Written by Ralph Sun
While the battle for GPU supremacy between NVIDIA and AMD captures the majority of headlines, a quieter, highly lucrative sub-sector of the AI hardware market is rapidly expanding. As the artificial intelligence industry shifts from the capital-intensive training phase to the high-volume inference phase, major cloud service providers — commonly referred to as hyperscalers — are aggressively seeking ways to reduce their operational costs. Their primary strategy is to design their own custom Application-Specific Integrated Circuits (ASICs), chips purpose-built for specific inference workloads rather than the general-purpose flexibility of a GPU. This trend is creating a massive tailwind for the companies that help design, build, and connect these custom chips. For investors, Broadcom (AVGO), Marvell Technology (MRVL), and Intel (INTC) represent distinct, compelling plays on the infrastructure required to sustain the AI revolution — each with a different risk profile, moat structure, and potential return trajectory.
The Rise of Custom Silicon
The economic imperative driving the shift to custom silicon is straightforward: inference at scale is prohibitively expensive using general-purpose GPUs. While GPUs are incredibly versatile and necessary for training complex models, they are often over-engineered for specific, repetitive inference tasks. Hyperscalers like Google, Amazon Web Services (AWS), Meta, and Microsoft have realized that by designing chips tailored specifically for their internal workloads, they can dramatically improve performance-per-watt and lower the total cost of ownership. The savings at the scale of billions of daily queries are enormous.
This has led to the proliferation of proprietary chips: Google’s Tensor Processing Units (TPUs), AWS’s Trainium and Inferentia accelerators, and Meta’s Training and Inference Accelerator (MTIA). Projections indicate that custom chips designed by cloud providers will capture up to 40% of the AI server market by 2030. However, these tech giants rarely build these chips entirely from scratch. They rely heavily on established semiconductor companies for intellectual property (IP), design expertise, and crucial interconnect technologies. This dependency creates a durable, recurring revenue stream for the companies that sit at the center of the custom silicon ecosystem.
Broadcom: The ASIC Kingpin
Broadcom stands as the undisputed leader in the custom AI chip market. Trading at approximately $422.76 with a market capitalization of $2.00 trillion, the company is projected to capture a staggering 60% market share in the AI server compute ASIC segment by 2027. That projection alone tells you everything about the scale of Broadcom’s competitive advantage in this space.
Broadcom’s business model is both highly profitable and capital-light. It operates primarily on an IP licensing model, providing the essential building blocks — custom accelerator architectures, high-speed SerDes interconnects, and Ethernet networking silicon — for hyperscalers’ custom designs. Rather than competing directly with its clients, Broadcom positions itself as the indispensable design partner, earning royalties and design fees on every chip that goes into production. This approach offers medium risk and medium elasticity, with the recurring nature of the revenue providing a degree of predictability uncommon in the semiconductor sector.
The company is deeply embedded with the biggest players in the industry. Broadcom is a crucial long-term partner for Google’s TPUs, and in April 2026 it expanded this collaboration to include new custom Tensor Processing Units and networking solutions through a renewed agreement with Google Cloud. That same month, Broadcom announced a multi-year agreement with Meta to develop custom AI accelerator chips through 2029. The company is also a key supplier for Microsoft’s Maia AI chip. These are not one-time contracts; they are multi-year, multi-generational partnerships that lock in Broadcom’s revenue for years to come.
Financially, Broadcom is delivering. In Q1 2026, the company reported revenue of $19.31 billion, a 29% year-over-year increase, with AI-related segments driving the growth. The consensus among 31 Wall Street analysts is a “Strong Buy” rating, with an average 12-month price target of $465.55 and a high estimate of $630. The primary risk for Broadcom is the constant race against its own clients’ maturing internal design teams. As hyperscalers accumulate more in-house semiconductor expertise, they may attempt to bring more of the design process in-house, potentially reducing their reliance on Broadcom’s IP. However, the complexity of leading-edge custom silicon design means this risk is more of a long-term consideration than an immediate threat.
Marvell Technology: The Interconnect Supercycle
If Broadcom is the king of custom compute, Marvell Technology is the lord of interconnects. As AI clusters scale to tens of thousands of chips, the bottleneck shifts from the processors themselves to the speed at which data can move between them. This requires advanced optical networking solutions, specifically optical Digital Signal Processors (DSPs) that translate between the electrical signals inside a server and the optical signals that travel between servers at the speed of light.
Marvell dominates this critical niche, controlling an estimated 70% to 80% of the 800G optical DSP market. This dominance makes Marvell a high-elasticity, medium-risk play on the AI infrastructure build-out. Trading at $164.31 with a market cap of $143.68 billion, Marvell recently reported Q4 fiscal year 2026 net revenue of $2.219 billion, representing a 22% year-on-year growth. The company is not merely riding the current wave; it is actively shaping the next one. In March 2026, Marvell announced a major expansion of its 1.6T optical DSP platform, aiming to redefine AI data center end-to-end connectivity for the next generation of cluster architectures.
The company is also expanding its custom silicon client base beyond its traditional stronghold with AWS. In April 2026, reports emerged that Google is in active talks with Marvell to produce new versions of its AI chips, which would represent a significant diversification of Marvell’s hyperscaler relationships. Furthermore, Marvell is solidifying its position in the broader AI ecosystem through a strategic partnership with NVIDIA to connect to the NVIDIA AI factory via NVLink Fusion, announced in March 2026. This partnership is particularly notable because it demonstrates that Marvell is not merely an alternative to NVIDIA’s ecosystem — it is becoming an integrated part of it.
With a dominant position in the optical interconnect supercycle, Marvell is projected to reach $15 billion in revenue by FY2028. Analyst price targets average around $128–$129, though recent upgrades from five-star analysts have pushed individual targets as high as $170. The forward P/E of 42.73x reflects the market’s confidence in Marvell’s growth trajectory, though investors should note that the stock’s premium valuation leaves it vulnerable to any slowdown in hyperscaler capital expenditure.
Intel: The Turnaround Wild Card
Intel presents the most complex, and potentially the most asymmetric, investment profile of the three. Trading at $82.57 with a market cap of $414.85 billion, Intel is currently struggling with profitability, as evidenced by its deeply negative trailing P/E ratio. The company has endured years of manufacturing setbacks, market share losses in CPUs, and a failed attempt to enter the discrete GPU market. Yet dismissing Intel entirely would be a significant analytical error.
The downside protection comes largely from the United States government. Intel is viewed as a critical national security asset, essential to maintaining domestic semiconductor manufacturing capabilities and reducing dependence on TSMC in Taiwan. This strategic importance has translated into substantial support, including up to $8.9 billion in CHIPS Act funding. In an era of escalating geopolitical competition over semiconductor supply chains, Intel’s role as America’s primary advanced foundry candidate provides a floor beneath the stock that few other companies can claim.
The upside potential hinges on two distinct catalysts. First, there is a surprising and underappreciated resurgence in demand for data center CPUs. As AI workloads shift to inference, robust CPUs are required for data scheduling, retrieval-augmented generation (RAG) pipelines, and security processing. Intel’s data center and AI group saw a 22% revenue increase in the most recent quarter, indicating that CPUs remain vital components in the AI server architecture. Furthermore, Intel’s Gaudi 3 AI accelerators are positioned to compete directly with NVIDIA in the inference market, with Intel claiming they are 50% faster for training tasks.
Second, and most importantly for the long-term thesis, is Intel’s foundry business. The company’s entire strategic narrative relies on successfully scaling its 18A manufacturing node by 2027. Recent reports indicate that Intel is making progress ahead of expectations with its 18A and 14A process nodes, with improving yields. If Intel can successfully establish itself as a viable, high-volume alternative to TSMC for advanced semiconductor manufacturing, the stock could see a dramatic re-rating. Analyst price targets range from $60 to $80, reflecting cautious optimism about the turnaround story, with New Street Research recently raising its target to $80. The forward P/E of 161x is not a valuation metric for a mature business; it is a bet on a transformation.
The Infrastructure Imperative
As the AI market matures, the easy money in hardware has arguably been made. The next phase of returns will come from identifying the companies that provide the essential, inescapable infrastructure for the inference era. Broadcom and Marvell offer highly profitable, entrenched positions in custom silicon and interconnects, benefiting directly from hyperscalers’ relentless drive for cost efficiency. Intel, while riskier and more complex, offers a compelling turnaround story backed by the strategic imperatives of the US government and a genuine technology roadmap. For investors looking beyond the GPU wars, these three companies represent the foundational pillars of the next decade of AI infrastructure growth — and they are far less crowded trades than the names that dominate the headlines.
This article is for informational and educational purposes only and does not constitute financial, investment, or trading advice. EquitiesOrbis.com and its contributors are not responsible for any financial losses or damages incurred as a result of relying on the information presented. Readers are strongly advised to conduct their own independent due diligence, consult with a qualified financial advisor, and carefully consider their risk tolerance before making any investment decisions. Past performance is not indicative of future results, and the value of investments can fluctuate significantly.
