据多家境外媒体报道，AI大模型独角兽DeepSeek正筹划向华为下达一份规模空前的采购订单——至少16万颗昇腾950DT AI加速器，将部署于内蒙古一座新建的大型数据中心。这一规模若属实，将成为迄今已知最大的国产AI芯片集群之一。
这不仅仅是一笔采购订单。在美国出口管制持续收紧、Nvidia高端芯片对华供应受阻的大背景下，这份订单的战略含义远超商业层面：它意味着中国头部AI公司正在以规模化部署的方式，将核心算力基础设施从"美系依赖"向"国产替代"切换。而华为作为国产AI芯片的核心供给方，也首次面临"产能瓶颈倒逼客户向监管层求助"的局面——这种供需关系的紧张本身，就是国产算力产业链走向成熟的关键信号。
事件背景：一场酝酿已久的算力迁移
据报道披露，DeepSeek此次采购的昇腾950DT芯片并不用于模型训练，而是面向推理场景。这一细节值得行业从业者高度关注——它揭示了当前大模型产业算力消耗结构正在发生的根本性变化。
过去几年，AI算力市场的重心几乎完全由训练侧主导。千亿参数、万亿参数级别的预训练任务，吞噬了全球最昂贵的GPU资源。但随着GPT-4级别的基础模型趋于收敛、模型推理价格战全面打响，推理侧的算力需求正在以更快速度膨胀。OpenAI CEO Sam Altman此前曾公开判断，未来推理算力的消耗将远超训练算力，这一判断在国内市场同样成立。
对国产芯片而言，这是一个战略机遇窗口。训练任务对芯片的绝对性能、网络互联带宽、显存容量提出了近乎苛刻的要求，国产芯片短期内难以全面替代；而推理任务更看重能效比、单位算力成本和规模化部署能力，这恰恰是华为昇腾系列已经具备的竞争力优势。
昇腾950DT是华为公布的昇腾系列路线图中的关键一环。按华为轮值董事长徐直军此前披露的节奏，950DT预计将于2026年第四季度开始出货。DeepSeek此次下单的规模，几乎等同于华为单年Ascend 950DT产量的一个量级。
报道指出，受高端存储芯片供应紧张等因素影响，昇腾950DT今年的产能被限制在"低数十万颗"区间。换言之，DeepSeek一家的订单，就可能吃掉华为几乎全部的初期产能。这意味着华为需要在其客户矩阵中做出优先级排序——而DeepSeek已经向北京方面寻求协调，希望获得更优先的供货地位。这一供需关系的紧张，既是国产AI芯片"卖方市场"形成的标志性事件，也是产业从"能用"迈向"够用且紧缺"的关键拐点。
DeepSeek此次部署地点选在内蒙古，绝非偶然。在国家"东数西算"工程框架下，内蒙古和林格尔、贵州贵安、甘肃庆阳等八大国家算力枢纽节点，已经成为AI大模型训练与推理基础设施的核心承载地。报道披露，这座数据中心规划在吉瓦级（Gigawatt）电力消耗规模——按行业平均换算，足以满足约75万户家庭的用电需求。这种超大规模电力消耗，已经远远超出了一般商业数据中心的范畴，更接近于国家级超算枢纽的能级。
行业影响：三条产业链的连锁反应
DeepSeek订单最直接的冲击波，落在了半导体产业链上游，尤其是与AI芯片配套的高端HBM（高带宽内存）领域。报道明确指出，制约昇腾950DT产能的核心瓶颈之一是"高端存储芯片短缺"。这一表述背后，是整个国产AI芯片产业链的共同痛点：HBM的国产化进程目前仍落后于逻辑芯片。长江存储、长鑫存储等国内存储厂商虽在DDR5等产品上取得突破，但在AI芯片所需的HBM2E/HBM3层面，与SK海力士、三星仍有显著差距。
DeepSeek订单带来的需求拉动，将进一步刺激国产存储厂商加速HBM研发与产能扩张。叠加此前多家国内AI芯片厂商公布的HBM国产化导入计划，未来12-24个月内，国产AI芯片"算力+存力"的全栈国产化有望迈出实质性步伐。与此同时，先进封装（如CoWoS等）产能也将持续紧张。昇腾950DT这类高算力芯片对先进封装的依赖度极高，中芯国际等代工厂的封装产能扩张速度，将直接决定国产AI芯片的供给上限。
DeepSeek的此次布局，折射出中国AI产业竞争重心正在发生深刻迁移。过去两年，中国AI产业的叙事主线是"模型能力对标GPT-4"——从百模大战到价格屠夫，行业关注的焦点集中在模型参数、benchmark分数、推理价格。但随着基础模型能力逐步趋同，真正的竞争壁垒正在向底层基础设施迁移：谁能以更低成本、更高效率、更稳定的算力供给支撑大规模推理服务，谁就能在下一阶段的商业化落地中占据主导地位。
DeepSeek选择国产芯片进行规模化推理部署，正是这一趋势的典型注脚。对整个AI行业而言，这一信号可能引发连锁反应——其他头部大模型公司、互联网巨头、垂直行业AI服务商，都将更积极地评估国产AI芯片在其推理基础设施中的占比，进而加速国产算力的市场渗透。而这一过程，也将反过来倒逼国产AI芯片厂商加速产品迭代。DeepSeek目前仍依赖Nvidia芯片进行模型训练的事实表明，国产芯片在训练侧的全面替代仍需时日；但推理侧的大规模商用，将为国产芯片积累宝贵的工程化经验，反哺下一代产品的设计与优化。
对通信行业而言，这一事件的影响更具结构性意义。国内三大运营商——中国移动、中国电信、中国联通——近年来均在加速向"算力服务商"转型。中国移动的"九天"人工智能平台、中国电信的"息壤"算力平台、中国联通的国产算力底座建设，都已经形成相当规模的国产化算力服务能力。DeepSeek此次大规模采购国产AI芯片，进一步验证了国产算力在商业级AI推理场景中的可行性。
从产业链角度看，运营商在算力服务市场中的角色将出现重要变化：其一，运营商有望成为国产AI芯片的"最大集成商"，随着国产AI芯片供给增加，运营商将能够以更具竞争力的成本提供AI推理即服务（AI IaaS/PaaS）；其二，承载网络与算力的深度融合将加速，内蒙古这类超大规模AI数据中心必然对承载网络的低延迟、高带宽提出极高要求；其三，"算力+网络+AI"的一体化解决方案将成为运营商的新增长曲线，在政务、金融、能源、交通等关键行业，国产化算力的合规性优势与运营商的渠道优势叠加，可能催生全新的商业模式。
趋势展望：国产算力闭环的形成路径
从更宏观的视角看，DeepSeek与华为的这次联手，是中国国产算力闭环从"概念"走向"现实"的标志性事件。所谓"闭环"，指的是覆盖芯片设计、代工制造、封装测试、存储配套、系统集成、模型训练、推理部署、行业应用的完整国产化链条。当前，这一闭环已经取得了显著进展：在芯片设计端，华为、寒武纪、海光、燧原、摩尔线程等厂商已经形成多层次产品矩阵；在模型端，DeepSeek、文心一言、通义千问等国产大模型已经具备国际竞争力；在应用端，金融、政务、能源等行业的国产化部署正在加速。
但客观来看，闭环仍存在几个关键缺口：一是HBM等高端存储的国产替代，这是DeepSeek订单中最明显的瓶颈，也是整个国产AI算力体系最脆弱的环节，预计未来两年这一领域将是国产半导体突围的重点方向；二是先进制程代工的产能瓶颈，昇腾950DT等高端AI芯片的制造对国内代工厂的工艺成熟度提出了极高要求，中芯国际等代工厂的产能扩张与良率提升进度，将直接影响国产AI芯片的整体供给能力；三是训练侧的全面替代仍需时间，在需要极致性能的大模型预训练场景中，国产芯片的竞争力仍有差距，这一缺口的填补需要芯片厂商与AI公司在更长的时间维度上协同迭代。
值得注意的是，DeepSeek此次请求北京方面出面协调华为供货优先级，揭示了国产算力产业链背后的政策逻辑——在战略性产业领域，政府始终扮演着"隐性兜底者"的角色。这种"政策托底+市场拉动"的双轮驱动模式，已经在多个国产化替代领域被验证有效。DeepSeek订单很可能只是开端，未来将有更多AI公司、互联网企业、央企客户加入到国产AI芯片的大规模采购队列中。
The API key used for this request has reached its budget. Please [raise the key budget](https://enter.pollinations.ai/edit-key?id=89idjaTSg2hI4YwZDuO8ZME5Ma6GrFps&ref=agent_key_budget), then try again.
Topping up the wallet does not raise this limit. If this isn’t your Pollinations account, contact whoever runs the app or service you’re using.
本文来源：https://www.toutiao.com/article/7505042475268060246/
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
According to multiple foreign media reports, the AI unicorn DeepSeek is planning to place an unprecedented order with Huawei—at least 160,000 Ascend 950DT AI accelerators, which will be deployed in a newly built large-scale data center in Inner Mongolia. If this scale is true, it will become one of the largest known domestically produced AI chip clusters to date.
This is not just a purchase order. Against the backdrop of tightening U.S. export controls and blocked supply of Nvidia's high-end chips to China, the strategic meaning of this order goes far beyond the commercial level: it means that leading Chinese AI companies are switching their core computing infrastructure from "U.S. dependence" to "domestic replacement" through large-scale deployment. Huawei, as the core supplier of domestic AI chips, is also facing the situation of "production bottleneck forcing customers to appeal to regulatory authorities" for the first time—the tension in this supply-demand relationship itself is a key signal of the maturity of the domestic computing power industrial chain.
Background: A long-planned computing power migration
According to the report, the Ascend 950DT chips purchased by DeepSeek this time are not used for model training, but for inference scenarios. This detail is highly valued by industry practitioners—it reveals that the current structure of computing power consumption in the large model industry is undergoing a fundamental change.
In the past few years, the center of gravity of the AI computing power market has been almost completely dominated by the training side. Pre-training tasks with hundreds of billions and trillions of parameters have consumed the most expensive GPU resources in the world. However, as the basic model of GPT-4 level converges and the price war of model inference is fully underway, the demand for computing power on the inference side is expanding at a faster rate. OpenAI CEO Sam Altman previously publicly judged that the consumption of inference computing power will far exceed that of training computing power in the future, and this judgment also holds true in the Chinese market.
For domestic chips, this is a strategic opportunity window. Training tasks place almost harsh requirements on the absolute performance, network interconnection bandwidth, and memory capacity of chips, which domestic chips cannot fully replace in the short term. Inference tasks pay more attention to energy efficiency ratio, unit computing power cost, and large-scale deployment capabilities, which are the competitive advantages that Huawei's Ascend series have already possessed.
Ascend 950DT is a key part of the Ascend series roadmap released by Huawei. According to the schedule previously disclosed by Huawei's Rotating Chairman Xu Zhigang, 950DT is expected to begin shipping in the fourth quarter of 2026. The scale of DeepSeek's order this time is almost equivalent to the order of magnitude of Huawei's annual output of Ascend 950DT.
The report pointed out that due to factors such as the tight supply of high-end storage chips, the production capacity of Ascend 950DT this year was limited to "low hundreds of thousands of pieces". In other words, DeepSeek's order alone may consume almost all of Huawei's initial production capacity. This means that Huawei needs to prioritize its customer matrix—and DeepSeek has already sought coordination with Beijing to obtain higher priority delivery status. This tension in the supply-demand relationship is both a landmark event marking the formation of a "seller's market" for domestic AI chips, and a key turning point for the industry to move from "usable" to "sufficient and scarce".
DeepSeek chose Inner Mongolia as the deployment location, which was not accidental. Under the framework of the national "East-to-West Data Transfer" project, eight national computing power hubs, including Wulanchabu in Inner Mongolia, Gui'an in Guizhou, and Qingyang in Gansu, have become the core carriers of AI large model training and inference infrastructure. According to the report, the data center plans to consume electricity at the Gigawatt level—according to industry average conversion, this is sufficient to meet the electricity demand of approximately 750,000 households. This ultra-large-scale power consumption has far exceeded the scope of a general commercial data center and is closer to the energy level of a national supercomputing hub.
Industry impact: a chain reaction of three industrial chains
The most direct impact of DeepSeek's order was on the upstream of the semiconductor industry chain, especially the high-end HBM (High Bandwidth Memory) field that is paired with AI chips. The report clearly pointed out that one of the core bottlenecks limiting the production capacity of the Ascend 950DT is the "shortage of high-end storage chips". Behind this statement is the common pain point of the entire domestic AI chip industry chain: the localization process of HBM is currently lagging behind that of logic chips. Although domestic storage manufacturers such as Yangtze Memory and Changxin Memory have made breakthroughs in products such as DDR5, there is still a significant gap with SK Hynix and Samsung in terms of HBM2E/HBM3 required by AI chips.
The demand driven by DeepSeek's order will further stimulate domestic storage manufacturers to accelerate the R&D and production expansion of HBM. Combined with the previous HBM localization introduction plans announced by many domestic AI chip manufacturers, the full-stack localization of domestic AI chips "computing power + storage power" is expected to take a substantial step forward in the next 12 to 24 months. At the same time, the production capacity of advanced packaging (such as CoWoS) will also remain tight. High-computing-power chips such as the Ascend 950DT are highly dependent on advanced packaging. The rate of production capacity expansion of the foundries such as SMIC will directly determine the upper limit of the supply of domestic AI chips.
DeepSeek's layout this time reflects the deep migration of the competitive center of the Chinese AI industry. In the past two years, the narrative main line of the Chinese AI industry has been "model capability benchmarking against GPT-4"—from the battle of a hundred models to the price butcher, the focus of industry attention has been on model parameters, benchmark scores, and inference prices. However, as the basic model capabilities gradually converge, the real competitive barrier is shifting to the underlying infrastructure: who can provide stable computing power supply with lower costs, higher efficiency, and higher stability to support large-scale inference services, who will dominate the commercialization landing in the next stage.
DeepSeek's choice to use domestic chips for large-scale inference deployment is a classic footnote of this trend. For the entire AI industry, this signal may cause a chain reaction—other leading large model companies, Internet giants, and vertical industry AI service providers will more actively evaluate the proportion of domestic AI chips in their inference infrastructure, thereby accelerating the market penetration of domestic computing power. And this process will in turn force domestic AI chip manufacturers to accelerate product iteration. The fact that DeepSeek still relies on Nvidia chips for model training shows that the full replacement of domestic chips on the training side still takes time. However, large-scale commercial use on the inference side will accumulate valuable engineering experience for domestic chips, feeding back the design and optimization of the next generation of products.
For the communications industry, the impact of this event is more structural. In recent years, the three major domestic operators—China Mobile, China Telecom, and China Unicom—have been accelerating their transformation into "computing power service providers". China Mobile's "Nineteen" artificial intelligence platform, China Telecom's "X壤" computing power platform, and China Unicom's domestic computing power base construction have all formed a considerable scale of localized computing power service capabilities. DeepSeek's large-scale purchase of domestic AI chips further verifies the feasibility of domestic computing power in commercial-level AI inference scenarios.
From an industry chain perspective, the role of operators in the computing power service market will undergo important changes: firstly, operators are expected to become the "largest integrator" of domestic AI chips. As the supply of domestic AI chips increases, operators will be able to provide AI inference-as-a-service (AI IaaS/PaaS) at more competitive costs; secondly, the deep integration of bearer networks and computing power will accelerate, and ultra-large-scale AI data centers such as Inner Mongolia will inevitably put extremely high requirements on the low latency and high bandwidth of the bearer network; thirdly, the complete solution of "computing power + network + AI" will become a new growth curve for operators, and the compliance advantages of localized computing power and the channel advantages of operators will叠加 in key industries such as government affairs, finance, energy, and transportation, which may give birth to new business models.
Trend outlook: the formation path of the domestic computing power loop
Looking at it from a more macroscopic perspective, the collaboration between DeepSeek and Huawei is a landmark event for the Chinese domestic computing power loop to move from "concept" to "reality". The so-called "closed loop" refers to the complete localization chain covering chip design, wafer fabrication, packaging and testing, storage matching, system integration, model training, inference deployment, and industry applications. At present, this closed loop has made significant progress: on the chip design side, manufacturers such as Huawei, Cambricon, Hygon, Su Yuan, and Moore Threads have formed a multi-level product matrix; on the model side, domestic large models such as DeepSeek, Wenxin Yiyan, and Tongyi Qianwen have already been internationally competitive; on the application side, localization deployments in industries such as finance, government, and energy are accelerating.
However, objectively speaking, there are still several key gaps in the closed loop: first, the localization of high-end storage such as HBM, which is the most obvious bottleneck in the DeepSeek order and also the weakest link in the entire domestic AI computing power system. It is expected that this field will be a key direction for the breakthrough of domestic semiconductors in the next two years; second, the production capacity bottleneck of advanced process manufacturing, the high-end AI chips such as the Ascend 950DT put extremely high demands on the process maturity of domestic foundries, and the production capacity expansion and yield improvement progress of foundries such as SMIC will directly affect the overall supply capacity of domestic AI chips; third, the full replacement on the training side still takes time, and in the pre-training scenarios of large models that require extreme performance, the competitiveness of domestic chips is still far behind, and the filling of this gap requires chip manufacturers and AI companies to collaborate and iterate over a longer time dimension.
It is worth noting that the fact that DeepSeek requested Beijing to coordinate the delivery priority of Huawei reveals the policy logic behind the domestic computing power industry chain—in strategic industries, the government has always played the role of an "implicit backer". This "policy backup + market pulling" two-wheel drive model has been proven effective in many localization replacement fields. The DeepSeek order may just be the beginning. In the future, more AI companies, Internet enterprises, and central enterprises will join the large-scale procurement queue of domestic AI chips.
In summary, DeepSeek's order is not only a commercial contract, but also a political and economic signal. It symbolizes that China's AI industry is entering a new stage of "independence and self-reliance". In the context of global technology decoupling, domestic computing power is no longer an option, but a necessity driven by reality. This will profoundly change the competitive landscape of the Chinese AI industry and may even trigger a round of domestic technology substitution reactions driven by the computing power layer.
Topping up the wallet does not raise this limit. If this isn’t your Pollinations account, contact whoever runs the app or service you’re using.