May 27, 2026—Marvell Technology (MRVL) released its Q1 FY2027 financial results, reporting quarterly revenue of $2.418 billion, up 28% year-over-year and 9% quarter-over-quarter, slightly surpassing market expectations of $2.41 billion. But what truly sent the market into a frenzy wasn’t just this earnings beat—it was what happened next. On June 2, at COMPUTEX 2026 in Taipei, NVIDIA CEO Jensen Huang appeared on stage with Marvell CEO Matt Murphy and made a bold declaration: "Ladies and gentlemen, this is the next trillion-dollar company."
That statement sent Marvell’s stock soaring over 30% in a single day. Since the start of 2026, Marvell’s share price has nearly doubled, up 95% year-to-date around the earnings announcement.
Behind these dramatic moves lies a deeper industry narrative: AI custom chips (ASICs) are emerging as an independent track running parallel to GPUs. Why are tech giants like Google (TPU), Amazon (Trainium), and Meta (MTIA) bypassing NVIDIA to invest in their own chips? What role does Marvell play—is it a GPU replacement, or a collaborator?
The Essence of AI Custom Chips (ASIC): A Paradigm Shift from General-Purpose to Specialized Hardware
To understand why tech giants are investing heavily in custom chip development, we first need to clarify a key concept: the fundamental difference between ASICs and GPUs lies in the trade-off between general-purpose and specialized design.
GPUs (Graphics Processing Units) are general-purpose AI compute chips. NVIDIA’s GPUs excel at a wide range of AI tasks—training, inference, vision, speech, recommendation systems, and more. However, this versatility comes at the cost of excess circuitry and a broad instruction set, leaving room for efficiency gains in specific scenarios.
ASICs (Application-Specific Integrated Circuits), on the other hand, are hardware tailored for particular AI tasks. Take Google’s TPU (Tensor Processing Unit) as an example: its core is deeply hardwired for matrix multiplication, delivering several times the throughput of a GPU for matrix operations at the same power consumption. Specifically:
- Energy Efficiency: For targeted AI inference workloads, ASICs can deliver 3–5 times the performance per watt compared to GPUs
- Cost Optimization: At hyperscale deployments (think millions of chips in cloud data centers), ASICs offer a significantly lower TCO (Total Cost of Ownership) than commercial GPUs
- System Integration: Custom ASICs can be tightly integrated with a cloud provider’s software stack, network architecture, and cooling systems for end-to-end optimization
The logic behind this paradigm shift is clear: AI workloads are moving from diverse training tasks toward massive-scale inference. As AI model architectures converge (e.g., Transformer models become mainstream) and inference workloads grow exponentially, deep optimization through specialized hardware becomes inevitable.
One analyst summed it up perfectly: "Marvell isn’t ‘replacing NVDA’—it’s opening a second main track in the AI market. Custom ASICs may be the most overlooked but fastest-growing segment in the coming years."
Why Are Tech Giants Building Their Own Chips? The Cost-Efficiency Logic Behind De-NVIDIA-ization
Microsoft, Amazon, Google, and Meta—the four cloud titans—are accelerating their custom chip initiatives at an unprecedented pace, driving the most critical long-term trend in the AI chip sector.
Google TPU (Tensor Processing Unit): Now in its 7th generation, co-designed with Broadcom, and the industry’s earliest and largest custom chip project. Counterpoint estimates Broadcom will command about 60% of the AI server compute ASIC design market by 2027.
Amazon Trainium / Inferentia: The Trainium series, co-designed with Marvell, is being rapidly deployed. Trainium 3 was fully rolled out in early 2026.
Microsoft Maia: In January 2026, Microsoft launched its second-generation custom AI chip, Maia 200, built on TSMC’s 3nm process and now deployed in data centers.
Meta MTIA (Meta Training and Inference Accelerator): Co-designed with Broadcom.
Three key drivers are fueling this trend:
| Level | Core Logic | Key Evidence |
|---|---|---|
| 1: Cost | Massive GPU procurement means high capex | Top cloud providers’ total capex in 2026 is projected at $660–700 billion; custom ASICs can cut per-inference chip costs to 30–50% that of commercial GPUs |
| 2: Energy Efficiency | Data center power draw is a bottleneck | ASICs deliver higher throughput at the same rack power envelope |
| 3: Strategy | Avoid dependence on a single supplier | Cloud giants want to avoid having their core business dictated by NVIDIA’s product roadmap and pricing |
The concept of an "Anti-NVIDIA Alliance" is widely discussed in this context. It’s not a formal organization, but rather a vivid description of the collective shift by tech giants toward custom chips. According to Morgan Stanley and Counterpoint, the AI ASIC market will grow from roughly $12 billion in 2024 to $30 billion in 2027—a 34% CAGR.
Goldman Sachs is even more bullish: they project ASICs will account for 40% of the AI chip market in 2026 and surpass 45% in 2027—almost on par with GPUs. Meanwhile, ASIC server shipments are expected to jump 44.6% year-over-year in 2026, while commercial GPUs will grow just 16.1%.
Marvell MRVL’s Dual Role: Replacement or Collaborator?
Within the de-NVIDIA-ization narrative, Marvell’s market role is often mischaracterized as a direct NVIDIA replacement. In reality, the industry landscape is far more nuanced.
First, the custom chip market has a clear hierarchy.
According to Counterpoint and others, the current AI custom ASIC design services market is a duopoly:
- Broadcom (AVGO): Holds about 55–60% market share, the clear global leader in custom ASICs, deeply embedded with Google, Meta, OpenAI, and others.
- Marvell (MRVL): Holds about 13–15% market share, ranking second, with major clients including Amazon, Microsoft, and Google.
Together, they control about 95% of the custom AI ASIC design market. Importantly, the overall AI ASIC market is still in rapid expansion, with all players benefiting from growth—this is more about collective expansion than a zero-sum fight for existing share.
Second, Marvell’s relationship with NVIDIA is not one of replacement, but of deep collaboration.
This relationship fundamentally shifted in 2026. In March, NVIDIA announced a $2 billion strategic investment in Marvell. The two companies began deep technical collaboration around NVLink Fusion, integrating Marvell’s custom chips and optical interconnect solutions into NVIDIA’s AI Factory and AI-RAN ecosystems.
At COMPUTEX 2026 in June, Jensen Huang gave a clear endorsement: Marvell’s data center switches are "critical for AI workloads."
Why would NVIDIA invest in a company also making custom chips? Here’s the logic:
As AI training clusters scale from thousands to hundreds of thousands or even millions of GPUs, connectivity becomes more scarce—and more valuable—than compute. Huang’s core message at COMPUTEX was exactly this: as AI computation is distributed across entire data centers, networking hardware becomes as important as the GPU itself. Marvell’s expertise in high-speed optical interconnects, Ethernet switching, and 1.6T DSPs is irreplaceable.
Thus, Marvell’s role is best described as a collaborator—not seeking to replace NVIDIA’s GPUs, but providing custom chip options outside the NVIDIA ecosystem while also serving as a critical interconnect infrastructure supplier within it. This dual positioning gives Marvell unique strategic value across the entire AI infrastructure stack.
Marvell Q1 FY2027 Earnings Breakdown: Data-Driven Validation
Have these industry dynamics translated into measurable financial results? Marvell’s latest earnings provide key validation.
Key Financial Metrics
| Metric | Value | YoY/QoQ |
|---|---|---|
| Q1 FY2027 Revenue | $2.418 billion | YoY +28% / QoQ +9% |
| Data Center Revenue | $1.833 billion | YoY +27% / 76% of total revenue |
| Q2 FY2027 Revenue Guidance (midpoint) | $2.70 billion | Implies YoY +35% |
| FY2027 Full-Year Revenue Target | ~$11.5 billion | YoY +~40% |
| FY2028 Revenue Target | ~$16.5 billion | +44% vs. FY2027 |
| Core AI Custom Chip Long-Term Target | $10 billion by 2029 | — |
Source: Marvell official earnings and FY2027 Q1 earnings call
Notable Highlights
Marvell’s Q1 FY2027 data center revenue hit a record $1.833 billion, now 76% of total revenue, underscoring its strategic pivot to AI data centers.
Even more significant is management’s upward revision of guidance: FY2027 revenue target raised from ~$11.0 billion to $11.5 billion, and the FY2028 target hiked from ~$15 billion to $16.5 billion. Morgan Stanley promptly updated its long-term outlook, forecasting data center revenue to grow ~50% YoY in FY2027 and accelerate to ~55% in FY2028.
A milestone not to be overlooked: On June 22, 2026, Marvell will officially join the S&P 500 Index, replacing Pool Corp with a market cap of ~$254 billion. This marks another milestone for semiconductor companies entering mainstream equity indices driven by AI demand.
Marvell’s Acquisition of Celestial AI: Strategic Depth from Compute to Optical Interconnect
A key acquisition underpins Marvell’s growth narrative. In December 2025, Marvell announced a ~$6 billion acquisition of optical interconnect specialist Celestial AI, completed in February 2026.
Celestial AI focuses on silicon photonics and optical interconnects, aiming to solve the growing "memory wall" bottleneck in AI data centers—the data transfer gap between compute and storage.
The strategic intent: Marvell is integrating its custom ASIC, Ethernet switching, and 1.6T DSP capabilities with Celestial AI’s optical interconnect technology to build a full-stack data link solution. According to J.P. Morgan analysts, Marvell is now the only vendor covering custom ASIC design, 1.6T optical DSP, silicon photonics (via Celestial AI), and CXL switching—a comprehensive technical moat unmatched by any single competitor.
On the commercialization front, Marvell expects Celestial AI’s initial revenue contribution to begin in the second half of FY2028, reaching a $500 million annual run rate in Q4.
Comparative Analysis: Marvell vs. NVIDIA and AMD—Structural Differences
In the AI chip value chain, Marvell, NVIDIA, and AMD have fundamentally different business models, shaping their respective growth trajectories and valuation logic. Before diving into the comparison, note that the valuation metrics below are for reference only—not investment advice. Investors should make independent decisions based on their own risk tolerance.
Core Business Model Differences
| Dimension | NVIDIA (NVDA) | Marvell (MRVL) | AMD (AMD) |
|---|---|---|---|
| Core Model | Sells general-purpose GPUs and complete AI compute systems | Custom ASICs + high-speed interconnect infrastructure | Diverse portfolio: general-purpose GPUs, CPUs, FPGAs |
| AI Product Form | Finished chips/systems (HGX/DGX) | Semi-custom chips and interconnect solutions for cloud providers | MI-series GPUs and APUs |
| Customer Relationships | Broad end-customer base | Deeply embedded with top cloud providers (Amazon/Microsoft/Google) | Server OEMs, supercomputing centers, some cloud providers |
| Core Moat | CUDA software ecosystem + system integration | Customization + optical/Ethernet technology expertise | Multi-architecture integration + value positioning |
Revenue Scale and Growth Rate Comparison
| Metric | NVIDIA (FY2026, as of Jan 2026) | Marvell (FY2026 full year + FY2027 outlook) | AMD (2025 full year) |
|---|---|---|---|
| Annual Revenue | ~$130 billion | FY2026 ~$8.2 billion / FY2027 target ~$11.5 billion | ~$25–28 billion |
| Latest Quarterly AI Revenue | Data center >$35 billion/quarter | Data center $1.833 billion/quarter | MI-series ~$1.5–2 billion/quarter |
| YoY Growth Rate | ~40–50% range | FY2027 target ~40% | ~20–30% range |
Source: Company earnings and public market data.
Investor Perspective
J.P. Morgan notes that while NVIDIA’s long-term expected earnings growth rate (51.7%) is higher than Marvell’s (39.4%), Marvell’s valuation is more elastic—its share price is more sensitive to order wins and new customer additions. This difference stems from their respective lifecycle stages: NVIDIA is in a mature expansion phase, while Marvell is at the inflection point of custom ASICs moving from incremental to exponential growth.
Following Marvell’s Celestial AI acquisition, NVIDIA’s strategic investment, and S&P 500 inclusion, Wall Street firm Stifel raised Marvell’s price target to $321 (from $230), reiterating its buy rating.
Potential Risks in the Custom Chip Track
Despite the market’s optimism, several risk factors deserve close attention:
Intensifying Market Share Competition
While Marvell ranks second in custom ASICs, market leader Broadcom (AVGO) has secured major wins with Google TPU and Meta MTIA. Marvell’s ability to grow its share remains uncertain. Counterpoint even projects Marvell’s design services share could fall to around 8% by 2027.
Customer Concentration Risk
Marvell’s custom ASIC business is highly dependent on a few top clients—Amazon, Microsoft, and Google. Any change in product roadmap or supplier by a single client could have a significant impact. While Marvell has AI ASIC design partnerships with over 20 clients, revenue remains concentrated among core customers.
Profit Margin Stability
Marvell’s current operating margin is around 15%, reflecting its traditional hardware design services model. Whether margins can improve as custom ASICs scale up is a key market focus.
Uncertainty from Ongoing NVIDIA GPU Iterations
NVIDIA’s GPU roadmap is still advancing rapidly. Performance leaps in new generations could delay some custom chip projects. The competitive landscape in AI hardware remains fluid.
Geopolitical and Supply Chain Risks
The global semiconductor supply chain faces geopolitical uncertainties, including export controls and de-globalization risks.
Valuation Risk
Marvell’s FY2026 revenue was about $8.2 billion, but its market cap is around $250 billion—reflecting high expectations for future growth. Recent analysis by AInvest notes Marvell’s current price may face valuation pressure. Any shortfall in performance or order momentum could trigger a correction.
Conclusion
Marvell’s Q1 FY2027 earnings beat, combined with Jensen Huang’s trillion-dollar prediction, signals that the AI custom chip track is moving from the industry’s periphery to center stage.
From a broader industry perspective, AI infrastructure is undergoing a structural transformation—from a GPU-centric, monolithic architecture to a diversified model combining GPU training with ASIC inference and interconnect collaboration.
The rise of custom chips like Google TPU, Amazon Trainium, Microsoft Maia, and Meta MTIA reflects a unified direction among global cloud leaders to reduce reliance on NVIDIA. But "de-NVIDIA-ization" doesn’t mean replacing NVIDIA. In fact, Marvell’s deep capital and technical partnership with NVIDIA reveals a deeper trend: the key to winning in AI data centers is shifting from compute to connectivity. As compute clusters scale to hundreds of thousands of chips, efficient interconnect becomes as critical as compute itself.
In this new, multi-polar landscape, Marvell is building a unique moat with its dual focus on custom ASIC design and high-speed interconnect infrastructure. This isn’t a path to replace GPUs, but a parallel, indispensable track in the full-stack AI infrastructure ecosystem.
Whether Marvell becomes the next trillion-dollar company will depend on order execution, market share evolution, and technology roadmaps over the coming years. But one thing is clear: the era of custom chips has already begun.
