Key Takeaways
- DeepSeek V4 engineered for Huawei chips at $0.30/MTok signals Chinese stack independence from NVIDIA/TSMC supply chains for first time with frontier-class models
- Project Glasswing's 40-entity consortium excludes all Chinese companies AND excludes OpenAI (Western competitor), signaling trust-based selection and defensive posture, not universal coordination
- White House AI framework makes zero mention of export controls or chip restrictions — explicit break from Biden-era policy and active reversal of geopolitical supply-chain strategy
- DeepSeek Apache 2.0 license means Chinese stack is distribution-independent — weights deployable globally without US commercial channel involvement, unlike US models tied to US cloud platforms
- By 2028, Western stack carries safety frameworks and regulatory compliance; Chinese stack carries radical price advantage and open-weight flexibility — enterprise procurement becomes geopolitical hedging
Bifurcation Is Complete, Not In Progress
The simultaneous announcements of DeepSeek V4 on Huawei and Project Glasswing's restricted consortium mark the moment the global AI stack became structurally bifurcated. This is not 'decoupling in progress.' It is decoupling completed.
Reuters and The Information (April 3, 2026) confirmed that DeepSeek V4 — a 1 trillion parameter mixture-of-experts model — is built for and deployed on Huawei's latest chips. This is the first frontier-class model (competing with Claude Opus 4.6 and Gemini 3.1 Pro) that is explicitly engineered to run outside NVIDIA/TSMC supply chains.
The prior constraint on Chinese AI development was chip availability. NVIDIA's H100 and H200 were restricted from export to China. Chinese labs faced bandwidth limits (acquiring GPUs through grey markets, using fewer chips, accepting longer training timelines). DeepSeek V4 on Huawei chips proves that the innovation pathway is no longer constrained by US hardware monopoly.
This is not a capability claim — Huawei's chips may have lower throughput than NVIDIA equivalents, and Huawei's production capacity remains below market demand. The structural change is geopolitical: for the first time, a Chinese lab can develop and deploy frontier-class models using indigenous semiconductor infrastructure.
Project Glasswing: Defensive, Not Universalist
Project Glasswing is a 40-entity consortium restricted to Apple, Amazon, Google, Microsoft, NVIDIA — and crucially, excludes all Chinese companies AND excludes OpenAI (a Western competitor). The exclusion of OpenAI is underexplored; it signals the coalition is not a universal industry standard but a defensive alliance around Anthropic and Google's safety framework.
If Anthropic and Google believed Project Glasswing would become an industry norm for responsible frontier model access, they would have included OpenAI. Instead, they built it defensively: 'We trust these 40 entities with dangerous capability because they share our safety posture, and we are withholding it from everyone else, including our commercial competitors.'
This is not a global governance framework. It is a proprietary safety consortium that fragments along competitive and geopolitical lines. OpenAI is building GPT-5.5/6 without an equivalent consortium framework. Google includes itself (obviously) but excludes OpenAI. Anthropic reclassifies what qualifies for ASL-4 as commercial pressure mounts.
By Q2 2027, the consortium either expands to become a broader industry standard (unlikely, given OpenAI's exclusion) or it remains a proprietary tool for Glasswing members while the open-weight market (DeepSeek, Mistral, open-source models) captures the majority of accessible capability.
White House Framework's Silence on Export Controls Is Intentional
The Trump Administration's White House National Policy Framework on AI (March 2026) makes zero mention of export controls, chip restrictions, or geopolitical semiconductor strategy. This is an explicit break from Biden-era AI supply-chain policy and active reversal of the 2024 executive orders restricting advanced semiconductor exports to China.
The silence is the policy. By not reaffirming export controls and chip restrictions, the administration signals it is treating the Chinese stack as a separate market to compete against rather than restrict. This allows NVIDIA to continue China market access where permitted, allows Huawei to continue developing chips without US legislative obstruction, and implicitly accepts that Chinese AI development will proceed independently.
The alternative interpretation — that export controls will be enforced through other mechanisms (Commerce Department, CFIUS review process) — is undermined by the framework's tone. A document that recommends federal preemption of state AI safety laws and no new federal regulatory bodies is not signaling willingness to enforce aggressive chip export restrictions. It is signaling innovation-friendly deregulation.
Apache 2.0 License: Distribution Independence
DeepSeek's strategy of releasing under Apache 2.0 license (same as prior R1 release) is asymmetrically advantageous to Chinese labs. Claude Opus 4.6, GPT-5, and Gemini 3.1 Pro are tied to US cloud platforms (Anthropic's API, OpenAI API, Google Vertex AI). Enterprises deploy them through US commercial channels subject to US legal jurisdiction.
DeepSeek V4 weights can be downloaded and deployed globally without any US channel involvement. This is not legally ambiguous — it is explicitly within the Apache 2.0 terms. Any enterprise with sufficient GPU capacity can run DeepSeek V4 locally without routing through US infrastructure or subject to US legal jurisdiction.
For regulated enterprises (healthcare, finance, critical infrastructure), this creates a genuine optionality that US models do not offer: develop AI systems on offshore compute infrastructure using Chinese models and avoid US regulatory exposure. This is strategically attractive to Indian, Middle Eastern, African, and Southeast Asian enterprises that face compliance pressure from multiple jurisdictions but prefer cost and technical flexibility over Western regulatory certainty.
By 2028: Two Stacks, Diverging on Regulatory Posture
The Western stack carries safety frameworks (ASL-4 equivalents, state law compliance, EU AI Act), enterprise integration support, and US capital markets access. The Chinese stack carries radical price advantage ($0.30/MTok vs $2-15/MTok), open-weight flexibility (deployable anywhere), and independence from Western export control.
Enterprise procurement becomes a geopolitical hedging exercise. Regulated sectors (healthcare, finance, defense) select Western stack for audit trails, compliance frameworks, and legal defensibility. Cost-sensitive and non-regulated workloads select Chinese stack for price and flexibility.
This mirrors the cloud market bifurcation post-2018 (AWS/Azure/GCP in West, Alibaba/Tencent Cloud in China) but moves faster because model weights are portable where cloud infrastructure was not. An enterprise can migrate from one model to another in days. Cloud infrastructure migration takes months.
Multinational enterprises face procurement decisions that are no longer purely technical ('which model performs best') but geopolitical ('which stack can our business deploy given jurisdictional footprint'). The contention for 2027-2028 is over the middle market: Indian, Middle Eastern, African, and Southeast Asian enterprises with no inherent allegiance to either stack. Whichever stack offers better price + compliance flexibility wins those markets.
Winners and Losers in the Bifurcation
Huawei and SMIC (China's semiconductor manufacturer) win: their chips gain reference-customer status for frontier-class workloads, accelerating their indigenous semiconductor ecosystem and reducing dependence on TSMC.
Middle-market enterprises in non-aligned geographies (India, UAE, Saudi Arabia, Brazil) win: they gain negotiating leverage between two competing stacks and can optimize for their specific compliance footprint.
Open-weight infrastructure providers (Together, Fireworks, Groq, Cerebras) win: DeepSeek V4 becomes the most valuable open-weight model to host at scale, and these providers are stack-neutral — they monetize the infrastructure layer regardless of model origin.
TSMC and US export control policy lose: TSMC's strategic leverage over Chinese AI development erodes as Huawei chips prove sufficient. Export control policy built on 2023-2024 chip restriction assumptions is overtaken by Chinese architectural innovation (sparse attention, efficient MoE) that reduces dependence on restricted chips.
Anthropic's international expansion loses: its safety-first positioning is less competitive in price-sensitive markets where DeepSeek offers 90% of capability at 2% of price. State-affiliated AI programs (UK AISI, EU AI Office) lose influence — their governance authority is limited to Western-stack development; they have no leverage over Chinese-stack deployment or development.