DeepSeek V4 Isn’t a Chip Story. It’s a CUDA Story.

Thirteen months after DeepSeek R1 erased nearly $600 billion from Nvidia’s market cap in a single trading session, the Hangzhou-based lab is preparing to do it again. On March 2, 2026, TechNode confirmed that DeepSeek V4 will release the week of March 3 — a trillion-parameter multimodal model deliberately built to run on hardware that has nothing to do with Nvidia.

Everyone will cover DeepSeek V4 as a chip story. The deeper story is about software. By optimizing V4 for Huawei’s CANN framework instead of Nvidia’s CUDA, DeepSeek is seeding a parallel AI software stack that could permanently split the global AI ecosystem — and that’s the threat Nvidia can’t counter with export control workarounds.

What DeepSeek V4 Actually Is

V4 is DeepSeek’s first flagship multimodal model — native text, image, and video generation in a single architecture. According to leaked code and community analysis, the model uses a Mixture-of-Experts design with a reported one trillion total parameters but only approximately 32 billion active parameters per token via Top-16 expert routing. That is a 50% jump in total parameters over V3’s 671 billion while actually reducing active parameters from 37 billion to 32 billion. Bigger model, more efficient routing.

The model also features a one-million-token context window, enabled by architectural innovations covered below. DeepSeek has been testing V4 internally under the codename “sealion-lite” with strict NDAs. A brief technical explanation will accompany launch; the full engineering report is expected approximately one month later.

The release timing is not accidental. China’s Two Sessions — the annual parliamentary meetings where economic and technology priorities are set — begin on March 4. The 2026 government work report mentions “large-scale AI models” for the first time, with DeepSeek as the implicit exemplar. This is a product launch and a political statement in the same breath.

Three Papers, One Architecture

V4 is not a scaled-up V3. It incorporates three architectural innovations that DeepSeek published between December 2025 and January 2026 — all with open-source code. Each paper addresses a distinct scaling bottleneck, and together they explain how a trillion-parameter model can be built without proportionally scaling compute.

Manifold-Constrained Hyper-Connections (mHC): Published December 31, 2025, led by Zhenda Xie with 19 co-authors. The paper solves training instability in Hyper-Connections at scale by projecting the residual connection space onto a specific manifold. Translation: it lets you train larger models without them falling apart during training.

Engram Conditional Memory: Published January 12, 2026. This introduces conditional memory as an entirely new axis of sparsity complementary to MoE. It moves static knowledge lookups from expensive GPU compute to cheaper host memory with O(1) lookup time. Published results: BBH +5.0, MMLU +3.4, HumanEval +3.0, and MATH +2.4 over iso-parameter MoE baselines.

DeepSeek Sparse Attention with Lightning Indexer: First shipped in DeepSeek’s V3.2-exp model in September 2025 and refined in the official V3.2 release that December. A two-stage pipeline — Lightning Indexer identifies relevant token groups in sub-linear time, then fine-grained selection narrows to critical tokens. This reduces attention complexity from quadratic O(L²) to linear O(Lk), and it is how V4 achieves a million-token context window with roughly 50% less computational overhead.

DeepSeek V4 Targets CUDA, Not Nvidia’s Chips

Here is the part the chip narrative misses. Chips are replaceable hardware. CUDA — Nvidia’s proprietary software framework that nearly every AI model in the world is optimized for — is the actual moat. And V4 is the first frontier model designed to run on something else.

DeepSeek deliberately denied Nvidia and AMD pre-release V4 optimization access while giving Huawei a multi-week head start. This breaks standard industry practice, where chip vendors receive models early for driver and kernel tuning. V4 reportedly supports Huawei’s CANN (Compute Architecture for Neural Networks) as a first-class replacement for CUDA, with Cambricon and Hygon also receiving priority support.

The China Academy, citing Reuters reporting from February 26, framed the strategy plainly: “By prioritizing Huawei’s Ascend chips for optimization, DeepSeek is accelerating the development of a parallel software ecosystem that reduces long-term dependency on US technology.” More than 15 Chinese enterprises — including Huawei and China Mobile — have announced plans to adapt their infrastructure for DeepSeek models.

This is where scale matters. China has over 4,500 AI companies and 230 million generative AI users as of mid-2024 — a market large enough to sustain an independent software ecosystem. GLM-5, which we covered in February, already approached Claude Opus-level performance on zero Nvidia chips, and V4 validates that trend at frontier scale. If CANN reaches critical mass in this market, it creates a bifurcation that export controls cannot reverse.

The $650 Billion Question V4 Reopens

US Big Tech is projected to spend $650 billion on AI infrastructure in 2026. The R1 shock demonstrated that Chinese labs can match US model quality at a fraction of the cost — DeepSeek V3 was reportedly produced for under $6 million. V4 reopens that question at larger scale: if a trillion-parameter multimodal model can be built with architectural innovations rather than raw compute, the $650 billion capex thesis propping up US tech valuations looks shakier than it did in January.

The market is already pricing this in. Chinese semiconductor stocks rose 8% in pre-market trading on V4 rumors alone. Sam Altman’s assessment of R1 — “an impressive model, particularly around what they’re able to deliver for the price” — was the understatement of 2025. The cost-efficiency trend has only intensified since: Qwen3.5, Alibaba’s 397B open-weight model, already put pricing pressure on GPT-5.2 in February.

And a footnote on R1’s “$5.6 million” training cost, which gets cited endlessly: as researcher Gregory Allen told NPR, “The cost of developing a new medication is not just the cost of the clinical trial that worked… It’s the same with this AI model training run.” The headline figure likely understates true R&D expenditure. V4’s cost disclosure will face the same scrutiny.

What V4 Can’t Tell Us Yet

The “built on domestic hardware” narrative has a critical gap: training and inference are not the same thing. Training a trillion-parameter model requires thousands of top-tier GPUs running for weeks, while inference can run on far less powerful hardware with optimization. Reports suggest V4 may have been trained on Nvidia Blackwell chips — potentially violating US sanctions — and only optimized for inference on Huawei Ascend, which changes what “hardware independence” actually means.

By contrast, GLM-5 was described as trained entirely on domestic chips. V4’s full engineering report is not expected for approximately one month after launch — meaning the training question is unanswerable at release. Other unknowns include independent benchmark verification (all current performance claims are unverified internal numbers), API pricing, and how V4’s video generation compares to Sora or Veo.

One more piece of context: on February 23, Anthropic publicly accused DeepSeek of industrial-scale distillation of Claude models — 24,000 fraudulent accounts, 16 million exchanges. OpenAI raised similar concerns. We covered that story in detail. It raises questions about V4’s intellectual lineage that the architecture papers alone cannot answer.

The Real Threat Isn’t the Model

The question the data cannot answer yet: was V4 trained on Nvidia Blackwell chips to get there? If so, the “China built frontier AI on domestic hardware” narrative — the one that the entire Two Sessions political staging depends on — becomes significantly more complicated. Independence at inference is not the same as independence at training.

But here is the second-order insight that matters more than any benchmark: export controls designed to keep China from building frontier AI may have succeeded at exactly one thing — forcing Chinese labs to build the software ecosystem that makes those controls irrelevant long-term. Every enterprise that rewrites its stack for CANN is one that will never rewrite it back for CUDA. Hardware restrictions created the conditions for software independence. That is not an outcome anyone in Washington intended.

DeepSeek’s full V4 technical report, expected approximately one month after the March launch, will either confirm or undermine the domestic-training claim. If training ran on Blackwell, the CANN independence story shrinks to an inference optimization play. If it didn’t, Washington has a much bigger problem than chips.