Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-20
Qualcomm and AWS finalized a massive $60 billion custom AI chip agreement, signaling a major shift in the hyperscaler silicon landscape. Meanwhile, J.P. Morgan projects that custom ASICs will surpass GPUs in shipment volume by 2027, driven by aggressive adoption from tech giants like Meta and Amazon.
Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-20
Top developments
Qualcomm and AWS Sign $60 Billion AI Chip Deal
On September 8, 2026, Qualcomm and AWS signed a landmark $60 billion custom AI chip deal. This agreement includes a warrant vesting through 2036 and represents a significant strategic pivot for AWS, which has historically relied on Marvell for Trainium design and Nvidia for general compute. The market reacted swiftly, with Qualcomm stock jumping 7% on the news, highlighting the immense scale of custom silicon commitments in the current AI infrastructure race.
J.P. Morgan Predicts Custom ASICs Will Outship GPUs by 2027
According to a new forecast from J.P. Morgan published this week, custom AI chips (ASICs/XPU) are expected to account for 54% of AI chip shipments in 2027, marking the first time they will exceed GPU shipments. The report estimates the custom AI chip market size at $60–70 billion in 2026, with an annual growth rate exceeding 40%. Broadcom and Marvell currently dominate this sector, holding approximately 90% of the market share combined, with Broadcom alone controlling over 80%.
Anthropic Targets 5GW Compute Capacity
Anthropic is reportedly aiming to secure approximately 5 gigawatts of computing power by the end of the year. This target represents more than three times its capacity from last year and underscores the intense pressure on hyperscalers to deliver specialized silicon like Google TPUs and AWS Trainium to meet inference demands. The move highlights the "capacity race" where execution speed is becoming as critical as raw performance metrics.
Meta’s Iris Chip Enters Production Phase
Meta’s next-generation "Iris" AI chip (part of the MTIA series) has entered production this month. This development is a key step in Meta’s four-generation plan to reach 14 gigawatts of AI compute capacity by 2027. The Iris chip is designed to optimize internal workloads, reducing reliance on external GPU suppliers and leveraging Meta’s deep integration with custom silicon partners.
Local view
Chinese media outlets are closely monitoring the "custom ASIC" trend among local giants. Reports indicate that ByteDance is exploring collaborations with Broadcom to design AI-specific chips to save costs, although ByteDance has officially denied these specific rumors. The broader sentiment in Chinese tech media is that domestic internet giants are increasingly seeking ASIC solutions to mitigate supply chain risks and reduce dependency on US-based GPU manufacturers.
In Japan, industry analysts are focusing on the "compute pricing power" of cloud providers. Recent briefs note that cloud rental price hikes are meeting bottlenecks in packaging and optical delivery, forcing Japanese enterprises to evaluate custom silicon alternatives like AWS Trainium more seriously as they look to stabilize long-term AI infrastructure costs.
Context & numbers
- Market Share: Broadcom holds ~80%+ of the custom AI ASIC design market, while Marvell holds the remaining ~10–20%, primarily through AWS Trainium and Microsoft Maia deals.
- Market Size: The 2026 custom AI chip market is valued at $60–70 billion.
- Adoption Volume: Anthropic has committed to up to 1 million Google TPUs, while AWS has deployed over 500,000 Trainium chips.
- Cost Efficiency: Trainium2 bills at $0.77 per TB/s-hour of memory bandwidth on AWS, compared to $1.54 for an Nvidia B200, highlighting the economic incentive for custom silicon.
On the radar
- Trainium 4 Availability: AWS Trainium 4 was announced in late 2025 with promises of 3x FP8 performance and 4x memory bandwidth over Trainium 3. General availability is targeted for late 2026 or early 2027, making it a critical watch item for enterprise buyers looking to upgrade their inference stacks.
- Broadcom-Meta Extension: The existing deal between Meta and Broadcom extends through 2029, covering 1GW of 2nm MTIA silicon. Future updates on the "Iris" chip's real-world performance benchmarks will be key to validating Meta's 14GW roadmap.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
