Prabowo Widodo
The move, confirmed by Amazon’s AI chief Peter DeSantis to Bloomberg, signals a significant evolution in Amazon Web Services’ (AWS) long-standing strategy of exclusively deploying its proprietary silicon within its own vast cloud infrastructure. DeSantis indicated that AWS is currently engaged in preliminary discussions regarding the sale of its Trainium AI training chips, though specific potential buyers were not disclosed. This development follows remarks made by Amazon CEO Andy Jassy in his annual shareholder letter earlier in April, where he openly contemplated the possibility of selling the company’s highly coveted homegrown AI chips externally. Jassy’s letter highlighted the immense internal demand for AWS’s custom silicon, estimating that if its chips business operated as a standalone entity selling to both AWS and other parties, it would command an annual run rate of approximately $50 billion. "There’s so much demand for our chips that it’s quite possible we’ll sell racks of them to third parties in the future," Jassy stated, underscoring the strategic shift under consideration.
AWS’s Strategic Pivot: From Internal Consumption to External Supply
For years, AWS has leveraged its custom silicon, including the Graviton series for general-purpose computing and the Inferentia and Trainium lines for AI workloads, as a differentiator within its cloud services. This approach allowed AWS to optimize performance, control costs, and offer specialized instances tailored for specific customer needs. The decision to potentially enter the merchant silicon market marks a notable departure from this integrated model. Historically, AWS has resisted requests to sell its chips directly, prioritizing their deployment within its own ecosystem to power its cloud offerings and enhance the value proposition of its platform. The "waterfall effect" of this model meant that while AWS charged for AI tokens processed by these chips, it also reaped revenue from a comprehensive suite of ancillary services essential for AI applications, such as storage, security, networking, and monitoring. This holistic approach has been a cornerstone of AWS’s profitability and market leadership in cloud computing.
However, the explosive growth in demand for AI compute, particularly driven by generative AI models, has created unprecedented pressure on the supply chain for high-performance chips. Nvidia, with its dominant market share in AI GPUs, has been a primary beneficiary, but even its production capacity has struggled to keep pace with global appetite. This insatiable demand has seemingly prompted AWS to reconsider its stance. By potentially selling Trainium chips directly, AWS could tap into a new revenue stream and solidify its position not just as a cloud provider, but as a critical hardware supplier in the AI ecosystem.
The Genesis of AWS’s Custom Silicon Endeavor
AWS’s journey into custom chip design began over a decade ago, driven by the need for greater control over its infrastructure performance and cost efficiency. The Graviton series, based on ARM architecture, revolutionized cloud computing by offering superior price-performance ratios for a wide range of workloads compared to traditional x86 processors. Building on this success, AWS ventured into specialized AI silicon with Inferentia for inference tasks (running trained AI models) and Trainium for training complex AI models.
The first generation of Trainium, launched in 2020, was designed to provide high-performance, cost-effective training for deep learning models. Its successor, Trainium2, and the upcoming Trainium4 (expected in over a year), represent continuous advancements in this specialized hardware, boasting improved performance, efficiency, and scalability. AWS has successfully integrated these chips into its EC2 instances, making them available to cloud customers who require massive computational power for developing and deploying AI models. High-profile customers, including Anthropic and OpenAI (which AWS formally added to its served models recently), have reportedly leveraged AWS’s Trainium infrastructure for their advanced AI research and development, underscoring the chips’ capabilities.
Nvidia’s Reign and the $50 Billion Gauntlet
Nvidia currently holds an estimated 80-90% share of the AI chip market, a testament to its pioneering work in GPU technology and the robust CUDA software platform. The company’s revenue run rate stands at an astonishing $326 billion, buoyed by successive record-breaking quarters driven by unprecedented demand for its H100 and upcoming B200 AI accelerators. Its market capitalization has soared, positioning it as one of the most valuable companies globally.
Against this backdrop, AWS’s estimated $50 billion annual run rate for its chips business, if it were a standalone entity, presents a substantial, albeit not immediately existential, challenge to Nvidia. To put this figure in perspective, it is comparable to the annual revenues of established semiconductor giants like Intel, which reported $54.2 billion in revenue for the full year 2025. While a $50 billion competitor would not "tank" Nvidia, it would undoubtedly introduce a formidable new player into a market that has largely been Nvidia’s domain. This potential entry signals a maturing market where major cloud providers, armed with significant R&D capabilities and deep pockets, are increasingly seeking to internalize and even monetize their hardware innovations.
The competition extends beyond just hardware specifications. Nvidia’s strength lies not only in its GPUs but also in its comprehensive CUDA software platform, which has become the de facto standard for AI development. AWS’s Trainium, while powerful, operates within a different software ecosystem. Success in the merchant silicon market would require AWS to not only demonstrate superior hardware performance and cost-efficiency but also to foster a developer ecosystem that can rival or at least coexist effectively with CUDA.
Manufacturing Realities and Supply Chain Hurdles
One of the most significant challenges facing AWS in this endeavor is the complex and capital-intensive world of chip manufacturing. Like most fabless semiconductor companies, AWS relies on third-party foundries, primarily Taiwan Semiconductor Manufacturing Company (TSMC), to produce its custom silicon. TSMC is the undisputed leader in advanced chip manufacturing, but its capacity, especially for leading-edge processes crucial for AI chips, is finite and heavily allocated.
The original article notes that AWS has touted its chip capacity as selling out faster than it can produce them. Jassy himself stated that current Trainium capacity, and even the future Trainium4 capacity, had sold out almost instantly. This indicates a severe supply constraint, even for internal AWS use. To pivot to external sales, AWS would need to secure substantially more manufacturing capacity from TSMC or other foundries. This task is monumental, given that Nvidia has recently supplanted Apple as TSMC’s largest customer, effectively commanding a significant portion of the foundry’s most advanced production lines. "Elbowing Nvidia out of the way" for TSMC capacity would be an immense undertaking, requiring strategic long-term agreements, substantial upfront investments, and potentially a reordering of TSMC’s production priorities, which is highly unlikely given Nvidia’s volume and strategic importance.
This inherent tension highlights a fundamental dilemma: how can AWS satisfy burgeoning internal demand for its chips while simultaneously carving out capacity for external sales? The answer likely lies in a multi-pronged approach: aggressive capacity bookings with TSMC, diversification of foundry partners (if feasible for advanced nodes), and potentially a tiered sales strategy where external sales are initially limited or focused on specific high-value customers.
Broader Market Implications and Future Outlook
The potential entry of AWS as a direct seller of AI chips could have profound implications across the industry:
- Increased Competition and Innovation: More players in the AI chip market could spur greater innovation, drive down costs, and offer customers more choices beyond Nvidia. This competition is crucial for preventing a single vendor monopoly in such a critical technology sector.
- Diversification for AI Developers: Third-party companies currently reliant solely on Nvidia’s GPUs or existing cloud infrastructure could gain access to alternative, potentially more cost-effective or specialized, hardware for their AI training needs. This could reduce vendor lock-in and foster more resilient AI development pipelines.
- Evolution of Cloud Provider Strategies: Other major cloud providers like Google (with its TPUs) and Microsoft (with its custom AI chips like Maia and Cobalt) might observe AWS’s move closely. If successful, it could encourage them to explore similar strategies, further blurring the lines between cloud service providers and semiconductor manufacturers.
- Impact on Traditional Chipmakers: While Nvidia is the primary target, this move also impacts traditional chipmakers like Intel and AMD, who are striving to gain a larger foothold in the AI market with their own GPU and accelerator offerings. A new formidable competitor like AWS could further intensify the competitive pressure.
- Supply Chain Resilience: While challenging in the short term, a more diverse set of AI chip suppliers could contribute to greater supply chain resilience in the long run, reducing the industry’s vulnerability to disruptions affecting a single dominant vendor.
AWS spokesperson Doron Aronson reiterated the company’s evolving stance, confirming, "While we’ve historically declined requests to sell chips directly, Andy noted it’s quite possible we’ll sell racks of them to third parties in the future." This statement, coupled with DeSantis’s confirmation of ongoing talks, solidifies the seriousness of AWS’s exploration into this new market segment.
Jensen Huang, Nvidia’s founder and CEO, recently declared a new $200 billion market opportunity for Nvidia in selling CPUs for AI, moving directly into territory traditionally held by Intel and AMD. Simultaneously, Andy Jassy’s ambition for a $50 billion chip market for AWS directly elbows into Nvidia’s world. This parallel strategic expansion by two tech giants into each other’s traditional domains underscores the intense competition and massive opportunities unfolding in the AI era. The coming years will undoubtedly witness a dynamic interplay of innovation, strategic alliances, and fierce competition as these titans vie for dominance in the foundational technology powering the future of artificial intelligence.
