ByteDance Is Burning $5.5 Billion a Day on AI—and It Just Cut Its Own Profit Margin to Prove It

ByteDance’s profit-for-future bet: AI spending and a national compute push

In the first half of 2026, ByteDance reported revenue of about $120 billion, up 30% year over year. Net profit fell by a single-digit percentage to roughly $20 billion. Net margin slid to 16.7%, down from 26% three years earlier and 21% in 2024.

On paper, that looks like a warning. In context, it looks like a deliberate trade. ByteDance has not responded to the reports. The numbers point in one direction: the company is converting current earnings into future capacity. It is buying chips and building data centers. It is financing a global AI race with debt rather than equity. The margin decline reflects that choice. The core business remains intact, and management is betting that the next three to five years will be decided now.

Overseas revenue passes 30%: TikTok becomes a core business

Most commentary on ByteDance’s profit decline focuses on AI spending. That misses another shift in the same financial statement. Overseas revenue is becoming a core revenue source.

TikTok-led international business rose from 25% of ByteDance’s overall revenue in 2024 to 30% in 2025. In the first half of 2026, it continued to climb. TikTok Shop grew fastest. Its global GMV reached $50.3 billion in H1 2026, up 92% year over year. Full-year GMV could pass $100 billion for the first time. The United States overtook Indonesia as TikTok Shop’s largest single market, with $11.8 billion in half-year GMV, up 103%.

The more important change is consumer behavior. In the U.S. market, TikTok Shop’s mall contributed 51.4% of GMV, up from 36% a year earlier. Short video’s share fell from 50% to 40.4%. Livestreaming fell from 14% to 8.2%. “See it in a video, search it in the mall, buy it” is becoming a mainstream conversion path in the United States. TikTok Shop is turning from an impulse-driven content platform into a search-and-shelf e-commerce platform. It is moving onto Amazon’s core turf.

This matters for AI in two ways. First, overseas revenue gives ByteDance more dollar cash flow to buy compute and build data centers globally. Second, it shows that Chinese internet products can still scale across markets under regulatory pressure. TikTok’s growth is not a one-off. It shows that recommendation algorithms and product design travel well.

The $5.5 billion-a-day compute bill

ByteDance’s 2026 AI infrastructure capex budget was raised from about RMB 160 billion to at least RMB 200 billion, a 25% increase. That is more than RMB 550 million a day, or about $5.5 billion a month at rough exchange rates. The annual number is large enough to buy a mid-sized listed company outright.

About RMB 85 billion is earmarked for AI chip purchases. The rest goes to data center construction, liquid cooling, and power infrastructure in places such as Inner Mongolia and Xinjiang. ByteDance is also increasing the share of domestic AI chips in its procurement mix.

Among Chinese peers, ByteDance is now the most aggressive spender. Alibaba’s quarterly capex reached RMB 67.7 billion, up 75%. Tencent’s second-quarter capex was RMB 52.8 billion, up 176%. Baidu’s quarterly capex was RMB 11.4 billion, up nearly 200%. ByteDance’s full-year RMB 200 billion plan sits at the top of that group.

Globally, however, the gap remains wide. Moody’s expects the six largest U.S. hyperscalers to spend more than $785 billion in 2026, roughly six times the top Chinese tech companies. Over the past four quarters, Chinese cloud service providers increased capex by only 30% year over year, while U.S. peers grew 76%.

Payback is the harder question. Alibaba has said its AI capex can roughly be recovered within three years at current gross margins, and within two to two and a half years as self-designed chips scale. ByteDance has not published a similar timetable. Its logic is similar: chips and data centers become fixed assets, depreciated over years rather than expensed in a single quarter. The profit hit shows up in the income statement, but it is stretched over time.

The $29.6 billion loan: why debt beats equity

ByteDance did not fund this entirely from cash flow. In September 2026, it closed a $29.6 billion syndicated loan with 28 banks, 48% larger than the $20 billion initially sought. It was the second-largest Asian dollar loan of the year, behind SoftBank’s $40 billion for OpenAI.

The loan terms matter. The interest rate was SOFR plus 68 basis points, lower than ByteDance’s 2024 offshore loan at SOFR plus 85 basis points. The term is three years, extendable to five. It is unsecured. ByteDance pledged no assets or shares. Chinese banks were the main lenders, taking more than 60% of the deal.

The choice to borrow rather than issue equity comes down to valuation. ByteDance’s private-market valuation is about 2.6 times annualized revenue. Meta trades at about 7.3 times. Issuing stock at that discount means permanently selling future earnings. Selling 5% of the company means forever giving up 5% of long-term cash flow. The dilution cost is larger than the interest cost.

For ByteDance, $29.6 billion is roughly one year of free cash flow. With the yuan strengthening and the dollar weakening, borrowing dollars and repaying with future operating yuan cash flow is attractive. Exchange-rate moves could offset a meaningful part of the interest bill. The arithmetic is simple.

America brakes, China accelerates

In September 2026, several U.S. AI leaders called for a pause. Anthropic CEO Dario Amodei published a long essay calling for a pause on frontier model capability growth. Elon Musk agreed on X. Sam Altman followed with support for slowing frontier capability. Companies that usually fear being outrun found common ground in “everyone slow down.”

There is a safety narrative, and there is a financial one. U.S. AI investment depends heavily on Big Tech capex. The five largest U.S. tech companies are expected to spend about $780 billion on AI capex in 2026. That exceeds their operating cash flow, leaving a gap of roughly $130 billion. When capex eats most of operating cash flow, “slow down” stops being only a safety issue and becomes a balance-sheet issue.

China’s situation is different. Deloitte has noted that China’s AI application scenarios are far richer than those in Europe and the United States. A Morgan Stanley survey found that 80% of Chinese respondents use AI at least once a week, compared with 54% in the United States. China’s advantage is making AI useful in ports, warehouses, electric vehicles, hospitals, logistics networks, classrooms, and factory floors.

That scenario advantage is becoming a cost advantage. DeepSeek’s V4 series runs on Huawei’s Ascend 950PR inference chip, which costs roughly one-third to one-quarter of Nvidia’s H200 while delivering about half the absolute compute. In deployments, one 8-card 910B server can replace four traditional 8-card machines for a full DeepSeek model, cutting hardware cost from RMB 6 million to RMB 1.5 million and halving compute spending. DeepSeek V4.1 Flash has pushed output pricing down to RMB 4 per million tokens in idle periods.

China is building a different AI path. In the first quarter of 2026, China’s daily token calls for large models surpassed the United States for the first time. Chinese Academy of Engineering data shows that in H1 2026, inference demand was eight times training demand, and inference accounted for 60% to 90% of total AI energy use. Demand is shifting from pulse-like training purchases to 24/7 steady-state inference. Chinese AI compute spending is producing continuous demand.

Net margin decline: a deliberate trade

ByteDance’s net margin fell from 26% to 16.7%. As a static number, it is a warning. On the AI timeline, it is something else. 2026 is the window when AI moves from model competition to deployment competition. In that window, compute infrastructure decides the competitive map for the next three to five years. ByteDance is choosing to spend now rather than protect short-term margin. That suggests management sees AI as the primary battlefield.

Alibaba’s payback estimate offers a reference: three years, then two to two and a half years with more self-designed chips. If ByteDance’s return cycle is similar, today’s profit decline is a front-loaded cost. ByteDance also has TikTok as a global deployment surface. E-commerce recommendation, content distribution, advertising, and customer service each provide a live AI testbed.

Three risks stand out. Global AI capex sustainability is under scrutiny. UBS has warned about stretched compute-chain valuations and financing sustainability. Chinese cloud capex is growing fast but remains far below U.S. levels. ByteDance’s RMB 200 billion budget still has to turn into competitive AI products and revenue.

In this race, not betting is the biggest risk. ByteDance is trading profit for capacity and debt for time. Overseas revenue pays for global compute. Whether the bet pays off will be clearer in two years. For now, it is a coherent strategy with numbers behind it. Net margin is a lagging indicator. The assets forming now will matter. So will the compute reserves and the business scenarios AI is rewriting.