Hyperscalers Accelerate Custom Silicon Shift, Broadcom AI Revenue Soars 221%

Major cloud providers are deploying purpose-built AI chips at scale, reshaping the semiconductor landscape and diversifying supply chains.

Tradesnaut Quant Research Desk · September 26, 2026 · 6 min read · Semiconductors

Hyperscalers Accelerate Custom Silicon Shift, Broadcom AI Revenue Soars 221%

Key takeaways

Hyperscale cloud providers are aggressively developing and deploying custom AI chips (ASICs) to optimize for specific AI workloads and reduce costs.
This shift is driven by the economic and power efficiency advantages of ASICs, particularly for inference, and a desire to diversify away from a concentrated GPU supply base.
Investors should monitor hyperscaler capex allocations, new ASIC development cycles, and the performance of key ASIC design and interconnect partners like Broadcom and Marvell, as well as Nvidia's response to an evolving market.

What changed

A fundamental reorientation in artificial intelligence infrastructure spending is underway, as hyperscale cloud providers increasingly pivot towards custom-designed silicon. Broadcom (AVGO) recently reported a staggering 221% year-over-year increase in its AI semiconductor revenue for Q3 FY2026, reaching

6.7 billion. This surge underscores a broader industry trend where Google (GOOGL), Amazon Web Services, and Meta Platforms are significantly ramping up their in-house AI chip programs. Google’s latest Tensor Processing Unit (TPU) iteration, the TPU7x “Ironwood” from the seventh generation, became generally available in March 2026, supporting large-scale AI training and inference. Amazon Web Services (AWS) unveiled its Trainium3 AI chip in April 2026, its first 3nm offering, with a computing performance of 2.52 petaflops, twice that of its predecessor. In a clear signal of acceleration, AWS notified suppliers to increase Q3 2026 Trainium 3 ASIC server shipment targets by 20% to 30%. Meta Platforms, for its part, plans to commence production of its in-house Meta Training and Inference Accelerator (MTIA) chips in September 2026, detailing four new MTIA chips in March 2026. The company expects its 2026 capital expenditures to range from

25 billion to

45 billion, with much of that directed towards AI infrastructure.

The mechanism

The drive towards custom silicon is a strategic imperative for hyperscalers operating at massive scale. General-purpose Graphics Processing Units (GPUs), while powerful, represent a compromise, optimized for a wide array of tasks. Custom Application-Specific Integrated Circuits (ASICs), however, are purpose-built for specific AI workloads, primarily inference, which now accounts for roughly two-thirds of all AI compute cycles. This specialization yields significant advantages: superior power efficiency, lower latency, and dramatically reduced operational costs. For instance, Broadcom's custom ASICs can save a hyperscaler upwards of billion in upfront capital expenditure for a 100,000-chip cluster and reduce electricity costs by 50%. This economic leverage, coupled with the desire for greater supply chain control and reduced reliance on a single vendor, fuels the transition. Hyperscalers aim to optimize performance per watt and cost per token, making custom silicon an increasingly attractive proposition as AI deployment scales exponentially. The total AI accelerator market is projected to reach $604 billion by 2033, with custom silicon claiming an accelerating share. TrendForce estimates custom ASIC shipments will grow 44.6% in 2026, nearly triple the 16.1% growth rate projected for merchant GPUs.

Who is exposed

The shift creates distinct winners and presents evolving challenges across the semiconductor industry. Companies like Broadcom (AVGO), which today trades at 352.81, up 0.70%, stand as primary beneficiaries. Broadcom has cemented its position as a core enabler of the AI ASIC ecosystem, co-developing Google's TPUs for seven generations and serving as a design partner for Meta's MTIA chips. The company boasts a $73 billion AI backlog and anticipates reaching

15 billion in AI revenue for fiscal year 2027, with commitments for

50 billion in AI semiconductors to six major customers through fiscal year 2028. Marvell Technology (MRVL), up 1.15% to 261.94 today, is another significant player, projecting fiscal 2028 revenue near

5 billion driven by custom silicon demand. Marvell, which partners with Amazon on Trainium and Microsoft on Maia, focuses on the critical interconnect layer, including optical DSPs and switching silicon for AI data centers. While Nvidia (NVDA), trading at 225.07, up 0.22% today, still commands an estimated 75% to 81% of the AI accelerator revenue market in 2026, this represents a decline from its peak of approximately 87% in 2024 as custom silicon gains traction. However, the overall AI market growth is so substantial that Nvidia's absolute revenue continues to expand despite this market share erosion. Hyperscale companies themselves, including Google (GOOGL), trading at 343.92, up 0.46% today, stand to gain greater control over their infrastructure costs and performance, reducing dependence on external vendors.

Quantitative Outlook

The quantitative signals suggest a sustained and accelerating migration towards custom silicon within hyperscale AI infrastructure. TrendForce estimates that global data center power demand capacity will reach 161 GW in 2026, a 31% year-over-year increase, highlighting the critical need for power-efficient solutions that ASICs provide. The momentum is clear: custom AI accelerator market growth is projected at a 44.6% compound annual growth rate through 2033. As of today, Broadcom (AVGO) has demonstrated robust performance, with its shares trading at 352.81, up 0.70%, reflecting investor recognition of its strong position in this evolving landscape. Marvell (MRVL) also shows positive movement, up 1.15% to 261.94, supported by its significant role in AI data center interconnects and custom silicon programs. Google (GOOGL), trading at 343.92, up 0.46%, continues to invest heavily in its TPU roadmap, with the TPU7x Ironwood already in general availability. Meanwhile, Nvidia (NVDA), up 0.22% to 225.07, faces a nuanced challenge; while its GPU market share percentage is projected to decline, the sheer expansion of the total AI accelerator market ensures continued absolute revenue growth. The crucial factors to watch are the pace of hyperscaler capital expenditure allocations towards custom chips, the success of new ASIC generations from Google, AWS, and Meta, and the ability of design partners like Broadcom and Marvell to convert their substantial backlogs into sustained revenue. Any significant shifts in the cost-performance ratio between merchant GPUs and custom ASICs, or unexpected supply chain bottlenecks beyond TSMC's 3nm capacity, could alter the current trajectory. However, the prevailing trend points to a semiconductor market increasingly defined by specialized, hyperscaler-driven innovation.

Tags: Broadcom, Marvell, Google, AWS, Meta, Nvidia, Custom ASICs, AI Chips