China's Sovereign AI Stack is Real
DeepSeek's V4 model adaptation for Huawei's Ascend chips creates the first viable, state-sanctioned, full-stack alternative to NVIDIA's dominance.
The News
On April 24, 2026, DeepSeek announced its V4 model was adapted to run on Huawei's Ascend 950 series AI chips. The collaboration signals a major step towards China's goal of AI self-reliance, creating a domestic full-stack alternative (chip, software, model) in direct response to US export controls on advanced silicon. While V4 was still trained on NVIDIA GPUs, its inference capabilities have been ported to the Huawei stack, with scaled deployment expected in H2 2026.
Layer Scoring
Sublayer Impact Map
Which of the 50 sublayers this move actually touches, the magnitude of impact, and who plays that slice today.
Intelligence Cube · 2D
The move's footprint across the three Cube axes, Functions, Verticals, Layers, flattened into two readable 2D projections.
Layers × Verticals
10 cells · 5×2
Layers × Functions
10 cells · 5×2
Two 2D projections of the Intelligence Cube (Functions × Verticals × Layers). Filled cells = this move occupies that intersection.
Why Now
This move is a direct product of two forces reaching critical mass. First, ever-tightening US export controls made a domestic, sanction-proof AI stack an existential necessity for China, shifting the calculus from "best performance" to "secure supply." Second, Huawei's Ascend 9-series chips and CANN software stack finally achieved a minimum viable maturity. DeepSeek faced a choice: wait for hypothetical sanctions relief or partner with the only domestic player capable of scalable production. Geopolitical reality forced their hand at the exact moment domestic hardware became "good enough."
The Structural Take
This partnership is a textbook case of building a deep, compounding stack in response to scarcity. The scarcest layer for China's AI ambitions is L0/L1 compute, walled off by US sanctions. This is where value now accrues. Huawei is not just building a chip; it is building a kingdom. By providing the Ascend chip (L0), the CANN software (L1), and the SuperNode cluster (L1), Huawei is creating a deep stack that mimics NVIDIA's CUDA moat. DeepSeek adds the model (L2) and inference optimizations (L3), making the stack's value compound. This validates the second law: thin wrappers get crushed, deep stacks compound. Other Chinese model providers cannot simply port to Ascend; they must deeply integrate. Finally, the move rewrites Law 3. Distribution doesn't just beat intelligence; here, a new distribution channel is being *created* by fiat. By controlling the only sovereign hardware stack, Huawei and its chosen partner, DeepSeek, become the default distribution for AI to all Chinese enterprises, regardless of model performance relative to global leaders. The moat is geopolitical, not just technical.
Second-Order Effects
The most immediate effect is a forced consolidation of the Chinese AI landscape. Other leading model labs like Zhipu AI and Moonshot are now on the back foot, facing immense pressure to adapt their models to Huawei's Ascend/CANN stack or risk being locked out of the only scalable domestic infrastructure. This will trigger a talent drain from the global CUDA ecosystem towards a new center of gravity around CANN development within China. For Chinese cloud providers like Alibaba and Tencent, this move marginalizes their existing NVIDIA-based offerings, forcing them into a dependent partnership with Huawei. Globally, this solidifies the great "forking" of the AI stack—a CUDA-based international standard and a CANN-based Chinese standard, creating immense friction for any multinational company operating across both.
- Who Wins
- Huawei. Becomes the NVIDIA of China, controlling the key chokepoint (L0/L1) for the nation's entire AI industry. It’s a state-sanctioned monopoly.
- DeepSeek. Gets first-mover advantage and preferred access to the only scalable, sovereign AI hardware, securing its distribution within China.
- Chinese state AI initiatives. Achieve a monumental win for technological self-sufficiency, creating a resilient domestic AI industry shielded from foreign sanctions.
- Who's Exposed
- NVIDIA. Loses the entire addressable market of China for high-end AI accelerators long-term. The sanctions ultimately catalyzed the creation of their replacement.
- Other Chinese AI model makers (Zhipu, Moonshot). They are now forced to follow Huawei's lead and adapt to the CANN stack, ceding platform control and first-mover advantage.
- Alibaba/Tencent Cloud. Their massive investments in NVIDIA-based infrastructure are now at risk of becoming a legacy liability in the face of a state-backed, vertically integrated competitor.
Deep Product Lens
The core product insight is prioritizing sufficiency and ecosystem control over raw performance parity with NVIDIA. The choice to focus on inference adaptation using low-precision formats (MXFP4) and custom kernels (MegaMoE) is a pragmatic admission of the hardware's current state. They are designing *for* the Ascend platform, not trying to shoehorn a CUDA-centric architecture onto it. The "product" that shipped is not the V4 model alone, but the *stack integration*. This is the wedge. The roadmap writes itself: 1) Use state-owned enterprises as design partners to harden the CANN software stack. 2) In parallel, begin training DeepSeek V5 natively on Ascend clusters, severing the final tie to NVIDIA. 3) Package the full stack—Ascend 950s + SuperNode + CANN + pre-optimized DeepSeek models—as a "Sovereign AI Cloud" appliance for private and government cloud deployment. The lock-in is CANN; once an enterprise a builds its AI practice on this CUDA-alternative, the cost of switching back becomes prohibitive.
Deep Strategy Lens
This is a masterful execution of counter-positioning, enabled by geopolitical force. Huawei has built a business that NVIDIA is legally barred from competing with. It claims the scarcest resource in the Chinese market: sanction-proof, high-performance compute. The beauty of the strategy is that the competitive barrier isn
The Horizontal Lens
While the stack is horizontal, the go-to-market is fundamentally driven by the government vertical and state-owned enterprises (SOEs). Consider the buyer journey for a top Chinese bank like ICBC. The CIO is operating under a national mandate to achieve "secure and controllable" technology. Previously, their only option was buying NVIDIA hardware via a domestic cloud vendor, a clear point of strategic risk. Now, Huawei’s sales team can approach them not with a technical spec sheet, but with a strategic partnership for national technological security. The sale isn't about teraflops; it's about sovereignty. This purchase won't come from the standard IT infrastructure budget that AWS or Alibaba would compete for. Instead, it taps into a larger, strategic "digital transformation" or "national security" budget, reframing the purchase from a cost center to a state-aligned investment. Huawei isn't just selling a chip; it's selling patriotic resilience, making a standard TCO/performance analysis irrelevant.
- Steelman: The Counter-Thesis
The strongest counter-thesis is that the performance and developer experience of the Ascend/CANN stack will be too poor to gain real traction. CUDA has a 15-year head start, and the global developer community is built around it. Chinese developers, accustomed to a mature and high-performing ecosystem, may resist the move, leading to slow adoption and stunted innovation. Furthermore, if other domestic chipmakers emerge with their own incompatible stacks, China's internal market could fragment, preventing Huawei from achieving the critical mass needed to create a true standard. The entire effort could collapse under its own weight, remaining a "Potemkin stack" that exists for political appearances but is shunned by real builders. I maintain my thesis, however, because the powerful state mandate will override developer friction and force consolidation around Huawei as the designated national champion.
What to Watch (Next 90 Days)
- 01Do Zhipu AI or Moonshot announce Ascend 950 series adaptation within the next 6 months?
- 02Publication of the first credible, third-party benchmarks comparing DeepSeek-on-Ascend vs. a similar model on NVIDIA H100.
- 03Announcement of a major Chinese bank or SOE deploying the full Huawei/DeepSeek stack at scale.
- 04Any signal that DeepSeek's next model (V5) is being *trained* from scratch on Ascend hardware.
What This Means for You
Product Leader
This is the layer pattern worth studying: own at least one of L1 (data), L3 (compliance), or L8 (memory) under your surface. A pure L7 alone tends to compress over time.
Investor
Durable layer ownership supports premium multiples. Underwrite the moat layer, not the ARR.
Operator
This is a reasonable stack to standardize on, switching cost is the feature, not the bug. Data and memory built here compounds for you.
Candidate Law
"Geopolitical constraints create market-defining moats faster than technical superiority."
Sources
Written by Supply Chain of Intelligence™ analysis engine, reviewed weekly. By Anand Arivukkarasu · Ex-Meta Product Leader.
Share kit
Take this to LinkedIn
Three artifacts, one argument. The image carries the diagram, the short post stops the scroll, and the detailed article copies as rich text, so headings, bold lead-ins, italic standfirsts, pull-quotes and bulleted lists land in LinkedIn's Pulse editor already styled. No markdown markers, no tables, nothing to reformat by hand.
Supply Chain of Intelligence™ · Battle Card
Apr 24, 2026
China's Sovereign AI Stack is Real
Territory taken: L0 Infra · L1 Data · L2 Models — Huawei provides the Ascend chips, creating a domestic, sanction-proof silicon layer.
- Huawei — Becomes the NVIDIA of China, controlling the key chokepoint (L0/L…
- DeepSeek — Gets first-mover advantage and preferred access to the only scala…
- NVIDIA — Loses the entire addressable market of China for high-end AI acce…
- Other Chinese AI model makers (Zhipu, Moonshot) — They are now forced to follow Huawei's lead and adapt to the CANN…
Expected counter-moveThe strongest counter-thesis is that the performance and developer experience of the Ascend/CANN stack will be too poor to gain rea…
Anand Arivukkarasu
supplychainofai.com
↑ hover the card and hit PNG to download
The global AI stack has forked. DeepSeek's adaptation of its V4 model for Huawei's Ascend chips isn't just another partnership. It marks the birth of China's first viable, full-stack, sovereign AI infrastructure. L0 Silicon: Huawei Ascend 950 L1 Infrastructure: Huawei CANN & SuperNode L2 Model: DeepSeek V4 L3 Inference: Custom-built kernels This isn't about beating NVIDIA on performance—it's about building a fortress in a market NVIDIA is forbidden from entering. By creating a deep, integrated stack based on the scarcest resource (sanction-proof compute), Huawei establishes itself as the gatekeeper for AI within China. The second-order effect is a forced consolidation. Other Chinese model makers must now adapt to Huawei's stack or risk obsolescence. This move proves a new rule: geopolitical constraints can create market-defining moats faster than technical superiority can. The "best" tech doesn't always win when politics redefines the playing field. Will this sovereign stack accelerate China's AI progress or will the friction of a CUDA-alternative ecosystem slow it down? #AI #Strategy #China Full breakdown, with the layer map: https://supplychainofai.com/live/chinas-sovereign-ai-stack-deepseek-huawei #AI #Strategy #SupplyChainOfIntelligence #ProductStrategy #VentureCapital
Get the next teardown in your inbox.
One issue when something structurally important happens, usually weekly. No spam, no filler, unsubscribe anytime.
Worth sharing? Pull-quote: "DeepSeek's V4 model adaptation for Huawei's Ascend chips creates the first viable, state-sanctioned, full-stack alternative to NVIDIA's dominance."