Amazon achieves a $25 billion annualized run rate for its custom AI chips, raising questions about Nvidia’s market dominance.

Amazon (AMZN) continues to set industry benchmarks. Its custom chip division has surpassed a $25 billion annual revenue run rate, fueled by surging demand for its Trainium AI accelerators and Graviton central processing units (CPUs) within cloud computing.

The driver is straightforward: AWS custom chips offer greater affordability, meeting customer demand for lower-cost alternatives to Nvidia (NVDA) hardware.

Amazon already dominates U.S. e-commerce with a 40.5% market share. Concurrently, AWS—the pioneer of the cloud-computing industry—remains the world’s largest cloud infrastructure provider, holding a 28% global market share as of midyear. The critical question remains whether sufficient value-driven customers exist for AWS to meaningfully erode Nvidia’s dominant 80% to 87% share of the AI chip market.

Affordability remains the primary advantage. High performance at competitive pricing appears achievable with Amazon’s silicon. According to Dave Brown, AWS vice president of compute and machine learning, AI developers leveraging AWS chips can achieve 30% to 40% cost savings without compromising performance. Speaking to Yahoo Finance, Brown noted, “That’s what our customers are looking for, is to constantly get more compute and more performance, and then super importantly, at a lower price.”

Amazon designs its silicon “from the ground up” to optimize specific workloads. This portfolio includes Trainium AI accelerators, Graviton cloud processors, and Nitro chips, which underpin AWS’s cloud infrastructure with dedicated storage, networking, and security capabilities. How do these custom AI chips differ from Nvidia’s graphics processing units (GPUs)? While GPUs are engineered for general-purpose workloads, Amazon’s accelerators are purpose-built to execute the specific computational tasks demanded by its clients.

Amazon is hardly the first tech giant to develop proprietary silicon. Alphabet (GOOGL) pioneered the space years ago with its Tensor Processing Units (TPUs), delivering GPU-equivalent performance at a lower cost. Cost efficiency is paramount given that global AI expenditure is projected to approach $2.6 trillion by 2026, per Gartner research. Every dollar saved on AI infrastructure directly benefits corporate bottom lines, especially as investors grow increasingly impatient with unchecked technology spending and its negative impact on valuations.

Marginal gains in AI performance and efficiency can translate into annual savings of millions. In 2025, CloudExpat, a cloud-infrastructure optimization firm, benchmarked AWS Trainium1 chips against Microsoft (MSFT) Azure’s ND H100 instances and Nvidia GPUs. (While Google’s TPUs were also evaluated, this analysis focuses exclusively on the AWS versus Nvidia comparison.) Through dozens of rigorous technical evaluations, CloudExpat determined that organizations training GPT-4-scale models could reduce cloud expenditures by millions by switching from Nvidia GPUs to AWS Trainium chips.

Amazon has consistently enhanced Trainium efficiency and price-to-performance ratios across generations. According to Tech Buzz, Trainium2 delivers four times the speed of Trainium1 while reducing costs by approximately 35% relative to Nvidia’s H100. Amazon states that its subsequent Trainium3 architecture improves price-performance by up to 40% over Trainium2.

These advantages extend beyond training. For inference workloads—where per-query costs accumulate rapidly—Amazon reports that Inferentia2 chips deliver up to 50% lower inference costs compared to GPU-based alternatives. Given that AWS processes billions of inference queries daily, this efficiency gap represents substantial operational savings.

Adoption does present hurdles. Migrating from Nvidia GPUs to AWS Trainium requires enterprises to transition away from Nvidia’s entrenched CUDA ecosystem and adopt AWS’s Neuron software stack—a shift that carries non-trivial engineering and integration costs. Nevertheless, the raw performance and economic advantages of the silicon often justify the migration.

Enterprise adoption is accelerating. In February, OpenAI pledged to utilize 2 gigawatts (GW) of Trainium capacity to power its frontier models beginning in 2027, expanding their existing multiyear partnership by $100 billion over eight years atop a previously announced $38 billion commitment. Anthropic followed in April, pledging over $100 billion across a decade for AWS technologies, heavily weighted toward Trainium and Graviton capacity. Amazon confirmed that Anthropic will “secure up to 5 [GW] of capacity to train and power their advanced AI models, including significant Trainium3 capacity.”

Momentum extends beyond hyperscalers. Startups including TwelveLabs, Neura Robotics, Odyssey, Poolside, Decart, Karakuri, NetoAI, Splash Music, and Metagenomi Therapeutics (MGX) are increasingly deploying AWS custom silicon as a cost-effective substitute for Nvidia hardware. Demand has effectively cleared Trainium3 capacity. Forward bookings for Trainium4—scheduled for release in 2027 or 2028—are already heavily secured.

Graviton processors are experiencing equally robust adoption. Amazon reports that over 130,000 customers now operate Graviton-based servers, accounting for more than half of all new AWS processing capacity deployed. In April, Meta Platforms (META) emerged as one of Graviton’s largest enterprise customers, contracting to deploy tens of millions of Graviton cores to support its agentic AI workloads. Similarly, Uber Technologies (UBER) leverages AWS Trainium3 and Graviton4 processors to optimize infrastructure and accelerate customer-facing applications.

Mirroring Trainium’s trajectory, Graviton adoption is driven by approximately 40% superior price-to-performance metrics and 45% cost reductions versus traditional x86 CPUs. Tech Buzz estimates that, based on cloud-management platform analyses, “a typical Fortune 500 company running 10,000 instances on AWS could save $8-12 million annually by migrating from Intel-based instances to Graviton3.”

This sustained momentum propelled Amazon’s custom silicon division past the $25 billion annual revenue milestone. Labeling this $25 billion run rate merely “impressive” understates its strategic significance. Amazon’s achievement is actively reshaping both the semiconductor and AI infrastructure sectors. AWS has fundamentally pivoted from reselling third-party silicon to designing, manufacturing, and commercializing its own processor capacity. By engineering custom hardware at a fraction of the cost of a single Nvidia GPU, Amazon can now optimize silicon for highly specific workloads—a transformative industry shift.

Equally significant is Amazon’s reduced reliance on external suppliers like Nvidia or Intel (INTC). While AWS will continue purchasing third-party silicon to meet diverse client requirements, its proprietary chips now position the division as a direct competitor in the broader semiconductor market.

AWS’s second-quarter 2026 financial results underscore the profound market impact of its custom silicon strategy. Key highlights include:

Custom chip sales further amplify AWS’s status as a high-margin revenue engine. Despite Amazon’s dominance in North American e-commerce, AWS generated approximately 60% of the corporation’s total operating profit in 2025.

How does Amazon’s silicon division compare to industry leaders like Nvidia, Advanced Micro Devices (AMD), and Intel? Notably, CEO Andy Jassy has indicated that AWS custom chip revenue would double from its current $25 billion baseline if the company shifted from leasing capacity to selling physical chips directly to external buyers. In his April shareholder letter, Jassy stated:

“If we were a standalone chip company, our chips would be generating over $50 billion in annual revenue.”

While Amazon’s proprietary silicon does not yet hit the $50 billion threshold, its effective run rate places it competitively adjacent to major rivals. Nvidia maintains its leadership position, reporting a record $89 billion in revenue from its AI-driven Data Center segment in its most recent quarter (fiscal second quarter of 2027). Amazon’s trajectory places it neck-and-neck with Advanced Micro Devices and Intel. Advanced Micro Devices’ projected annual data-center run rate of roughly $26.9 billion (based on its actual $6.72 billion second-quarter revenues) sits