Cloud giants initiate the "AI Cost Revolution," marking the arrival of the ASIC era! Maweier Technology (MRVL.US) is about to see a performance surge

Focusing on large AI data centers and customized AI chips (AI ASICs), Marvell Technology (MRVL.US), one of the largest partners in Amazon AWS’s Trainium series AI ASICs, will release its earnings report after the US stock market closes on March 5th Eastern Time. Wall Street analysts unanimously expect that, under the wave of AI inference and the trend of embedding AI large models into enterprise operations through “micro-training,” more cost-effective AI ASICs will pose a strong challenge to NVIDIA’s nearly 90% market share in AI chips. Therefore, analysts predict that Marvell and the larger market leader in ASICs, Broadcom (AVGO.US), will both achieve strong financial growth, and management is likely to provide robust earnings guidance.

In its recent fiscal third quarter 2026 earnings report (for the period ending November 1, 2025), Marvell achieved approximately $2.075 billion in net revenue, up about 37% year-over-year and slightly exceeding market expectations. Adjusted earnings per share also surpassed Wall Street forecasts. This strong quarterly performance reflects the explosive growth in demand for customized AI ASICs driven by cloud computing leaders’ new builds and expansions of AI data centers.

According to data compiled by Zacks Investment Research, Wall Street analysts expect Marvell’s fourth fiscal quarter adjusted EPS to be around $0.79, representing a potential 31.7% increase from the same period last year. Revenue is forecasted to be about $2.21 billion, a significant year-over-year increase of 21%. For the full fiscal year, analysts generally expect EPS of $2.84, an 80.9% increase from the previous year. Revenue projections for this year and next are $8.18 billion and $10 billion, respectively, indicating growth of 41.8% and 22.3%.

Additionally, after completing the acquisition of optical interconnect technology company Celestial AI, Marvell will further enhance its technical capabilities in high-bandwidth, low-latency AI data center infrastructure. This acquisition is expected to contribute to revenue growth over the coming years and help expand the company’s share in the AI ecosystem. In its earnings report, besides strong Q3 results and optimistic outlook for the current quarter, the company also announced it will acquire the AI chip startup Celestial AI for $3.25 billion to strengthen its networking product portfolio.

Marvell CEO Matt Murphy stated during the earnings call that Celestial’s technology will be integrated into Marvell’s next-generation silicon photonics infrastructure hardware, which is expected to create a new, potentially $10 billion super blue ocean market.

Murphy and other executives also indicated that they expect to start generating significant revenue from Celestial AI’s business from the second half of fiscal 2028, reaching an annualized revenue of about $500 million by Q4 2028, and doubling that to $1 billion by Q4 2029.

Market Concerns About NVIDIA’s Future Are Justified

The global surge in generative AI has accelerated the development of AI chips by cloud giants and chip leaders, all competing to design the fastest and most energy-efficient AI compute clusters for advanced large AI data centers. Marvell and its main competitor Broadcom focus on leveraging their advantages in high-speed interconnects and chip IP to collaborate with cloud giants like Amazon, Google, and Microsoft to create tailored AI ASIC clusters based on specific data center needs. This ASIC business has become a vital part of both companies’ portfolios; for example, Google’s TPU AI compute clusters are a prime example of this AI ASIC technology route.

Amazon’s new head of AI infrastructure, Peter DeSantis, said in a recent media interview: “If we can build models on our own self-developed AI chips, we can do so at a fraction of the cost of pure AI large model providers.”

DeSantis added, “Building ultra-large AI data centers does come with certain cost challenges. If we want AI to change everything, the costs must be different.”

The market widely believes that NVIDIA (NVDA.US), the “AI chip superpower,” still dominates the core AI infrastructure market—specifically, the AI chip segment—holding the majority of market share. The chip giant led by Jensen Huang recently announced quarterly and next-quarter guidance that significantly exceeded expectations, but its stock fell sharply by 5% on Thursday. This decline was mainly driven by increasing concerns over the recent announcements from hyperscalers—large-scale cloud providers—about launching more cost-effective AI ASICs based on self-developed designs, which signals potential risks to NVIDIA’s long-standing dominance in the most critical AI infrastructure sector.

Undoubtedly, as Anthropic, dubbed a “strong rival to OpenAI,” plans to spend hundreds of billions of dollars to purchase 1 million TPU chips, and Meta (Facebook’s parent company) considers spending billions later in 2026 or 2027 to buy Google’s TPU AI infrastructure for its massive AI data centers, along with Amazon’s announcement to develop large models using Trainium and Inferentia, the market’s concern about NVIDIA’s future is justified. These moves indicate that cloud giants are initiating an “AI compute cost revolution” to promote self-developed AI ASICs at scale.

The Wave of AI Inference Is Coming, and NVIDIA’s “Monopoly Share” Faces Intense Disruption

Undoubtedly, major constraints related to cost and power are pushing Microsoft, Amazon, Google, and Meta to develop their own AI ASICs for cloud systems, aiming for better cost-performance and energy efficiency in AI compute clusters.

Building ultra-large AI data centers, like the “Stargate” project, is extremely expensive. As a result, tech giants increasingly demand AI compute systems that are more economical and energy-efficient, striving to optimize “cost per token” and “performance per watt.” The prosperity of AI ASIC technology is now a reality.

Furthermore, the long-standing demand for NVIDIA’s advanced AI GPU clusters based on the Blackwell architecture remains high, with costs and supply chain bottlenecks limiting availability. Self-developed AI ASICs can provide “second-curve capacity,” giving companies more leverage in procurement negotiations, pricing, and cloud service margins. Additionally, cloud giants like Google and Microsoft can co-design “chips—interconnects—systems—compilers/runtimes—scheduling—monitoring/reliability,” improving infrastructure utilization and reducing TCO.

NVIDIA’s AI GPUs almost monopolize AI training, requiring more powerful, versatile AI compute clusters and rapid iteration of the entire system. On the inference side, after the deployment of cutting-edge AI models at scale, the focus shifts to unit token cost, latency, and energy efficiency. For example, Google positions its Ironwood TPU as “designed for the inference era,” emphasizing performance, energy efficiency, and scalability. Meanwhile, Amazon’s latest actions suggest that AI ASICs have strong potential for training large models.

The AI ASIC compute ecosystem will, in the medium to long term, continue to erode NVIDIA’s monopoly premium and some market share—not through linear replacement of GPUs, but by reshaping industry profit pools and customer procurement structures. The core reason is that inference’s competitive focus has shifted from “peak compute” to “cost per token,” power consumption, memory bandwidth utilization, interconnect efficiency, and total cost of ownership after hardware-software co-design. In these metrics, ASICs tailored for specific workloads—featuring optimized data flow, compilers, and interconnects—are inherently more cost-effective than general-purpose GPUs.

However, for NVIDIA and AMD, this largely means that marginal pressure is real, likely manifesting as reduced bargaining power, market share erosion, and valuation compression, rather than absolute demand collapse. AI ASICs will undoubtedly continue to impact NVIDIA’s GPU dominance in AI inference, but more as a reshaping of industry profit pools and customer procurement patterns rather than invalidating GPU expansion logic.

AWS explicitly positions Trainium and Inferentia as dedicated accelerators for generative AI training and inference, with Trainium 2 offering approximately 30–40% better price-performance than its AI GPU cloud instances. Google has also announced that Gemini 2.0’s training and inference will run entirely on TPUs. This indicates that “large cloud providers using self-developed ASICs for core model training and inference” is no longer just a concept but has entered a reproducible industrial phase.