Pular para o conteúdo principal
Close
AI

Silicon Curtain: The Mirrored Wall of AI Between the US and China

Gabriel Ferraresi· CEO | Tech86July 12, 20264 min
aigeopoliticssilicon-curtainai-sovereigntydeepseekopen-sourceexport-controlnvidia

In June 2026, the US Commerce Department required a license to export Anthropic's Fable 5 and Mythos 5 to any foreigner. In July, according to Reuters, China's MOFCOM and NDRC gathered Alibaba, ByteDance and Z.ai to discuss restricting overseas access to advanced models. Qwen, Doubao, GLM-5.2. The same wall, from opposite sides. We track Silicon Geopolitics as a discipline, and the signal is clear: each side treats frontier AI as a strategic asset. Whoever is in the middle needs an answer that depends on neither.

The wall descends in layers: hardware, accusation, mirroring

The wall did not fall at once — it descended in layers. The US moved first on hardware: October 2022, semiconductor controls. Nvidia responded with modified chips at each change. On January 27, 2025, DeepSeek R1 proved that efficiency circumvents brute force. Trained with H800, cost of $5.6 million. Nvidia lost $589 billion in one day, a US market record.

The next layer was accusation. The shock changed the narrative. In February, OpenAI accused distillation in a memo to Congress. Anthropic identified DeepSeek, Moonshot and MiniMax in industrial campaigns: 24,000 accounts, 16 million interactions. On June 12, the Commerce Department transformed cloud-based model access into controlled export. The Harvard Law Review questioned the legal concept of release.

China mirrors the structure. SAMR promulgated Order 126 in February 2026, in force since June 1: algorithms, code and data as trade secrets for the first time. MOFCOM discusses restricting models. No timeline, no formal rules published. DeepSeek was not cited. The mirroring is structural — each side treats frontier AI as a strategic asset, and each side builds its own wall.

The twist: open source is the breach

The twist is that open source is the breach. Already-published models cannot be revoked. DeepSeek R1 and Qwen-2.5 are on Hugging Face, permanent. Restrictions affect only future releases. According to Hugging Face data, Chinese models are 41% of downloads on the platform. Global use of open-source LLMs reached 30% for Chinese models, versus 1.2% at the end of 2024.

This changes the reading of the Silicon Curtain. The wall controls future flow, not past stock. Whoever already downloaded Qwen-2.5 or DeepSeek R1 has the weights. Whoever depends on an API for the next release is subject to authorization. The breach is structural: once the weights are public, they are irreversible. That is why diversification of model sources became a compliance strategy, not a technical preference.

The impact for those in the middle

For those in the middle, the implication is direct. According to market estimates, Brazilian companies use Chinese models in 1-3% of AI and 10-20% in technology. The DeepSeek API costs $0.14 per million input tokens, 100x cheaper than GPT-4o. If restrictions materialize, future releases may arrive slower. Diversification became a compliance strategy.

The point is not to pick a side. The point is that both sides are building walls simultaneously, and whoever depends on a single provider — American or Chinese — is exposed. Concentrated dependency is the risk. Sovereign inference on-prem is the structural answer.

The defense: sovereign inference on-prem

The defense is sovereign inference on-prem. Open models like DeepSeek, Qwen and Llama running inside the perimeter via NVIDIA NIM, not subject to a blockable API. Infrastructure as Code with Terraform ensures multi-cloud portability. If one cloud restricts, you deploy elsewhere. The architecture is layered because the wall is layered too.

At Tech86, AI Engineering and Security delivers this architecture, and Elite DeepTech treats Silicon Geopolitics as a discipline. We do not sell a technical preference — we sell the architecture that survives both sides of the wall. The endpoint is in your perimeter, the data is in your infrastructure, and inference does not depend on anyone authorizing access.

Conclusion: whoever is in the middle chooses neither side

Whoever is in the middle of the wall chooses which side controls inference. The answer is neither. The Silicon Curtain is not a one-off event — it is a layered structure descending over hardware, model and API simultaneously, from both sides. The only structural defense is sovereign inference on-prem with open models, multi-cloud portability via Terraform, and diversification treated as compliance.

We help companies build this architecture — before the next release arrives slower, before the next API is blocked, before the wall closes on the side you did not choose.

Need expert guidance?

Schedule a consultation with our specialists.

Sovereign Inference and AI Strategy

Frequently Asked Questions

The Silicon Curtain is the mirrored wall of export control that the US and China are building simultaneously over frontier AI. In June 2026, the US Commerce Department required a license to export Anthropic's Fable 5 and Mythos 5 to any foreigner. In July, according to Reuters, China's MOFCOM and NDRC gathered Alibaba, ByteDance and Z.ai to discuss restricting overseas access to Qwen, Doubao and GLM-5.2. The same wall, from opposite sides. For AI consumers, the implication is that access to frontier models may become conditional on government authorization from either side.

Already-published models cannot be revoked. DeepSeek R1 and Qwen-2.5 are on Hugging Face, permanent. Restrictions affect only future releases. According to Hugging Face data, Chinese models are 41% of downloads on the platform, and global use of open-source LLMs reached 30% for Chinese models, versus 1.2% at the end of 2024. Open source is the breach because the weights, once published, are irreversible.

On January 27, 2025, DeepSeek R1 proved that efficiency circumvents brute force. Trained with H800, cost of $5.6 million. Nvidia lost $589 billion in one day, a US market record. The shock changed the narrative: in February, OpenAI accused distillation in a memo to Congress, and Anthropic identified DeepSeek, Moonshot and MiniMax in industrial campaigns — 24,000 accounts, 16 million interactions. On June 12, the Commerce Department transformed cloud-based model access into controlled export. The Harvard Law Review questioned the legal concept of release.

Sovereign inference is running open models inside your own perimeter, via NVIDIA NIM, instead of consuming a vendor API that can be blocked. According to market estimates, Brazilian companies use Chinese models in 1-3% of AI and 10-20% in technology. The DeepSeek API costs $0.14 per million input tokens, 100x cheaper than GPT-4o. If restrictions materialize, future releases may arrive slower. Sovereign inference on-prem removes the dependency on a blockable API.

The defense is sovereign inference on-prem with open models (DeepSeek, Qwen, Llama) running inside the perimeter via NVIDIA NIM, complemented by Infrastructure as Code with Terraform for multi-cloud portability. If one cloud restricts, you deploy elsewhere. Diversification became a compliance strategy. At Tech86, AI Engineering and Security delivers this architecture, and Elite DeepTech treats Silicon Geopolitics as a discipline. Whoever is in the middle of the wall chooses which side controls inference — the answer is neither.

Blog — Get in Touch

Have a question about our articles or services? Our team is ready to help.

Schedule a Meeting

Book a time slot.

Schedule Now

Email

Send us a message.

[email protected]

WhatsApp

Quick conversation.

Address

Avenida Paulista, 1636 - São Paulo - SP - 01310-200

Tech86 Specialist

Online now

Hello! How can we help scale your business today?

Tech86 Engineering

We Value Your Privacy

We use cookies and similar technologies to optimize your experience, analyze site traffic, and personalize content. By clicking "Accept All", you agree to the use of all cookies. Read our Privacy Policy.