Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-11
Qualcomm has secured a massive custom AI silicon deal with Amazon for AWS infrastructure, marking a significant shift in the hyperscaler supply chain beyond Broadcom and Marvell. Meanwhile, Meta’s Iris AI chip (MTIA) has entered production, targeting 14GW of compute by 2027, and OpenAI is diversifying its manufacturing risk by adding Samsung Foundry to its silicon program alongside TSMC.
Hyperscaler Custom Silicon: TPU, Trainium, Maia, MTIA — 2026-09-11
Top developments
Qualcomm Secures Major AWS AI Silicon Contract
In a significant expansion of its data center strategy, Qualcomm announced a deal with Amazon to develop custom AI inference chips and 1.6T optical connectivity for AWS infrastructure. The agreement includes a warrant tied to up to $60 billion in purchases, signaling Amazon's intent to diversify its accelerator suppliers beyond its existing Trainium partners. This move pressures the established duopoly of Broadcom and Marvell in the custom ASIC market, as AWS seeks to leverage Qualcomm’s expertise in efficient inference and connectivity.
Meta’s Iris AI Chip Enters Production
Meta’s next-generation MTIA chip, codenamed "Iris," has officially entered production this month, according to reports from Tech Insider. This chip is a critical component of Meta’s four-generation roadmap to achieve 14 gigawatts of dedicated AI compute capacity by 2027. The production start confirms that Meta is accelerating its internal silicon deployment to reduce reliance on external GPU vendors for its massive inference workloads.
OpenAI Adds Samsung Foundry to Silicon Program
OpenAI is actively diversifying its manufacturing supply chain by adding Samsung Foundry to its custom silicon program, confirmed by OpenAI Korea GM Harrison Kim. This partnership involves joint production and research for next-generation AI accelerators and operates through Broadcom’s $200 billion manufacturing MOU. By securing a second foundry option beyond TSMC, OpenAI aims to mitigate single-vendor risk in the increasingly constrained advanced node market.
Local view
Japan: Japanese financial media are closely monitoring the Qualcomm-Amazon deal, with Zaikei Shimbun highlighting the "purchase-linked warrant" structure that could grant Amazon up to 25 million shares of Qualcomm, underscoring the depth of the strategic alliance. Additionally, Kobaran notes that despite Alphabet’s stock sliding toward $300, Google Cloud leadership continues to tout their TPU lead as a key differentiator against Nvidia.
Context & numbers
- AWS Backlog: Amazon Web Services currently sits on a $496 billion backlog, driving the urgent need for diversified AI silicon supply chains.
- Meta Compute Target: Meta is targeting 14GW of AI compute capacity by 2027, with the newly produced Iris MTIA chip as the current cornerstone.
- Broadcom Valuation: Despite strong performance metrics, analysts note that Broadcom (AVGO.US) is trading at a valuation lower than peers, positioning it as a potential "value play" amidst the AI chip boom.
On the radar
- Google TPU Availability: Google Cloud Chief Thomas Kurian continues to emphasize TPU lead, but market watchers are waiting for further details on third-party availability of Ironwood TPUs beyond select operators.
- Nvidia Alternatives: With Trainium2 billing at $0.77 per TB/s-hour compared to $1.54 for B200s, cost-conscious enterprises are increasingly evaluating Trainium and TPU alternatives for inference workloads.
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
