North America Equity Research 28 July 2025 Hardware & Networking Scale-Up AI Networking in Numbers: Key Forecasts from 650 Group for Scale-Up AI and Technology of Choice We recently hosted a conference call with Alan Weckel of 650 Group and Harlan Sur, J.P. Morgan's Semi & Semi Cap Equipment / IT Hardware analyst, to discuss Scale-up AI Networking. Please see Harlan Sur's report here on the takeaways from the call on Scale-up AI Networking. Below we walk through the key forecasts in relation to AI Networkin ...

Core Viewpoint - InfiniBand has been a dominant structure for high-performance computing (HPC) and AI applications, but its market position is challenged by Broadcom's new low-latency Ethernet switch, Tomahawk Ultra, which aims to replace InfiniBand and NVSwitch in AI and HPC clusters [3][5][26]. Group 1: InfiniBand and Its Evolution - InfiniBand gained traction due to Remote Direct Memory Access (RDMA), allowing direct memory access between CPUs, GPUs, and other processing units, which is crucial for AI model training [3]. - Nvidia's acquisition of Mellanox Technologies for $6.9 billion was driven by the anticipated growth of generative AI, necessitating InfiniBand for GPU server connectivity [3][4]. - The rise of large language models and generative AI has propelled InfiniBand to new heights, with NVLink and NVSwitch providing significant advantages for AI server nodes [4]. Group 2: Broadcom's Tomahawk Ultra - Broadcom's Tomahawk Ultra aims to replace InfiniBand as the backend network for HPC and AI clusters, offering low-latency and lossless Ethernet capabilities [5][6]. - The development of Tomahawk Ultra predates the rise of generative AI, targeting applications sensitive to latency [5]. - Tomahawk Ultra's architecture allows for shared memory clusters, enhancing communication speed among processing units compared to traditional InfiniBand or Ethernet [5][6]. Group 3: Performance Metrics - InfiniBand's packet size typically ranges from 256 B to 2 KB, while Ethernet switches often handle larger packets, impacting performance in HPC workloads [7]. - InfiniBand has historically demonstrated lower latency compared to Ethernet, with significant improvements in latency metrics over the years, such as 130 nanoseconds for 200 Gb/s HDR InfiniBand [10][11]. - Broadcom's Tomahawk Ultra boasts a port-to-port jump latency of 250 nanoseconds and a throughput of 77 billion packets per second, outperforming traditional Ethernet switches [12][28]. Group 4: Competitive Landscape - InfiniBand's advantages in latency and packet throughput have made it a preferred choice for HPC workloads, but Ethernet technologies are rapidly evolving to close the gap [6][10]. - Nvidia's NVSwitch is also under threat from Broadcom's Tomahawk Ultra, which is part of a broader strategy to enhance Ethernet capabilities for AI and HPC applications [26][29]. - The introduction of optimized Ethernet headers and lossless features in Tomahawk Ultra aims to improve performance and compatibility with existing standards [15][16].

Core Insights - Broadcom's Tomahawk Ultra Ethernet switch is a significant driver for its AI revenue growth, with a reported 46% year-over-year increase in AI revenues for Q2 FY25 [1][10] Group 1: Product and Technology - The Tomahawk Ultra chip is designed to meet the increasing demand for high-bandwidth, low-latency interconnects necessary for scaling AI workloads [2] - It boasts impressive specifications, including a throughput of 51.2 terabits per second, the ability to handle 77 billion packets per second, and ultra-low latency of 250 nanoseconds, which are essential for connecting numerous GPUs and AI accelerators [3] - The chip supports up to 1,024 accelerator connections, providing a competitive advantage over Nvidia's NVLink [3] - Its pin compatibility with the previous generation (Tomahawk 5) facilitates easier upgrades for customers, promoting faster adoption among hyperscalers and OEMs [4] Group 2: Market Position and Competition - Broadcom is positioning itself as a major player in AI infrastructure by integrating Tomahawk Ultra with its broader portfolio, including Jericho routers and custom AI chips [5] - The company faces competition from NVIDIA, which leads the AI chip market with its GPU performance and CUDA-NVLink ecosystem, benefiting from strong demand from cloud and internet giants [6] - Advanced Micro Devices (AMD) is also a competitor, supplying high-performance EPYC CPUs and Instinct GPUs to major hyperscalers, with a reported 57% year-over-year increase in AI data center revenues [7] Group 3: Financial Performance and Valuation - Broadcom's shares have returned 20.9% year-to-date, outperforming the Zacks Electronics - Semiconductors industry, which grew by 15.6% [8] - The company is currently trading at a forward price-to-sales ratio of 18.41X, which is significantly higher than the industry average of 8.75X, indicating a potentially overvalued position [11] - The Zacks Consensus Estimate for Q3 FY25 earnings is $1.66 per share, reflecting a 33.87% year-over-year growth, while the estimate for fiscal 2025 earnings is $6.64 per share, suggesting a 36.34% growth compared to fiscal 2024 [14][15]

Summary of NVIDIA's Q1 Fiscal 2026 Conference Call Company Overview - **Company**: NVIDIA - **Fiscal Quarter**: Q1 of Fiscal 2026 - **Date of Call**: May 28, 2025 Key Industry Insights - **Data Center Revenue**: Reached $39 billion, a 73% year-on-year growth driven by AI workloads transitioning to inference and AI factory build-outs [2][3] - **Export Controls Impact**: New U.S. export controls on the H20 GPU, specifically designed for the China market, resulted in a $4.5 billion inventory write-down and a loss of $2.5 billion in potential revenue for Q1 [2][13] - **China Market**: The loss of access to the China AI accelerator market, projected to grow to nearly $50 billion, poses a significant risk to NVIDIA's business [2][19] Financial Performance - **Revenue Breakdown**: - Q1 recognized $4.6 billion in H20 revenue before export controls [2] - Anticipated total revenue for Q2 is $45 billion, with a significant decline in China data center revenue expected [11][12] - **Gross Margins**: GAAP gross margin at 60.5%, non-GAAP at 61%. Excluding the $4.5 billion charge, non-GAAP gross margins would have been 71.3% [11] - **Shareholder Returns**: NVIDIA returned a record $14.3 billion to shareholders through share repurchases and dividends [11] Product and Technology Developments - **Blackwell Architecture**: Contributed nearly 70% of data center compute revenue, with significant improvements in manufacturing yields and ramp-up rates [3][4] - **Inference Demand**: Strong demand for inference, with Microsoft processing over 100 trillion tokens in Q1, a five-fold increase year-on-year [4] - **AI Factory Deployments**: Nearly 100 NVIDIA-powered AI factories in operation, doubling year-on-year, with significant growth in GPU usage per factory [5] Strategic Partnerships and Market Position - **Collaborations**: Partnerships with major companies like Microsoft, OpenAI, and Yum Brands to enhance AI capabilities across various sectors [6][10] - **Networking Solutions**: Revenue from networking grew 64% quarter-over-quarter to $5 billion, with significant adoption of Spectrum X among major cloud service providers [7][28] Future Outlook - **Guidance for Q2**: Expected revenue decline in China data center revenue, with a loss of approximately $8 billion in H20 revenue anticipated [11][18] - **Long-term Growth**: NVIDIA's roadmap extends through 2028, with a focus on AI infrastructure, enterprise AI, and industrial AI [4][30] - **AI as Infrastructure**: The company emphasizes the importance of AI as essential infrastructure, similar to electricity and the internet, with a significant build-out expected globally [22][25] Additional Insights - **Export Control Concerns**: The U.S. export restrictions are seen as detrimental to American competitiveness in the global AI market, potentially benefiting foreign competitors [13] - **Emerging AI Technologies**: The introduction of reasoning AI models is driving a surge in inference demand, with significant implications for compute requirements [14][19] - **Investment in Manufacturing**: NVIDIA is investing in onshore manufacturing capabilities to strengthen its supply chain and support AI infrastructure development [15][26] This summary encapsulates the critical points discussed during NVIDIA's Q1 Fiscal 2026 conference call, highlighting the company's performance, strategic direction, and the broader implications for the AI industry.

Investment Rating - The report does not explicitly state an investment rating for the industry or specific companies within it Core Insights - Capital expenditure (Capex) for AI infrastructure is expected to remain strong, driven by tier-2 clouds and large enterprises diversifying their spending on data center equipment [2][3] - The transition to 800G networking is anticipated to dominate through 2026, with Ethernet potentially scaling up in networking opportunities [2][5] - Data center equipment spending is diversifying across more cloud players, with a forecasted growth rate of approximately 18% CAGR from 2024 to 2029, reaching around $925 billion [3] Summary by Sections Data Center Equipment Spending - Cloud Provider capex is projected to grow significantly, with more cloud players expected to spend over $5 billion annually, enhancing vendor customer base diversity [3] - Hyperscalers are likely to continue favoring best-of-breed networking providers, while tier-2 cloud providers will prefer full-stack solutions [3] Networking Trends - The transition to 800G is ongoing, with AI players at various stages, and the revenue recognition timeline for networking vendors has lengthened from 1-2 quarters to 2-4 quarters [5] - Vendor diversity for switching silicon is becoming increasingly important for AI workloads, with Cisco and Juniper expected to gain market share [5] Market Dynamics - The report highlights relative strength in Asia and Middle Eastern sovereign AI initiatives, indicating a robust market for data center equipment [3] - Concerns regarding data center switching market share, particularly for Arista, are viewed as somewhat exaggerated [2]

Core Viewpoint - The article discusses the competitive landscape in the GPU and high-performance computing (HPC) market, focusing on the advancements of Broadcom's Tomahawk Ultra technology compared to Nvidia's NVLink and the emerging UALink protocol [3][4][5]. Group 1: Technology Comparison - AMD and other chip suppliers are narrowing the performance gap with Nvidia in terms of GPU FLOPS, memory bandwidth, and HBM capacity, but lack high-speed interconnects like NVLink and NVSwitch, limiting their scalability [3]. - Broadcom is promoting its Scale-up Ethernet (SUE) technology, claiming it can support systems with at least 1024 accelerators on any Ethernet platform, while Nvidia's NVLink can support up to 576 accelerators [4][5]. - The Tomahawk Ultra switch from Broadcom offers a bandwidth of 51.2 Tbps, which is higher than Nvidia's 28.8 Tbps for its fifth-generation NVLink switch, allowing for a vertical scaling architecture with 128 accelerators [7]. Group 2: Performance and Features - Tomahawk Ultra is designed for low latency, achieving as low as 250 nanoseconds, and is optimized for handling smaller data packets common in HPC systems [6]. - The switch includes congestion control mechanisms and supports collective operations, enhancing network efficiency compared to Nvidia's NVLink [6][7]. - Broadcom's Tomahawk Ultra ASIC has begun shipping to customers, and its compatibility with existing switch chassis is expected to facilitate adoption [7]. Group 3: Market Dynamics - The UALink protocol, while still in development, is being integrated into AMD's Helios rack systems, which will utilize both UALink and Ethernet for their expansion architecture [9]. - There are concerns regarding the feasibility of achieving UALink's target latency of 100-150 nanoseconds when transmitted over Ethernet, which may hinder AMD's competitive position against Nvidia's advanced systems [10].

Core Insights - NVIDIA Corporation's networking business is experiencing significant growth, with revenues increasing 64% sequentially to approximately $5 billion in Q1 FY26, driven by AI infrastructure demand [2][11] - The company's full-stack approach, including high-performance connectivity solutions, is enhancing its ability to support complex AI workloads [3][6] - New products like NVLink Fusion and the expansion of the Spectrum-X Ethernet networking platform are attracting major cloud providers, indicating strong early adoption and potential for future revenue growth [4][5][11] Networking Business Performance - Networking revenues surged to about $5 billion in Q1 FY26, marking a 64% sequential increase [2][11] - NVLink shipments exceeded $1 billion in the last quarter, showcasing strong early adoption of new offerings [4][11] - Spectrum-X is projected to generate over $8 billion in annualized revenues, with deployments from major companies like Microsoft, Meta, Oracle, and Google Cloud [5][11] Competitive Landscape - Broadcom and Marvell Technology are key competitors in the AI networking space, but both lack the comprehensive end-to-end solutions offered by NVIDIA [7][8][9] - Broadcom focuses on custom networking chips and high-speed interconnects but relies on partnerships rather than a unified platform [8] - Marvell Technology is expanding its presence in data center connectivity but also lacks the software and system-level integration that NVIDIA provides [9] Valuation and Earnings Estimates - NVIDIA shares have increased approximately 22.2% year-to-date, outperforming the Zacks Computer and Technology sector's gain of 7.4% [10] - The company trades at a forward price-to-earnings ratio of 33.81, higher than the sector average of 27.39 [12] - The Zacks Consensus Estimate for NVIDIA's fiscal 2026 earnings indicates a year-over-year increase of about 41.8%, while fiscal 2027 estimates show a 31.9% increase [14]

Group 1 - Nvidia is planning to launch an AI chip specifically for the Chinese market as early as September, despite U.S. export restrictions [1] - The new AI chip will be based on Nvidia's Blackwell RTX Pro 6000 processor, modified to comply with existing AI chip restrictions [2] - The upcoming chips will not include high-bandwidth memory or NVLink, which are features of Nvidia's advanced AI chips [2] Group 2 - Nvidia CEO Jensen Huang stated that the company would no longer include the Chinese market in its revenue and profit forecasts, although this may be a temporary change [3]

Core Insights - NVIDIA Corporation is experiencing significant sales losses due to U.S. export restrictions, particularly from the trade conflict with China, resulting in a $2.5 billion loss in Q1 of fiscal 2026 and an anticipated $8 billion loss in Q2 [1][10] - The company is focusing on sovereign AI projects to mitigate risks associated with the trade war, establishing partnerships and infrastructure in various regions [1][5] Group 1: Sovereign AI Initiatives - Sovereign AI involves countries developing their own AI infrastructure to protect data and maintain control, with NVIDIA playing a crucial role in these efforts [2] - NVIDIA has announced major AI factory projects in Saudi Arabia, planning to deliver 18,000 AI chips, and has received strong support for its initiatives in Europe from leaders in France and Germany [2][3] - The European Union is investing $20 billion to build four AI gigafactories, and NVIDIA is collaborating with firms like Foxconn and local governments in Taiwan and the UAE to enhance AI infrastructure [3] Group 2: Competitive Landscape - Competitors like Advanced Micro Devices (AMD) and Intel are also trying to expand in the AI infrastructure space but are lagging behind NVIDIA in sovereign AI projects [6] - AMD is gaining traction with its Instinct MI300X chips, but lacks a full-stack solution comparable to NVIDIA's offerings [7] - Intel is promoting its Gaudi 3 AI chips as a cost-effective option, but its AI ecosystem is still developing and does not provide a complete package [8] Group 3: Financial Performance and Valuation - NVIDIA's shares have increased by approximately 19.1% year-to-date, outperforming the Zacks Computer and Technology sector, which gained 7% [9] - The company trades at a forward price-to-earnings ratio of 33.13, higher than the sector average of 27.37 [11] - The Zacks Consensus Estimate for NVIDIA's fiscal 2026 and 2027 earnings indicates a year-over-year increase of about 41.8% and 31.9%, respectively, with recent upward revisions for fiscal 2027 estimates [12]

Summary of U.S. Semiconductors & Semiconductor Capital Equipment Conference Call Industry Overview - The conference call focused on the U.S. Semiconductors and Semiconductor Capital Equipment industry, particularly the competition among scale-up technologies: UALink (UAL), Scale Up Ethernet (SUE), and NVLink [1][2][3]. Core Points and Arguments - **Importance of Interconnects**: Interconnect technology is critical for the success of XPU (cross-processor unit) efforts, especially in scale-up designs where competition is intense among UAL, SUE, and NVLink [1]. - **Current Adoption**: Most hyperscalers currently utilize NVDA's NVLink for AI deployments, with some using Ethernet and PCIe for ASIC programs [3]. - **Future Decisions**: Hyperscalers and Tier 2 companies must decide on chip designs for future volumes by 2027, with NVLink being a dominant choice due to its proven deployment [3]. - **Technology Development**: UAL was ratified in April 2025, with the first design expected to be implemented with AMD Helios in mid-2026. However, readiness issues may push this to 2027 [6]. - **Switching Partners**: UAL has limited backing from established switch vendors, raising concerns about its adoption outside AMD. SUE, led by AVGO, has a more established technology portfolio [6][8]. - **Latency and Performance**: UAL claims to offer lower latency and an open ecosystem, while SUE is based on standard Ethernet but has higher latency. NVLink is noted for its low latency and proven reliability [8][10]. Key Comparisons - **Latency**: - UAL: <1 microsecond RTT - SUE: <2 microseconds RTT - NVLink: ~0.3 microseconds RTT [11]. - **Architecture**: - UAL uses a custom protocol stack optimized with PCIe technology. - SUE is based on Ethernet MAC/packet-based architecture. - NVLink employs a proprietary stack from NVDA [11]. - **Availability**: NVLink is currently available, while UAL and SUE are expected to be broadly available by late 2026/2027 [9]. Additional Insights - **Ecosystem Considerations**: The success of these technologies will depend on the availability of interconnect vendors. UAL has support from ALAB and MRVL, while SUE is backed by AVGO [20]. - **Future Specifications**: UAL plans to release a 128G Specification in July 2025, which will utilize a PCIe-based PHY, enhancing its capabilities [24]. - **Market Dynamics**: The competition among UAL, SUE, and NVLink will focus on reliability, latency, bandwidth, and power efficiency, with each technology having its advantages and disadvantages [7][10]. Conclusion - The U.S. semiconductor industry is at a pivotal point with the emergence of new interconnect technologies. The competition among UAL, SUE, and NVLink will shape the future of AI and high-performance computing, with significant implications for hyperscalers and semiconductor companies alike. The readiness and adoption of these technologies will be crucial for maintaining competitive advantages in the market [2][3][6].