The era of AI custom chips: Can MRVL become the next trillion-dollar company through the collaborative evolution of ASICs and GPUs?

May 27, 2026, Marvell Technology (MRVL) released its FY2027 first quarter financial report—single-quarter revenue of $2.42B, up 28% year-over-year, up 9% sequentially, slightly exceeding market expectations of $2.41 billion. But what truly set the market ablaze was not just this better-than-expected report card, but what happened afterward: On June 2, NVIDIA CEO Jensen Huang appeared alongside Marvell CEO Matt Murphy at Taipei COMPUTEX 2026, publicly declaring: "Ladies and gentlemen, this is the next company with a trillion-dollar market cap."

This statement caused Marvell's stock price to surge over 30% in a single day. Around the time of the earnings release, Marvell's stock price has nearly doubled since 2026, with a year-to-date increase of 95%.

Behind these fluctuations, a deeper industry proposition emerges: AI custom chips (ASICs) are becoming an independent track parallel to GPUs. Why are tech giants (Google TPU, Amazon Trainium, Meta MTIA) all bypassing NVIDIA to invest in self-developed chips? What role does Marvell play in this—an alternative to GPUs or a collaborator?

The Essence of AI Custom Chips ASIC: From General-Purpose to Specialized Paradigm Shift

To understand why tech giants are investing in self-developed chips, first clarify a key concept: the fundamental difference between ASICs and GPUs lies in the trade-off between versatility and specialization.

GPUs (Graphics Processing Units) are general-purpose AI computing chips. NVIDIA’s GPUs excel in training, inference, vision, speech, recommendation systems, and various AI tasks, but they also carry the overhead of redundant circuits and universal instruction sets, leaving room for efficiency optimization in specific scenarios.

ASICs (Application-Specific Integrated Circuits) are hardware tailored for specific AI tasks. Take Google’s TPU (Tensor Processing Unit) as an example: its core is deeply hard-coded for matrix multiplication operations, enabling matrix computation throughput several times that of GPUs under the same power consumption. Specifically:

• Efficiency advantage: In certain AI inference tasks, ASICs can achieve 3-5 times the performance per watt of GPUs
• Cost optimization: In large-scale deployment scenarios (e.g., cloud data centers deploying millions of units), ASICs’ total cost of ownership (TCO) is significantly lower than commercial GPUs
• System integration advantage: Self-developed ASICs can be end-to-end optimized with cloud providers’ software stacks, network architecture, and cooling systems

This paradigm shift is based on the logic that AI workloads are shifting from diversified training to large-scale inference. As AI model architectures converge (e.g., Transformer becoming mainstream) and inference scale exponentially grows, deep hardware optimization becomes an inevitable trend.

An analyst’s comment is very precise: “Marvell is not ‘replacing NVIDIA,’ but opening a second major track in the AI market. Custom ASICs may be the fastest-growing but most overlooked track in the next few years.”

Why Do Tech Giants Self-Develop Chips? The Cost-Performance Logic of Moving Away from NVIDIA

Microsoft, Amazon, Google, and Meta are accelerating their self-developed chip programs at unprecedented speed, forming the most critical long-term trend in AI chip industry in recent years.

Google TPU (Tensor Processing Unit): now in its 7th generation, designed with Broadcom’s assistance, it is the earliest and largest custom chip project in the industry. Counterpoint estimates Broadcom will hold about 60% of the AI server ASIC design market share in 2027.

Amazon Trainium / Inferentia: The Trainium series, assisted by Marvell, is accelerating deployment. Trainium 3 was fully deployed in early 2026.

Microsoft Maia: In January 2026, Microsoft announced the second-generation self-developed AI chip Maia 200, manufactured with TSMC’s 3nm process, and has begun deployment in data centers.

Meta MTIA (Meta Training and Inference Accelerator): designed with Broadcom’s assistance.

The driving logic chain behind this trend includes three levels:

| Level | Core Logic | Key Evidence | | --- | --- | --- | | First: Cost | High capital expenditure for large-scale GPU procurement | Top cloud providers’ total capital expenditure in 2026 reaches around $660-700 billion; self-developed ASICs can reduce per-inference chip costs to 30-50% of commercial GPUs | | Second: Efficiency | Data center power consumption becomes operational bottleneck | ASICs can deliver higher throughput under the same rack power consumption | | Third: Strategy | Break free from dependence on a single supplier | Cloud giants avoid being constrained by NVIDIA’s product roadmap and pricing strategies for core business |

The concept of an anti-NVIDIA alliance is widely discussed in this context. It’s important to emphasize that this is not an official organizational structure, but a figurative description of the industry trend of tech giants collectively shifting toward custom chips. According to Morgan Stanley and Counterpoint forecasts, the AI ASIC market size will grow from about $12 billion in 2024 to $30 billion in 2027, with a compound annual growth rate (CAGR) of 34%.

Goldman Sachs’s more aggressive forecast predicts that ASIC’s share in the AI chip market will rise to 40% in 2026, surpass 45% in 2027, nearly evenly split with GPUs. Meanwhile, ASIC server shipments are expected to grow 44.6% year-over-year in 2026, compared to only 16.1% for commercial GPUs.

Marvell MRVL’s Dual Positioning: Replacer or Collaborator?

In the narrative of moving away from NVIDIA, Marvell’s market role is often misunderstood as an NVIDIA substitute. But the full industry landscape is far more complex than this label.

First, there is a clear tiered distribution in the custom chip market.

According to data from Counterpoint and others, the current AI custom ASIC design service market shows a duopoly:

• Broadcom (AVGO): approximately 55-60% market share, the absolute leader in the global custom ASIC market, deeply tied to clients like Google, Meta, OpenAI.
• Marvell (MRVL): about 13-15% market share, second in the industry, with key clients including Amazon, Microsoft, Google.

Together, they control about 95% of the custom AI ASIC design market. It’s worth noting that the overall AI ASIC market is still rapidly growing, with all parties sharing the incremental benefits; the current pattern is closer to joint expansion rather than competition for existing market share.

Second, Marvell’s relationship with NVIDIA is not substitution but deep collaboration.

This relationship saw a structural change in 2026. In March 2026, NVIDIA announced a $2 billion strategic investment in Marvell. The two sides are collaborating on NVLink Fusion, integrating Marvell’s custom chips and optical interconnect solutions into NVIDIA’s AI factory and AI-RAN ecosystem.

Later, at COMPUTEX 2026 in June, Jensen Huang gave a clearer endorsement: (Marvell’s data center switches) are critical for handling AI workloads.

Why would NVIDIA invest in a company also making custom chips? The logic chain is as follows:

As AI training clusters scale from thousands of cards to hundreds of thousands or even millions, connectivity becomes a scarcer resource than computing. Huang’s core message at COMPUTEX reflects this—when AI computation is distributed across the entire data center, networking equipment’s importance is on par with GPUs themselves. Marvell’s expertise in high-speed optical interconnects, Ethernet switches, and 1.6T DSPs is irreplaceable.

Therefore, Marvell’s role can be defined as a collaborator—it does not aim to replace NVIDIA’s GPUs but provides an alternative in custom chips outside NVIDIA’s ecosystem, while also serving as a key interconnection infrastructure supplier within NVIDIA’s system. This dual positioning grants it a unique strategic value across the entire AI infrastructure stack.

Dissection of Marvell Q1 FY2027 Financials: Validating the Logic

Is the industry logic above reflected in quantifiable financial results? Marvell’s latest earnings report provides key validation.

Core financial data

| Metric | Value | YoY / QoQ | | --- | --- | --- | | Q1 FY2027 Revenue | $2.42B | +28% YoY / +9% QoQ | | Data Center Revenue | $1.833 billion | +27% YoY / 76% of total revenue | | Q2 FY2027 Revenue Guidance Median | $2.7 billion | implied +35% YoY | | FY2027 Full-Year Revenue Target | about $11.5 billion | roughly +40% YoY | | FY2028 Revenue Target | about $16.5 billion | +44% over FY2027 | | Long-term AI Custom Chip Business Goal | $10 billion by 2029 | — |

Source: Marvell official financial report and FY2027 Q1 conference call

Noteworthy Indicators

Marvell’s FY2027 Q1 data center revenue hit a record $1.833 billion, accounting for 76% of total revenue, fully demonstrating its strategic focus on AI data centers.

More importantly, management raised the outlook: the FY2027 full-year revenue target was revised upward from about $11 billion to $11.5 billion, and the FY2028 target was significantly increased from about $15 billion to $16.5 billion. Morgan Stanley (MS) immediately updated its long-term outlook, forecasting FY2027 data center revenue YoY growth of about 50%, accelerating to about 55% in FY2028.

A milestone event not to be overlooked: Marvell will be officially included in the S&P 500 index on June 22, 2026, replacing Pool Corp with a market cap of about $254 billion. This marks another symbolic case of AI-driven semiconductor companies entering mainstream stock indices.

Marvell’s Acquisition of Celestial AI: From Computing Power to Optical Interconnects

In interpreting Marvell’s growth narrative, one acquisition warrants deep analysis—In December 2025, Marvell announced the acquisition of optical interconnect company Celestial AI for about $6 billion, completing the deal in February 2026.

Celestial AI specializes in silicon photonics and optical interconnect technology, aiming to address the increasingly severe memory wall bottleneck in AI data centers—the data transfer bottleneck between compute and storage.

The core strategic intent of this acquisition is to integrate Marvell’s expertise in custom ASICs, Ethernet switches, and 1.6T DSPs with Celestial AI’s optical interconnect capabilities, building a full-stack technology capability covering data links. JPMorgan analysts note that Marvell has become the only company covering custom ASIC design, 1.6T optical DSP, silicon photonics (via Celestial AI), and CXL switches—this full set of technical barriers is currently unmatched.

From a commercialization pace, Marvell expects Celestial AI’s initial revenue contribution to start from the second half of FY2028, reaching an annualized run rate of $500 million in Q4.

Comparative Analysis: Structural Differences Between Marvell, NVIDIA, and AMD

In the AI chip industry chain, the business models of Marvell, NVIDIA, and AMD are fundamentally different, which determines their growth paths and valuation logic. Before comparison, note that valuation metrics vary significantly due to differences in business structure, scale, growth rate, and profit margins. The figures listed are for reference only and do not constitute investment advice. Investors should make independent judgments based on their risk tolerance.

Key Business Model Differences

| Dimension | NVIDIA (NVDA) | Marvell (MRVL) | AMD (AMD) | | --- | --- | --- | --- | | Core Model | Selling general-purpose GPUs and complete AI compute systems | Custom ASIC + high-speed interconnect infrastructure | General-purpose GPUs + CPUs + FPGAs diversified layout | | AI Product Form | Finished chips/systems (HGX/DGX) | Semi-custom chips and interconnect solutions for cloud providers | MI series GPUs and APU products | | Customer Relationship | Broad end-user coverage | Deep binding with top cloud providers (Amazon/Microsoft/Google) | Server manufacturers, supercomputing centers, some cloud providers | | Core Barriers | CUDA software ecosystem + system integration | Customization capability + optical interconnect/Ethernet tech accumulation | Multi-architecture integration + cost-performance positioning |

Revenue Scale and Growth Rate Comparison

| Metric | NVIDIA (FY2026, as of Jan 2026) | Marvell (FY2026 full year + FY2027 outlook) | AMD (full year 2025) | | --- | --- | --- | --- | | Annual Revenue | ~ $130 billion level | FY2026 approx. $8.2 billion / FY2027 target ~ $11.5 billion | ~$25-28 billion | | Latest AI-related Quarterly Revenue | Data center business over $35 billion/quarter | Data center Q1 revenue $1.833 billion | MI series quarterly ~$1.5-2 billion | | YoY Growth Rate | ~40%-50% range | FY2027 target about 40% | ~20%-30% range |

Data sources: company financial reports and market data.

Investor Perspective on Differences

JPMorgan points out that, compared to NVIDIA, Marvell’s long-term expected profit growth (51.7%) is higher than Marvell’s (39.4%), but Marvell’s valuation elasticity is greater, with its stock price more sensitive to order breakthroughs and new customer additions. This difference mainly stems from their current lifecycle stages: NVIDIA is in a mature expansion phase, while Marvell is at a critical point where custom ASICs are shifting from quantitative to qualitative growth.

After Marvell’s acquisition of Celestial AI, securing NVIDIA’s strategic investment, and entering the S&P 500, Wall Street investment bank Stifel raised Marvell’s target price sharply to $321 (previously $230), reaffirming a buy rating.

Potential Risks in the Custom Chip Track

In the highly optimistic market atmosphere, the following risk factors should be incorporated into the framework:

Intensified Market Share Competition

Although Marvell ranks second in the custom ASIC market, the leading player Broadcom (AVGO) has secured key large orders from Google TPU and Meta MTIA. Whether Marvell can expand its market share remains uncertain. Counterpoint even predicts that by 2027, Marvell’s design service share may decline to around 8%.

Customer Concentration Risk

Marvell’s custom ASIC business heavily depends on a few top clients like Amazon, Microsoft, and Google. Changes in product roadmaps or supplier switches by these clients could significantly impact the business. Currently, Marvell has design collaborations with over 20 clients, but core revenue remains concentrated among key customers.

Profit Margin Stability

Marvell’s operating profit margin is about 15%, mainly reflecting its traditional hardware design service nature. As the volume of custom ASIC production scales up, whether profit margins can steadily improve remains a key concern.

Uncertainty from NVIDIA GPU Iterations

NVIDIA’s GPU product line continues rapid iteration; performance improvements in new generations may delay some potential custom chip projects. The competitive landscape in AI hardware remains dynamically evolving.

Geopolitical and Supply Chain Risks

The global layout of the semiconductor industry faces geopolitical uncertainties, including export controls and de-globalization of supply chains.

Valuation Risks

Marvell’s FY2026 full-year revenue is about $8.2 billion, with a current market cap of around $250 billion, indicating that the market has already priced in high growth expectations for the coming years. Recent analyses by AInvest also suggest that Marvell’s current stock price may face valuation pressure. Any underperformance in order growth or revenue momentum could trigger valuation adjustments.

Conclusion

Marvell’s stellar FY2027 Q1 performance, combined with Huang’s trillion-dollar prophecy, signals that the AI custom chip track is moving from industry periphery to the center stage.

From a macro industry perspective, AI computing infrastructure is undergoing a structural transformation—from a unipolar architecture centered on NVIDIA GPUs to a diversified architecture combining GPU training with ASIC inference and interconnect collaboration.

The collective emergence of custom chips like Google TPU, Amazon Trainium, Microsoft Maia, and Meta MTIA reflects a common trend among leading cloud providers to move away from NVIDIA. But moving away from NVIDIA does not mean replacing NVIDIA. In fact, the deep integration of NVIDIA and Marvell’s capital and technology collaborations reveals a deeper pattern: the key to winning in AI data centers is extending from computing power to connectivity. When the scale of compute clusters exceeds hundreds of thousands of cards, how to efficiently interconnect massive chips becomes as important as each chip’s own computing capability.

In this new pattern of multi-polar competition and cooperation, Marvell, with its dual layout in custom ASIC design and high-speed interconnect infrastructure, is building a unique moat. This is not a path to replace GPUs, but a parallel track—an indispensable independent track within the complete AI infrastructure ecosystem.

Whether the next trillion-dollar company’s prophecy will come true depends on future order execution, market share evolution, and technological evolution over the coming years. But one thing is certain: the era of custom chips has already begun.