Amazon Could Spend $60 Billion on Qualcomm AI Chips as Big Tech Looks Beyond Nvidia
Amazon may spend upwards of $60 billion on Qualcomm AI chips- a move that speaks volumes about where the market is headed. While Nvidia is still the standard, the biggest cloud companies are building backup options before AI infrastructure gets even more expensive.
Amazon has spent years buying Nvidia hardware and making it available to AWS customers that need serious computing power.
At the same time, Amazon has been building its own processors, expanding its custom silicon business and now entering a long-term agreement that could involve as much as $60 billion in Qualcomm AI data-center technology.
This substantial dollar figure is a major indicator for not just Amazon, but also for Microsoft, Google, and Meta - who are all spending heavily on alternatives to Nvidia hardware.
Nvidia still has a commanding position in AI computing. Its largest customers just seem increasingly reluctant to depend on one vendor for every layer of future growth.
Amazon and Qualcomm Are Making a Very Large Bet on Inference
Qualcomm announced the agreement with Amazon on September 8. The companies plan to collaborate across multiple generations of hardware for large AI data centers, with a particular focus on AI inference.
Inference is the work that happens after a model has already been trained. Every time someone asks an assistant a question, generates an image, summarizes a document or uses an AI coding tool, computing resources are needed to produce that response.
Under the new agreement, Amazon could buy up to $60 billion of Qualcomm AI data-center chips and related products over time. Qualcomm also granted Amazon warrants worth about $4 billion that vest with product purchases.
The partnership extends beyond processors. Qualcomm and Amazon are also developing high-speed optical connectivity for data-center networks, with solutions extending up to 1.6 terabits per second.
That matters because modern AI infrastructure is not simply a collection of powerful chips. Thousands of processors may need to communicate with memory, storage and one another with very little delay. As these systems grow, the networking between them becomes a major part of performance.
For Qualcomm, the agreement is a significant move beyond the smartphone market where the company is best known. For Amazon, it adds another potential source of AI hardware to a supplier list that is already expanding.
Nvidia's Lead Took Years to Build
Nvidia did not become the dominant name in AI computing overnight.
Its graphics processors were originally designed for gaming and visual computing. GPUs also happen to be very good at performing large numbers of calculations at the same time, which made them useful for machine learning long before the current generative-AI wave.
Nvidia recognized that opportunity early and built around it.
One of the most important pieces was CUDA, Nvidia's software platform for programming its GPUs. Over time, researchers, universities, cloud providers and software developers built tools, libraries and workflows around CUDA.
That gave Nvidia an advantage that goes well beyond the physical chip.
A company can replace one processor with another. Replacing years of software development, engineering experience, libraries and infrastructure is much harder.
Today, Nvidia sells the hardware, software and networking technology that make up a large part of the modern AI data center. That combination has helped push the company into the ranks of the world's most valuable businesses.
It has also created a situation where many of the largest technology companies rely heavily on a supplier that has considerable influence over pricing, availability and the direction of the hardware they use.
Why Amazon Wants More Than One Option
AWS has no obvious reason to stop offering Nvidia hardware. Customers want it, and Nvidia continues to produce some of the most capable systems available.
What AWS is trying to avoid is dependence on a single supplier.
If a large portion of AWS expansion relies on one chipmaker, then supply depends partly on that company's production capacity. Costs are influenced by its pricing. New services may need to wait for its next generation of hardware.
That matters more as companies spend tens of billions of dollars building AI infrastructure.
Amazon benefits from having several alternatives even if Nvidia remains the market leader. More suppliers can improve negotiating power. Different processors can be used for different workloads. Amazon can also design some hardware specifically around the needs of AWS rather than adapting everything to an outside supplier's roadmap.
The goal does not have to be replacing Nvidia. Having credible alternatives is valuable on its own.
AWS Has Already Been Building Its Own Chips
Qualcomm is not Amazon's first move in this direction.
AWS has spent years developing its own processors. Trainium is designed for machine-learning workloads, while Inferentia was created specifically for inference.
Those chips are now being used at significant scale.
AWS has described Project Rainier as a Trainium2 deployment that will use nearly one million Trainium2 chips to support Anthropic workloads.
At the same time, AWS continues offering Nvidia GPUs.
That mix says more about Amazon's strategy than any single chip announcement does.
Amazon does not need one processor to win every workload. It can use Nvidia hardware where Nvidia makes sense, Trainium where its own hardware provides better economics and potentially Qualcomm chips for workloads where they become competitive.
Cloud customers may not care which processor handled every request. They care about price, performance and reliability. Amazon cares because those hardware decisions affect its margins, capacity and ability to scale.
Inference Could Become a Much Bigger Part of the Market
Most of the attention around AI chips has focused on training very large models. Training can require enormous clusters of processors running for weeks or months.
Once those models are available to the public, another type of demand begins.
Every AI request requires inference.
A few requests from one person barely matter. Millions of people using AI throughout the day do.
Now add AI features to search engines, office software, customer-service systems, smartphones, development tools and business applications. The amount of computing required to serve those requests becomes substantial.
This is where Qualcomm is trying to build a position.
The company is emphasizing efficiency and the cost of producing AI output at scale. Those issues are becoming increasingly important as data-center operators run into constraints involving electricity, cooling and available capacity.
The fastest chip is not necessarily the best chip for every job.
A processor that can perform a common workload using less power or at a lower cost can be extremely valuable when thousands of those processors are operating continuously. That gives specialized hardware room to compete.
Big Tech Wants Flexibility, Not a Single Winner
Amazon is not alone.
Microsoft has been developing its Maia AI accelerators. Google has spent years building its Tensor Processing Units. Meta is developing custom silicon for internal workloads. AMD continues to compete directly with Nvidia in the accelerator market.
The strategies are different, but the motivation is similar.
These companies expect computing demand from AI to remain large. Depending too heavily on one supplier creates financial and operational risk.
Not every custom processor needs to outperform Nvidia across the entire market.
A chip can still be useful if it handles one important class of workloads more cheaply or efficiently. At the scale of AWS, Microsoft Azure or Google Cloud, even relatively small efficiency improvements can add up to enormous amounts of money.
There is also evidence of how difficult Nvidia is to avoid completely. Even companies building alternative AI hardware still have to account for the software ecosystems, deployment tooling and infrastructure patterns Nvidia has already established.
Nvidia Is Still Strong, but the Market Is Broadening
None of this means Nvidia is suddenly losing the market.
Nvidia remains deeply embedded in AI research, software development and cloud infrastructure. Its hardware continues to be heavily demanded, and its software ecosystem gives companies another reason to stay with the platform.
What may change is the assumption that nearly every major workload should automatically run on Nvidia hardware.
Training a large model may favor one type of processor. Serving huge numbers of everyday AI requests may favor another. Robotics, video generation and other workloads could have different needs again.
The likely result is not one processor replacing Nvidia. It is a market with several different architectures operating side by side. That is a much more useful position for cloud providers.
What This Means for AWS
Cloud computing became successful partly because customers stopped worrying about the exact physical server running their software.
AI hardware may move in the same direction.
AWS may increasingly decide which processor makes the most sense for a particular workload. Nvidia GPUs may remain the best choice for some jobs. Trainium may make more sense for others. Qualcomm could become competitive for high-volume inference.
From the customer's perspective, the underlying hardware could become less important. From Amazon's perspective, it becomes extremely important.
The company with the most hardware options has more control over cost, availability and performance.
Bottom Line
Amazon's Qualcomm agreement fits a broader change across the cloud industry. The largest technology companies are no longer treating one processor architecture as the default answer for every workload.
Nvidia remains in an exceptionally strong position. Its hardware is still heavily demanded, and its software ecosystem gives it an advantage competitors have not been able to quickly reproduce. At the same time, Amazon, Microsoft, Google and Meta are investing heavily in alternatives because the stakes have become too large to rely on one supplier for everything.
That does not mean Nvidia has to lose for Qualcomm, Trainium, Maia or other processors to gain ground. Different chips can serve different purposes. Training large models may favor one platform, while high-volume inference could favor another based on cost, efficiency or power consumption.
For AWS, that flexibility could become a major advantage. Amazon can continue buying Nvidia systems, expand its own Trainium hardware and add Qualcomm technology where it makes sense. The more viable options it has, the more control it gains over cost, supply and performance.
Market test: if hyperscalers keep spreading AI demand across Nvidia, custom silicon and specialist inference chips, the future winner may not be the company with one dominant architecture, but the platform operator with the best mix of price, supply and performance choices.
Sources
- Reuters, September 8, 2026: Qualcomm strikes AI chip deal with Amazon, offers right to buy about $4 billion in stock
- Qualcomm, September 8, 2026: Qualcomm announces multi-generational product collaboration with Amazon
- Amazon Web Services: AWS Trainium customers and Project Rainier deployment with Anthropic
- Reuters, August 10, 2026: Microsoft plans to unveil its new Maia 300 AI chip this fall
- NVIDIA Developer: CUDA and NVIDIA accelerated-computing platforms and tools
Key Takeaways
- Amazon's Qualcomm agreement is less about abandoning Nvidia and more about reducing single-vendor dependence as AI demand scales.
- Inference is becoming a major battleground because the economics of serving AI output at scale are different from training models.
- Custom silicon, specialist accelerators and Nvidia systems can coexist if each wins a different workload or cost profile.
- Cloud providers with more hardware choices gain leverage over pricing, capacity planning and long-term AI margins.