Qualcomm Technologies announced on 8 September 2026 a multi generation collaboration with Amazon to build customised silicon for Amazon Web Services (AWS) data centres. The agreement covers Qualcomm AI chips for AI inference and high speed optical connectivity, with qualifying Amazon purchases of up to $60 billion through September 2036. The Qualcomm Amazon partnership signals a major shift toward power efficient, custom AI chips as an alternative to dominant suppliers in cloud infrastructure.
What Is Qualcomm?
Qualcomm is an American semiconductor and wireless technology company founded in 1985 and headquartered in San Diego, California. Qualcomm Technologies Inc, a subsidiary of Qualcomm Incorporated, designs chips and software, while the parent company holds most of the patents and runs the licensing business under President and Chief Executive Officer Cristiano Amon.
Qualcomm is best known for Snapdragon, which is its brand of processors for smartphones, laptops, cars and connected devices. A Snapdragon processor is a complete system on a chip. It combines the central processor, graphics, artificial intelligence engine, camera support and modem for mobile network connection in one package. This integration helped Qualcomm power a large share of Android phones and Windows on Snapdragon laptops worldwide.
In recent years, Qualcomm has pushed beyond phones. The company now reports strong growth in automotive and Internet of Things (IoT) products, which connect vehicles, industrial machines and smart devices to the internet. Qualcomm has offices in India in Bengaluru and Hyderabad, and its diversification plan targets $40 billion in non handset revenue by fiscal 2029, with data centres as the largest part of that goal.
AI Chips: What They Are and Why They Matter
An artificial intelligence chipset is a special processor designed to run AI tasks quickly and with less power than a general processor. AI chips handle the heavy maths behind chatbots, image recognition, recommendations and voice assistants, and they matter because every AI response depends on fast, low cost and energy efficient computing.
AI work has two stages. Training is the learning stage, where a model studies large amounts of data over days or weeks and adjusts its internal settings. Inference is the working stage, where the finished model answers new questions in real time. Training happens once per model version, while inference happens billions of times each day, so inference decides the daily cost of running AI services.
The table below captures the difference in simple terms.
| Feature | Training Chips | Inference Chips |
|---|---|---|
| Main job | Teach a new model from data | Run the trained model to answer users |
| Key measure | Raw computing power and memory speed | Cost per answer, speed and power use |
| How often it runs | Once per model, for weeks | Continuously, for every user request |
| Typical hardware | Large graphics units costing about $30,000 to $40,000 | Graphics units, custom accelerators and chips like AWS Inferentia |
| Share of lifetime cost | About 10 to 20 percent | About 80 to 90 percent |
Industry estimates place inference at 60 to 70 percent of a nearly $400 billion AI accelerator market in 2026. This explains why Qualcomm, Amazon, Google and others now focus on custom AI chips for inference. The winner will not only be the fastest chip, but the chip that delivers each answer at the lowest power and cost.
What Is the Qualcomm Amazon Partnership for AWS?
The Qualcomm Amazon infrastructure partnership is a multi generation product collaboration announced on 8 September 2026 to build next generation AI data centre infrastructure for Amazon Web Services (AWS). AWS is Amazon’s cloud computing division, which rents computing power, storage and AI tools to companies and governments through large data centres around the world. Under the agreement, Qualcomm Technologies will develop customised silicon at scale for large AI data centres, with work centred on AI inference.
The collaboration has two technical tracks. The first is custom AI inference silicon built for AWS needs, rather than off the shelf chips. The companies have not disclosed the chip design, factory, process size, quantities or launch dates. The second is high performance optical connectivity reaching up to 1.6T, which means 1.6 terabits per second, with future generations to follow. Faster links between chips, servers and memory are critical because large AI models split work across thousands of chips that must constantly exchange data.
Qualcomm said the connectivity work will use its SerDes and optical DSP technologies. SerDes stands for serializer deserializer, which are circuits that convert data for high speed travel over wires and fibre. DSP stands for digital signal processor, which cleans and manages light signals inside fibre cables. Qualcomm strengthened this portfolio through its purchase of Alphawave Semi, a connectivity company completed in late 2025.
The financial structure is important to understand clearly. Qualcomm issued a warrant to Amazon.com NV Investment Holdings LLC, an Amazon affiliate, to buy up to 25 million Qualcomm shares at $161.26 per share. At that fixed price, the total value is about $4.03 billion. Only 3.75 million shares vested at issue based on initial purchase commitments. The remaining 21.25 million shares will vest in stages linked to commercial agreements, binding orders and actual purchases of Qualcomm server chip products, technology and systems.
Qualcomm disclosed in its regulatory filing that qualifying Amazon payments of up to $60 billion can count toward vesting during the warrant period ending on 3 September 2036. The $60 billion figure is therefore a ceiling for the incentive, not a confirmed order or guaranteed revenue. Qualcomm leadership said the work brings power efficient computing to AI infrastructure, while AWS leadership said the joint work aims to deliver more efficient and cost effective infrastructure for cloud customers.
As part of the expanded ties, Qualcomm plans to deepen its use of AWS tools, including Amazon Bedrock, for electronic design automation (EDA) workloads. EDA refers to the software used to design and test chips before manufacturing. Qualcomm expects this step to shorten chip design cycles.
Qualcomm AI Chips Portfolio and Data Centre Entry
Qualcomm AI chips began with the Cloud AI 100 family, which was built for low power AI inference in cloud and edge data centres. The original Cloud AI 100 card delivered up to 400 TOPS, which means trillion operations per second, at only 75 watts. The newer Cloud AI 100 Ultra offers up to 870 TOPS, 576 MB of on chip memory and 128 GB of device memory, and is tuned for generative AI and large language models. AWS has offered cloud instances powered by Qualcomm AI 100 accelerators since late 2023, so the Amazon relationship builds on an existing technical base.
In June 2026, Qualcomm unveiled a wider data centre roadmap under the Dragonfly brand at its Investor Day in New York. Dragonfly is Qualcomm’s platform for rack scale AI infrastructure, which means complete rows of servers designed to work as one large AI computer. The company set a target of more than $15 billion in annual data centre revenue by fiscal 2029.
The table below summarises the main Dragonfly building blocks disclosed by Qualcomm.
| Product | Type | Key Detail |
|---|---|---|
| Dragonfly C1000 | Data centre central processor | Planned for production in the second half of 2028, built for power efficient scale out computing |
| Dragonfly AI200 | Inference accelerator | Earlier rack level inference card for AI serving |
| Dragonfly AI250 | Inference accelerator with HBC Gen 1 | Sampling expected in mid 2027, with sharply higher memory speed per card |
| Dragonfly AI300 | Inference accelerator with HBC Gen 2 | Sampling expected in 2028, designed for large language and multimodal models |
| High Bandwidth Compute (HBC) | Memory and compute technology | Stacks compute close to memory to reduce data movement and energy per answer |
| Custom silicon | Tailor made chips for cloud buyers | Bespoke designs for hyperscalers, which are very large cloud operators like AWS |
Qualcomm has also moved on software and partnerships to support this push. Qualcomm completed the purchase of software company Modular in July 2026 to improve tools that let AI programs run across different chips. Qualcomm named Meta as the launch customer for the Dragonfly C1000 central processor for its next generation server fleet, and Microsoft endorsed its High Bandwidth Compute design. Reports in May 2026 also linked ByteDance, the owner of TikTok, to millions of custom Qualcomm chips for recommendation and AI agent workloads, though neither company disclosed final commercial terms.
How Custom Silicon and Optical Connectivity Will Work
Custom AI silicon means chips designed for one buyer’s exact workload instead of general sale. In the Qualcomm Amazon deal, Qualcomm will shape its power efficient processor designs around AWS requirements for AI inference, including model size, response time and energy limits. This approach differs from buying standard graphics units. It can lower the total cost of ownership, which is the full cost of buying, powering and running servers, and improve token economics, which is the cost of generating each unit of AI output.
Most modern data centre chips use the Arm instruction set or similar energy efficient designs. Arm is a British chip design company whose blueprints are licensed by Qualcomm, Amazon and many others. Qualcomm’s strength in low power mobile design comes from decades of work on Arm based Snapdragon systems, and the company now applies the same focus on performance per watt to servers. Performance per watt measures how much AI work a chip completes for each unit of electricity, a key limit in large data centres where power and cooling decide capacity.
The networking side is equally important. As AI clusters grow to hundreds of thousands of chips, data must move quickly between compute, memory and storage. Qualcomm will contribute serializer deserializer circuits and optical digital signal processors to lift link speeds to 1.6 terabits per second and beyond. These parts help AWS move large volumes of AI data inside its data centres with lower delay and power use. AWS already buys connectivity parts from suppliers such as Marvell and Broadcom, so Qualcomm becomes an additional source rather than a sole replacement.
The Qualcomm Amazon work does not replace Amazon’s in house chips. AWS has invested for years in Graviton central processors for general cloud tasks, Trainium accelerators for AI training and Inferentia accelerators for AI inference. Trainium2 powers EC2 Trn2 instances for large generative models, while the newer Trainium3 powers Trn3 UltraServers for advanced reasoning and video generation tasks. Qualcomm’s custom inference silicon will sit alongside these programs and give AWS customers another hardware option for serving models at lower cost.
Qualcomm vs Nvidia in AI Chips: Competitive Landscape
Nvidia AI chips dominate AI computing, with estimates placing Nvidia at more than 80 percent of AI accelerator revenue in 2026. Nvidia won the training phase with powerful graphics units, fast NVLink connections that join thousands of chips into one cluster, and CUDA, which is its widely used software platform for writing AI programs. Flagship training chips can cost $30,000 to $40,000 each, and CUDA creates strong loyalty because engineers and tools are built around it.
Inference changes the terms of competition. Inference does not always need a giant cluster. It needs many efficient endpoints that answer quickly at low cost. This lowers switching costs and lets cloud operators test specialised chips from Qualcomm, Cerebras, Groq and in house programs. Hyperscalers now build their own silicon to cut dependence on a single supplier. Google uses Tensor Processing Units, Amazon uses Trainium and Inferentia, Meta builds MTIA and Microsoft builds Maia.
The table below places the Qualcomm Amazon deal in this wider race.
| Player | Strength | Recent Move Relevant to the Deal |
|---|---|---|
| Nvidia | Leader in training and full AI systems | Continues to supply most training and inference graphics units |
| Qualcomm | Power efficiency and mobile scale | Custom inference silicon and 1.6T optics for AWS, plus Meta central processor deal |
| Broadcom and Marvell | Custom chip design services | Marvell signed a custom AI chip arrangement with Google, while Broadcom works with OpenAI on custom inference chips |
| AMD and Intel | Server processors and accelerators | Compete for data centre share as clouds diversify suppliers |
| AWS, Google, Meta, Microsoft | In house cloud chips | Reduce cost per answer and secure supply through own designs |
For Qualcomm, the Amazon name matters more than the near term volume. Qualcomm faces pressure as Apple shifts to in house modems, with Apple related chip revenue reported to fall sharply in late 2026. A multi generation AWS engagement helps Qualcomm prove that its inference roadmap can move from mobile heritage to hyperscale data centres, a step the company needs to reach its 2029 revenue goals.
Market Reaction and Broader Significance for India
The Qualcomm Amazon announcement drew close attention to the Qualcomm share price and the Amazon share price, two of the most tracked technology stocks globally. Qualcomm shares moved higher in the days after the 8 September 2026 disclosure, as investors viewed a hyperscale AWS win as validation for the data centre plan outlined in June 2026. The warrant price of $161.26 also became a reference point for the market, since Amazon can buy shares at that fixed level as purchases progress. Amazon investors focused on a different gain. More efficient inference and networking can lower the cost of running AI on AWS and support the cloud unit, whose custom chip business was reported at an annualised revenue run rate of more than $25 billion at the end of the June 2026 quarter.
For India, the significance lies in cloud cost, capacity and chip skills. AWS operates data centre regions in Mumbai and Hyderabad, and Indian startups, banks, schools and public agencies increasingly run AI applications on AWS. Cheaper and more power efficient inference can lower the price of chatbots, translation, fraud checks and video services that depend on continuous AI responses. AWS has reported a global average power usage effectiveness of 1.14, which measures how much total power a data centre uses for each unit delivered to servers, and more efficient chips and optics can improve that score further.
The deal also connects to India’s semiconductor and data centre growth. Qualcomm employs engineers in Bengaluru and Hyderabad, and wider adoption of Arm based, power efficient designs supports demand for chip design, testing and software roles. As AI use rises, Indian cloud regions will need more inference capacity that stays within power and cost limits. A successful Qualcomm AWS deployment would show how custom AI chips can expand that capacity without a matching rise in energy bills.
Key Takeaways
- Qualcomm Technologies and Amazon announced a multi generation AI infrastructure collaboration on 8 September 2026 for custom silicon for AWS data centres.
- Qualifying Amazon purchases of up to $60 billion through 3 September 2036 can trigger vesting under the warrant agreement, which sets a ceiling rather than a guaranteed order.
- Amazon received a warrant for up to 25 million Qualcomm shares at $161.26, with 3.75 million shares vested at issue and the rest tied to binding orders and purchases.
- The deal focuses on AI inference chips and optical connectivity up to 1.6 terabits per second using SerDes and optical DSP technologies.
- Qualcomm targets more than $15 billion in annual data centre revenue by fiscal 2029 through its Dragonfly portfolio, including the C1000 CPU and AI250 and AI300 accelerators.
- AWS will continue its in house Trainium, Inferentia and Graviton programs, while AWS regions in Mumbai and Hyderabad make cheaper inference directly relevant for Indian cloud users.
