# SpaceX与开源AI初创Reflection签署GPU计算租赁协议

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-06-22 23:58
- AIHOT 分数：73
- AIHOT 链接：https://aihot.virxact.com/items/cmqpeward00mwsldkn70ku98a
- 原文链接：https://x.com/rohanpaul_ai/status/2069087523679428623

## AI 摘要

SpaceX完成857亿美元IPO后，与开源AI初创Reflection签署GPU计算租赁协议。Reflection自2026年7月起每月支付1.5亿美元，使用SpaceX Colossus数据中心的Nvidia GB300集群，若延续至2029年总支付约63亿美元，协议可于3个月后提前90天取消。Reflection采用B2B/B2G开放权重模式，构建前沿模型并向政府、国家实验室和企业出售定制化部署。该集群用于训练基础模型并适配科学和国家安全场景。Anthropic切断Fable和Mythos服务后，Reflection强调关键AI客户不应依赖可被撤销或受限的闭源供应商。

## 正文

CNBC： SpaceX just landed a new compute lease deal with Reflection， an open-source AI startup.

Reflection will pay $150M/month from 07-26 for access to Nvidia GB300s， If the deal runs to 2029， total payments would be around $6.3 B.

The deal can be cancelled by either company， with 90 days' notice after the first 3 months.

SpaceX is now selling its GPU capacity like a cloud provider， using Colossus as infrastructure for outsiders.

Reflection's current business model appears to be B2B/B2G frontier open-weight AI： build powerful models， then sell access， customization， and controlled deployment to governments， national labs， enterprises.

Reflection needs SpaceX's GB300 cluster because its product is the model itself， and it still has to train a frontier-class base model， adapt it for scientific and national-security customers.

Anthropic cutting off Fable and Mythos gave Reflection a sharper sales argument： critical AI customers should not build core workflows on closed models that can be revoked， changed， or restricted by a vendor they cannot inspect or control.

---
cnbc. com/2026/06/22/spacex-ai-colossus-data-center-reflection.html

### 引用推文

> Rohan Paul：SpaceX just announced its first investment-grade bond sale after its record $85.7B IPO last week. The company is issuing senior unsecured notes, which means big...