A $60 billion Qualcomm-Amazon AI chip deal is the clearest sign yet that hyperscalers are racing to reduce their dependence on Nvidia. Here's how the AI chip war is shifting from training to inference — and why Nvidia's biggest customers are becoming its biggest competitors.
TL;DR
TL;DR — Qualcomm and Amazon announced a deal worth up to $60 billion for AI data-center chips, the latest sign that hyperscalers (Amazon, Google, Microsoft, Meta, even OpenAI) are all building custom silicon to reduce dependence on Nvidia — particularly for inference, not training. Nvidia's real moat is its CUDA software ecosystem, not just raw chip performance, and it's responding by opening its infrastructure to third-party chips via NVLink Fusion. The likely outcome isn't "Nvidia disappears" — it's "Nvidia stops being the only answer," as the AI chip market fragments into training, inference, edge, and specialized categories.
For the past few years, one company has been at the center of the artificial intelligence revolution: Nvidia. Its GPUs became the foundation on which many of the world's largest AI models were trained. OpenAI, Google, Microsoft, Meta, Anthropic, and countless AI startups have spent enormous amounts of money buying or renting Nvidia hardware.
But something important is changing. The biggest technology companies in the world are increasingly asking: why should we depend entirely on Nvidia?
Amazon is developing its own AI chips. Google has its TPUs. Microsoft is developing custom silicon. Meta is working on custom AI hardware. Qualcomm is aggressively entering data-center AI. Broadcom is helping major technology companies design custom accelerators. And now even OpenAI is developing its own AI chips.
The result: the AI-chip war is becoming much bigger than Nvidia vs AMD. It is becoming Nvidia vs an entire ecosystem of companies trying to reduce dependence on Nvidia. And the latest Amazon–Qualcomm deal shows just how serious that challenge is becoming.
On September 8, Qualcomm announced a major long-term partnership with Amazon. Under the agreement, Amazon could purchase up to $60 billion worth of Qualcomm AI data-center chips and related products. The deal also includes approximately $4 billion in Qualcomm stock warrants that Amazon can exercise under specified conditions. (Reuters)
This is not a small experiment. It's a huge signal that hyperscalers are actively looking for alternatives to Nvidia's dominant AI accelerators. And there's an important detail: the chips are particularly aimed at AI inference. That distinction matters.
When an AI company builds a model, it needs enormous computing power to train it: massive dataset → AI model → millions/billions of calculations → training → finished model.
But once the model exists, users start asking questions. Every time ChatGPT generates an answer, an AI coding assistant writes code, or an AI voice agent responds to someone, the model is performing inference. In simple terms: training teaches the model, inference uses the model. And inference could eventually represent an enormous portion of total AI computing demand — that's why companies are designing specialized chips specifically for it.
To understand the threat, you first need to understand Nvidia's advantage. Nvidia didn't simply build powerful GPUs — it built an entire ecosystem: GPUs, networking, CUDA, libraries, development tools, optimized AI frameworks, server platforms, software, and developer support.
The result is a powerful feedback loop: more developers → more Nvidia software adoption → more AI applications → more Nvidia demand → more investment → better hardware and software → more developers. This is why competing with Nvidia is so difficult — a competitor doesn't just need a faster chip, it needs an ecosystem.
One of Nvidia's strongest advantages is CUDA, its software layer that lets developers use Nvidia GPUs for general-purpose computing and AI workloads. Over many years, an enormous ecosystem of libraries, frameworks, kernels, tools, and optimization techniques has developed around it.
So if a company says "we'll build a chip that's 20% faster," that isn't necessarily enough — developers may still ask "can my existing software run on it?" This is why Nvidia's software moat is so important.
Amazon, Google, and Microsoft don't necessarily need to sell chips to everyone. They can build chips for themselves. That's a completely different strategy: instead of buying Nvidia hardware, deploying it, and paying Nvidia's margins, a hyperscaler can design specialized hardware, manufacture it, deploy it in its own cloud, and optimize its own software stack. If the chip is cheaper or more efficient, it can capture those savings across its enormous infrastructure.
Amazon has already developed its own AI accelerators under the Trainium family, plus Inferentia, designed specifically for inference. Reuters reported that Amazon's custom chip business had reached an annualized revenue run rate of more than $25 billion by the end of the June quarter. (Reuters) Amazon isn't experimenting with custom chips as a side project — it's becoming a major part of the company's AI infrastructure strategy.
Amazon isn't necessarily trying to do everything itself. Instead, the industry is moving toward a model where hyperscalers combine internal chip design with specialized external semiconductor partners. Qualcomm brings expertise in chip architecture, AI acceleration, power efficiency, networking, and advanced semiconductor design. Amazon brings AWS, data centers, customers, software, and enormous workloads. Put them together and you get custom AI infrastructure optimized for Amazon's cloud.
Qualcomm historically became famous for smartphone processors — its Snapdragon chips power many mobile devices. But it's aggressively expanding into PCs, automotive, edge computing, robotics, industrial systems, and data-center AI. The Amazon deal is particularly important because Qualcomm has said it wants to reach approximately $15 billion in data-center chip revenue by 2029. (Reuters) Qualcomm doesn't want to remain primarily a smartphone-chip company — it wants to become an AI infrastructure company.
Amazon isn't the only hyperscaler pursuing custom silicon. Google developed its own TPU (Tensor Processing Unit), specialized processors designed for machine-learning workloads, and uses them extensively within its own infrastructure — meaning it doesn't have to depend entirely on Nvidia to run Gemini and other AI workloads.
The custom-chip ecosystem is expanding further: in August, Google entered a major partnership with Marvell to develop custom AI chips, with Google receiving an option that could ultimately be worth approximately $12.2 billion in Marvell shares. (Reuters)
Microsoft has developed custom chips including Maia for AI workloads and Cobalt for general-purpose cloud computing. The strategic idea: if Microsoft can run certain workloads more efficiently on its own silicon, it can reduce infrastructure costs and improve performance. It doesn't need to replace every Nvidia GPU — only where custom hardware makes economic sense.
Meta is also developing custom AI infrastructure, investing heavily in AI accelerators and data-center hardware to serve Facebook, Instagram, WhatsApp, Meta AI, recommendation systems, and advertising systems — where even small efficiency improvements can translate into billions of dollars of infrastructure savings.
One of the most interesting developments is that OpenAI itself is also developing custom AI chips. Reuters reported in June that OpenAI unveiled a custom AI chip called Jalapeño, designed with Broadcom and manufactured by TSMC, targeting AI inference workloads. (Reuters)
This is strategically important: OpenAI is one of Nvidia's largest customers, yet even OpenAI is preparing to control more of its hardware stack. That tells you something about the direction of the industry.
If Nvidia is the obvious AI-chip winner, Broadcom may be one of the most interesting companies behind the scenes. Broadcom helps major technology companies design custom AI chips — its customers and partners include Google, Meta, OpenAI, and Anthropic. Instead of trying to beat every hyperscaler directly, Broadcom helps them build their own chips: the more companies that want alternatives to Nvidia, the more valuable custom-chip design becomes.
Nvidia isn't sitting still. One of its responses is particularly interesting: NVLink Fusion, technology that allows third-party/custom AI chips to connect directly into Nvidia's data-center systems. A recent example is d-Matrix, which plans to use Nvidia's NVLink Fusion to integrate its inference processors into Nvidia server systems. (Reuters)
This creates a fascinating possibility: Nvidia doesn't necessarily have to own every chip. It could become the infrastructure platform connecting the chips — much like Microsoft doesn't manufacture every component inside a PC but provides the operating system and software ecosystem. If that happens, Nvidia could remain extremely influential even if customers increasingly use non-Nvidia accelerators.
Instead of one universal AI processor, we could eventually have specialized hardware for training (huge models, large-scale computation), inference (chatbots, AI agents, search, voice), edge AI (phones, laptops, cameras, cars), robotics (real-time vision, motion control), and specialized AI (healthcare, scientific computing, finance, autonomous systems). This fragmentation could create opportunities for many semiconductor companies.
It's important not to overstate the threat. None of this means "Nvidia is finished" — quite the opposite. Nvidia remains one of the most important companies in AI infrastructure, with advantages in software ecosystem, developer adoption, performance, networking, scale, manufacturing relationships, and massive installed base. The custom-chip movement is better described as diversification — hyperscalers don't necessarily want to eliminate Nvidia, they want to avoid being completely dependent on it.
This mirrors what happened with CPUs: for decades Intel dominated PC processors, then companies developed alternatives and specialized architectures, and the market became much more diverse. AI chips could follow a similar trajectory — instead of one dominant architecture, a collection of architectures optimized for different workloads.
Ultimately, the AI-chip war will be decided by economics. If Nvidia hardware costs $X and a custom accelerator can perform the same workload for 0.6X, a hyperscaler has a huge incentive to switch. But if the custom chip costs 0.9X and requires completely different software infrastructure, the savings may not justify the transition.
Note
The critical metric isn't always "which chip is fastest?" It's how much does it cost to produce one useful AI response? That's the real battlefield.
This is perhaps the biggest shift. The early AI boom was dominated by models, then applications, and now increasingly infrastructure — chips, memory, networking, data centers, electricity, cooling, semiconductor manufacturing. Reuters recently reported that major hyperscalers including Alphabet, Amazon, Meta, Microsoft, and Oracle had issued roughly $220 billion in bonds over the previous year to help finance massive data-center expansion. (Reuters) AI is no longer just a software story — it's becoming an industrial-scale infrastructure story.
This is the most interesting strategic problem. Nvidia's biggest customers are also some of the world's biggest technology companies — Amazon (Trainium/Inferentia), Google (TPU), Microsoft (Maia), Meta (custom accelerators), OpenAI (custom inference chips) — and those same customers are building alternatives, giving them an incentive to reduce their Nvidia spending.
But Nvidia has one enormous advantage: those customers still need AI compute. Even if they build custom chips, the total amount of AI compute may grow so rapidly that Nvidia can continue expanding. The market can therefore grow even while Nvidia's share decreases.
It may be Nvidia + custom chips rather than Nvidia OR custom chips. A data center of the future could contain Nvidia GPUs, Google-style accelerators, Amazon accelerators, Qualcomm inference chips, specialized AI processors, and custom networking, all connected together — which is why interconnect technology is becoming increasingly important.
There is another bottleneck that deserves attention: AI processors aren't useful without fast memory, and large AI models require enormous amounts of memory bandwidth. That's why companies such as SK Hynix, Samsung, and Micron have become increasingly important to the AI ecosystem. The next AI hardware bottleneck may not always be the processor — it could be memory + networking + power. This is one reason OpenAI is also deepening relationships with Samsung and SK Hynix for AI memory supply. (Reuters)
This global chip competition is particularly important for India. As India develops its semiconductor ecosystem, it has an opportunity to participate in AI accelerator design, chip verification, advanced packaging, memory, networking, edge AI, and semiconductor software. India doesn't necessarily need to build a direct Nvidia competitor immediately — it could become a major part of the AI hardware supply chain, a much more realistic (and potentially enormous) opportunity.
Nvidia isn't facing a single "Nvidia killer." It is facing something potentially more complicated: an industry-wide movement toward AI hardware diversification. Amazon is building custom accelerators. Google has TPUs. Microsoft and Meta are developing their own silicon. Qualcomm is entering data-center AI with a potentially $60 billion Amazon opportunity. Broadcom is becoming a key partner for custom chip development. OpenAI is developing its own inference hardware. And Nvidia itself is adapting by building an ecosystem in which custom chips can work alongside its technology. (Reuters)
So the future probably isn't "Nvidia disappears." It's more likely: "Nvidia stops being the only answer." And that may actually be good for AI — more competition could mean cheaper AI, more computing capacity, more AI applications, more AI agents, and more innovation.
The next stage of the AI revolution may therefore depend less on who builds the smartest model — and more on who can make that intelligence cheap enough to run everywhere. The AI race is becoming a chip race. And the chip race has only just begun.
Nextgen Blog — YouTube Channel Owner
IIT Mandi CSE graduate and Full Stack Engineer with 6+ years of experience. Creator of the Nextgen Blog YouTube channel, covering AI, web development, and modern software engineering.
Discussion
Comments powered by Giscus
To enable comments, fill in GISCUS_REPO, GISCUS_REPO_ID, and GISCUS_CATEGORY_ID in components/blog/GiscusComments.tsx
