DeepSeek V4-Pro: China's 1. 6 Trillion-Parameter Answer to US Export Controls - FourWeekMBA
DeepSeek released V4-Pro, a 1. 6 trillion-parameter MoE model trained on Huawei Ascend 950 chips — no Nvidia. China's domestic semiconductor stack just produced a frontier AI model despite US export controls.
China trained a 1. 6 trillion-parameter frontiermodelon domestic chips. The export-control thesis just took its sharpest test yet — and the results are not what Washington expected. DeepSeek V4-Pro — By The Numbers1. 6TParameters (Mixture-of-Experts)0Nvidia chips used in training$94. 5BHBM memory supercycle forecast (assumes Nvidia dominance)What HappenedDeepSeek has released V4-Pro, a Mixture-of-Expertsmodelwith 1. 6 trillion total parameters — placing it squarely in the tier of GPT-4o and Gemini Ultra byscale. The critical detail is not the parameter count. It is the hardware stack: V4-Pro was trained entirely on Huawei Ascend 950 chips, with no Nvidia H100s, H200s, or Blackwell GPUs in the stack. This is not a research preview. It is a production-grade frontiermodelfrom China’s most capable AI lab, built on China’s most capable domestic accelerators. The US exportcontrolregime — which has blocked advanced Nvidia chips from Chinese buyers since October 2022 — did not prevent this. The timing lands against a busy week for AI geopolitics. At G7, ASML’s CEO said Europe is “quite behind” on AI sovereignty. The US government blocked Anthropic’s Fable 5 from foreign nationals on national security grounds. McKinsey pegged Europe’s AI prize at $1. 9 trillion by 2030. And memory analysts arepricinga $94. 5B HBM supercycle — a forecast that assumes Nvidia continues to dominate the training infrastructure market. DeepSeek V4-Pro is the most direct challenge yet to that assumption. The Core Question“If China can train a 1. 6 trillion-parametermodelon Huawei silicon, what exactly are export controls controlling?”The Structural ReadThe US exportcontrolstrategyoperated on a specific theory: restrict access to the best hardware, and you restrict access to frontier capability. Nvidia GPUs were the chokepoint. Cut off the H100, and China’s labs hit a ceiling. That theory has now produced a testable prediction — and V4-Pro is the data point that challenges it. Huawei’s Ascend 950 is not equivalent to an H100 chip-for-chip. Theperformance-per-chip gap is real. What DeepSeek appears to have done is compensate with architectural efficiency — the Mixture-of-Expertsdesignactivates only a fraction of total parameters for any given inference call, dramatically reducing compute per token. At 1. 6 trillion parameters, MoE architecture means the active parameter count during training runs is a fraction of the headline number. Efficiency, not raw compute, closed the gap. This matters for the broader AI supply chain thesis. The HBM supercycle — the $94. 5B memory forecast that has priced SK Hynix and Samsung into stratospheric valuations — rests on the assumption that Nvidia-architecture training clusters will dominate global AI infrastructure through 2028. An alternative stack gaining credibility in China changes that risk profile. It does not collapse the thesis overnight, but it introduces a scenario that most supercycle models do notprice. Capability Signal1. 6T Parameters on Domestic SiliconMoE architecture activates a fraction of params per forward pass. Efficiency-firstdesigncompensates for per-chipperformancegaps between Ascend 950 and H100. China’s labs are training at GPT-4 tierscalewithout US export-controlled hardware. Economics Signal$94. 5B HBM Supercycle Assumes One StackMemory analysts’ supercycle forecast assumes Nvidia GPU dominance through 2028. A proven alternative training stack — even if China-exclusive — compresses the monopoly premium and introduces downside scenarios for HBMpricingthat consensus models don’t currentlyprice. Geopolitical SignalSemiconductor Sovereignty Is Now BilateralThe US has chip sovereignty via TSMC and Nvidia. China now has a credible domestic stack: SMIC for fabrication, Huawei Ascend for accelerators, DeepSeek formodeldevelopment. Export controls accelerated this — they didn’t prevent it. Three ImplicationsIMPLICATION 1 — Huawei Ascend Is the Alternative StackBefore V4-Pro, Huawei Ascend was a credible alternative in theory. Now it is a demonstrated training substrate for a frontier-class MoEmodel. That changes how every non-US country thinks about its AI infrastructure options — particularly in Southeast Asia, the Middle East, and Latin America, where Nvidia access is restricted or expensive. IMPLICATION 2 — Export Controls Need a New Theory of ChangeThe currentcontrolarchitecture targets hardware. DeepSeek’s efficiency innovations — MoE architectures, inference-time compute, distillation techniques — aresoftwareand research strategies that reduce hardware dependency. Restricting chips while Chinese labs publish the algorithmic techniques that circumvent the restriction is a leaky embargo. Washington will need to rethink the playbook. IMPLICATION 3 — Europe’s Sovereignty Gap Just Got More UrgentASML’s CEO told G7 that Europe is “quite behind” on AI sovereignty. The subtext of V4-Pro is that the window for building an independent AI stack is closing. The US has one. China now has one. Europe’s $1. 9Topportunity(McKinsey) depends on whether it can build — or credibly access — a sovereign infrastructure layer. Dependence on either Washington or Beijing’s stack is a strategic liability at thisscale. This dynamic maps directly onto what we’ve analyzed inAI business model structures— the shift from closed, proprietary AI to open, distributed alternatives is accelerating, and the geopolitical layer is now an explicit variable in every AI infrastructure investment decision. It also connects to thecompetitive moat framework: Nvidia’s moat has always been the CUDAsoftwareecosystem, not just the silicon. If China’s labs can train frontier models on Ascend, the question is whether CUDA dependency follows — or whether DeepSeek’s stack develops its own ecosystem gravity. The key insight:Export controls were designed to maintain US AIleadershipby restricting hardware access. DeepSeek V4-Pro demonstrates that efficiency-first architecture — specifically MoE atscale— can partially decouple capability from the hardware generation gap. The policy lever worked less well than modeled. Business Engineer FrameworkThe AI Supercycle — Who Really Controls the StackDeepSeek V4-Pro is a case study in how the AI supercycle plays out when a second hardware stack enters the game. The Business Engineer AI Supercycle framework maps the fullvaluechain — from silicon tomodelto application — and shows which players hold durable leverage when the infrastructure layer bifurcates. Explore the AI Supercycle Framework →The Bottom LineThe US bet that restricting Nvidia would slow China’s AI trajectory. DeepSeek V4-Pro — 1. 6 trillion parameters, Huawei Ascend 950, no export-controlled hardware — is the clearest evidence yet that the bet has not paid out on its original timeline. China didn’t hit a ceiling; it built around the wall. The more important question now is not whether export controls failed, but what comes next: a stricter regime targeting algorithms and data, or an acceptance that AI capability is already too distributed to contain. Either answer reshapes the AI investment map for the next five years. Sources: DeepSeek official release;ASMLG7 remarks;McKinsey Global Institute— Europe AI Report 2026; analyst HBM supercycle estimates viaindustryreports. Published June 19, 2026.