Analysis of the AI Chip Market by Nvidia

In a recent report released by Nomura, a comprehensive analysis delves into the AI server chip market, specifically focusing on NVIDIA's GB200 and GB300 seriesThis in-depth exploration presents a landscape rich with both opportunities and challenges amid the evolving tech industry.

The report, released on February 12 by Nomura analyst Anne Lee, highlights the tumultuous journey NVIDIA's GB200 AI server chip has faced since entering the market

Advertisements

While there have been some improvements in production yield and stability since the Lunar New Year, overall shipment figures have not met expectationsDespite this, the market's adaptability has proven impressive, significantly mitigating risks associated with the slow GB200 shipments anticipated in the first half of 2025. Consequently, some analysts have recalibrated their forecasts, reducing their previous optimistic outlook for GB shipments in 2025 from a range of 35,000 to 50,000 units down to a more conservative estimate of 20,000 to 25,000 unitsThis adjustment reflects a pragmatic understanding of GB200's current status and provides a more realistic basis for future market decisions.

Despite the fluctuations in GB200 shipments, Nomura maintains an optimistic outlook for the global server market's revenue performanceThe firm projects a remarkable 46% year-on-year increase in global server revenue for 2025, followed by a steady growth rate of 22% in 2026. AI servers are expected to be the primary growth driver, with revenue anticipated to rise by 75% in 2025 and 31% in 2026. These forecasts align closely with the recent capital expenditure guidance revisions by major U.S. cloud service providers, indicating that Nomura's projections are grounded in a sound understanding of market trends and dynamicsIncreased capital spending by cloud service providers signals their confidence in the AI server market's prospects, further fostering technological innovation and expansion in AI applications.

As the market for the GB200 stabilizes, attention is gradually shifting to the upcoming GB300 architectureNomura anticipates an acceleration in NVIDIA's chip and module supply chain transition to B300 and related modules starting in the second quarter of 2025. They project that by then, over 50% of NVIDIA's chips and modules will be based on this new architecture

Advertisements

This rapid technological advancement underscores NVIDIA's commitment to continuous innovation and breakthroughs in the AI chipset sectorThe report also notes a relatively smooth progression for the downstream rollout of the B300 and GB300, with no significant changes observedLarge-scale production of B300 OAM (or SXM) modules is expected to commence in Q2 2025, while the launch of the B300 HGX is projected for mid-2025, and small-scale availability of the standard GB300 system may occur in Q3 2025. However, the rollout of GB300 faces its own set of challengesThe bill of materials design for the B300 UBB and GB300 compute boards remains unresolved, as NVIDIA seeks to reduce costs and diversify its supplier base by introducing a substantial number of new component vendorsWhile this influx of new suppliers may enhance cost efficiency and stabilizes supply chains, it complicates the coordination and finalization of the bill of materials, requiring additional time and resources.

Within NVIDIA's AI server supply chain, distinct disparities exist at different levels, particularly pertaining to the supply-demand gap associated with upstream CoWoS chip supplyNomura has adjusted its Hopper supply estimates, projecting that all CoWoS - L materials will be utilized for the production of B200/300 in 2025. On the demand side, the firm estimates that only 4.4 million and 5.5 million NVIDIA GPUs will be transformed into modules and systems in 2024 and 2025 respectively, a significant shortfall compared to the ideal number required for GB systemsNomura believes that during the period from 2024 to 2026, there will be an approximate 20-23% supply-demand gap in NVIDIA GPUsSuch a substantial gap not only hampers the production timeline of NVIDIA's AI servers but also injects a level of uncertainty into the overall development of the industry supply chain.

Moreover, Nomura has reiterated its optimistic stance on the ASIC (Application-Specific Integrated Circuit) AI sector, predicting that Amazon Web Services (AWS) will significantly expand its proprietary ASIC AI server capacity in 2025. The firm expects the supply of AWS's Trainium 2 chips to peak in the second quarter of 2025, followed by a rapid decline in the fourth quarter as AWS rushes to transition to Trainium 3, aiming to complete this migration by late 2025. Nomura notes that the market may focus on the beneficiaries of Google's TPU and firms that can capitalize on the advantages stemming from AWS's Trainium 3. This shifting market dynamic will further transform the competitive landscape of the AI chip sector, presenting new opportunities and challenges for relevant enterprises.

The market performance of NVIDIA's GB200 and GB300 AI server chips, alongside the overarching trends in the AI chip industry, remains pattern with uncertainty and possibility

Advertisements

Advertisements

Advertisements