--- title: "Breakthrough in NVIDIA's computing power monopoly! Amazon's AI ASIC welcomes a structural turning point: starting to gain industry favor" type: "News" locale: "en" url: "https://longbridge.com/en/news/287022087.md" description: "Amazon's Trainium artificial intelligence accelerator is beginning to gain favor among developers, gradually breaking Nvidia's market monopoly. As Nvidia GPU supply tightens, interest in Trainium has risen among developers, with some users shifting workloads to the second-generation Trainium chips, achieving cost reductions of up to 35%. Amazon CEO Andy Jassy stated that if operated independently, the chip business could generate annual revenue of $50 billion" datetime: "2026-05-20T07:18:03.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/287022087.md) - [en](https://longbridge.com/en/news/287022087.md) - [zh-HK](https://longbridge.com/zh-HK/news/287022087.md) --- # Breakthrough in NVIDIA's computing power monopoly! Amazon's AI ASIC welcomes a structural turning point: starting to gain industry favor According to reports, Amazon's (AMZN.US) Trainium artificial intelligence accelerator has begun to win favor among some AI developers who have traditionally relied on NVIDIA's (NVDA.US) products. NVIDIA's GPUs are widely regarded as dominating the AI accelerator market, and their supply has been in a state of shortage due to strong demand from hyperscale data centers, AI frontier model laboratories, and other buyers. Although there are alternatives to NVIDIA products on the market, including those from AMD, Amazon, Google, and other custom application-specific integrated circuits (ASICs), it has been reported that an increasing number of developers are beginning to recognize the appeal of Amazon's Trainium products. The media outlet The Information cited interviews with six individuals who use or collaborate with the chip. Daniel Svonava, CEO of Superlinked, told The Information: "We have always believed that insufficient software support was a barrier. But this has changed in the past few months, and that barrier has been removed." Another developer, Bojan Jakimovski, head of machine learning at Loka, also noted that interest in Trainium has increased in recent months, partly due to the tight supply of NVIDIA GPUs. He added that one client switched their inference workload to Trainium's second-generation chip after testing showed that the cost could be reduced by up to 35% compared to NVIDIA's H100 series chips. However, Jakimovski added that he still recommends training large language models on NVIDIA's products. Amazon CEO Andy Jassy recently stated that the company's chip business could generate $50 billion in revenue annually if operated independently. Jassy wrote in a recent letter to shareholders: "Our custom chip business is now one of the top three data center chip businesses in the world." ## Why are developers moving from "no choice" to "actively embracing"? NVIDIA's GPUs are widely regarded as the king of the AI accelerator market, and its CUDA software ecosystem has built a moat that competitors find difficult to cross. However, due to NVIDIA's overly strong position, its product supply has long been tight—hyperscale cloud service providers, AI frontier model laboratories, and other buyers have an insatiable demand, leading to a structural shortage of NVIDIA GPUs. This supply-demand imbalance has created a rigid demand for alternatives in the market. While there are alternative options such as AMD, Google TPU, and other custom ASICs available, Trainium is gaining actual adoption from developers at a pace that exceeds market expectations. **Software Ecosystem: A Qualitative Change from "Barrier" to "Elimination"** Daniel Svonava, CEO of Superlinked, succinctly summarized this turning point in a statement to The Information: "We have always believed that insufficient software support is a barrier. But this has changed in the past few months, and that barrier has been eliminated." The weight of this statement lies in the fact that, in the competition for AI chips, hardware parameters often only determine the lower limit of a product, while the software ecosystem determines the upper limit. The transformation of Trainium from "barrier" to "elimination" at the software level means it is no longer just an alternative option for limited testing but has become a productivity tool with conditions for large-scale commercial use. **Cost Advantage: The New Generation of "Cost Reduction and Efficiency Improvement" Weapon** Bojan Jakimovski, head of machine learning at Loka, has also observed that the appeal of Trainium is significantly rising, supported by solid economic logic. The direct reason some customers are turning to Trainium is the difficulty in obtaining NVIDIA GPUs; however, more importantly, one customer decisively switched their inference workload to Trainium after testing revealed that the cost of Trainium's second-generation chips could be reduced by up to 35% compared to NVIDIA's H100 series. As AI inference workloads increasingly become the main force in computing power consumption (currently accounting for about two-thirds of all AI computing), a 35% cost advantage means that a medium-sized AI company could save millions to tens of millions of dollars in computing power expenses each year. This is not a slight shift in a zero-sum game but a structural advantage sufficient to change procurement decisions. **First-Mover Advantage in Technical Architecture: The Unique MoE Inference Moat** Gavin Baker's judgment is particularly sharp and technically insightful. He pointed out that the current mainstream cutting-edge AI models all adopt a Mixture of Experts (MoE) architecture, and running inference tasks for such models requires the infrastructure of a Switched Scale-up Network. Currently, only two companies globally have operational Switched Scale-up Networks: one supports NVIDIA's GPU clusters, and the other drives Amazon's Trainium. This means that in the rapidly growing key track of MoE model inference, Trainium is not merely a follower but a first mover with unique technical barriers. Baker further noted that Google's TPU does not possess equivalent capabilities in the same field and revealed that although Google invented the MLPerf benchmark, it has never submitted TPU test results. This detail undoubtedly reinforces the market's reassessment of Trainium's technical uniqueness. Baker expects that after the large-scale production of Trainium 3 in the second half of this year, Trainium's market position in 2026 will be comparable to that of TPU in 2025 ## Customer Ecosystem: The Leap from "Ten Thousand" to "One Hundred Thousand" Threshold The breakthrough of Trainium is not only reflected in technological aspects but also in the large-scale validation of its customer base. According to Amazon's disclosure during the deepening of its strategic cooperation with Anthropic in April, both Trainium and Graviton have over 100,000 customers each, and most inference tasks on Amazon Bedrock currently run on Trainium. The figure of 100,000 customers marks a transition from quantitative to qualitative change in Trainium's customer base since the second half of 2025—it is no longer a niche product tested only in a few laboratories but has become a systematic alternative with large-scale commercial validation. **Anthropic and OpenAI: The Biggest "Quality Proof"** At the key customer level, Trainium has gained deep ties with two of the world's most important AI model companies. On April 20, Amazon announced the deepening of its strategic cooperation with Anthropic: Amazon will invest up to $25 billion more in Anthropic, while Anthropic commits to investing over $100 billion in AWS-related technologies over the next decade and purchasing up to 5 gigawatts of current and future generations of Trainium chip computing power. Anthropic's flagship model, Claude, runs on over 1 million Trainium2 chips. The partnership with OpenAI is equally significant. In February of this year, OpenAI established a multi-year strategic partnership with Amazon, with Amazon investing $50 billion and providing OpenAI with 2 gigawatts of Trainium computing capacity. OpenAI has committed to using Trainium 3 and the next generation Trainium 4 chips to support its extensive advanced AI workloads. For chip products, the quality of customers often signals more than the quantity. When the world's most technically discerning AI frontier laboratories choose to run their core workloads on Trainium, it is itself the strongest endorsement of the chip's performance and ecological maturity. **From "Leasing Computing Power" to "Direct Selling Chips": The Blueprint for a $50 Billion Empire** What is even more noteworthy is the strategic elevation of Trainium's business model. In April of this year, Amazon CEO Andy Jassy disclosed in a letter to shareholders that the company is considering changing its previous strategy of only internal use to directly selling its self-developed chips and complete machine racks to third parties—if this department operates independently and fully opens up, the annual revenue scale is expected to reach $50 billion. Jassy further pointed out that this figure exceeds the levels of AMD and Intel during the same period, stating, "Our custom chip business is now one of the top three data center chip businesses in the world." This is not just talk. As of the disclosure, Amazon has already secured $225 billion in Trainium chip revenue commitments, covering strategic customers such as Anthropic and OpenAI The cost-performance ratio of Trainium2 has surpassed that of similar GPU products by 30%, and it is basically sold out; Trainium3 just started shipping in 2026, with a cost-performance ratio improved by 30% to 40% compared to Trainium2, and it has almost all been reserved; even Trainium4, which is still about 18 months away from large-scale supply, has most of its production capacity locked in. The complete sell-out of two generations of products, along with the significant reservations for the next-generation chip that has not yet been mass-produced, is an extremely rare demand signal in the history of the semiconductor industry. It indicates that the appeal of Trainium is not just a short-term hype, but a long-term strategic commitment made by customers after thorough evaluation. ## Structural Turning Point for ASICs The rise of Trainium is reshaping the deepest industrial relationships in the AI chip field—specifically, the long-standing "supplier and customer" model between Amazon and NVIDIA. This relationship was originally clear: NVIDIA was responsible for designing and manufacturing the most powerful AI chips, while Amazon, as one of the largest cloud service providers, made large-scale purchases. However, as Amazon began to design and deploy its own AI accelerators, the roles of both parties underwent subtle changes. Recent data shows that the number of Trainium servers currently deployed by Amazon has surpassed that of NVIDIA servers, and the company estimates that self-developed chips can save billions of dollars in capital expenditures compared to using purchased GPUs. However, this relationship is not a simple replacement. Amazon has neither abandoned the procurement of NVIDIA chips—recently signed procurement commitments are still expanding—nor has it stopped its heavy investment in Trainium. The two currently present a complex "competitive coexistence" pattern: Trainium is rapidly expanding its share in inference workloads, while NVIDIA GPUs still dominate in training large-scale foundational models. From a broader industry perspective, custom ASICs are experiencing a structural turning point. Data shows that by 2026, custom AI chips from Google, Microsoft, Amazon, and Meta are expanding at a compound annual growth rate of 44.6%, while the compound annual growth rate for general-purpose GPUs is only 16.1%. The growth of custom ASICs primarily targets the inference market—currently, inference accounts for about two-thirds of all AI computing. Although NVIDIA currently holds over 90% of the AI accelerator market, analysts expect that by 2028, its share in the inference field may drop from over 90% to 20% to 30%. Trainium is one of the most important variables in this wave of custom ASICs. As industry reports conclude: 2026 marks the moment when "custom ASICs are no longer just experimental projects, but become a scalable alternative to NVIDIA GPU monopoly." **Reality Check: How far is Trainium from "full replacement"?** Despite Trainium experiencing significant user growth and performance upgrades, it is essential to remain objective and calm when fully assessing its market positioning. One key point that needs clarification is that for most cutting-edge AI laboratories, Trainium is currently more suitable for inference rather than training Although Bojan Jakimovski confirmed the cost advantages of Trainium in the inference phase, he still suggested that clients continue to use NVIDIA's products for large language model training. This reflects a reality: the advantages established by the NVIDIA CUDA ecosystem in terms of flexibility for large-scale model training, the completeness of the operator ecosystem, and the depth of community support remain significant. Additionally, it is worth noting that the strong demand for Trainium is somewhat disconnected from Amazon's recent stock performance. Despite the growing interest from developers in Trainium AI chips, Amazon's stock performance has not been as strong as that of other tech giants recently. There is an overall revaluation process in the market regarding the intensified competition in the AI chip sector—NVIDIA, AMD, Google TPU, Microsoft's Maia, and Meta's MTIA are all competing on the same stage. Although Gavin Baker holds a positive stance on Trainium, he also emphasized, "I will never short Google, nor will I short Broadcom," indicating that this is a win-win market rather than a zero-sum game. Furthermore, all mainstream AI chips—whether custom ASICs or NVIDIA GPUs—are manufactured using TSMC's 3nm process. This means that Google, Microsoft, Amazon, Meta, and NVIDIA are all competing for the limited capacity of the same foundry. Capacity constraints apply equally to all players, and any rapid expansion by a chip designer may encounter physical delivery limits ### Related Stocks - [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [SOXX.US](https://longbridge.com/en/quote/SOXX.US.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [AMZZ.US](https://longbridge.com/en/quote/AMZZ.US.md) - [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [AMZW.US](https://longbridge.com/en/quote/AMZW.US.md) - [SMH.US](https://longbridge.com/en/quote/SMH.US.md) - [XLK.US](https://longbridge.com/en/quote/XLK.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [NVDU.US](https://longbridge.com/en/quote/NVDU.US.md) - [AMZU.US](https://longbridge.com/en/quote/AMZU.US.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [CLOU.US](https://longbridge.com/en/quote/CLOU.US.md) - [07788.HK](https://longbridge.com/en/quote/07788.HK.md) - [07388.HK](https://longbridge.com/en/quote/07388.HK.md) - [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md) ## Related News & Research - [Cerebras's stock set for blast off, as early indications point to a near doubling](https://longbridge.com/en/news/286434594.md) - [Joseph P. Lucia & Associates LLC Has $10.21 Million Holdings in NVIDIA Corporation $NVDA](https://longbridge.com/en/news/286531492.md) - [Nvidia earnings on May 20 could sway AI market sentiment](https://longbridge.com/en/news/286642553.md) - [Nvidia's $78 billion earnings test could reveal the real AI demand story](https://longbridge.com/en/news/286447895.md) - [Up 110% YTD, Why Nvidia’s Backing Could Keep Nokia Stock Rallying](https://longbridge.com/en/news/287094557.md)