A deal that goes beyond a simple chip purchase

Nvidia will sell one million GPU chips to AWS by the end of 2027, and the contract will not begin in the distant future, but as early as 2026. At the same time, it’s not just about the GPUs themselves, as the deal also includes other elements from Nvidia’s portfolio, including Spectrum networking chips, ConnectX technology, and the new Groq chips, which Nvidia unveiled this week as part of a broader strategy for mainstream AI operations.[1]

That is precisely where the real significance of this news lies. When AWS buys not only accelerators but also the networking layer and the layer for processing AI requests, it creates more than just a larger volume of cloud capacity. An architecture is being created to support the mass deployment of AI on a large scale, and from an investment perspective, this is far more important than the size of the order itself.

The market is shifting from training to routine AI operations

According to Reuters, Nvidia raised its estimate of the total market opportunity for its Blackwell and Rubin AI chips to at least $1 trillion by 2027 at the GTC conference, which is significantly higher than the previous estimate of $500 billion by 2026. At the same time, the company clearly communicated that the next phase of growth will be driven by inference, the process in which models answer questions in real time, generate outputs, and perform tasks for end users. This is also key for Amazon. Training large models was the first chapter of the AI boom, but the truly massive market will emerge only when these systems begin to be used daily by hundreds of thousands of companies and hundreds of millions of users. Nvidia further plans to split inference into a prefill phase, to be handled by Vera Rubin chips, and a decode phase, where Groq chips will play a role, indicating an increasingly specialized and multi-layered architecture for AI infrastructure.[2]

Amazon is not relying solely on Nvidia, but on a hybrid infrastructure

Importantly, with this agreement, Amazon is not capitulating to Nvidia’s dominance but rather expanding its own strategy. AWS plans to integrate Nvidia’s NVLink Fusion technology into the upcoming Trainium4 chip, with the goal of building larger AI servers capable of communicating faster and scaling more efficiently when training large-scale models. Amazon has also introduced new servers built on Trainium3, with each server containing 144 chips, offering more than four times the computing power of the previous generation while consuming 40% less energy.[3]

Nvidia itself confirmed in its blog that AWS is designing Trainium4 to integrate with NVLink and the MGX architecture, marking the first phase of a multi-generation collaboration between the two companies. From a market perspective, this means that Amazon won’t be putting all its eggs in one basket. It will combine its own silicon where economics and control over the platform make sense, while simultaneously purchasing the most powerful solutions from Nvidia where speed of deployment and technological advantage are key.

The biggest question is no longer whether demand is strong, but who will profit the most from it

In its March report, Reuters also highlighted one detail that is extremely important from an investment perspective. AWS has been building its own network architecture in its data centers for years, yet it will collaborate with Nvidia on deploying ConnectX and Spectrum X technologies for its most critical AI workloads. In other words, Nvidia is no longer just selling chips. It is gradually pushing deeper into the infrastructure layer of data centers, where hyperscalers have so far sought to maintain more control.1

Nvidia’s stock price performance over the past five years*

This topic is therefore crucial for the stock market for two reasons. The first is confirmation that demand for AI computing capacity remains extremely strong even after months of doubts about the sustainability of AI valuations. The second is that further growth will increasingly hinge on margins, return on capital employed, and which companies can turn the AI boom into stable cash flow, not just revenue growth and another round of massive investments.

InvestingFox is a trademark of CAPITAL MARKETS, o.c.p., a.s., with its registered office at Slávičie údolie 106, Bratislava – Staré Mesto district, 811 02. The company is registered in the Commercial Register of the Municipal Court Bratislava III, Section: Sa, File No.: 4295/B, ID No.: 36 853 054, VAT No.: 2022505419.

CAPITAL MARKETS, o.c.p., a.s. is a securities dealer pursuant to Section 55(1) of Act No. 566/2001 Coll. on Securities and Investment Services and on Amendments and Supplements to Certain Acts, as amended (hereinafter the “Securities Act”). On October 30, 2007, CAPITAL MARKETS, o.c.p., a.s. was granted, by Decision No. OPK-2297/2007 of the National Bank of Slovakia-PLP, a license to provide investment services pursuant to Section 54(2) in conjunction with Sections 59(2) and (3) of the Securities Act, which was extended in accordance with the provisions of the Securities Act by Decision No. OPK-1830/2008-PLP dated April 21, 2008, Decision No. OPK-11601-1/2008 dated January 28, 2009, Decision No. ODT-5059-3/2012 dated July 23, 2012, and Decision No. ODT-9332/2014-1 dated October 21, 2014.

* Past performance is no guarantee of future results

Warning! This marketing material is not and should not be construed as investment advice. Historical data is not a guarantee of future returns. Investing in foreign currencies may affect returns due to fluctuations. All securities transactions may result in both gains and losses. Forward-looking statements represent assumptions and current expectations that may not be accurate or are based on the current economic environment, which may change. These statements do not guarantee future performance. InvestingFox is a trademark of CAPITAL MARKETS, o.c.p., a.s., regulated by the National Bank of Slovakia.

[1]https://www.reuters.com/business/retail-consumer/nvidia-sell-1-million-chips-amazon-by-end-2027-cloud-deal-2026-03-19/

[2]https://www.reuters.com/world/asia-pacific/nvidia-ceo-set-reveal-new-chips-software-ai-megaconference-gtc-2026-03-16/

[3]https://www.reuters.com/business/retail-consumer/amazon-use-nvidia-tech-ai-chips-roll-out-new-servers-2025-12-02/