1.5 万亿美元的豪赌：全球 AI 基础设施军备竞赛全景——ByteDance 一家就要烧掉 700 亿美元
The $1.5 Trillion Bet: Inside the Global AI Infrastructure Arms Race — ByteDance Alone Plans to Burn Through $70 Billion
> 📌 TL;DR
> 2026 年全球 AI 基础设施支出正以令人窒息的速度膨胀：美国四大科技巨头合计逼近 7000 亿美元，ByteDance 单家就考虑投入 700 亿美元，全球前九大云服务商总计将突破 8300 亿美元。这不是泡沫不泡沫的问题——这是一场没有人敢退出的军备竞赛。
一、ByteDance 的 700 亿美元震撼弹
5 月 27 日，Bloomberg 独家报道：ByteDance 正在讨论将 2026 年资本支出提高至 700 亿美元——这个数字相当于越南全年 GDP 的约 15%，是去年支出的两倍多。
这不是空穴来风。ByteDance 2025 年净利润约 500 亿美元，有底气烧钱。更惊人的是，如果 2026 年一切顺利，他们已经在讨论 2027 年把预算拉到 1000 亿美元。
具体花在哪里？
| 项目 | 预估金额 |
|------|----------|
| NVIDIA AI 芯片（H200 等） | ~140 亿美元（1000 亿元人民币） |
| 华为昇腾芯片 | ~56 亿美元 |
| 数据中心建设 | 剩余部分 |
值得注意的是，ByteDance、阿里巴巴和腾讯获得了中国政府批准，可以合计采购 40 万颗 NVIDIA H200 芯片。在美国芯片出口管制的大背景下，这个数字本身就是一个地缘政治信号。
二、美国四大巨头：7000 亿美元的「基建狂魔」
如果你觉得 ByteDance 的数字吓人，看看美国这边：
| 公司 | 2026 年资本支出 | 同比增长 | Capex/收入比 |
|------|----------------|----------|-------------|
| Microsoft | ~1900 亿美元 | 超过分析师预期 | ~47% |
| Alphabet/Google | ~1850 亿美元 | 大幅增长 | ~46% |
| Meta | 1250-1450 亿美元 | 上调指引 | ~54% |
| Amazon | 陡峭攀升中 | AWS 15 个季度最快增速 | — |
| 合计 | ~6500-7000 亿美元 | — | — |
Meta 的 Capex/收入比达到 54%——意味着每赚 2 块钱，就有 1 块多砸进了 AI 基础设施。这在科技史上前所未有。
为什么疯狂加码？因为谁都不敢不加。微软 CFO Amy Hood 说得很直白：即便花了 1900 亿，产能依然不够（capacity-constrained through at least 2026）。Google Cloud 收入同比暴涨 63%，投资者用 7% 的盘后涨幅投了赞成票。
三、中国 AI 军团：制裁下的「弯道烧钱」
中国科技巨头的 AI 支出同样在急剧攀升，尽管面临芯片供应限制：
| 公司 | 2026 AI 基建投入 | 策略 |
|------|-----------------|------|
| ByteDance | 300-700 亿美元（审议中） | NVIDIA + 华为双轨并行 |
| 阿里巴巴 | ~560 亿美元（3 年） | 利润暴跌，全力 All-in AI |
| 腾讯 | ~50 亿美元（2026） | 下半年加速，等待国产芯片 |
| 百度 | 跟进投入 | 聚焦自动驾驶+大模型 |
Goldman Sachs 估算，中国头部互联网公司 2026 年数据中心支出合计将超过 700 亿美元，同比增长 48%。
但中国面临的挑战是美国没有的：芯片供应受限。虽然获批采购 H200，但数量有限。这迫使中国企业双线作战——一边抢购 NVIDIA 芯片，一边大量采购华为昇腾芯片。ByteDance 今年仅华为昇腾订单就超过 56 亿美元。
中国最大芯片制造商 SMIC 的高管甚至公开警告：仓促建设的 AI 数据中心产能可能会闲置——建起来容易，用起来难。
四、全球图景：8300 亿美元的疯狂
把中美放在一起看，全球 AI 基建支出的规模令人咋舌：
- 全球前 9 大云服务商 2026 年资本支出：~8300 亿美元（TrendForce 估算）
- Goldman Sachs 预测 2026-2031 年累计：~7.6 万亿美元
- 2031 年年度 AI 资本支出：预计达到 1.6 万亿美元/年
这些数字意味着什么？做个类比：
- 8300 亿美元 ≈ 荷兰全年 GDP
- 7.6 万亿美元 ≈ 日本 + 德国 GDP 之和
- 全球 AI 年度 Capex 已超过大多数国家的主权投资计划
五、泡沫还是理性？一场「不敢下桌」的赌局
看多派的逻辑
Wells Fargo 分析师 Ohsung Kwon 的态度最直白：这是泡沫，但你应该买入。理由很简单——涌入 AI 的资本量太大了，大到你无法忽视。
支撑看多逻辑的硬数据：
- Microsoft AI 业务年化收入 370 亿美元，同比增长 123%
- Google Cloud 收入同比增长 63%，达到 200 亿美元/季度
- AWS 增速创 15 个季度新高
看空派的担忧
但数学很残酷：
- 要维持历史水平的资本回报率，这些公司需要创造超过 1 万亿美元的年利润——是 2026 年共识预期（~4500 亿美元）的两倍多
- Amazon 2026 年自由现金流可能为 -170 亿到 -280 亿美元
- Alphabet 自由现金流可能暴跌 90%，从 733 亿降到 82 亿美元
- Capex 强度已达收入的 34%——超过 1990 年代互联网泡沫峰值（15%）的两倍
OpenAI 的 200 亿美元 ARR 只占 2026 年超级大厂资本支出的约 3%。Anthropic 的 90 亿 + 所有 AI 独立公司加起来，2026 年收入预计也不到 350 亿美元。
关键区别：这次有底气
不过，和 2000 年互联网泡沫有一个本质区别：这轮投资来自历史上最强的企业资产负债表，而不是债务融资。这些公司有真实的营收引擎、真实的客户需求。
> ⚠️ 值得警惕的信号
> 如果 AI 采用速度不及预期，或者效率提升减少了每个工作负载所需的算力，回报可能会令人失望。基础设施建设和收入实现之间的 18-36 个月时间差，是最大的风险窗口。
六、这场军备竞赛的赢家和输家
确定的赢家：
- NVIDIA：不管谁建数据中心，都得买它的芯片。仅 ByteDance 一家今年就要花 140 亿美元买 NVIDIA 芯片
- 华为：中国企业的「Plan B」，昇腾芯片订单暴涨
- 电力基础设施：数据中心吃电量惊人，电网升级和可再生能源投资跟着受益
- 半导体设备商：台积电 2nm 产线、ASML 光刻机——需求确定性极高
面临风险的：
- 过度投入的科技公司：Meta 54% 的 Capex/收入比率是在走钢丝
- 中小 AI 创业公司：当巨头的基础设施价格下降，独立 AI 公司的差异化空间被压缩
- 依赖廉价算力的企业：短期内算力仍然供不应求，价格不会便宜
七、我的判断
这场军备竞赛不会在 2026 年结束。事实上，它可能才刚刚开始。
Goldman Sachs 预测到 2031 年 AI 年度 Capex 将达到 1.6 万亿美元——意味着未来 5 年的投入总量将是今天的翻倍。每家公司都在做同一个计算：错过 AI 基础设施窗口的代价，远大于过度投资的代价。
ByteDance 的 700 亿不是疯狂——它是在用资本换时间。在 NVIDIA 芯片供应受限的窗口期，谁先建成算力堡垒，谁就在下一轮 AI 竞争中占据不可逆转的优势。
但风险也是真实的。当 Capex 强度超过互联网泡沫峰值的两倍时，容错空间已经很小了。一旦 AI 的商业化速度跟不上基建速度，2027-2028 年可能出现「产能过剩+估值修正」的双杀。
> ✨ 底线判断
> 这不是一个「AI 是不是真的」的问题——技术是真的。这是一个「资本周期是否过热」的问题。答案很可能是：短期过热，长期不够。就像 2000 年铺下的光纤，泡沫破了，但基础设施留下来了，最终成就了移动互联网时代。AI 基础设施的故事，大概率也会这样收场。
---
数据来源：Bloomberg（2026-05-27）、Goldman Sachs、TrendForce、Tom's Hardware、Fortune、CNBC、Wells Fargo Research。文中财务数据截至 2026 年 5 月。
> 📌 TL;DR
> Global AI infrastructure spending in 2026 is expanding at a breathtaking pace: US hyperscalers are approaching $700 billion combined, ByteDance alone is considering $70 billion, and the world's top nine cloud providers will exceed $830 billion in total. This isn't a question of bubble or no bubble — it's an arms race nobody dares to quit.
1. ByteDance's $70 Billion Bombshell
On May 27, Bloomberg broke the news: ByteDance is discussing raising its 2026 capital expenditure to as much as $70 billion — roughly 15% of Vietnam's entire annual GDP and more than double last year's spending.
This isn't idle talk. ByteDance earned approximately $50 billion in net profit in 2025, giving it the war chest to back up such ambitions. Even more staggering: if 2026 goes well, they're already discussing pushing the budget to $100 billion in 2027.
Where's the money going?
| Category | Estimated Amount |
|----------|-----------------|
| NVIDIA AI chips (H200, etc.) | ~$14 billion (100B yuan) |
| Huawei Ascend chips | ~$5.6 billion |
| Data center construction | Remainder |
Notably, ByteDance, Alibaba, and Tencent received Chinese government approval to collectively purchase 400,000 NVIDIA H200 chips. Under the backdrop of US chip export controls, this number itself is a geopolitical signal.
2. US Hyperscalers: $700 Billion Worth of "Infrastructure Mania"
If ByteDance's numbers seem scary, look at the American side:
| Company | 2026 Capex | YoY Growth | Capex/Revenue Ratio |
|---------|-----------|------------|---------------------|
| Microsoft | ~$190B | Exceeded analyst estimates | ~47% |
| Alphabet/Google | ~$185B | Major increase | ~46% |
| Meta | $125-145B | Guidance raised | ~54% |
| Amazon | Steep climb | AWS fastest growth in 15 quarters | — |
| Combined | ~$650-700B | — | — |
Meta's capex-to-revenue ratio hits 54% — meaning for every $2 earned, more than $1 goes into AI infrastructure. This is unprecedented in tech history.
Why the frenzy? Because nobody dares to stop. Microsoft CFO Amy Hood was blunt: even after spending $190 billion, they remain capacity-constrained through at least 2026. Google Cloud revenue surged 63% year-over-year, and investors voted with a 7% after-hours price bump.
3. China's AI Battalion: Spending Through Sanctions
Chinese tech giants are ramping up AI spending just as aggressively, despite chip supply constraints:
| Company | 2026 AI Infra Spending | Strategy |
|---------|----------------------|----------|
| ByteDance | $30-70B (under review) | Dual-track: NVIDIA + Huawei |
| Alibaba | ~$56B (over 3 years) | Profits plunging, all-in on AI |
| Tencent | ~$5B (2026) | Accelerating in H2, awaiting domestic chips |
| Baidu | Following suit | Focus on autonomous driving + LLMs |
Goldman Sachs estimates China's top internet companies will collectively spend over $70 billion on data centers in 2026 alone, a 48% year-over-year jump.
But China faces a challenge the US doesn't: chip supply restrictions. While approved for H200 purchases, quantities are limited. This forces Chinese companies to fight on two fronts — racing to buy NVIDIA chips while simultaneously placing massive orders for Huawei Ascend chips. ByteDance's Huawei Ascend orders alone exceed $5.6 billion this year.
SMIC, China's largest chipmaker, has even publicly warned: hastily built AI data center capacity may sit idle — easy to build, hard to utilize.
4. The Global Picture: $830 Billion of Frenzy
Zooming out to the combined US-China picture, the scale of global AI infrastructure spending is staggering:
- Top 9 global cloud providers' 2026 capex: ~$830 billion (TrendForce estimate)
- Goldman Sachs projection for 2026-2031 cumulative: ~$7.6 trillion
- 2031 annual AI capex: Expected to reach $1.6 trillion/year
To put these numbers in perspective:
- $830 billion ≈ The entire GDP of the Netherlands
- $7.6 trillion ≈ Japan's + Germany's GDP combined
- Global AI annual capex already exceeds most nations' sovereign investment programs
5. Bubble or Rational? A Game Nobody Dares to Leave
The Bull Case
Wells Fargo analyst Ohsung Kwon put it most bluntly: This is a bubble, but you should buy it. The reasoning is simple — the amount of capital flowing into AI is too massive to ignore.
Hard data supporting the bull case:
- Microsoft AI business annualized revenue: $37 billion, up 123% YoY
- Google Cloud revenue: up 63% YoY, hitting $20 billion/quarter
- AWS growth rate: fastest in 15 quarters
The Bear Case
But the math is brutal:
- To maintain historical return on capital, these companies would need to generate over $1 trillion in annual profit — more than double the 2026 consensus estimate (~$450 billion)
- Amazon's 2026 free cash flow may be -$17 billion to -$28 billion
- Alphabet's free cash flow may plunge 90%, from $73.3 billion to $8.2 billion
- Capex intensity has reached 34% of revenue — more than double the 1990s dot-com bubble peak (15%)
OpenAI's $20 billion ARR represents only about 3% of 2026 hyperscaler capex. Anthropic's $9 billion plus all standalone AI companies combined are projected to generate less than $35 billion in 2026 revenue.
The Key Difference: This Time They Can Afford It
However, there's one fundamental difference from the 2000 dot-com bubble: this round of investment comes from the strongest corporate balance sheets in history, not debt financing. These companies have real revenue engines and real customer demand.
> ⚠️ Warning Signs Worth Watching
> If AI adoption progresses more slowly than anticipated, or if efficiency gains reduce the compute required per workload, returns could disappoint. The 18-36 month gap between infrastructure buildout and revenue realization is the biggest risk window.
6. Winners and Losers in This Arms Race
Clear Winners:
- NVIDIA: No matter who builds data centers, they all need NVIDIA chips. ByteDance alone plans to spend $14 billion on NVIDIA chips this year
- Huawei: China's "Plan B," with Ascend chip orders surging
- Power infrastructure: Data centers consume enormous electricity, driving grid upgrades and renewable energy investment
- Semiconductor equipment makers: TSMC 2nm production lines, ASML lithography machines — demand certainty is extremely high
At Risk:
- Over-investing tech companies: Meta's 54% capex-to-revenue ratio is walking a tightrope
- Smaller AI startups: As hyperscaler infrastructure costs come down, differentiation space for independent AI companies gets compressed
- Businesses relying on cheap compute: In the short term, compute remains supply-constrained and won't be cheap
7. My Take
This arms race won't end in 2026. In fact, it may have just begun.
Goldman Sachs projects annual AI capex reaching $1.6 trillion by 2031 — meaning total investment over the next five years will double today's levels. Every company is making the same calculation: the cost of missing the AI infrastructure window far exceeds the cost of overinvesting.
ByteDance's $70 billion isn't crazy — it's trading capital for time. During the window when NVIDIA chip supply is constrained, whoever builds their compute fortress first gains an irreversible advantage in the next round of AI competition.
But the risks are real. When capex intensity exceeds twice the dot-com bubble peak, the margin for error is already razor-thin. If AI commercialization can't keep pace with infrastructure buildout, 2027-2028 could see the double hit of "overcapacity plus valuation correction."
> ✨ Bottom Line
> This isn't a question of "is AI real" — the technology is real. This is a question of "has the capital cycle overheated." The answer is likely: overheated in the short term, not enough in the long term. Just like the fiber optic cables laid in 2000 — the bubble burst, but the infrastructure remained, ultimately enabling the mobile internet era. The AI infrastructure story will most likely end the same way.
---
Data sources: Bloomberg (2026-05-27), Goldman Sachs, TrendForce, Tom's Hardware, Fortune, CNBC, Wells Fargo Research. Financial data as of May 2026.