2026 年 3 月 2 日 行业研究 GTC 大会前瞻：重视 LPU 对 PCB 设备和钻针带来的增量需求 ——PCB 设备系列跟踪报告（三） 要点 事件：据新浪财经报道，英伟达计划在 2026 年 3 月的 GTC 开发者大会上发布 一款整合了 Groq"语言处理单元"（LPU）技术的全新推理芯片。 LPU 具有低时延高带宽特点，与 GPU 在 AI 工作流中形成互补。 LPU（Language Processing Unit）是一种专为 AI 推理，特别是低延迟实时交互（如对话）设计 的专用处理器，其核心是通过"编译器驱动"的静态调度实现确定性执行，并依 赖高速片上 SRAM（带宽可达 80TB/s）来消除内存瓶颈，从而将首词延迟降至 约百毫秒内，在主流大模型（以 Llama2-70B 模型为例）推理上比 H100 GPU 快约 10 倍，综合能效可提升约 10 倍。相比之下，GPU（如 H100）是通用高吞 吐架构，依赖大容量 HBM 显存，擅长大规模并行计算，是大模型训练和高吞吐 量任务的主力，但在单序列、实时生成的场景中会受限于内存带宽和运行时调度， 难以突破低延迟瓶颈。LPU 与 GPU 在 A ...

Core Insights - Nvidia is shifting the AI computing competition focus from training to inference, with plans to unveil a new inference chip integrated with Groq's LPU technology at the upcoming GTC developer conference [1] - OpenAI has agreed to become a major customer for Nvidia's new processor, indicating a strong demand for dedicated inference capacity [1] - The report from Shenwan Hongyuan highlights four key trends in inference computing: increased deployment of pure CPU scenarios, the rise of specialized architectures like LPU, accelerated breakthroughs in domestic computing chips, and a shift in demand structure towards mass token consumption [2] Inference Demand Explosion - The demand for inference has surged, driven by the monetization of large models and the rapid deployment of agents in real-world applications, requiring substantial inference computing power [3] - Data shows a significant increase in inference volume during the Chinese New Year, with major models reaching record token consumption [3] LPU's Emergence - Nvidia's acquisition of Groq's core technology for $20 billion signifies the growing importance of pure inference chips, with LPU architecture offering efficiency advantages in inference scenarios [6] - The future AI chip landscape is expected to differentiate between training and inference, with training continuing to use GPU-HBM combinations while inference evolves towards ASIC+LPU-SRAM+SSD configurations [6] System-Level Innovations - The upgrade in inference computing also involves a shift from single chips to system-level innovations, with a three-layer network architecture emerging to meet the demands of low latency and high throughput [7] - Nvidia is expanding its collaboration with Meta Platforms to support large-scale pure CPU deployments, moving beyond a single GPU sales model [7] Domestic Chip Breakthroughs - Domestic inference chips are experiencing significant technological upgrades, with new designs supporting low-precision data formats and enhanced interconnect bandwidth [9] - The supply chain for domestic chips is also improving, as evidenced by the rapid growth in revenue from high-performance computing chip packaging services [9]

英伟达整合LPU（语言处理单元）技术、OpenAI多线押注推理芯片，正在将AI算力竞争的主战场从训练切换至推理。申万宏源研究认为，2026年 算力产业的核心关键词将是推理，Token消耗总量与技术范式均将围绕这一主题深度重构。 2月28日，据《华尔街日报》报道，英伟达计划在下月的GTC开发者大会上发布一款整合了Groq"语言处理单元"（LPU）技术的全新推理芯片，英 伟达首席执行官黄仁勋称其为"世界从未见过"的全新系统。OpenAI已同意成为该处理器的最大客户之一，并将向英伟达购买大规模"专用推理产 能"。 与此同时，OpenAI上月还与初创公司Cerebras达成数十亿美元计算合作，后者称其推理芯片速度已超越英伟达GPU（图形处理器）。这一系列动 向表明，AI巨头正在从训练算力的军备竞赛，转向推理算力的多线布局。 申万宏源报告指出，Token经济时代，推理算力正迎来四大趋势：一是纯CPU（中央处理器）部署场景增多，低成本推理需求加速算力下沉；二 是LPU等专用架构崛起，挑战GPU在推理环节的主导地位；三是国产算力芯片加速突破，供应链多元化趋势明确；四是推理算力的需求结构从"单 次训练"向"海量Token消耗 ...

英伟达凭借 Hopper 和 Blackwell 架构绝对主导了 AI 模型训练市场，并通过 Rubin CPX 架构的注意力加速引擎，覆盖了推理的"预填充"（Prefill）阶段，但 在对延迟极度敏感的"解码"（Decode）环节，公司亟需引入 Groq 的技术来确立行业标杆。 战略布局方面，黄仁勋强调 Groq 将补齐 AI 推理阶段的短板，实现超低延迟的解码能力。AI 行业目前正加速迈向多智能体协同（Agentic AI）时代，应用层 要求极低的延迟和超快的响应速度。 IT之家 2 月 27 日消息，科技媒体 Wccftech 昨日（2 月 26 日）发布博文，报道称在 2026 财年第 4 财季（截至 2026 年 1 月）财报会议上，英伟达 CEO 黄仁 勋透露了收购 Groq 后的核心整合计划。 技术实现方面，英伟达希望全面释放 Groq 的硬件潜力。Groq 的语言处理单元（LPU）采用片上 SRAM（静态随机存取存储器），能够提供每秒数十 TB 的 内部超高带宽。 重要性方面，黄仁勋将此次价值 200 亿美元（IT之家注：现汇率约合 1370.47 亿元人民币）的非授权收购，和当年收购 Mel ...

Core Viewpoint - Nvidia announced a significant deal worth $20 billion to acquire technology licensing from AI chip startup Groq, marking its largest transaction in history, comparable to the total of all previous acquisitions [1][3]. Group 1: Transaction Structure - The deal is structured as a non-exclusive technology licensing agreement rather than a full acquisition, which is a strategic move to avoid antitrust scrutiny [3][4]. - Nvidia's market capitalization is approaching $3.5 trillion, making it a target for regulatory oversight on major actions [4][6]. Group 2: Strategic Rationale - The $20 billion investment not only secures technology but also the expertise and patents of Groq's team, particularly its founder, a key figure in AI chip architecture [6][8]. - By attracting Groq's talent, Nvidia effectively removes a critical competitor from the market while gaining access to advanced technology [8][22]. Group 3: Technology Insights - Groq's core product, the Language Processing Unit (LPU), is designed specifically for AI inference, distinguishing it from Nvidia's GPUs, which dominate the training market [9][11]. - Groq claims its LPU offers significantly faster inference speeds and lower costs compared to Nvidia's H100, which could disrupt Nvidia's current market position [11][13]. Group 4: Competitive Landscape - The AI chip market is becoming increasingly competitive, with major players like Google, Amazon, and AMD aggressively pursuing market share in inference technology [19][27]. - Nvidia's acquisition of Groq can be seen as a strategic insurance policy to maintain its competitive edge in the evolving AI landscape [22][29]. Group 5: Market Implications - The integration of Groq's LPU technology into Nvidia's existing product line could enhance its distribution capabilities and accelerate market penetration [25][27]. - This transaction reflects Nvidia's urgency to adapt to a rapidly changing market where it faces significant competition, indicating a shift in the AI chip industry dynamics [27][29].

Core Viewpoint - Nvidia's acquisition of Groq's technology and talent for $20 billion raises questions about the strategic rationale behind the deal, especially given the potential for antitrust scrutiny and the actual benefits derived from Groq's technology [1][2]. Group 1: Nvidia's Acquisition Details - Nvidia paid $20 billion for a non-exclusive license of Groq's intellectual property, including its Language Processing Unit (LPU) and associated software libraries [2]. - Groq will continue to operate independently, retaining its high-performance inference-as-a-service product, despite significant talent loss to Nvidia [2]. - The acquisition is seen as a move to eliminate competition, but the justification for the $20 billion price tag remains debatable [2]. Group 2: Technology Insights - Groq's LPU utilizes Static Random Access Memory (SRAM), which is significantly faster than the High Bandwidth Memory (HBM) used in current GPUs, potentially offering 10 to 80 times the speed [3]. - Groq's chip achieved a token generation speed of 350 tok/s in tests, and even higher at 465 tok/s when running mixed expert models [3]. - However, SRAM's low space efficiency means that running medium-sized language models would require hundreds or thousands of Groq's LPUs, raising questions about its practicality [4]. Group 3: Architectural Innovations - The key innovation from Groq is its "dataflow architecture," designed to accelerate linear algebra operations during inference, which could provide Nvidia with a competitive edge in chip performance [5][6]. - This architecture allows for continuous processing of data without waiting for memory, potentially overcoming bottlenecks that slow down GPU performance [6][7]. - Groq's LPU can theoretically achieve performance levels comparable to high-end GPUs, but practical performance may vary [7]. Group 4: Future Implications - Nvidia's collaboration with Groq could lead to new technology options for enhancing chip performance, particularly in inference optimization, an area where Nvidia has previously lacked a strong offering [8]. - The upcoming Rubin series chips from Nvidia are designed to optimize the inference pipeline, indicating a shift in architecture that could leverage Groq's technology [9]. - Groq's existing chip designs may not serve as excellent decoders, but they could be useful for speculative decoding, which enhances performance by predicting outputs from smaller models [9]. Group 5: Market Context - The $20 billion price tag for Groq's technology is substantial but manageable for Nvidia, given its recent operating cash flow of $23 billion [10]. - The acquisition may not immediately impact Nvidia's current chip production, as the company could be positioning itself for long-term strategic advantages [12].

Core Viewpoint - Wall Street analysts are optimistic about NVIDIA's acquisition of AI inference chip company Groq, viewing it as a strategic move that combines both offensive and defensive elements [1][4][7] Group 1: Acquisition Details - NVIDIA has signed a non-exclusive licensing agreement with Groq, allowing NVIDIA to use Groq's inference technology, with Groq's key personnel joining NVIDIA to enhance the implementation of this technology [3][4] - The acquisition is valued at approximately $20 billion, focusing on Groq's intellectual property and talent [3][4] Group 2: Analyst Ratings - Cantor has reiterated NVIDIA as a "preferred stock," maintaining a "buy" rating with a target price of $300, emphasizing the dual strategic significance of the acquisition [4][5] - Bank of America has also maintained a "buy" rating for NVIDIA with a target price of $275, acknowledging the high cost of the acquisition but recognizing its strategic value [6][7] Group 3: Strategic Implications - The acquisition is seen as a way for NVIDIA to convert potential threats from ASIC technology into competitive advantages, thereby strengthening its market position in AI infrastructure, particularly in real-time workloads like robotics and autonomous driving [5][10] - Analysts highlight that Groq's low-latency, high-efficiency inference technology will be integrated into NVIDIA's complete system stack, potentially enhancing compatibility with CUDA and expanding NVIDIA's share in the inference market [5][10] Group 4: Groq's Background and Technology - Groq, founded in 2016 by Jonathan Ross, a key developer of Google's TPU, focuses on AI inference chips and has developed a language processing unit (LPU) that significantly outperforms NVIDIA's GPUs in inference speed [10][11] - Groq's partnerships with major companies like Meta and IBM, as well as its involvement in the U.S. government's "Genesis Project," position it as a strong competitor in the AI chip market [11]

Core Insights - Groq has entered into a $20 billion technology licensing agreement with NVIDIA, which is not a legal acquisition but a non-exclusive technology licensing deal allowing NVIDIA to use Groq's hardware and architecture designs [2][3] - Groq's CEO Jonathan Ross and nearly all key members will join NVIDIA, while Groq will continue to operate as an independent company, retaining its core intellectual property [2] - The deal's significance extends beyond the monetary value, as it reflects NVIDIA's strategic positioning in the AI landscape amid regulatory challenges and market pressures [3][28] Group 1: Transaction Details - The transaction is valued at approximately $20 billion, which is enough to acquire GlobalFoundries entirely or represents a quarter of Intel's market value [3] - NVIDIA will acquire all of Groq's physical assets but not its intellectual property, indicating a focus on talent acquisition rather than a traditional acquisition [2] - Groq has raised a total of $1.8 billion in funding, with significant investments from entities like the Saudi sovereign wealth fund [5][27] Group 2: Groq's Technology and Market Position - Groq's core product, initially named TSP and later LPU, utilizes a unique architecture with 144-wide VLIW design, offering advantages in speed and efficiency [5][9] - The architecture's reliance on on-chip SRAM instead of external memory allows for fast access speeds but limits storage capacity, posing challenges for deploying larger AI models [6][7] - Groq's architecture is distinct from traditional ASICs and TPUs, focusing on deterministic system behavior and low latency, which are appealing for real-time inference scenarios [10][11] Group 3: Industry Context and Strategic Implications - The deal is seen as a strategic move by NVIDIA to solidify its position in the AI infrastructure market, especially in light of increasing regulatory scrutiny and competition [28][31] - The transaction may also serve as a means for NVIDIA to gain favor with U.S. and Middle Eastern stakeholders, potentially easing export restrictions on AI products [28][30] - The broader context includes a trend of large tech companies engaging in high-value agreements with promising startups to secure technology and talent without formal acquisitions [26][27]

Core Insights - Groq, an AI inference chip startup founded in 2016, has entered a non-exclusive licensing agreement with Nvidia, where Nvidia pays approximately $20 billion for Groq's core AI inference technology and related assets [2][5] - Groq's technology is seen as a significant competitor to Nvidia's GPUs, particularly in the AI inference market, where Groq claims its chips can achieve up to 10 times the inference speed compared to Nvidia's offerings [1][5] - The transaction reflects a growing trend among tech giants to utilize "quasi-acquisitions" to acquire technology and talent while avoiding full ownership and regulatory scrutiny [4][5] Company Overview - Groq was founded by Jonathan Ross, a key member of Google's TPU project, to address inefficiencies in traditional computing architectures for modern AI tasks [1] - The company has recently partnered with major firms like Meta and IBM to enhance its AI inference capabilities [3] Financial Aspects - The $20 billion deal significantly exceeds Groq's previous valuation of $6.9 billion, indicating a strong market interest in its technology [7][8] - Groq's recent revenue forecast was lowered by approximately 75%, highlighting challenges in scaling its operations and the competitive landscape [7] Strategic Implications - Nvidia aims to integrate Groq's low-latency processors into its AI factory architecture to enhance its platform capabilities for AI inference and real-time workloads [3][5] - The acquisition strategy allows Nvidia to strengthen its position in the AI inference market while maintaining Groq's operational independence, which could lead to faster commercialization of Groq's technology [8]

Core Insights - Groq, an AI chip startup, has raised $750 million in funding, exceeding the initial expectation of $600 million, bringing its valuation to $6.9 billion [1][4][5] - The company's valuation has more than doubled in one year, from $2.8 billion to $6.9 billion [2][4][5] - Groq's CEO, Jonathan Ross, emphasizes the importance of inference in the current AI era and the company's goal to build infrastructure for high-speed, low-cost delivery [3][4] Funding and Valuation - The recent funding round was led by Disruptive, with significant investments from BlackRock, Luminus Management, and Deutsche Telekom Capital Partners, among others [6][9] - Groq has raised over $3 billion in total funding to date [6][9] Company Strategy and Operations - Groq plans to use the new funds to expand its data center capacity, including announcing its first Asia-Pacific data center location this year [7][9] - The company has received requests from clients for higher capacity that it currently cannot meet [8] Product and Technology - Groq is known for producing AI inference chips optimized for pre-trained models, with a founding team that includes many former Google TPU engineers [9][10] - The company has developed the world's first Language Processing Unit (LPU) and refers to its hardware as "inference engines," designed for efficient AI model operation [12] - Groq claims its inference acceleration solution is ten times faster than NVIDIA's GPUs while reducing costs to one-tenth [14]