directly fuels this competitiveness, providing the compute necessary for these companies and its own research teams.
The advantage of large-scale deployment
A high-performing model in the lab is worthless without deployment. And this is where ByteDance's infrastructure creates a decisive advantage: with over a billion users on TikTok and Douyin, the company has a distribution channel that even OpenAI and Anthropic envy.
This is precisely the dynamic driving Anthropic and OpenAI to each launch their $10 billion enterprise JV: the need to deploy AI in enterprises at scale. ByteDance already has this channel, the infrastructure that goes with it, and the users to validate and improve its models in production.
The risks: bubble, profitability, and dependency
An investment of this magnitude is not without risks. Several factors could turn this offensive into a trap.
The risk of an infrastructure bubble
The SCMP, citing Macquarie, estimates that an AI infrastructure bubble is unlikely before 2027. But this reasoning applies to the United States, where cloud revenues justify the spending. For ByteDance, the equation is different: the company is burning 70% of its profits on AI chips according to OneHack.
If generative models do not generate sufficient additional revenue, ByteDance's profitability could deteriorate rapidly. TikTok generates massive profits, but margins could compress if capex continues to increase at the same pace.
Dependency on TikTok
The financing for this offensive relies on TikTok's profits. However, the app faces constant regulatory risks in the United States, Europe, and other markets. A ban or significant restriction in a key market would reduce revenue and, through a domino effect, the ability to fund AI capex.
This is ByteDance's paradox: AI is its future, but the present that finances it is under threat.
Technical execution
Migrating an infrastructure of this magnitude between Nvidia and Huawei has never been done at this scale. The 950PR's CUDA compatibility reduces friction, but the reality of production deployment always reveals problems invisible in the lab. Bugs, library incompatibilities, differences in numerical precision: so many technical challenges that can delay the return on investment by several quarters.
❌ Common mistakes
Mistake 1: Confusing total capex and chip spending
The $70 billion includes all investment expenditures: data centers, cooling, cabling, energy, and chips. Only about $20 to $25 billion would be directly allocated to semiconductors. Reducing this plan to a simple GPU purchase is an analytical error.
Mistake 2: Thinking ByteDance is alone
Goldman Sachs projects more than $70 billion in data center spending for all Chinese companies combined, according to CryptoBriefing. Alibaba, Tencent, Baidu, and others are also investing massively. ByteDance is the most aggressive, but it is part of a national movement.
Mistake 3: Underestimating Huawei's software stack
The Ascend 950PR is not just a chip. It is a software ecosystem with CUDA compatibility. Ignoring this aspect and judging the chip solely on its raw PFLOPS is making the same mistake as those who underestimated CUDA against OpenCL fifteen years ago. The software stack is often more important than the silicon.
❓ Frequently Asked Questions
Can ByteDance really spend 70 billion in a year?
The upper limit of 70 billion is an internal estimate cited by Bloomberg, not a formal budget. The official plan is for 200 billion yuan (~$30 billion) according to AI Haberleri. Reality will likely fall somewhere between the two, depending on import authorizations and the speed of data center construction.
Why would Beijing block the H200 when ByteDance wants to buy them?
US restrictions limit exports to China. Beijing could refuse approval to avoid increased dependence on Nvidia, or conversely, grant it to allow ByteDance to remain competitive. It is a geopolitical lever, not a purely commercial decision.
Are Huawei chips really a viable alternative?
In raw performance, the 950PR (1.56 PFLOP) is inferior to the H200. But CUDA compatibility and production volume (750,000 units) make substitution realistic for model training. The results of Kimi K2.6 and DeepSeek V4 Pro concretely demonstrate this.
What is the link between this capex and models like Kimi K2.6?
Moonshot AI, the publisher of Kimi K2.6, indirectly benefits from this infrastructure. The more giants like ByteDance invest in data centers and demand for Chinese chips, the more the supply ecosystem strengthens, reducing costs for all Chinese AI players.
Will this cause a global GPU shortage?
If the $14 billion in H200 orders are approved, the additional demand could create tensions on supply, including for customers outside China. But TSMC's production and Nvidia's allocations have factored in these scenarios since 2024.
✅ Conclusion
ByteDance's envisioned $70 billion AI capex is not just a simple corporate budget. It is a geopolitical act that redefines the AI infrastructure war between China and the United States. By balancing Nvidia, Huawei, and Qualcomm, ByteDance is building the most diversified supply chain in the history of Chinese AI. The question is no longer whether China can compete on infrastructure, but when the local ecosystem will become completely autonomous. Everything points to the answer arriving in 2026.