Hyperscalers Pivot to Custom Silicon, Driving Marvell's 227% Annual Surge

Cloud giants commit billions to proprietary AI chips, challenging GPU dominance and reshaping the semiconductor landscape.

Tradesnaut Quant Research Desk · September 28, 2026 · 6 min read · Semiconductors

Hyperscalers Pivot to Custom Silicon, Driving Marvell's 227% Annual Surge

Key takeaways

Hyperscale cloud providers are aggressively developing and deploying custom AI accelerators, migrating away from general-purpose GPUs for specific workloads.
This shift is propelled by the economic imperative of optimizing total cost of ownership (TCO) for inference, enabling tailored performance, and securing supply chain independence, particularly in advanced packaging.
Investors should monitor hyperscaler capital expenditure allocations, the capacity and technological advancements at foundries like TSMC, and the evolving software ecosystems that support specialized silicon deployments.

What changed

A fundamental reallocation of capital is underway in the semiconductor industry, as hyperscale cloud providers dramatically accelerate their pivot to custom silicon for artificial intelligence workloads. Custom ASIC shipments are projected to grow 44.6% year-over-year in 2026, nearly three times the 16.1% growth rate forecast for merchant GPUs over the same period, according to TrendForce in January 2026. This structural shift is visibly impacting market participants. Marvell Technology (MRVL) shares, for instance, have surged 227.75% over the past year, reflecting investor confidence in its role in the custom silicon boom. The company recently showcased industry-first 2nm optical interconnect technology demonstrations at ECOC 2026 in September, aiming to enhance bandwidth and power efficiency for AI data centers. Broadcom (AVGO), another key player, reported

6.70 billion in Q3 FY2026 AI semiconductor revenue, marking a 221% year-over-year increase, and has guided for 1.7 billion in Q4 AI revenue, a 236% rise. The company holds a disclosed $73 billion AI backlog and has secured commitments for

50 billion in AI semiconductors to six frontier customers through fiscal year 2028. Meanwhile, major hyperscalers are rolling out their in-house designs. Google (GOOGL) continues to advance its Tensor Processing Unit (TPU) line, with its v6e, known as Trillium, delivering up to 4x better training performance and 3x higher inference throughput system-wide compared to v5e, with a 67% improvement in energy efficiency. Amazon Web Services (AWS) launched its Trainium3 AI chip in December 2025 and its Graviton4 processor in April 2026. The AWS chips business, including Graviton, Trainium, and Nitro, hit a 0 billion annual run rate in Q1 2026, growing triple digits year-over-year. Meta (META) detailed four new MTIA chips in March 2026, which will enter production in September 2026, designed for training, recommendation algorithms, and inference workloads. The company projects capital expenditures between

25 billion and

45 billion in 2026, with a significant portion dedicated to AI infrastructure.

The mechanism

This migration is not a casual diversification but a strategic imperative driven by the economics of large-scale AI deployment. Hyperscalers aim for a 40-65% Total Cost of Ownership (TCO) advantage by deploying custom silicon like TPUs, Trainium, and MTIA over general-purpose GPUs for high-volume inference workloads. Inference, the process of running trained AI models, now constitutes approximately two-thirds of all AI compute, a significant and growing share. Custom ASICs offer superior power efficiency and performance tailored to specific algorithms, an advantage that scales into billions of dollars in annual savings across millions of accelerators. The shift also addresses supply chain vulnerabilities and the desire for greater control over hardware roadmaps. Companies like Broadcom and Marvell act as crucial co-design partners, translating hyperscalers' specific workload requirements into manufacturable silicon. This dynamic creates a distinct bifurcation in the AI compute market: general-purpose GPUs largely remain dominant for frontier model training, while purpose-built ASICs increasingly handle inference at hyperscale. The entire custom silicon ecosystem, however, is heavily reliant on advanced manufacturing capacity. TSMC's 3nm process, critical for these chips, is running at 100% capacity utilization, with demand outstripping supply by roughly three times. Similarly, advanced packaging technologies, such as CoWoS, represent a binding supply constraint across the industry.

Who is exposed

The implications of this shift are far-reaching. Broadcom (AVGO) and Marvell (MRVL) are direct beneficiaries, together controlling an estimated 95% of the custom AI ASIC co-design market. Broadcom's extensive partnerships with Google, Meta, OpenAI, and Anthropic for custom XPU accelerators and Ethernet switching silicon solidify its position. OpenAI, for example, is partnering with Broadcom for its first custom AI chip, targeting deployment in 2027. Marvell's strong data center growth, fueled by custom silicon and optical interconnects, positions it well, with Google reportedly in talks for new AI chip development with Marvell. Hyperscalers like Google (GOOGL), Amazon, and Meta are reducing their reliance on external GPU vendors for a portion of their AI infrastructure, enhancing their control over performance and costs. Google's Anthropic deal, committing to hundreds of thousands of Trillium TPUs in 2026 scaling to one million by 2027, illustrates this internal pivot. Amazon's Trainium3 is nearly fully subscribed for 2026, with commitments from major AI players like OpenAI and Anthropic. While Nvidia (NVDA) still holds approximately 70% of the AI chip market share, analysts project its inference market share could fall from over 90% to 20-30% by 2028 as hyperscalers scale their custom solutions. Nvidia's recent acquisition of Hugging Face for approximately

2.9 billion in September 2026 reflects its strategic move to hedge against custom silicon inroads by building out its software ecosystem.

Quantitative Outlook

The divergence in growth rates between custom ASICs and merchant GPUs points to a rebalancing of the AI accelerator market. While the total AI accelerator market is projected to reach $604 billion by 2033, custom silicon is capturing an accelerating share of spending. Combined hyperscaler capital expenditure is expected to reach $660-690 billion in 2026, with approximately 75% directed at AI-specific infrastructure. This heavy investment underpins the opportunities for companies enabling custom silicon. For Broadcom (AVGO), its AI semiconductor revenue growth has been explosive, with Q3 FY2026 showing a 221% year-over-year increase. However, its stock performance, up only 4.75% over the past year, lags this operational strength. This disconnect, as noted by 24/7 Wall St. on September 17, 2026, is partly due to investor concerns over margin compression (Q3 gross margin fell to 75%, Q4 guided to 73%) as its XPU mix rises, and risks associated with hyperscaler deployment schedules. Conversely, Marvell (MRVL) shares have seen robust appreciation, up 227.75% over the last year, driven by strong data center segment performance and strategic partnerships, including a new custom silicon warrant deal with Google. The ability of TSMC to meet the surging demand for advanced process nodes and packaging remains a critical choke point, influencing the pace of custom silicon deployment across the industry. The ongoing capital commitments by Google, Amazon, and Meta to their internal chip programs underscore a long-term strategic shift that will continue to reshape supplier relationships and competitive dynamics in the semiconductor sector. The market will closely watch hyperscaler capex reports for indications of sustained investment in custom solutions and the impact on overall AI compute deployment.

Tags: Broadcom, Marvell, Google, AWS, Meta, Nvidia, ASICs, Semiconductors