Huawei’s AI Master Plan Isn’t About Beating OpenAI. It’s About Becoming the Next Nvidia.

Huawei's AI Strategy: Sell the Base, Not the Strongest Model

The headline "Huawei executive's internal 10,000-word memo exposed, ICT business target is to become Nvidia" grabs attention. It is inaccurate. Huawei's Xinsheng Community published a summary of a discussion between Guo Ping, chairman of Huawei's supervisory board, and new employees. No secret file leaked. No employee risked a job to smuggle it out. Huawei used the session to explain what it will do in AI, what it will not do, and how it plans to close gaps.

The route matters more than the slogan. Huawei does not want to be the next OpenAI or a copy of Nvidia. It wants to sell AI infrastructure: chips, clusters, cloud, toolchains and industry entry points that let models run.

1. "No Business Expansion Plan" Sets a Boundary

Guo Ping said Huawei's core strategy is "focus," deepening work in connectivity and computing, with "no business expansion plan."

That sounds like retreat. It works as boundary management. Huawei will not enter every terminal market. It will try to keep the valuable parts of terminals close: connectivity, computing, operating systems and intelligent driving.

Automotive shows the pattern. Huawei does not build cars. Through Yinwang, it supplies intelligent driving. Yinwang has automaker shareholders. Huawei will lead its development for a long time. Huawei gives up the heavy assets and geopolitical risk of car manufacturing. It keeps the high-value intelligent driving entry point. It has not left the auto industry. It sits at the industry's intelligent layer.

6G follows the same logic. Guo said Huawei will "make a difference" in "zero-to-one" exploration. It will keep executing with speed, quality, efficiency and economy. It will remain an important contributor to global unified standards. This is not a reset. It is more work on standards, connectivity and underlying capabilities along the main communications track.

"No business expansion plan" is a boundary statement. By skipping some businesses, Huawei can go deeper in the ones it keeps.

2. Becoming Nvidia Means Selling the Base

Guo said Huawei's goal in ICT and computing is to become "Nvidia." The goal is not to own a self-developed large model. The goal is to help customers build strong models and let any global model run efficiently on Ascend, Kunpeng, SuperNode and clusters.

The second half matters. Nvidia is hard to catch because of more than the GPU. CUDA, NVLink, communications, servers, development tools, model adaptation and the code customers have already written form the moat. Huawei will not close the gap by shipping one chip with similar performance. The harder test comes after a customer migrates a model. Can it run? Can it run stably? Can it be modified easily? Is support available when something breaks?

Huawei wants to sell the foundation that runs all large models. It does not want to sell one large model. If that position holds, its competitor is not a model company. Its competitor is the standard-setter for AI infrastructure. The next fight is over the chips, clusters, cloud and industry entry points needed to run models.

3. Large Models: No Single Faith, Calculate by Business

Guo's approach to large models is practical. He did not ask all of Huawei to believe in Pangu. He split the question by business line.

ICT and computing do not need their own large model. The focus is letting global models run efficiently on Ascend, Kunpeng, SuperNode and clusters. Terminals and intelligent driving should have their own models. The data loop and product experience stay in-house. Huawei Cloud must have a competitive self-developed model. Without one, why would customers choose its cloud?

The logic splits by business line. A large model is no longer one technology company's unified face. Compute sellers want more models. Phone and intelligent driving teams want models that understand their users. Cloud sellers use models to drive compute, storage and services. Many big companies claim "full stack," then try to do everything and prove they are strongest in everything. Huawei has made its rule clear: self-develop when the business needs it. If it does not, build the base.

Guo also admitted Huawei started early in large models and did not dominate. Anthropic, Kimi and DeepSeek made the breakthroughs. He concluded that big companies have resource advantages and organizational disadvantages. For new employees, this is also a question for Huawei's management: when technology moves faster than a big company's decision cycle, how do you avoid "rising early and arriving late"?

4. If Single Chips Lose, Absorb the Gap with Systems

Chips are another area where Huawei uses strengths and avoids weaknesses. Guo said Huawei is constrained in advanced processes. It needs design and manufacturing integration. It must use combined advantages to cover process shortcomings. He mentioned "Tao's Law," proposed by He Tingbo. The idea replaces "geometric scaling" with "time scaling." Process and manufacturing bind more tightly to improve competitiveness under constraints.

On the Ascend 950 series, the strategy is direct. Do not fight over single-chip parameters. Fight over system-level advantages.

Public comparisons show the gap. A single Ascend 950DT delivers around 2 PFLOPS FP4. Nvidia's Vera Rubin is at the 17.5 PFLOPS level in dense FP8/FP4, with higher sparse inference numbers. On HBM capacity, the 950DT has 144GB. Rubin has 288GB HBM4. On bandwidth, the 950DT has 4TB/s. Rubin has 19.2TB/s, with an architectural peak of 22TB/s. In a direct single-chip fight, the gap comes from process. Process alone cannot close it quickly.

Huawei knows the gap and does not compete on that line. Its play is SuperNode and clusters. The system starts at 1024 cards and scales to 8192 cards. It uses unified memory addressing and low-latency all-optical interconnect. The Atlas 950 shown at WAIC was a 1024-card version: 1 EFLOPS FP8, 2 EFLOPS FP4, 256TB global unified memory space and about 3 microseconds RTT. The full 8192-card version is planned for 8 EFLOPS FP8, 16 EFLOPS FP4, 160 racks and about 1000 square meters.

Nvidia's NVL72 uses 72 GPUs. The full Atlas 950 has higher total system compute but uses 114 times as many chips. It is like an army. Individual soldiers may be weaker. Communications, formation and command systems organize thousands into one body. Huawei spent thirty years in communications and connectivity. Its advantage sits there.

Nvidia's moat is NVLink plus CUDA. On interconnect, Nvidia has hit a wall. Its electrical interconnect loses sharply beyond about one meter. Single-rack scale has long been stuck around 72 GPUs. It can push efficiency inside one rack to the limit. It cannot expand outward indefinitely. Huawei goes the opposite way with all-optical interconnect. Electricity runs inside the rack. Optics run between racks. The aim is to bypass the physical limit and stack larger scale. As models grow and MoE and long context make communication the bottleneck, the deciding factor shifts from "whose single card is faster" to "who can organize ten thousand cards to work as one."

That play has costs. Scaling SuperNode from 1024 to 8192 cards brings diminishing returns. More cards mean more optical modules, connectors, memory stacks and failure points. Communication complexity, energy consumption and fault-tolerance costs all rise. Aggregate bandwidth numbers are total-system figures. They do not mean any two cards can use that bandwidth exclusively. Microsecond latency is a single-hop metric. Once a task leaves the SuperNode and returns to the outer network, latency jumps to another order of magnitude. Actual training capacity must be multiplied by parallel efficiency and available time. The 82% to 85% cluster utilization and "several times Nvidia" claims come mainly from Huawei's own disclosures. No third-party measurement under a unified benchmark exists.

Hardware parameters can pass a competitor overnight. Moving the world's developers from CUDA code to CANN takes time and software-stack work. CANN's open-source community has around 3,000 monthly active developers. Against CUDA, that number is almost negligible. Guo said the goal is to let "any global large model run efficiently." More than half of that pressure sits on migration. The chip problem is not the biggest one. Developer habits, toolchain maturity and migration cost are harder.

Guo recommended reading Liang Wenfeng's exchange notes. He cited a judgment: rapid AI iteration can weaken CUDA-type moats or replace them with new technology. A company currently choked by another's software stack choosing to believe "moats will be diluted by time" contains a bet. It also has logic. As model layers and compilation layers become more abstract, upper-layer developers may care less about whose hardware sits underneath. The question is how long that trend takes and whether Huawei can wait.

5. Bottlenecks: HBM, Power, Software Stack

The first hard constraint on Huawei's system route is domestic HBM capacity.

Some analysis predicts that if only domestic HBM is used, Huawei's AI compute output by 2028, measured in H100-equivalent products, may be only about 1% of Nvidia's. Bloomberg previously reported that DeepSeek ordered 160,000 Ascend 950DT chips from Huawei. Whether the numbers are precise, the direction is clear. No matter how strong the Ascend SuperNode is, it needs enough memory stacks and advanced packaging capacity. The chip problem is not the largest. Memory and manufacturing are.

Another variable sits across the Pacific. The US does not lack compute chips now. It lacks power more than memory. Reports say US AI data centers are queuing for power. 700GW of power applications contain large amounts of "ghost demand." Musk said about 15GW of AI compute may fail to come online by 2027 because power cannot keep up. In China, if you have cards, you can generally power them. The gap exists. It will not widen indefinitely. The US lacks power. China lacks memory. It depends on whether domestic capacity moves faster or overseas power infrastructure is built faster.

The software stack is the third bottleneck. CANN's open-source community has around 3,000 monthly active developers. Against CUDA, that number is almost negligible. Guo said the goal is to let "any global large model run efficiently." More than half of that pressure sits on migration. The chip problem is not the biggest one. Developer habits, toolchain maturity and migration cost are harder.

Guo recommended reading Liang Wenfeng's exchange notes. He cited a judgment: rapid AI iteration can weaken CUDA-type moats or replace them with new technology. A company currently choked by another's software stack choosing to believe "moats will be diluted by time" contains a bet. It also has logic. As model layers and compilation layers become more abstract, upper-layer developers may care less about whose hardware sits underneath. The question is how long that trend takes and whether Huawei can wait.

6. Talent and Organization: Do Not Compete with Machines on Standard Answers

Guo also discussed AI talent and organization. He said that in the AI era, humans have no advantage competing with AI on standard answers. What matters more is describing problems precisely, continuing to ask questions and solving practical problems. Continuous learning remains the basic ability for survival and development at work.

Take carrier networks. From 2G to 5G and future networks, network elements are complex and diverse. O&M costs are high. Future work needs AI-driven automated network O&M. That requires large numbers of applied AI talent. Huawei does not need only algorithm researchers who publish papers. It needs people who can embed AI into complex engineering scenarios and solve practical problems.

He concluded that AI may be the last technological revolution in human society. Huawei needs to gain competitive advantage in this revolution. Organization, process, capability and talent all need to match it. The sentence admits that technological revolution is not only a technology problem. It is an organizational problem. Big companies have resource advantages and organizational burdens. Guo saying "big companies have organizational disadvantages" is a warning bell for Huawei's management system.

7. Closing: Gaps, Bottlenecks, Bets

Huawei's AI route is clear. Single-chip process cannot catch up, so it relies on system capability. Self-developed large models are not leading, so it repositions as the base that runs all models. It started early and did not become first tier. That is no shame. It shows big companies have advantages and burdens.

Its next goal is to fight for the chips, clusters, cloud and industry entry points needed to run models. It does not want to be the next OpenAI or simply copy Nvidia. It wants to become another kind of infrastructure giant in the AI era. It may not own the strongest model, but it wants all models to run on its foundation. It may not have the fastest single chip, but it wants to organize ten thousand cards to work as one. It may not do everything, but it wants connectivity and computing to be hard to avoid.

Whether this road works depends in the short term on domestic HBM capacity, in the medium term on CANN migration speed and in the long term on whether the organization can stay agile in a fast technology cycle. There are gaps, bottlenecks and bets. Huawei has not avoided the gaps. It has stated the approach: use strengths, avoid weaknesses and change the rules of competition.

The US lacks power. China lacks memory. The single-chip gap is several times. The system uses 114 times as many chips. CUDA has 3,000 monthly active developers. The goal is global models running efficiently. Those numbers define the bet.