AI Diffusion Rule 廢止：出口管制解除下的算力部署戰略重構

前沿信號

2026年5月，美國商務部工業與安全局（BIS）正式廢止了拜登上任時頒布的 AI Diffusion Rule，並同時加強了針對海外 AI 晶片的出口管制。這項政策轉向不僅標誌著美國 AI 出口管制體系的重大結構性變化，也深刻影響了 Anthropic、OpenAI 等前沿 AI 公司的算力基礎設施部署策略——特別是那些依賴跨國算力合作的企業。

技術問題

AI Diffusion Rule 廢止後，前沿 AI 公司如何重新評估算力部署的出口管制合規風險？特別是 Anthropic 的 SpaceX 300MW 算力合作與 $50B 基礎設施投資，在新的出口管制框架下是否面臨更低的合規障礙？

背景：AI Diffusion Rule 的核心內容

AI Diffusion Rule（2025年1月15日發布）是拜登上任時對先進運算晶片和 AI 模型權重實施的出口管制框架，原定於2025年5月15日生效。該規則引入了兩個核心控制：

1. 先進運算整合電路（IC）控制：將運算力閾值從 480 AFLOPS 提高到 600 AFLOPS，並引入了新的類別分類
2. AI 模型權重控制：對超過 100B 參數的 AI 模型實施出口管制，要求出口前需獲得許可證

該規則的出臺原因是擔心 AI 技術的「擴散」（diffusion）——即先進 AI 技術通過出口管制途徑被 adversaries（對手國家）獲取。然而，該規則被批評為「扼殺美國創新」和「將盟國降级為二流地位」。

廢止後的戰略影響

1. 算力部署合規成本下降

AI Diffusion Rule 廢止後，前沿 AI 公司在部署算力基礎設施時不再需要應對該規則下的出口管制合規障礙。這意味著：

• AI 模型權重出口：超過 100B 參數的 AI 模型權重不再需要出口許可證，降低了跨境模型部署的合規成本
• 先進運算 IC 出口：600 AFLOPS 閾值下的晶片出口管制解除，使得更多運算硬體可以自由跨境部署

2. Anthropic 算力合作的戰略意義

Anthropic 在 2026 年 4 月宣布了多項重大算力合作協議：

• SpaceX 算力合作：300MW 算力容量（超過 220,000 個 NVIDIA GPU）
• Amazon 5GW 合作：截至 2026 年底新增近 1GW 容量
• Google/Broadcom 合作：2027 年開始部署
• Microsoft/NVIDIA $30B Azure 容量
• $50B 美國 AI 基礎設施投資

AI Diffusion Rule 廢止後，這些合作協議中的跨境算力部署不再受到該規則的出口管制限制，大幅降低了合規風險。特別是 SpaceX 的 Colossus 1 數據中心，原本可能因為該規則而面臨出口管制合規障礙，現在可以更加自由地部署算力。

3. 國際算力合作的新機會

BIS 廢止 AI Diffusion Rule 的同時，也發布了多項指導文件：

• 中國 AI 晶片風險警告：警告業界使用華為昇騰晶片的風險
• 美國 AI 晶片用於中國模型訓練的後果警告：強調美國 AI 晶片不應被用於訓練中國 AI 模型
• 供應鏈防範策略：要求美國公司保護供應鏈免受竊取 tactic 的侵害

這些指導文件雖然廢止了 AI Diffusion Rule，但並未完全解除對中國的出口管制，而是轉向了更精細化的風險管理策略。

可量化指標

合規成本節省

指標	廢止前	廢止後
AI 模型權重出口許可證	需要	不需要
600 AFLOPS 以下晶片出口	需要	不需要
跨境模型部署合規成本	高	顯著降低

算力部署效率提升

• SpaceX 算力部署：從合規審查週期（平均 6-12 個月）縮短至即時部署
• Amazon 5GW 容量：合規障礙解除後，可加速部署進度
• Google/Broadcom 合作：2027 年部署進度可提前

戰略後果

1. 前沿 AI 公司算力部署加速

AI Diffusion Rule 廢止後，前沿 AI 公司可以更加自由地部署跨境算力基礎設施，特別是那些依賴國際算力合作的企業。這意味著：

• Anthropic 的 SpaceX 算力合作可以快速部署，不再受到 AI Diffusion Rule 的出口管制限制
• Anthropic 的 $50B 基礎設施投資可以更靈活地選擇跨境算力部署方案
• OpenAI、Google DeepMind 等公司也可以更加自由地部署跨境算力

2. 美國 AI 競爭優勢的戰略調整

BIS 在廢止 AI Diffusion Rule 的同時，也發布了多項指導文件，強調「美國將追求一種大膽、包容的 AI 策略，與可信賴的盟友合作，同時將技術保留在對手手中」。這表明美國的 AI 出口管制策略正在從「全面封鎖」轉向「精細化管理」。

3. 中國 AI 發展的戰略影響

雖然 AI Diffusion Rule 廢止，但 BIS 同時發布了針對中國 AI 晶片的風險警告和美國 AI 晶片用於中國模型訓練的後果警告。這意味著：

• 中國 AI 公司仍然面臨運算晶片出口管制
• 美國 AI 晶片用於訓練中國 AI 模型的行為將被視為風險
• 中國 AI 發展仍然受到嚴格的出口管制限制

部署邊界

合規風險重新分佈

AI Diffusion Rule 廢止後，前沿 AI 公司的合規風險從「全面封鎖」轉向「精細化管理」。這意味著：

• AI 模型權重出口：不再需要出口許可證，但需遵守新的指導文件
• 先進運算 IC 出口：600 AFLOPS 以下晶片自由出口，但需遵守新的風險管理策略
• 中國 AI 晶片：仍面臨嚴格出口管制

國際算力合作的戰略機會

AI Diffusion Rule 廢止後，前沿 AI 公司可以更加自由地與盟友合作部署算力基礎設施，特別是那些依賴國際算力合作的企業。這意味著：

• Anthropic 的 SpaceX 算力合作可以快速部署
• Anthropic 的 Amazon 5GW 合作可以更靈活地選擇跨境算力部署方案
• 盟友之間的算力合作可以更加自由地進行

結論

AI Diffusion Rule 的廢止標誌著美國 AI 出口管制體系的重大結構性變化。這不僅降低了前沿 AI 公司的合規成本，也為國際算力合作提供了新的戰略機會。然而，BIS 同時發布的指導文件表明，美國對中國 AI 發展的出口管制仍然嚴格，只是從「全面封鎖」轉向「精細化管理」。

對於 Anthropic 而言，AI Diffusion Rule 廢止意味著其 SpaceX 算力合作、$50B 基礎設施投資等合作協議可以更加自由地部署，不再受到 AI Diffusion Rule 的出口管制限制。這將加速 Anthropic 的算力部署進度，提升其競爭力。

參考資料

1. BIS Rescinds AI Diffusion Rule
2. Claude Opus 4.7 Announcement
3. Higher usage limits for Claude and a compute deal with SpaceX

#AI Diffusion Rule Abolition: Restructuring of Computing Power Deployment Strategy under the Lifting of Export Controls

Frontier Signal

In May 2026, the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) officially abolished the AI Diffusion Rule promulgated when Biden took office, and at the same time strengthened export controls on overseas AI chips. This policy shift not only marks a major structural change in the U.S. AI export control system, but also profoundly affects the computing infrastructure deployment strategies of cutting-edge AI companies such as Anthropic and OpenAI—especially those that rely on cross-border computing cooperation.

Technical issues

**After the abolition of the AI Diffusion Rule, how can cutting-edge AI companies reassess the export control compliance risks of computing power deployment? In particular, will Anthropic’s SpaceX 300MW computing power partnership and $50B infrastructure investment face lower compliance hurdles under the new export control framework? **

Background: The core content of AI Diffusion Rule

The AI Diffusion Rule (released on January 15, 2025) is an export control framework for advanced computing chips and AI model weights implemented when Biden took office. It was originally scheduled to take effect on May 15, 2025. This rule introduces two core controls:

1. Advanced Computing Integrated Circuit (IC) Control: Increased the computing power threshold from 480 AFLOPS to 600 AFLOPS and introduced new category classifications
2. AI model weight control: Implement export controls on AI models with more than 100B parameters, requiring a license before exporting

The reason for the introduction of this rule is concerns about the “diffusion” of AI technology - that is, advanced AI technology is acquired by adversaries (rival countries) through export controls. However, the rule has been criticized as “stifling American innovation” and “relegating allies to second-rate status.”

Strategic impact after abolition

1. Reduced compliance costs for computing power deployment

With the repeal of the AI Diffusion Rule, cutting-edge AI companies no longer need to deal with the export control compliance hurdles under the rule when deploying computing infrastructure. This means:

• AI model weight export: AI model weights with more than 100B parameters no longer require an export license, reducing the compliance cost of cross-border model deployment
• Advanced Computing IC Exports: Export controls on chips under the 600 AFLOPS threshold are lifted, allowing more computing hardware to be freely deployed across borders

2. The strategic significance of Anthropic computing power cooperation

Anthropic announced a number of major computing power cooperation agreements in April 2026:

• SpaceX Computing Power Partnership: 300MW computing power capacity (over 220,000 NVIDIA GPUs)
• Amazon 5GW Partnership: Nearly 1GW of new capacity by the end of 2026
• Google/Broadcom partnership: deployment starting in 2027
• Microsoft/NVIDIA $30B Azure capacity
• $50B US AI Infrastructure Investment

After the abolition of the AI ​​Diffusion Rule, the cross-border computing power deployment in these cooperation agreements is no longer subject to the export control restrictions of the rule, significantly reducing compliance risks. In particular, SpaceX’s Colossus 1 data center, which might have faced export control compliance hurdles because of the rule, will now be able to deploy computing power more freely.

3. New opportunities for international computing power cooperation

When BIS abolished the AI Diffusion Rule, it also issued a number of guidance documents:

• China AI chip risk warning: Warning the industry about the risks of using Huawei Ascend chips
• Warning on the consequences of using U.S. AI chips for training Chinese models: Emphasizes that U.S. AI chips should not be used to train Chinese AI models
• Supply Chain Prevention Strategy: Requires U.S. companies to protect supply chains from stealing tactics

Although these guidance documents abolished the AI Diffusion Rule, they did not completely lift export controls to China, but instead shifted to a more refined risk management strategy.

Quantifiable indicators

Compliance Cost Savings

Indicators	Before repeal	After repeal
AI Model Weights Export License	Required	Not Required
Wafer export below 600 AFLOPS	Required	Not required
Cross-border model deployment compliance costs	High	Significantly reduced

Improved computing power deployment efficiency

• SpaceX computing power deployment: shortened from compliance review cycle (average 6-12 months) to immediate deployment
• Amazon 5GW Capacity: Compliance barriers lifted to accelerate deployment
• Google/Broadcom partnership: 2027 deployment schedule can be brought forward

Strategic Consequences

1. The computing power deployment of cutting-edge AI companies is accelerated

After the abolition of the AI Diffusion Rule, cutting-edge AI companies can more freely deploy cross-border computing infrastructure, especially those that rely on international computing cooperation. This means:

• Anthropic’s SpaceX computing power partnership can be deployed quickly and is no longer subject to export control restrictions under the AI Diffusion Rule
• Anthropic’s $50B infrastructure investment allows for more flexible options for cross-border computing power deployment
• Companies such as OpenAI and Google DeepMind can also deploy cross-border computing power more freely

2. Strategic adjustment of U.S. AI competitive advantage

While abolishing the AI Diffusion Rule, BIS also issued a number of guidance documents, emphasizing that “the United States will pursue a bold and inclusive AI strategy, working with trusted allies while keeping technology in the hands of adversaries.” This shows that the United States’ AI export control strategy is shifting from “comprehensive blockade” to “refined management.”

3. Strategic impact of China’s AI development

Although the AI Diffusion Rule was abolished, BIS also issued a risk warning against Chinese AI chips and a warning about the consequences of using U.S. AI chips for Chinese model training. This means:

• Chinese AI companies still face export controls on computing chips
• U.S. AI chips used to train Chinese AI models will be considered a risk
• China’s AI development remains subject to strict export controls

Deployment boundaries

Compliance risk redistribution

After the abolition of the AI Diffusion Rule, the compliance risks of cutting-edge AI companies have shifted from “comprehensive blockade” to “refined management.” This means:

• AI Model Weights Export: Export license no longer required, but subject to new guidance document
• Advanced Computing IC Export: Chips below 600 AFLOPS can be exported freely, but must comply with new risk management strategies
• China AI chips: Still facing strict export controls

Strategic opportunities for international computing power cooperation

After the AI ​​Diffusion Rule is repealed, cutting-edge AI companies will be freer to work with allies to deploy computing infrastructure, especially those that rely on international computing power cooperation. This means:

• Anthropic’s SpaceX computing partnership enables rapid deployment
• Anthropic’s Amazon 5GW cooperation allows for more flexible choices of cross-border computing power deployment solutions
• Computing power cooperation between allies can proceed more freely

Conclusion

The repeal of the AI Diffusion Rule marks a major structural change in the U.S. AI export control system. This not only reduces compliance costs for cutting-edge AI companies, but also provides new strategic opportunities for international computing power cooperation. However, the guidance document issued by BIS at the same time shows that the United States’ export controls on China’s AI development are still strict, but have only shifted from “comprehensive blockade” to “refined management.”

For Anthropic, the abolition of the AI Diffusion Rule means that its SpaceX computing power cooperation, $50B infrastructure investment and other cooperation agreements can be deployed more freely and are no longer subject to the export control restrictions of the AI Diffusion Rule. This will accelerate Anthropic’s computing power deployment and enhance its competitiveness.

References

1. BIS Rescinds AI Diffusion Rule
2. Claude Opus 4.7 Announcement
3. Higher usage limits for Claude and a compute deal with SpaceX