Serenity解析AMD MI500与CPO路线图：若明确承诺光学互联将成行业级利好
Related
Serenity: SIVE has the potential to grow into the next LITE.
According to BlockBeats, on July 6th, Serenity published an article expressing its long-term optimism about Swedish optical company Sivers Semiconductors (SIVE), believing it has the potential to become "the next Lumentum (LITE)." In recent months, Sivers has made several key advancements in the AI optical interconnect field, including collaborating with O-Net to advance ELS to mass production, supporting the mass production of Jabil (JBL) 1.6T LRO optical modules, partnering with GlobalFoundries (GFS) to develop continuous wave (CW) lasers for hyperscale cloud vendors, and entering the CPO (co-packaged optics) ecosystem with NVIDIA NVLink, AMD, Amazon, Alchip, and GUC through Ayar Labs. Furthermore, the company is collaborating with Lightium to develop TFLN+CW lasers and is expected to support optical interconnect solutions from Marvell Celestial, Lightelligence, and Lightmatter, while continuing to expand its customer base for undisclosed pluggable optical modules. A recent TrendForce report on cloud vendors like AMD actively securing long-term supply agreements (LTAs) for CW lasers further validates the long-term demand prospects for independent CW laser suppliers. Serenity stated that as AI data center construction continues, various optical interconnect architectures, including 1.6T optical modules, LRO, NPO, and CPO, will develop in parallel over the next few years, providing Sivers with sustained revenue growth opportunities. The company recently completed an oversubscribed institutional financing round, with funds expected to be used to expand laser production capacity and strengthen wafer foundry partnerships. It also plans to complete its Nasdaq listing in the coming quarters to provide financing support for subsequent acquisitions. Currently, the market is overly focused on the commercialization pace of CPO, neglecting the fact that laser products can be widely used in various architectures such as pluggable optical modules, NPO, and CPO. With 1.6T optical modules entering mass production in the second half of 2026, CPO demand accelerating further in 2027, and the Nasdaq listing facilitating the implementation of its acquisition strategy, Sivers is expected to usher in a new round of performance and valuation improvements.
Serenity responds to the sharp drop in CPO stocks: The market has overreacted; the value of related stocks is supported by real income.
BlockBeats reported on July 3rd that Serenity responded to the recent collective decline in photonics stocks, stating that the previously mentioned CPO stocks have actual revenue support, contrasting sharply with the near-zero revenue of quantum computing. Among them, LITE capacity is sold out until 2029, and AAOI is entering the market with its US-made independent CW capacity, projecting quarterly revenue of $1.4 billion by the first half of 2027, yet its current market capitalization is a ridiculously low $9.3 billion. Serenity reiterated its bullish outlook on the optical module market, predicting that TAM will grow ninefold over the next two years, reaching $154 billion (according to a GS report), especially with dollar-denominated content expanding/upgrading by 16x/45x. "Despite short-term volatility, supply and demand fundamentals will drive stock prices back to normal."
Serenity: The US-China AI race has shown characteristics of a "modern Cold War," and the interdependence of supply chains may make it difficult to prevent the escalation of confrontation.
According to BlockBeats, on June 30th, Serenity published an analysis stating that it's hard to imagine a significant drop in capital expenditure by large-scale US AI companies, from Google to Meta, because Chinese companies like 360 have claimed to possess "AI-era cyber nuclear weapons," capable of attacking Western corporate and government systems (previously, 360 founder Zhou Hongyi likened Anthropic's Mythos model to an "AI cyber nuclear weapon," subsequently claiming to have developed a Chinese version with equivalent capabilities). Serenity believes we may be witnessing a "modern Cold War," but the competition is no longer about nuclear arsenals, but rather about superintelligence with both offensive and defensive capabilities, and multiple fronts are unfolding simultaneously, including a supply chain export control game between large-scale Chinese, Japanese, and American companies. She also points out a paradox: all parties remain deeply interdependent—the US relies on China for rare earth and raw material supplies, while China relies on Europe and the US for EUV lithography technology, EDA tools, and precision substrates. This is precisely why Nvidia, AMD, and other AI chip exporters have been able to maintain an "exchange relationship" with China in their rare earth magnet supply chain. It also underscores the urgency for the US to establish its own rare earth supply chain as soon as possible, while avoiding provoking key supply chain nodes of allies like ASML of the Netherlands and Ulvac of Japan through tariffs, lest it lose its negotiating leverage. Serenity warns that as China's self-reliance continues to grow, the "deterrent effect" of supply chain interdependence in containing escalating confrontation is becoming increasingly fragile, potentially becoming a tipping point for the situation to spiral out of control. --------------------------------- Click the original link below to join the Beating · Lark AI news channel for 24/7 monitoring of global AI hotspots and news.
JPMorgan Chase: AI custom chip shipments may surpass GPU shipments in 2027, with Broadcom and Marvell poised for takeoff.
According to Mars Finance, on June 22nd, JPMorgan Chase stated that the custom ASIC market is entering a new growth cycle as large cloud computing companies and tech giants seek to reduce AI computing costs, improve energy efficiency, and move away from a single path dependence on general-purpose GPUs. Broadcom and Marvell are expected to be the biggest beneficiaries of this trend. In a recent semiconductor industry research report, JPMorgan analysts Harlan Sur and Mayur Ramdhani estimated that the digital AI ASIC market will reach approximately $60 billion to $70 billion by 2026, maintaining a compound annual growth rate of over 40% to 50% in the coming years. The report states that Broadcom currently holds approximately 80% to 85% of the high-end ASIC market share, with Marvell ranking second with approximately 10% to 12%. The rapid growth in AI computing demand is changing chip procurement structures. JPMorgan believes that customers such as Google, Amazon, Meta, Microsoft, OpenAI, and SoftBank/Arm are accelerating the development of their own or custom AI processors to achieve better performance, power consumption, and total cost of ownership. Unlike Nvidia and AMD's general-purpose GPUs, ASICs are typically designed for specific customers, software stacks, or platforms, making them more suitable for hyperscale cloud vendors with large-scale internal workloads. The report projects that Broadcom's AI revenue will grow significantly from approximately $20 billion in fiscal year 2025 to over $60 billion in fiscal year 2026, and track to over $150 billion in fiscal year 2027. Its project pipeline includes Google TPU, Meta MTIA, ByteDance AI video and network chips, OpenAI XPU, SoftBank/Arm XPU, and Anthropic's related TPU rack-mount solutions. For Marvell, JPMorgan Chase projects its data center revenue to grow from approximately $6.1 billion in 2025 to approximately $9.3 billion in 2026, reaching approximately $14.6 billion in 2027. Growth drivers include Amazon Trainium 3 and Trainium 4, Microsoft Maia, Google SmartNIC/DPU, CXL controllers, and 800G/1.6T optical DSPs, coherent lite, and early CPO solutions. The report also makes a key prediction: by 2027, the unit shipment of AI ASICs/XPUs will surpass that of GPUs. JPMorgan Chase projects total AI accelerator shipments to reach 23.3 million units in 2027, with GPUs accounting for 10.9 million units (47%) and ASICs/XPUs accounting for 12.5 million units (53%). This means that while GPUs will continue to grow, custom chips are likely to capture a larger share of new AI computing power deployments. JPMorgan Chase, citing Google/Broadcom's TPU7x Ironwood and Nvidia's Blackwell as examples, argues that AI ASICs are competitive in terms of cost-effectiveness and power efficiency. The report shows that the TPU7x Ironwood's FP8 computing power is close to that of the Nvidia B200/B300, but its estimated price is around $13,000, lower than the B200's $35,000 and the B300's $40,000; its computing power per dollar and computing power per watt are also superior to comparable GPUs. This assessment does not imply a rapid decline in Nvidia demand. Instead, it points to a divergence in AI infrastructure investment: GPUs continue to serve general training and inference needs, while cloud vendors' self-developed ASICs will achieve higher penetration rates in large-scale, stable, and predictable internal workloads. For investors, JPMorgan Chase's report reinforces the logic of the AI hardware chain diversifying from GPUs to ASICs, advanced packaging, HBM interfaces, SerDes, optical interconnects, and CPOs. If the predictions in the report come true, Broadcom and Marvell will not just be suppliers of AI networks or connectivity chips, but will become core platform companies in the next phase of AI computing architecture migration.
Serenity: XFAB upgraded to "Buy"; AI and photonic chips may unlock long-term growth potential.
BlockBeats reported on July 6th that Serenity reposted Kelper's latest research report, upgrading semiconductor foundry XFAB's rating to "Buy." The report argues that Melexis demand is continuously improving, and the demand for silicon carbide (SiC), gallium nitride (GaN), and photonic chips driven by AI is expected to offset cyclical fluctuations in the automotive industry. Furthermore, the report notes that specialized foundries, scarce in the West, are receiving higher valuation premiums. Data shows that XFAB's current 2027 EV/EBITDA is only 4.8x, significantly lower than the industry median of 14.5x. Serenity stated that this indicates XFAB's core fundamentals are improving, but they are more optimistic about the growth potential of photonic chips, which the market has not yet fully priced in. They pointed out that XFAB leads the EU's PhotonixFAB project, which aims to build a European silicon photonics industry chain and promote the large-scale manufacturing of photonic chips. In the future, XFAB is expected to benefit from the growth in data center optical communication demand driven by AI giants such as Nvidia. However, since related commercial orders have not yet been implemented on a large scale, traditional valuation models still cannot reflect this long-term growth logic.
SemiAnalysis: NVIDIA Delays or Adjustments to Multiple AI Rack-Scale Architectures, Limiting Rubin Ultra's Expansion Path
According to Mars Finance, on July 6th, SemiAnalysis reported that NVIDIA's latest rack-mount interconnect architecture, Kyber NVL144, has undergone a major adjustment just three months after its release, being delayed by more than 12 months to 2028. The main reason is the ongoing challenges in PCB planar design regarding manufacturing feasibility. Simultaneously, the NVL72x2 back-to-back rack architecture has been cancelled. This solution was originally intended to enhance the scalability of pure copper NVLink by deploying two Oberon racks back-to-back, but its complex structure and high operational burden on hyperscale cloud providers (CSPs) led to strong market skepticism and its eventual abandonment. Furthermore, due to the immaturity of CPO (Co-packaged Optics) technology, NVIDIA's larger-scale scaling solutions based on CPO NVSwitch (such as NVL576) may also continue to be delayed, or limited to small-batch trial production. This means that NVIDIA lacks a stable large-scale scale-up solution until CPO matures. Furthermore, the Rubin Ultra product roadmap has also changed: the originally planned "four-computing-chip" version has been cancelled, with only a "dual-computing-chip" version remaining, and the overall system-level performance is expected to drop to about half of the original solution. These adjustments mean that Nvidia's scale-up capabilities in the Rubin Ultra generation are limited. With the CPO NVSwitch unable to be deployed before the Feynman architecture, competitors such as the AMD MI500X or Google TPU v8i may have a relative window of opportunity in terms of scaling capabilities for large-scale training clusters. Meanwhile, Nvidia is expected to fill market demand during the product transition period and maintain overall supply chain pacing by shipping large quantities of Oberon Rubin racks and their "Ultra" versions.