PWW

Portfolio AI Insights

pww.comCisco Systems, Inc., New Street Research LLP - Special Call - SEHK:4333

SEHK:4333

Ahmed Sami Badri [Head of Investor Relations] 💬

Ahmed Sami Badri, the Head of Investor Relations at Cisco, made the following statements during the special call:

  1. Introduction:

    • Greeted attendees and thanked them for joining the tech talk on Silicon One and AI.
    • Provided his contact information for any follow-up questions.

  2. Housekeeping Items:

    • Mentioned that the Cisco team would present a slide deck, followed by questions from analyst moderator Pierre Ferragu of New Street Research.
    • Instructed attendees to submit questions through the Q&A chat box in the Webex application.

  3. Safe Harbor Statement:

    • Read out a disclaimer regarding the forward-looking statements made during the presentation.
    • Noted that actual results may differ materially from projections due to risks detailed in Cisco's SEC filings.

  4. Handover to Pierre Ferragu:

    • Handed over the call to Pierre Ferragu of New Street Research to introduce himself and the Cisco executives.

  5. Closing Remarks:

    • Thanked the speakers (Eyal Dagan and Rakesh Chopra) and attendees for participating in the call.

    • Provided contact information for further questions through the Cisco Investor Relations portal.

    • Offered to address unanswered questions on a case-by-case basis after the call.

Rakesh Chopra [Fellow of Common Hardware Group] 💬

Rakesh Chopra, a Fellow of the Common Hardware Group at Cisco, provided extensive insights during the special call. Here’s a detailed summary of his contributions:

Introduction

  • Role and Excitement: Rakesh expressed excitement about discussing Cisco Silicon One and its relevance in the context of the rise of artificial intelligence (AI).
  • Acknowledgment: He thanked Pierre Ferragu, Sami Badri, and Heather Whitfield for their introductions.

Presentation Overview

  • Historical Context: Rakesh highlighted the realization that the industry had been making the same mistakes repeatedly, leading to the development of Cisco Silicon One.
  • Approach: To address these issues, Cisco created a new organization with a focus on building a single silicon architecture that can be used across different network segments and business models.
  • Technology Innovations:
    • Data Plane Architecture: Developed a new data plane architecture using slices containing packet processors, Ethernet ports, and SerDes.
    • Packet Processor: Created a run-to-completion, purpose-built engine coupled with hardware-accelerated databases.
    • Memory Architecture: Designed a new memory architecture to unify packet buffers, improving burst absorption and reducing power consumption.
  • Scalability: Each component of the architecture is linearly scalable, enabling continuous improvement.
  • Portfolio Expansion: Cisco Silicon One’s portfolio has expanded from service provider cores to web-scale data centers, covering various network segments.
  • Customer Experience: Customers receive a consistent experience without compromise.

Impact and Success

  • Availability: Cisco Silicon One is available across Cisco’s top product classes, including the Cisco 8000, Catalyst, and Nexus platforms.
  • Alternate Business Models: Cisco offers technology through third-party hardware supporting various operating systems.
  • Power Efficiency: An example of power savings achieved by Deutsche Telekom (DT) through the adoption of Cisco Silicon One, resulting in a 92% reduction in power bills.
  • Web-Scale Penetration: Successfully penetrated five of the six global Tier 1 web scalers through silicon-only and white-box models.

Relevance in AI

  • Interpretation of AI: Distinguished between using AI to improve Cisco products (AI-enhanced products) and selling products to enable customers to build AI networks.
  • Network Types:
    • Front-End Network: Traditional Ethernet network connecting generic compute and servers.
    • Back-End Network: Historically InfiniBand, now transitioning to Ethernet for AI/ML networks.
  • Market Opportunity: Net new opportunity for Cisco in the back-end network, selling silicon, systems, and optics.
  • Ethernet Transition:
    • Reasons: Multiple sources of GPUs, web-scale customers’ desire for open standards, and the need for multi-sourcing technology.
  • Why Cisco in AI/ML:
    • Technology Leadership: Best silicon, system, and optics architectures.
    • Business Model Flexibility: Works with customers in various business models, meeting them where they want to be met.

Technical Differentiators

  • Networking Challenges: Load balancing, congestion reaction, and reliability at scale.
  • Options for Customers:
    • Standard Ethernet: True multisource interoperability.
    • Scheduled Ethernet: Ultimate performance and reliability.
    • Enhanced Ethernet: A middle ground, combining standard Ethernet with advanced features.
  • Performance Benefits:
    • Job Completion Time (JCT): Reduction by almost a factor of 2, improving network efficiency.

Cisco Silicon One G200

  • Characteristics: Ultra-low latency, ultra-high performance, twice the power efficiency of previous generations, and a full radix of 512.
  • Advantages: Reduced optics, switches, and an entire layer of networking, saving 1 megawatt of power for a single AI/ML cluster.

Conclusion

  • Competitive Advantage: Right technology, investments, scale, cost points, and business models, making Cisco uniquely positioned to succeed in the AI/ML space.

Q&A Session

  • Commercial Reality: Discussed the cost implications of different silicon production models (fabless, ASIC, and merchant) and how they influence the market.
  • Market Size and Planning: Addressed questions about the total addressable market (TAM) for AI networking, emphasizing a conservative estimate of $10 billion by 2027.
  • Flavors of Ethernet: Explained the three flavors of Ethernet offered by Cisco Silicon One (standard, enhanced, and scheduled) and the importance of standardization efforts like Ultra Ethernet.
  • Commercial Traction: Provided insights into the $1 billion pipeline of orders, highlighting the role of trusted partnerships with web-scale customers and the applicability of Cisco Silicon One to AI.
  • Market Opportunity: Discussed the potential for enterprise adoption of AI clusters, emphasizing the importance of data sovereignty and the scale of the enterprise market.
  • Cross-Pollination of Business Models: Described the dynamic relationship between silicon-only, white box, and full system deployments among web-scale customers.
  • Comparison with NVIDIA Spectrum-X: Commented on the inevitability of the transition from InfiniBand to Ethernet and the adaptability of Cisco Silicon One to different customer requirements and network functions.

Rakesh Chopra’s insights covered a comprehensive overview of Cisco Silicon One’s technological advancements, market positioning, and strategic approach to leveraging AI.

Eyal Dagan [Executive Vice President of Strategic Projects] 💬

Eyal Dagan, the Executive Vice President of Strategic Projects at Cisco, discussed several topics during the special call:

  1. Commercial Reality and Cost Models:

    • Eyal explained three different models for producing silicon: fabless or COT (Customer Owned Technology), ASIC (Application-Specific Integrated Circuit) model, and merchant silicon.
    • He detailed the cost implications of each model, noting that the fabless or COT model is the most cost-effective, while the merchant silicon model is the most expensive.

  2. Impact on Market Dynamics:

    • Eyal discussed how these models influence the market, specifically mentioning that Cisco used to operate under the ASIC model and also used merchant silicon in some cases.
    • He highlighted Cisco's shift towards a fabless COT model over the past five to six years.

  3. Branded Box Model:

    • Eyal analyzed the branded box model, noting that the cost of hardware and switching silicon are becoming equal.
    • He compared the total bill of materials (BOM) costs for vendors using merchant silicon versus those using their own silicon, illustrating that the total BOM cost is higher for those using merchant silicon.

  4. White Box Model:

    • Eyal analyzed the white box model, emphasizing the trend of web scalers moving towards white boxes due to the efficiency and scale of these boxes.
    • He compared the BOM costs for white boxes using merchant silicon versus those using proprietary silicon, indicating that the margins are significantly reduced for those using merchant silicon.

  5. Market Participation and Total Addressable Market (TAM):

    • Eyal noted that Cisco has built the capability to participate in the full TAM of the market, including silicon, white boxes, and branded boxes.
    • He mentioned that Cisco has a full fabless semiconductor operation to support participation in both AI and front-end network markets.

  6. Historical Evolution of Technology:

    • Eyal provided insight into his experience starting Dune and Leaba, and the eventual sale of these companies to Broadcom and Cisco, respectively.
    • He highlighted the development of a single architecture capable of serving diverse market segments, from campus access to service providers, web scalers, and now AI applications.

  7. Market Size and Business Planning:

    • Eyal discussed the total addressable market (TAM) for AI networking, estimating it to be $10 billion by calendar year 2027.
    • He explained that Cisco takes a conservative approach to TAM estimation, considering both top-down and bottom-up methodologies to validate market size estimates.
    • Eyal mentioned that Cisco sees substantial interest and orders from web scale customers related to AI networking.

  8. Web Scale Market Opportunity:

    • Eyal acknowledged that Cisco missed the initial web scale market opportunity but has since refocused on participating in this market.
    • He highlighted Cisco's readiness and technological capabilities, including optics, DSPs, silicon photonics, and advanced packaging, which position the company well for the current AI networking opportunity.

  9. Line of Sight for Revenue Growth:

    • Eyal confirmed that Cisco has line of sight for over $1 billion in revenues related to AI networking by fiscal year 2025.
    • He mentioned that Cisco has agreements, pilots, and other indicators of customer interest to support this revenue projection.

  10. Enterprise AI Cluster Opportunity:

    • Eyal and Rakesh Chopra discussed the potential for an enterprise AI cluster market beyond the hyperscaler space, suggesting that enterprises will likely focus on retraining pre-trained models or inference clusters.

  11. Cross-Pollination of Business Models:

    • Eyal and Rakesh mentioned the cross-pollination between different business models, such as silicon-only, white box, and full system, among web scale customers.
    • They noted that customers often start with one model and then expand into others based on their experiences and needs.

  12. Comparison with NVIDIA's Approach:

    • Eyal and Rakesh briefly addressed the comparison between Cisco's Silicon One and NVIDIA's Spectrum-X approach, highlighting that both recognize the transition from InfiniBand to Ethernet.
    • They emphasized Cisco's ability to accommodate different customer preferences and requirements, whether intelligence is placed in the network or endpoints.

These points summarize the key contributions made by Eyal Dagan during the special call.

Feedback