In [Jensen Huang’s] view, competition within the AI industry is no longer a battle over a single technology stack, but rather a contest of systems engineering. In the future, the barrier to entry will shift from “whether one possesses computing power” to “whether one possesses system-level capabilities.” 在他看来，AI产业的竞争不再是单一技术栈之争，而是系统工程的比拼。未来门槛将从“是否拥有算力”转向“是否具备系统级能力”。
Hu Jiaqi (胡嘉琦), “Nvidia’s Next Battle: Not Chips, but the Computing Power Ecosystem” (英伟达的下一战：不是芯片，是算力体系), Business Management Review (商学院), April 19, 2026
Journalist Hu Jiaqi’s article in Business Management Review highlights an often unrecognized but strategically significant point. The AI industry, from chip makers to think tanks, is beginning to converge on a system-level understanding of computing power. Notably, it is a view that Beijing has already elevated to the level of national strategy. This is not simply parallel development, but the gradual alignment of AI industry practice with a model of organization that China has spent the past decade constructing. The AI Industry is discovering what China has already systematized.
A translation of Hu Jiaqi’s “Nvidia’s Next Battle: Not Chips, but the Computing Power Ecosystem” follows below. My summary and analysis of the article can be found here.
Nvidia’s Next Battle: Not Chips, but the Computing Power Ecosystem
英伟达的下一战：不是芯片，是算力体系
Author: Hu Jiaqi | 胡嘉琦
Source: Business Management Review, April Issue | 《商学院》杂志4月刊
Publication Date: April 19, 2026
Translation: This is a lightly-edited Google Translate neural machine translation.
Text color and highlighting, as well as photograph placement, approximate the original document.
An advantage in computing power no longer automatically translates into an advantage in efficiency. The key to this competition is no longer “who is stronger,” but “who can more efficiently define infrastructure. | 算力优势，不再自动等同于效率优势。竞争的关键已不在于“谁更强”，而在于“谁更高效地定义基础设施”。
With a flurry of tech launch events taking place in the spring of 2026, San Jose, California, has once again emerged as the epicenter of the global competition for computing power. Amidst the concentrated launch of a new generation of chips, interconnect technologies, and system-level products, NVIDIA founder and CEO Jensen Huang has raised market projections for the 2025–2027 period to $1 trillion, a figure reflecting the growth opportunities emerging from the massive shift in AI logic from “training” to “inference.”
2026年春季科技发布会密集召开，美国加州圣何塞再次成为全球算力竞争的风暴中心。随着新一代芯片、互连技术与系统级产品集中发布，英伟达（NVIDIA）创始人兼CEO黄仁勋将2025—2027年的市场预期上调至1万亿美元，体现了当AI逻辑从“训练”向“推理”大迁移后带来的发展机遇。
NVIDIA has not stopped at chips; rather, it continues to push the boundaries. In the context of actual industrial operations, the chip serves merely as the entry point; centered around computing power lies a far more expansive ecosystem of expenditures, encompassing high-bandwidth memory, server clusters, data centers, electrical power systems, liquid cooling solutions, and the high-speed interconnect networks that weave through them all.
英伟达并未止步于芯片，而是不断扩展边界。在真实的产业运行中，芯片只是入口，围绕算力展开的，是一整套更为庞大的支出体系，包括高带宽内存、服务器集群、数据中心、电力系统、液冷散热以及贯穿其间的高速互连网络。
A journalist from from Business Management Review learned from NVIDIA that the company has launched Spectrum-X Photonics, an integrated optoelectronic networking switch, designed to facilitate the scaling of “AI factories” to the level of millions of GPUs. Concurrently, NVIDIA unveiled the NVIDIA Spectrum-X and NVIDIA Quantum-X silicon photonics networking switches, enabling AI factories to interconnect millions of GPUs across different regions. This innovation significantly reduces energy consumption while optimizing operational costs, thereby achieving a deep fusion of electronic circuits and optical communications.
《商学院》记者从英伟达方面获悉，英伟达推出了光电一体化封装网络交换机Spectrum-X Photonics，助力“AI工厂”扩展至数百万GPU级别。同时，英伟达发布NVIDIA Spectrum-X与NVIDIA Quantum-X硅光网络交换机，使AI工厂能够跨区域连接数百万GPU，在大幅降低能耗的同时优化运营成本，实现电子电路与光通信的深度融合。
This series of developments indicates that AI is no longer a market for point solutions, but a systemic competition centered on infrastructure. Chips, networks, and storage are being reorganized and deployed into the industry as a holistic capability.
这一系列动作都表明，AI不再是一个单点产品市场，而更像是一场围绕基础设施展开的系统性竞争。芯片、网络与存储被重新组织，作为整体能力落地到产业。
01
The Era of Inference: Reconstructing the Compute System and Performance Explosion | 推理时代：算力体系的重构与性能爆发
Computing power supply models characterized by centralization and scale can efficiently support the training and inference requirements of large-scale models. | 以集中化、规模化为特征的算力供给方式，能够高效支撑大模型训练与推理需求。
Amidst this concentrated wave of technology and product launches, a more intuitive signal emerges: the way AI computing power is organized is being redefined.
在这一轮技术与产品的集中发布中，一个更直观的信号是：AI算力的组织方式正在被重新定义。
At this year’s GTC, the Groq 3 LPU made its debut, expanding the Rubin platform’s core compute chip lineup to seven units. Concurrently, NVIDIA packaged these chips into five rack-scale systems, creating a complete AI supercomputer solution.
在本届GTC上，Groq 3 LPU首次亮相，使Rubin平台的核心计算芯片扩展至7颗。与此同时，英伟达将其与5个机架级系统打包，形成一整套AI超级计算机方案。
This shift is no mere coincidence; rather, it serves as a direct response to the most critical demands of the present moment. The training and inference of large-scale models are now entering a phase of massive scaling; as model parameters surge into the trillion-scale era, traditional server architectures are increasingly struggling to keep pace, compelling the industry to pivot toward system-level integration characterized by higher density and lower latency.
这种变化并非偶然，而是直接回应了当下最核心的需求。大模型训练与推理正在进入规模化阶段，随着模型参数步入万亿级时代，传统服务器架构逐渐难以承载，行业不得不转向更高密度、更低延迟的系统级整合。
In contrast, however, Google’s TPUs, Cerebras’s wafer-scale chips, and the specialized inference architectures from vendors such as Groq are able to offer more direct advantages in terms of energy efficiency, response speed, and cost control. This disparity is currently influencing customers’ decision-making logic. Large-scale model companies, exemplified by OpenAI and Meta, are shifting away from their past reliance on single compute providers toward a multi-vendor portfolio strategy, aiming to achieve optimal efficiency amidst continuously rising inference costs.
但是相比之下，Google的TPU、Cerebras的晶圆级芯片以及Groq等厂商的专用推理架构，能够在单位能耗、响应速度和成本控制上形成更直接的优势。这种差异正在影响客户的选择逻辑。以OpenAI和Meta为代表的大模型公司，正在从过去对单一算力供应商的依赖，转向多供应商组合策略，以在推理成本不断攀升的背景下实现效率最优。
In other words, NVIDIA’s problem lies not in a lack of performance, but rather in the fact that its general-purpose architecture no longer possesses an absolute efficiency advantage in the “inference-first” era. As AI enters the phase of large-scale commercial application, the focal point of industry competition is shifting from “who can provide greater computing power” to “who can consistently deliver computing power at lower costs and with higher efficiency.” It is precisely against this backdrop that NVIDIA’s role has evolved: no longer merely a GPU supplier, the company now seeks to define the fundamental architecture of AI infrastructure through a comprehensive ecosystem of products and frameworks, thereby securing a more elevated position within the industry value chain.
换言之，英伟达的问题并不在于性能不足，而在于其通用架构在“推理优先”时代不再具备绝对的效率优势。当AI进入大规模商业化应用阶段，行业竞争的焦点正在从“谁能提供更强算力”，转向“谁能以更低成本、更高效率持续提供算力”也正是在这一背景下，英伟达的角色发生了变化，它不再只是GPU供应商，而是试图通过完整的产品与架构体系，定义AI基础设施的基本形态，并在这一过程中，占据产业链更高的位置。
Against this backdrop, Rubin represents more than just a product upgrade. As the successor to the previous-generation Blackwell GPU architecture, Rubin is, in fact, more akin to a comprehensive restructuring of the system architecture. It first debuted at Computex Taipei 2024, with more complete technical details disclosed at GTC 2025; named after American astronomer Vera Rubin, it continues NVIDIA’s tradition of naming its architectures after female scientists.
在这样的背景下，Rubin的意义不只是一次产品升级。作为上一代Blackwell GPU架构的继任者，Rubin更像是一轮系统架构的重构。它最早在2024年台北电脑展上亮相，并在2025年GTC上披露更完整的技术细节，名称则来自美国天文学家薇拉·鲁宾，延续了英伟达以女性科学家命名架构的传统。
In terms of product cadence, Rubin has clearly accelerated the pace of development. At CES 2026, Jensen Huang revealed that this architecture has entered the mass production phase, with the first batch of products expected to be delivered in the second half of 2026. Cloud providers such as Microsoft, Amazon, Google, and Oracle will be the first to deploy it, while AI companies like OpenAI have also joined this ecosystem.
从产品节奏来看，Rubin推进明显加快。在2026年CES上，黄仁勋就透露，该架构已进入量产阶段，首批产品预计在2026年下半年交付，微软、亚马逊、谷歌、甲骨文等云厂商将率先部署，OpenAI等AI公司也已加入这一体系。
In terms of performance, Rubin delivers a significant leap forward. Training efficiency has increased by approximately 3.5 times, while inference performance has improved by roughly 5 times. Furthermore, thanks to a more integrated architecture, the cost per unit of computing power has declined significantly; in certain scenarios, the cost per inference token is projected to drop to one-tenth of that of the previous generation.
性能层面，Rubin带来的是一次明显跃升。训练效率提升约3.5倍，推理性能提升约5倍，同时由于架构更为集成，单位算力成本显著下降，在部分场景中，推理Token成本有望降至上一代的十分之一。
At the architectural level, Rubin is collaboratively constituted by GPUs, a custom CPU (Vera), and network switching components, and is equipped with HBM4 high-bandwidth memory, significantly enhancing data throughput capabilities. Computing, storage, and networking are integrated within a single architecture, bringing it closer to a complete system rather than a chip product in the traditional sense.
架构层面，Rubin由GPU、定制CPU（Vera）以及网络交换组件协同构成，并配备HBM4高带宽内存，显著提升数据吞吐能力。计算、存储与网络被整合在同一体系中，使其更接近一个完整的系统，而非传统意义上的芯片产品。
However, higher performance also brings new engineering challenges. As compute density increases, power consumption rises rapidly, reaching upwards of 2,000 watts per card, imposing more stringent demands on cooling systems. Liquid cooling is gradually transitioning from an optional solution to a standard feature, and the design logic of data centers is evolving accordingly.
但更高性能也带来了新的工程挑战。随着算力密度提升，功耗迅速上升，单卡可达2000瓦以上，对散热系统提出更高要求。液冷逐渐从可选方案变为标配，数据中心的设计逻辑也随之改变。
In larger-scale deployments, Rubin will be integrated into NVIDIA’s supercomputing ecosystem, such as the DGX SuperPod, while also being available as modular products that customers can flexibly combine. For complex tasks, such as those involving ultra-long contexts, NVIDIA has designed specialized GPU variants that operate collaboratively within massive clusters, delivering single-node computing power approaching “data center-grade” levels.
在更大规模的应用中，Rubin将被纳入英伟达的超算体系，例如DGX SuperPod，同时也可以拆分为模块化产品供客户灵活组合。针对超长上下文等复杂任务，英伟达还设计了专用版本GPU，并通过大规模集群协同运行，提供接近“数据中心级”的单机算力。
Zhang Xinyuan, a veteran expert in corporate strategy and technology innovation management, as well as the head of Kefangde Consulting, pointed out in an interview with Business Mangagement Review magazine that the “AI Factory” and large-scale GPU cluster models championed by NVIDIA are emerging as one of the key forms of future AI infrastructure. This method of computing power provision, characterized by centralization and scale, can efficiently support the training and inference requirements of large-scale models; however, its sustainability hinges on whether the demand for computing power continues to grow, whether the supply of resources such as electricity and land remains commensurate, and whether the software and hardware ecosystem possesses sufficient openness and compatibility.
资深企业战略和技术创新管理专家、科方得咨询机构负责人张新原在接受《商学院》记者采访时指出，英伟达所推动的“AI工厂”与大规模GPU集群模式，正在成为未来AI基础设施的重要形态之一。 这种以集中化、规模化为特征的算力供给方式，能够高效支撑大模型训练与推理需求，但其可持续性取决于算力需求是否持续增长、电力与土地等资源供给是否匹配，以及软硬件生态是否具备足够的开放性与兼容性。
Under this logic, NVIDIA’s initiatives have also begun to extend further upstream and across a broader scope.
在这一逻辑下，英伟达的动作也开始向更上游和更广范围延伸。
02
Extending into Optical Communications and the Network Layer | 向光通信与网络层延伸
Against the backdrop of the accelerating development of generative AI, competition is shifting from models and algorithms to computing power supply and infrastructure capabilities. | 在生成式人工智能加速发展的背景下，竞争正在从模型与算法，转向算力供给与基础设施能力。
In early March 2026, NVIDIA announced partnerships with and investments in Lumentum Holdings and Coherent, extending its strategic footprint from computing chips into the realms of optical communications and networking. This move points to an emerging bottleneck: bandwidth and latency. As model scale expands, the flow of data across GPUs, racks, and even entire data centers is gradually becoming a critical factor influencing performance; relying solely on increased computing power is no longer sufficient.
2026年3月初，英伟达宣布与Lumentum Holdings和Coherent建立合作并进行投资，将布局从计算芯片延伸至光通信与网络层。这一动作指向一个正在显现的瓶颈，即带宽与延迟。随着模型规模扩大，数据在GPU、机架乃至数据中心之间的流动，逐渐成为影响性能的关键因素，仅依靠算力提升已经不再足够。
Behind this shift lies a transformation in industrial logic. Against the backdrop of the accelerating development of generative AI, competition is shifting from models and algorithms to computing power supply and infrastructure capabilities.
这一变化背后，是产业逻辑的转向。在生成式人工智能加速发展的背景下，竞争正在从模型与算法，转向算力供给与基础设施能力。
NVIDIA’s silicon photonics switch, through the innovative integration of optical components, reduces the number of lasers by approximately four-fold, boosts energy efficiency by roughly 3.5 times, enhances signal integrity by about 63 times, improves reliability in large-scale networking by approximately 10 times, and increases deployment efficiency by roughly 1.3 times, thereby driving the evolution of large-scale AI infrastructure toward greater efficiency and stability.
英伟达推出的硅光交换机通过创新性集成光器件，将激光器数量减少约4倍，能源效率提升至约3.5倍，信号完整性提升约63倍，大规模组网可靠性提升约10倍，部署效率提升约1.3倍，可以推动大规模AI基础设施向更高效、更稳定的方向演进。
Zhang Xinyuan posits that the significantly increased demand for data exchange between nodes has caused network bandwidth and latency to gradually emerge as critical factors constraining overall efficiency. This trend stems from model parameter scales reaching the trillion level, the continuous expansion of training datasets, and the heightened requirements of distributed training for high-frequency synchronous communication. Against this backdrop, silicon photonics technology is garnering heightened expectations; however, it remains in a transitional phase, moving from the laboratory toward large-scale commercial deployment. It continues to face challenges regarding manufacturing yields, cost control, and compatibility with existing electrical interconnect systems. While the technology is expected to see initial adoption in select high-performance applications within the next 3 to 5 years, a complete replacement of existing systems will require a significantly longer timeframe.
张新原认为，节点之间的数据交换需求显著增加，使网络带宽与延迟逐渐成为制约整体效率的关键因素，这一趋势源于模型参数规模进入万亿级、训练数据持续扩大，以及分布式训练对高频同步通信的更高要求。在此背景下，硅光技术被寄予更高期待，但仍处于从实验室走向规模化应用的过渡阶段，在制造良率、成本控制及与现有电互联体系的兼容性方面仍面临挑战，未来3—5年有望在部分高性能场景率先落地，但全面替代仍需更长周期。
Meanwhile, data center architectures are undergoing a transformation; a “network-centric” design is gradually driving systems away from the separation of compute and storage, toward a deep fusion of compute, networking, and storage. The network is no longer merely a connectivity conduit; it is gradually evolving into a central hub for resource scheduling and task orchestration. Consequently, the value of optical interconnects is being re-evaluated; their energy-efficiency advantages in long-distance and high-bandwidth transmission are continuously expanding their scope of application within high-performance computing and AI clusters. The industrial chain is also undergoing restructuring during this process; a trend toward optoelectronic synergy is gradually taking shape, presenting an holistic landscape characterized by the coexistence of vertical integration and specialized division of labor.
与此同时，数据中心的架构正在发生转变，“以网络为核心”的设计逐渐推动系统从计算与存储分离，走向计算、网络与存储深度融合。网络不再只是连接通道，而是逐步成为资源调度与任务编排的核心中枢。与之对应，光互连的价值被重新评估，其在长距离与高带宽传输中的能效优势，使其在高性能计算与AI集群中的应用空间持续扩大。产业链也在这一过程中发生重构，光电协同趋势逐渐形成，整体呈现出垂直整合与专业分工并存的格局。
Given this evolution, the entry of GPU manufacturers into the fields of networking and optical communications is also inevitable. Zhang Xinyuan pointed out that as the scale of AI clusters expands, network performance has emerged as a critical variable for system efficiency. By mastering full-stack capabilities, GPU vendors can achieve further optimization at the system level; this, in turn, is driving a shift in the competitive landscape, moving away from a focus on standalone hardware toward a competition centered on full-stack solutions. However, at the same time, the latency and consistency issues inherent in cross-regional deployment persist. Due to physical constraints, these challenges can, in the short term, only be mitigated through methods such as asynchronous training, gradient compression, and network optimization.
在这一演进下，GPU厂商进入网络与光通信领域也具有必然性。张新原指出，随着AI集群规模扩大，网络性能已成为系统效率的关键变量，通过掌握全栈能力，GPU厂商能够在系统层面实现进一步优化，这也推动产业竞争从单一硬件，转向全栈解决方案的竞争格局。但与此同时，跨区域部署带来的延迟与一致性问题依然存在，受物理约束影响，短期内只能通过异步训练、梯度压缩及网络优化等方式缓解。
As AI competition shifts from models to infrastructure, the role of hardware transforms accordingly; it is no longer merely a vehicle for computing power, but rather a core variable that influences system efficiency and cost structure. What NVIDIA is doing is consolidating distributed computing power into an industrial-grade capability that is both schedulable and scalable and Rubin serves as a pivotal link in this process.
当AI竞争从模型走向基础设施，硬件的角色也随之改变，不再只是算力载体，而是影响系统效率与成本结构的核心变量。英伟达正在做的，是将分散算力整合为可调度、可规模化运行的工业级能力，而Rubin正是这一进程中的关键一环。
03
Reconstructing the “Five-Layer Architecture” of AI Systems | 重构AI系统的“五层架构”
AI infrastructure exhibits a clearer layered structure: the bottom layer consists of chips and interconnections; the middle layer, systems and networks; and the top layer, models and applications. | AI基础设施呈现出更清晰的分层结构：底层是芯片与互连，中间是系统与网络，上层是模型与应用。
If computing power is merely one component, how is the entire AI system organized? Nvidia’s answer is a “five-layer architecture.”
如果算力只是其中一环，那么整套AI系统如何被组织？ 英伟达给出的答案，是“五层架构”。
Along this ever-expanding path, the question it seeks to answer is not merely the magnitude of computing power, but rather how that power can be utilized more efficiently.
在这条不断延展的路径中，它试图回答的不只是算力强弱，而是算力如何被更高效地使用。
As model scale increases, simply stacking GPUs rarely yields linear improvements; instead, performance becomes constrained by data scheduling, communication latency, and system bottlenecks. Consequently, the competition for computing power has shifted from “single-point performance” to “systemic coordination capabilities.” Task decomposition, data flow, and resource scheduling have emerged as new key variables.
随着模型规模增长，单纯堆叠GPU难以带来线性提升，反而受到数据调度、通信延迟与系统瓶颈的限制。算力竞争因此从“单点性能”转向“系统协同能力”。任务拆分、数据流动与资源调度，成为新的关键变量。
This also gives rise to a more distinct layered structure for AI infrastructure: the bottom layer consists of chips and interconnects; the middle layer, systems and networks; and the top layer, models and applications. NVIDIA is attempting to extend into every layer and integrate them into a cohesive whole through a holistic architecture.
这也让AI基础设施呈现出更清晰的分层结构：底层是芯片与互连，中间是系统与网络，上层是模型与应用。 英伟达正试图向每一层延伸，并通过统一架构将其整合为一个整体。
This structure was summarized by Jensen Huang as a “five-layer cake.” In his view, competition within the AI industry is no longer a battle over a single technology stack, but rather a contest of systems engineering. In the future, the barrier to entry will shift from “whether one possesses computing power” to “whether one possesses system-level capabilities.” Single-point advantages will be diminished, while systemic gaps will be amplified.
这一结构被黄仁勋概括为“五层蛋糕”。在他看来，AI产业的竞争不再是单一技术栈之争，而是系统工程的比拼。未来门槛将从“是否拥有算力”转向“是否具备系统级能力”。单点优势会被削弱，而系统差距将被放大。
From this perspective, what NVIDIA has built is not merely a product portfolio, but an infrastructure paradigm. As computing power is reorganized and scheduled at scale, AI gradually evolves from a modeling problem into an engineering problem.
从这个角度看，英伟达所构建的并不仅是产品组合，而是一种基础设施范式。当算力被重新组织并规模化调度后，AI逐渐从模型问题，演变为工程问题。
On this basis, this “five-layer cake” has been further broken down into a more specific industrial structure.
在此基础上，这一“五层蛋糕”也被进一步拆解为更具体的工业结构。
Jensen Huang posits that, when viewed from an industrial perspective, AI can be understood as a “five-layer stack”: the first layer, the very bottom, is energy. Intelligence is a process continuously shaped through real-time generation, a process that relies on stable electricity. The generation of every token involves electron flow, thermal management, and the conversion of energy into computation. This process involves no abstract space, nor are there any shortcuts. Therefore, energy emerges as the first principle of AI, determining the upper limit of intelligence.
黄仁勋提出，从工业视角审视AI，可以理解为一个“五层栈”：第一层，也是最底层，是能源。 智能是在实时生成中被持续塑造的过程，而这一过程依赖稳定的电力。 每一个Token的生成，都涉及电子流动、热量管理以及能量向计算的转化。 这一过程没有抽象空间，也不存在捷径。 因此，能源成为AI的第一性原理，决定智能的上限。
The second layer are chips. A chip is a factory that transforms energy into compute. As AI drives increasing demand for parallel computing, high-bandwidth memory, and high-speed interconnects, the evolution of chip architecture directly impacts system efficiency and the frontiers of intelligence.
第二层是芯片。 芯片是将能源转化为计算的工厂。 随着AI对并行计算、高带宽内存与高速互联的需求提升，芯片架构的演进直接影响系统效率与智能边界。
The third layer is infrastructure. Encompassing land, power, cooling and networking systems, as well as complex engineering infrastructure, it organizes thousands of processors into a single, collaboratively operating machine. Jensen Huang calls it an “AI factory,” the goal of which is no longer to store information, but to continuously produce intelligence.
第三层是基础设施。 包括土地、电力、冷却与网络系统以及复杂的工程体系，它将成千上万颗处理器组织为一台协同运转的机器。黄仁勋称之为“AI工厂”，其目标不再是存储信息，而是持续生产智能。
The fourth layer is the model. AI is gradually coming to understand more of the “languages of the world”: human languages, biological languages, chemical languages, physical languages, and even financial logic. Language models represent just one category; even more transformative advancements are taking place in protein prediction, chemical simulation, physical modeling, as well as in robotics and autonomous systems.
第四层是模型。 AI正在逐步理解更多“世界语言”：人类语言、生物语言、化学语言、物理语言乃至金融逻辑。 语言模型只是其中一类，更具变革性的进展正在蛋白质预测、化学模拟、物理建模以及机器人和自主系统中发生。
The fifth layer is the application layer. True value is unleashed at this level: drug discovery, industrial robots, legal assistants, autonomous driving, and humanoid robots, among others. Every successful application serves to drive, in a feedback loop, the models, infrastructure, chips, and even the energy system, thereby fostering a continuously reinforcing virtuous cycle.
第五层是应用。 真正的价值在这一层释放：药物研发、工业机器人、法律助手、自动驾驶与人形机器人等。 每一个成功应用，都会反向推动模型、基础设施、芯片乃至能源体系形成持续强化的正向循环。
Gao Chengyuan, a renowned financial writer and President of the Tiaoyuan Influence Academy, points out that computing power, networks, and energy collectively constitute the underlying constraints; the true focal point of competition lies in who can most efficiently integrate these elements and construct a sustainably scalable system architecture. Throughout this process, the importance of full-stack capabilities and systems engineering expertise continues to rise, and the industrial landscape will be reshaped around the “right to define the system.”
知名财经作家、眺远影响力研究院院长高承远则指出，算力、网络与能源共同构成底层约束，真正的竞争焦点在于谁能够更高效地整合这些要素，并构建可持续扩展的系统架构。 在这一过程中，全栈能力与系统工程能力的重要性持续上升，产业格局也将围绕“系统定义权”展开重塑。
From a long-term perspective, NVIDIA’s advantage within its full-stack strategy is reflected primarily in its capabilities for hardware-software synergy and its developer ecosystem. Its CUDA platform and AI software ecosystem are deeply integrated with its GPUs and networking hardware, forming a complete technical closed loop spanning from chips to systems and applications; this not only delivers performance advantages but also establishes significant ecosystem barriers. Today, AI infrastructure is evolving toward heterogeneity, distribution, and sustainability. The supply of computing power is becoming more diversified, and network architectures are placing greater emphasis on low latency and high bandwidth, while energy efficiency and system coordination capabilities are set to become the critical variables in future competition.
从长期来看，英伟达在全栈布局中的优势，更多体现在软硬件协同能力与开发者生态之上。 其CUDA平台与AI软件体系，与GPU及网络硬件深度绑定，形成从芯片到系统再到应用的完整技术闭环，不仅带来性能优势，也构建了较高的生态壁垒。 如今，AI基础设施正在走向异构化、分布式与绿色化，算力供给更加多元，网络架构更加注重低延迟与高带宽，而能源效率与系统协同能力，将成为未来竞争的关键变量。
As AI competition shifts from model capabilities to infrastructure, computing power is no longer the end goal, but rather a component of the overall system. NVIDIA is attempting to define the standards for this ecosystem through chips, networking, and systems by reconfiguring distributed computing power into sustainably operating “intelligent factories.”
当AI竞争从模型能力转向基础设施，算力不再是终点，而是体系的一部分。 英伟达正试图通过芯片、网络与系统，定义这一体系的标准，将分散算力重构为可持续运行的“智能工厂”。
However, in a phase where inference has become dominant, and cost and efficiency have emerged as core variables, the limitations of general-purpose architectures are becoming apparent. An advantage in computing power no longer automatically translates into an advantage in efficiency. The key to this competition is no longer “who is stronger,” but “who can more efficiently define infrastructure.”
但在推理成为主导、成本与效率成为核心变量的阶段，通用架构的边界正在显现。 算力优势，不再自动等同于效率优势。这场竞争的关键，已不在于“谁更强”，而在于“谁更高效地定义基础设施”。
I use AI tools to support my editing, research, and translation process. Learn more on my AI Transparency Page.