Analysis suggests Huawei’s strategic focus is shifting toward dominating the AI accelerator hardware market to become the next Nvidia, rather than competing directly with OpenAI on models.
The headline “Huawei executive’s internal 10,000-word memo exposed, ICT business target is to become Nvidia” grabs attention, but it is misleading. Huawei’s Xinsheng Community published a summary of a discussion between Guo Ping, chairman of Huawei’s supervisory board, and new employees. There was no leaked confidential document, nor did any employee risk their career to smuggle one out. Instead, Huawei used the session to clarify its AI strategy: what it will pursue, what it will avoid, and how it intends to close existing technology gaps.
The route matters more than the slogan. Huawei does not aim to replicate OpenAI or copy Nvidia. It intends to sell AI infrastructure: chips, clusters, cloud platforms, toolchains, and industry entry points that enable models to run.
1. “No Business Expansion Plan” Sets a Boundary
Guo Ping stated that Huawei’s core strategy is “focus,” emphasizing deeper investment in connectivity and computing while maintaining “no business expansion plan.” While this may sound like a retreat, it functions effectively as boundary management. Huawei will not attempt to dominate every terminal market. Instead, it aims to retain control over the most valuable components: connectivity, computing, operating systems, and intelligent driving.
The automotive sector illustrates this approach. Huawei does not manufacture vehicles. Through its HarmonyOS-based ecosystem (Yinwang), it supplies intelligent driving solutions. Yinwang includes automaker shareholders, and Huawei will continue to lead its development for the foreseeable future. By avoiding the heavy capital requirements and geopolitical risks of vehicle manufacturing, Huawei preserves a high-value entry point into intelligent driving. It has not abandoned the automotive industry; rather, it positions itself at the industry’s intelligent layer.
The same logic applies to 6G. Guo noted that Huawei will “make a difference” in “zero-to-one” research and development. The company will continue to execute with speed, quality, efficiency, and cost-effectiveness, remaining a key contributor to global unified standards. This is not a strategic reset, but rather a continuation of efforts focused on standards, connectivity, and foundational capabilities along Huawei’s primary communications trajectory.
Ultimately, “no business expansion plan” is a deliberate boundary statement. By declining certain markets, Huawei can concentrate resources and achieve greater depth in the sectors it chooses to serve.
2. Becoming Nvidia Means Selling the Base
Guo Ping clarified that Huawei’s objective within ICT and computing is to become “Nvidia.” However, this does not mean developing or owning a proprietary large language model. Instead, the goal is to enable customers to build robust models and ensure that any global model runs efficiently on Huawei’s Ascend, Kunpeng, SuperNode, and cluster architectures.
The real challenge lies in the ecosystem. Nvidia’s dominance extends far beyond its GPUs. CUDA, NVLink, networking hardware, servers, development toolchains, model adaptation frameworks, and the vast repository of existing customer code create a formidable moat. Huawei cannot bridge this gap simply by releasing a single chip with comparable performance. The true test begins after a customer migrates a model: Will it run? Will it run stably? Can it be modified without friction? Is reliable support available when failures occur?
Huawei aims to provide the foundational infrastructure that powers all large models, rather than competing to sell one specific model. If this positioning succeeds, Huawei’s competition will not be another model developer, but rather the entity that sets the standard for AI infrastructure. The coming battle will center on the chips, clusters, cloud platforms, and industry gateways required to deploy and scale models.
3. Large Models: No Single Faith, Calculate by Business
Guo Ping’s approach to large models is pragmatic. He did not mandate universal adoption of Pangu across the company. Instead, he evaluated the role of large models on a per-business-line basis.
The ICT and computing division does not require its own proprietary model. Its priority is ensuring that global models operate efficiently on Ascend, Kunpeng, SuperNode, and cluster systems. Conversely, the consumer terminals and intelligent driving divisions should develop their own models to maintain closed-loop data pipelines and optimize product experiences internally. Huawei Cloud, meanwhile, must offer a competitively self-developed model; otherwise, customers have little reason to prefer its platform.
This segmentation reflects a broader industry shift. A large model is no longer the singular identity of a tech conglomerate. Infrastructure providers seek environments that support diverse models. Device and autonomous driving teams require models tailored to user behavior. Cloud providers leverage models to stimulate demand for compute, storage, and related services. While many major firms tout “full-stack” capabilities and attempt to excel everywhere, Huawei has drawn a clear line: develop proprietary models only when the business case demands it. Otherwise, invest in the underlying infrastructure.
Guo also acknowledged that Huawei entered the large-model race early but failed to capture market leadership, ceding breakthroughs to competitors like Anthropic, Kimi, and DeepSeek. He noted that large enterprises possess resource advantages but suffer from organizational inertia. For new hires, this raises a critical management question: when technological iteration outpaces corporate decision cycles, how does Huawei avoid “rising early and arriving late”?
4. If Single Chips Lose, Absorb the Gap with Systems
Chip architecture represents another domain where Huawei leverages strengths while circumventing weaknesses. Guo Ping acknowledged that Huawei remains constrained by limitations in advanced semiconductor manufacturing. To compensate, the company must integrate design and fabrication more closely, using systemic advantages to offset process-node shortcomings. He referenced “Tao’s Law,” proposed by researcher He Tingbo, which advocates replacing traditional “geometric scaling” with “time scaling.” Under this framework, tighter coupling between process engineering and manufacturing yields competitive gains despite physical constraints.
Applied to the Ascend 950 series, the strategy is explicit: abandon the pursuit of single-chip parameter supremacy and instead compete on system-level performance.
Independent benchmarks highlight the disparity. A single Ascend 950DT delivers approximately 2 PFLOPS in FP4 precision. By contrast, Nvidia’s Vera Rubin achieves roughly 17.5 PFLOPS in dense FP8/FP4, alongside superior sparse inference metrics. Memory capacity differs significantly: the 950DT offers 144 GB of HBM, while Rubin provides 288 GB of HBM4. Bandwidth similarly trails, with the 950DT at 4 TB/s compared to Rubin’s 19.2 TB/s (architectural peak of 22 TB/s). In a head-to-head single-chip contest, the gap stems directly from manufacturing node limitations—a deficit that cannot be rapidly overcome through silicon alone.
Recognizing this reality, Huawei refuses to compete on that front. Its alternative strategy centers on SuperNode architectures and large-scale clusters. These systems begin at 1,024 accelerators and scale up to 8,192 cards, utilizing unified memory addressing and low-latency all-optical interconnects. The Atlas 950 unveiled at the World Artificial Intelligence Conference (WAIC) featured a 1,024-card configuration delivering 1 EFLOPS in FP8, 2 EFLOPS in FP4, 256 TB of global unified memory, and approximately 3 microseconds of round-trip time (RTT). The complete 8,192-card deployment targets 8 EFLOPS in FP8, 16 EFLOPS in FP4, spanning 160 racks across roughly 1,000 square meters.
Nvidia’s NVL72 chassis houses 72 GPUs. While the fully scaled Atlas 950 surpasses it in aggregate compute, it requires 114 times more individual chips. The analogy is military: individual units may be less powerful, but communications, formation discipline, and command structures unify thousands into a cohesive force. Huawei’s three-decade expertise in telecommunications and networking provides the foundation for this approach.
Nvidia’s defensive moat rests on NVLink paired with CUDA. However, interconnect scalability has reached a physical ceiling. Electrical interconnects degrade sharply beyond distances of approximately one meter, leaving single-rack deployments capped near 72 GPUs. Nvidia can maximize intra-rack efficiency, but it cannot scale outward indefinitely. Huawei pursues the inverse path: electrical signaling within racks, optical transmission between them. The objective is to circumvent physical interconnect limits and enable massive scaling. As models expand and techniques like Mixture-of-Experts (MoE) and extended context windows turn communication into the primary bottleneck, competitive advantage will shift from “whose single card is faster” to “who can organize ten thousand cards to work as one.”
This approach carries inherent trade-offs. Scaling a SuperNode from 1,024 to 8,192 cards introduces diminishing returns. Additional accelerators increase the volume of optical modules, connectors, memory hierarchies, and potential failure points. Communication overhead, power draw, and fault-tolerance requirements escalate accordingly. Published aggregate bandwidth figures represent system-wide totals, not per-card allocations. Reported microsecond latency reflects single-hop performance; once workloads traverse external networks, latency increases by an order of magnitude. Real-world training throughput must be adjusted for parallel efficiency and operational uptime. Claims of 82% to 85% cluster utilization and performance “several times Nvidia” originate primarily from Huawei’s internal disclosures. No independent, standardized benchmark has yet validated these figures.
Hardware specifications can be matched overnight. Transitioning the global developer ecosystem from CUDA to Huawei’s CANN software stack requires sustained effort and architectural refinement. CANN’s open-source community currently hosts approximately 3,000 monthly active developers—a figure that remains marginal compared to CUDA’s entrenched base. Guo emphasized the objective of letting “any global large model run efficiently.” Overcoming this hurdle relies heavily on migration support. The silicon challenge is secondary; developer workflows, toolchain maturity, and transition costs present steeper obstacles.
Guo advised new employees to study exchange notes from Liang Wenfeng. He highlighted a forward-looking assessment: rapid AI iteration may erode or supplant legacy software moats like CUDA. A company currently dependent on a rival’s stack that bets on “moats will be diluted by time” is making a calculated wager. Yet the logic holds. As model abstraction and compiler layers grow more sophisticated, application developers may increasingly decouple from underlying hardware vendors. The critical variables are timeline and patience—whether Huawei can sustain its infrastructure play until the ecosystem shifts.
5. Bottlenecks: HBM, Power, Software Stack