6.7 billion in the third quarter of fiscal 2026, largely driven by custom silicon design wins with hyperscale clients. The company is now on track to deliver $58 billion in AI revenue for fiscal 2026, exceeding its original forecast of $56 billion. This surge underscores a significant shift in how major cloud providers are building their AI infrastructure. Meta Platforms has committed a substantial capital expenditure of
30 billion to
45 billion in 2026, primarily directed towards AI infrastructure, nearly doubling its $72.22 billion spending in 2025. A central part of this investment is the ongoing deployment of its in-house Meta Training and Inference Accelerator (MTIA) chips, with four new generations — MTIA 300, 400, 450, and 500 — scheduled for deployment or already in production between 2026 and 2027. Meanwhile, Amazon Web Services (AWS) unveiled its Trainium3 AI chip in April 2026, designed for advanced AI applications, and anticipates its Trainium4 to deliver six times the performance and four times the memory bandwidth. Google (GOOGL), currently trading at
42.75, rolled out its seventh-generation TPU7x (Ironwood) in March 2026, engineered for large-scale AI training and inference. The broader market now projects custom ASIC shipments to grow 44.6% in 2026, almost triple the 16.1% growth rate forecast for merchant GPUs. While NVIDIA (NVDA) saw its shares climb 1.68% today to 28.86 and are up 17.51% over the last 90 days, analysts project its inference market share, which previously exceeded 90%, could decline to 20-30% by 2028 as hyperscaler ASICs reach production scale.
The mechanism
The strategic pivot by hyperscalers is not arbitrary; it is driven by fundamental economic and operational advantages. Custom ASICs offer a total cost of ownership (TCO) advantage of up to 65% over general-purpose GPUs for AI inference at production scale. This is critical as inference workloads, the act of running trained AI models to make predictions, now account for two-thirds of all AI compute. Training large foundation models often occurs once, but serving those models to millions of users happens billions of times daily, making inference efficiency paramount. Hyperscalers also seek to reduce their reliance on a single supplier and gain greater control over their supply chains, a concern heightened by recurring GPU and high-bandwidth memory (HBM) shortages. By designing their own silicon, companies like Google, AWS, and Meta can optimize chips specifically for their proprietary models and software stacks, creating a vertically integrated ecosystem that enhances performance and differentiates their services. This customization provides a competitive moat, ensuring their infrastructure aligns precisely with their evolving AI strategies.
Who is exposed
The migration to custom ASICs creates distinct exposures across the semiconductor and cloud industries. Broadcom (AVGO), currently trading at
49.57, is a significant beneficiary, with its custom AI processor shipments jumping 3.5 times year-over-year in Q3 fiscal 2026. The company holds a $73 billion AI backlog and targets
00 billion in annual AI chip revenue by 2027. Marvell (MRVL), whose shares have gained 201.21% over the past year, is also a key player, projecting up to
1 billion in AI ASIC revenue for 2026. Together, Broadcom and Marvell control an estimated 95% of the custom AI ASIC co-design market. Marvell's data center business now accounts for 76% of its total revenue, with projected 50% year-on-year growth this fiscal year. Hyperscalers such as Google (GOOGL), Amazon, and Meta are the primary drivers of this trend, aiming for greater efficiency and strategic control over their AI infrastructure. For NVIDIA (NVDA), which trades at 28.86, the shift to custom ASICs presents a challenge to its dominance in inference workloads. However, NVIDIA still holds approximately 70% of the overall AI chip market share and remains critical for training the largest AI models. Furthermore, AWS announced plans in August 2026 to deploy an additional 2 million NVIDIA Blackwell Ultra, Rubin, and Rubin Ultra GPUs in 2027–2028 across its global infrastructure. NVIDIA's robust CUDA software ecosystem also continues to be a formidable moat, deeply embedded within the AI research and engineering community.
Quantitative Outlook
The structural divergence in the AI accelerator market is clear, with custom ASIC shipments projected to grow at a compound annual rate of 44.6% through 2033, significantly outpacing the 16.1% growth forecast for GPU-based solutions over the same period, according to Bloomberg Intelligence. This rapid expansion is reflected in the strong performance of key enablers: Broadcom's projected AI revenue of $58 billion in fiscal 2026, and its target of
00 billion by 2027, alongside Marvell's projected
1 billion in AI ASIC revenue for 2026. The substantial increase in Meta's capital expenditure guidance to
30 billion-
45 billion for 2026, nearly doubling its 2025 spending, underscores the significant financial commitment hyperscalers are making to this strategic pivot. While NVIDIA faces a likely reduction in its inference market share, its enduring strength in AI model training and its comprehensive software stack, including CUDA, provide substantial competitive advantages. The market is evolving towards a bifurcated landscape where specialized ASICs excel in inference, while advanced GPUs likely retain a critical role in training and more general-purpose AI tasks. Investors should monitor Broadcom and Marvell's reported ASIC backlogs and design wins as direct indicators of continued hyperscaler adoption. Additionally, further updates on Google's TPU roadmap, AWS's Trainium deployments, and Meta's MTIA advancements will be crucial for assessing the ongoing impact of these custom silicon strategies.
Tags: Broadcom, Marvell, Hyperscalers, AI Chips, Semiconductors
