Research Finder
Find by Keyword
Is AWS's $50B Silicon Business Potentially the Most Undervalued Asset in Tech?
Amazon's three-chip portfolio could represent a standalone semiconductor giant hiding inside a cloud company's balance sheet.
04/14/2026
Key Highlights
- Andy Jassy's shareholder letter indicates that Amazon's silicon business (Graviton, Trainium, Nitro) generates an annual revenue run rate exceeding $20 billion, and is growing at triple-digit YoY percentages.
- Trainium3, AWS's first 3nm AI chip manufactured by TSMC, started shipping in early 2026 and is described as nearly fully subscribed, while Trainium4 (~18 months from wide availability) has already seen significant pre-reservations.
- Amazon's hypothetical standalone chips valuation of approximately $50 billion assumes external third-party sales that have not yet materialized, making this one of the most consequential "what if" disclosures in recent shareholder letter history.
- Graviton5, with 192 cores at 3nm and a five-times larger L3 cache than its predecessor, now powers more than half of all new CPU capacity added to AWS for the third consecutive year.
- The integration of Graviton, Trainium, and Nitro onto unified compute sleds within Trainium3 UltraServers represents a vertically integrated silicon architecture with no direct analog among merchant chip vendors.
The News
In his 2025 Letter to Shareholders, Amazon CEO Andy Jassy discusses a wide range of the company’s lines of business. Key among these was a disclosure that AWS's custom silicon portfolio, including the Graviton CPU, Trainium AI accelerator, and Nitro networking controller, now generates annual revenue run rate exceeding $20 billion, with triple-digit year-over-year growth. He notes that a hypothetical external sales model would imply a run rate approaching $50 billion. The Letter revealed that Trainium3, the company's first 3nm AI accelerator built by TSMC, began shipping in early 2026 and is nearly fully subscribed. The follow-on, Trainium4, is targeting a release in approximately 18 months, and has already seen significant capacity reserved by customers. The letter frames Amazon's custom chips as a structural operating margin lever, projecting the potential to save tens of billions of dollars in annual capex and deliver several hundred basis points of margin advantage over relying on third-party silicon for inference workloads. The full letter is available at aboutamazon.com.
Analyst Take
Reading Andy Jassy's 2025 shareholder letter through the lens of a semiconductor analyst, the passage that stopped us cold had nothing to do with retail, logistics, or even the OpenAI partnership. It was a single sentence about chips. Amazon, Jassy argued, is not really in the infrastructure rental business anymore. It is in the semiconductor business, and it has been for years. The company's custom silicon portfolio now carries a stated annual revenue run rate exceeding $20 billion, growing at triple-digit percentages year-over-year. Most analysts have been treating Amazon's chips as a cost-reduction tool. Jassy is telling us they are a business. That is a meaningful distinction, and we do not see that the market has fully priced the implications.
It is worth remembering how Amazon arrived here, because the origin story is instructive. Everything traces to the 2015 acquisition of Annapurna Labs, an Israeli chip startup that Amazon purchased for a reported $350 million. That acquisition gave Amazon the foundational design capabilities that became Graviton, then Nitro, then Trainium. A decade of compounding that initial investment has produced a portfolio now generating more than $20 billion in annual run-rate revenue. If the $50 billion standalone hypothesis Jassy floats is defensible at all, the Annapurna acquisition would rank among the highest-ROIC technology transactions of the past 20 years, and provides the most useful historical lens through which to evaluate what Amazon's silicon strategy may look like in the next decade.
The contrarian read on the $50 billion figure is deliberate: the letter acknowledges it as a calculation, not a strategy. But the calculation itself is the strategy. By publishing that number alongside Trainium4 pre-reservation news, Jassy could well be signaling intent without announcing a product. That is exactly how Amazon launched AWS. Nobody believed the internal IT shop would become a cloud giant. The same cognitive error may now be applied to silicon.
Market Analysis
The competitive backdrop amplifies why the Jassy disclosures matter at this moment. According to TrendForce, custom ASIC shipments from cloud providers are projected to grow 44.6% in 2026, compared to 16.1% growth for GPU shipments, a divergence that would have seemed implausible three years ago. NVIDIA remains the dominant infrastructure vendor by revenue, and we do not see any scenario in the near term where that changes for frontier model training. However, inference economics are a different calculation, and inference is where scale economics favor proprietary silicon. Amazon Bedrock already runs most of its inference workloads on Trainium, and the letter indicates that Bedrock processed more tokens in Q1 2026 than in all prior years combined. That is the signature of a platform in production, not a product in evaluation.
Incidentally, a solid comparison of Jassy’s hypothetical is AMD. AMD's full-year 2025 data center semiconductor revenue, its highest-growth and most strategically important segment, came in at $16.6 billion. Amazon's current chips run rate, at over $20 billion, already exceeds that figure, and Amazon is not selling a single chip on the open market. AMD has a full sales organization, a partner ecosystem, a CUDA-alternative software stack, and decades of brand equity in the market. Amazon has none of those things externally, yet is already generating more silicon revenue, purely through internal EC2 attribution. That gap would close quickly if Amazon ever moved to external sales, but the current asymmetry is worth noting analytically, even before jumping the number to the hypothetical $50B.
Deloitte's analysis of enterprise AI adoption patterns consistently identifies total cost of ownership as the primary gating factor for production AI deployments. This dynamic plays directly into Amazon's silicon thesis. By designing Trainium to run most of Bedrock's inference internally, Amazon is effectively self-funding the optimization loop. Each inference workload tunes the Neuron SDK, which improves the next generation of chips, which reduces costs for the next wave of Bedrock customers. McKinsey's research on vertical integration economics in technology industries suggests that companies with end-to-end silicon-to-service control typically sustain operating margin advantages of several hundred basis points over peers relying on third-party components, which is precisely the framing Jassy uses in the letter.
The most significant demand concentration dynamic in Amazon's silicon story involves Anthropic. Through Project Rainier, AWS assembled a cluster of more than 500,000 Trainium2 chips, described at the time as the largest AI compute cluster in the world, purpose-built for Anthropic's model training workloads. Anthropic has since been publicly cited as an early Trainium3 customer, reporting up to 50% cost reductions relative to comparable GPU instances. This creates a structure in which Anthropic is simultaneously a financial investment recipient, a cloud infrastructure customer, and an independent performance validator for Amazon's silicon claims. The analytical risk worth monitoring is demand concentration. Anthropic's recent use of Google's TPU infrastructure for some Claude inference workloads suggests that even Amazon's closest AI partner is running a multi-cloud, multi-silicon strategy. Should Anthropic diversify its training compute at scale toward Google Cloud or another provider, the demand signal that Project Rainier represents would be materially diminished, and Trainium's sold-out narrative would require reexamination.
The competitive positioning against Google is also worth examining. Google's TPU v7 (Ironwood) is widely regarded as technically capable and is now offered on Google Cloud, but it operates within a walled garden. The Trainium4 design, with its NVIDIA NVLink Fusion support, is architected to lower migration barriers for customers currently running NVIDIA workloads, which is a fundamentally different market strategy. Google competes by optimizing its own stack; Amazon appears to be competing by making its stack interoperable with the dominant existing ecosystem. We have not seen this distinction receive adequate attention from the broader market.
The Arm AGI CPU, launched in March 2026 with Meta and OpenAI as lead partners, marks the first time in Arm's 35-year history that the company has entered production silicon, moving beyond IP licensing into direct chip sales for AI data centers. That transition is architecturally significant for AWS: Graviton is built on Arm IP, and Arm's decision to sell finished silicon validates the same performance-per-watt thesis that has driven Graviton's displacement of x86 across AWS's fleet for three consecutive years. The more nuanced observation is that Arm's incentives have shifted. A company now selling its own chips has a financial interest in customers choosing its silicon over designing their own, which introduces a subtle but structurally new dynamic into the Amazon-Arm relationship, even as AWS's public posture toward the AGI CPU launch has remained collegial.
Looking Ahead
Based on what we are observing, the most significant forward signal in the Jassy letter is not the $20 billion run rate. It is the $50 billion hypothetical. Amazon has historically disclosed strategic intentions through financial constructs before announcing products. The original AWS was positioned as an internal IT infrastructure tool before it was monetized as a service. Graviton was an internal CPU before it became the basis of 98% of top EC2 customer deployments. The $50 billion framing, paired with Jassy's mention that third-party chip sales are "quite possible," reads as deliberate preparation of investor expectations. The Trainium4 compatibility with NVIDIA NVLink Fusion is the architectural bridge that makes external sales structurally feasible, and HyperFRAME will be monitoring whether Amazon begins offering Trainium availability as a discrete commercial offering rather than an EC2 instance type.
One supply-side risk deserves explicit attention as we track this story. Both Trainium3 and Graviton5 are manufactured on TSMC's N3P process node, and TSMC's 3nm capacity was reported at near 100% utilization through the first half of 2026. That constraint is not specific to Amazon, but it is material context for interpreting Jassy's language around Trainium3 being "nearly fully subscribed." The characterization may reflect exceptional demand, constrained supply, or both simultaneously, and distinguishing between those scenarios matters for projecting when Trainium4 capacity could be brought online at scale. If TSMC's 3nm allocation does not expand materially by the time Trainium4 enters mass production, the pre-reservation demand Jassy highlights may prove harder to fulfill on the timelines customers are assuming.
Stephen Sopko | Analyst-in-Residence – Semiconductors & Deep Tech
Stephen Sopko is an Analyst-in-Residence specializing in semiconductors and the deep technologies powering today’s innovation ecosystem. With decades of executive experience spanning Fortune 100, government, and startups, he provides actionable insights by connecting market trends and cutting-edge technologies to business outcomes.
Stephen’s expertise in analyzing the entire buyer’s journey, from technology acquisition to implementation, was refined during his tenure as co-founder and COO of Palisade Compliance, where he helped Fortune 500 clients optimize technology investments. His ability to identify opportunities at the intersection of semiconductors, emerging technologies, and enterprise needs makes him a sought-after advisor to stakeholders navigating complex decisions.