In the early hours of October 26, Musk confirmed on X that SpaceX had completed $40 billion(approximately RMB 26.76 billion) in debt financing, primarily to cover payments for GPU hardware procurement scheduled for delivery in Q1–Q2 2027. His exact words were “this funding is primarily redeemable debt,” and he emphasized, “it is not a major factor for this quarter.”
This is not just ordinary financing news. It means:Buying GPUs has shifted from “buying with profits” to “buying with borrowed money.”And the entity borrowing is not an AI company—it’s a rocket company.
How exactly will this money be spent?
Reuters reported earlier this month that SpaceX is negotiating with multiple banks and asset management firms to raise $40 billion to purchase NVIDIA AI chips. Musk’s statement is now official confirmation.
Several key details are worth noting:
- The funding is debt—not equity.“Redeemable debt” means the company retains the right to redeem it early, preserving control without dilution—but repayment pressure is very real.
- Delivery is scheduled for Q1–Q2 2027.Chips haven’t arrived yet—but the money has already been borrowed: this is a prepayment for 6 to 9 months of future compute capacity.
- The stated purpose is direct: buying GPUs.There is no vague phrasing like “for general corporate purposes.”
Why does a rocket company need to stockpile so much compute?
SpaceX and xAI both belong to Musk’s ecosystem, with extensive overlap in compute resources, data, and talent. Training Grok, optimizing Starlink network scheduling, Starship simulation, and Tesla FSD training clusters all require GPUs. Committing $40 billion to GPUs essentially redirects cash flow generated from rockets and satellites into the AI compute battlefield.
More importantly, it has strong financial footing. In Q3 FY2026 (July–September 2026), SpaceX’s revenue $14.575 billion, up 161% year-on-year; gross profit $9.097 billion, up 216.86% year-on-year, gross margin 62.42%; net income attributable to shareholders $4.404 billion.
In other words:It is not short of cash—it simply chooses not to spend it.Using debt to acquire assets with rapid depreciation and faster refresh cycles—while preserving cash on hand—is a mature company’s financial maneuver, indicating that, in Musk’s view, GPUs are no longer “equipment procurement” but “strategic reserves.”
The compute arms race has fully entered the “debt-financing phase.”
When viewed together, recent events over the past two months reveal a clear trend:
| Officially, it will enter Model Garden “in the coming weeks,” with Antigravity support following shortly thereafter | Incident | scale | Nature |
|---|---|---|---|
| 2026-10 | SpaceX debt financing for GPU purchases | $40 billion | Redeemable debt |
| 2026-10 | Anthropic–Akamai cloud agreement | Typically, you must link and authorize your Adobe account with Gemini to access tools such as Photoshop and Lightroom; free accounts offer limited functionality, while full capabilities require an Adobe subscription. | Long-term compute procurement |
| 2026-10 | OpenAI’s new funding round | $30 billion | Equity financing |
| 2026-10 | DeepSeek funding | $15 billion | Equity financing |
| 2026 | Microsoft Maia 300 in-house chip | +40% energy efficiency | In-house substitution |
两条路同时在走：有钱的直接借债囤卡，有能力的自己造芯。中间夹着的，是想买买不到、想造造不出的公司——这就是 3 亿美元 AI 服务器走私案 为什么会发生。
对被卡在中间的人来说，这意味着什么
1. 算力不会变便宜
当 SpaceX 这种量级的买家愿意用债务锁定 2027 年的产能，市面上可流通的高端 GPU 只会更紧。Amazon 推 TorchNeuron 抢 Trainium 生态、各家自研 ASIC，都是在给「买不到卡」找后路。
2. API 价格的下降会更依赖「效率」而不是「产能」
过去一年模型 API 大幅降价，主要靠推理优化（MoE、稀疏激活、量化）而不是硬件降价。这个逻辑会继续。Nemotron 3.5 Super VL 这类 120B 开源模型能在企业私有环境跑起来，就是效率红利的体现。
3. 「算力折旧」会变成财务风险
GPU 的折旧周期通常按 3–5 年算，但技术迭代可能 18 个月就换代。用 5 年期债务买 3 年就落后的资产，赌的是「推理需求增长快过折旧速度」。这一注押对了是护城河，押错了就是资产负债表上的窟窿——OpenAI 年化收入被曝比预期少 200 亿美元，就是这类风险的预警。
Frequently Asked Questions (FAQ)
SpaceX 为什么要借钱买 GPU，而不是用利润？
两个原因。一是金额太大——400 亿美元接近其单季营收的 3 倍，用现金会抽干运营资金；二是「可赎回债务」保留提前还款权利，不稀释股权，对未上市的 SpaceX 尤其重要。
这些 GPU 是给 Grok 用的吗？
官方没有明确分配方案。但 SpaceX 与 xAI 在算力、数据上高度协同，Grok 训练、星链网络调度、星舰仿真、FSD 训练都可能共享这批硬件，具体比例未公开。
可赎回债务是什么意思？风险大吗？
指发行方在约定条件下可以提前偿还的债务。对借款人来说灵活度高，但通常利率略高于不可赎回债。风险在于：如果 AI 需求增长不及预期，这批 GPU 产生的现金流可能覆盖不了利息和折旧。
普通用户能感受到什么变化？
短期几乎无感。中期看两件事：一是 Grok 系列模型迭代速度是否提速，二是 Starlink 与 AI 服务是否出现绑定套餐。GPU 到位时间在 2027 年上半年，真正的产品变化大概率出现在 2027 年下半年。
Summarize
400 亿美元借债买卡，是这一轮 AI 基建竞赛里最有象征意义的一笔交易：算力从「成本项」变成了「资本项」。它证明 AI 投入已经大到连现金流最好的硬件公司都不愿独自承担。对普通用户，这意味着未来两年的 AI 服务竞争，胜负手不在谁的模型更聪明，而在谁先把卡插上电。
想持续跟踪各家模型与算力动态？查看我们的 AI Model Library and Tool Comparison Engineor continue reading:NVIDIA Nemotron 3.5 Super VL Review · Amazon TorchNeuron 解析 · Anthropic 116 亿美元云协议.