It sounds like the famous Cole Porter song “Let’s Do It, Let’s Fall in Love.” Google does it. Amazon does it. Microsoft does it. Like Porter’s birds, bees, and “educated fleas,” AI cloud producers are falling in love with designing new semiconductors to help power their data centers.
Their goal is clear: to reduce dependence on NVIDIA, which has become the world’s most valuable company by producing the semiconductors that power AI’s creation of images, videos, and 3D graphics. Yet when headlines shout about how Google’s new Tensor chips or Amazon’s custom silicon represent a “shot at NVIDIA,” they miss the mark. A close look reveals that these in-house chips will complement, not replace, the world’s most powerful semiconductor company.
AI computing depends on orchestrating a wide variety of chips. The cloud companies are adding their own general-purpose central processing units (CPUs) on top of NVIDIA’s graphics processing units (GPUs). As AI develops, that hybrid model will deepen. The cloud companies will co‑own parts of the AI layer, while NVIDIA continues to build the central engines, at least for the foreseeable future.
The lucrative GPU market is booming, estimated to grow from around $36 billion this year to a staggering $811 billion in 2035. NVIDIA holds well over 80% of that market, and its data center revenue is projected to rise from roughly $115 billion in 2025 to around $483 billion by 2030, even after factoring in growing use of complementary cloud provider chips. The only serious GPU competitor is AMD, another American company.
The cloud companies aim to lower costs and gain leverage, not to replace NVIDIA. Google is the clearest example. Its latest TPU 8t and 8i chips run Google’s own models and select partners’ workloads efficiently on Google Cloud; they are integrated with Google’s software stack and priced to undercut some NVIDIA chips, especially for inference. With TPU 8t/8i, Google gains a powerful cost‑control and bargaining tool and reduces its dependence on NVIDIA at the margin. But NVIDIA chips, enveloped in a proprietary software suite called CUDA, remain unavoidable.
Amazon Web Services is building perhaps the most complete custom chip portfolio. Its Arm-based Graviton CPUs do general compute, its Trainium trains AI models, and its Inferentia specializes in inference, using a trained model to make predictions, decisions, or generate outputs based on new, unseen data. Amazon designs these chips in‑house, while Taiwan Semiconductor Manufacturing Company (TSMC) manufactures them. AWS’s own materials stress that these chips complement, not replace, NVIDIA: they are offered alongside NVIDIA GPUs, and AWS explicitly presents custom silicon as a way to “push our custom silicon edge” while continuing partnerships with NVIDIA, AMD, and Intel.
Microsoft follows a similar pattern. Its Maia AI Accelerator and Arm-based Azure Cobalt CPU were developed to optimize the company’s Azure internal and customer AI workloads, using a “systems approach” that tailors everything from silicon to service. Microsoft is explicit that Maia “is not powerful enough to replace GPUs from NVIDIA for the purposes of developing large language models.” Instead, Maia chips are designed for inference and efficiency of specific high-volume tasks, not the full spectrum of AI workloads, reducing but not eliminating Microsoft’s need for NVIDIA and AMD accelerators.
GPUs will remain central for large‑scale parallel computation, while CPUs and “agent‑style” accelerators orchestrate and plan workflows. AI data centers count one CPU for every four GPUs; that ratio will move toward parity in the coming decades, according to a TrendForce analysis. Demand for CPUs and other AI chips will rise, but without a decline in the absolute number of GPUs deployed.
Although headlines may claim that Google is “taking a shot at NVIDIA,” the underlying reality is that the AI chip market is becoming layered and collaborative, with custom silicon acting as cost and bargaining levers that coexist with, rather than replace, NVIDIA’s dominant GPUs. AI chips of various colors and shapes will need to be mixed and matched to achieve optimal outcomes.
Christopher Cytera CEng MIET is a senior fellow with the Tech Policy Program at the Center for European Policy Analysis and a technology business executive with over 30 years’ experience in semiconductors, electronics, communications, video, and imaging.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
Tech 2030
A Roadmap for Europe-US Tech Cooperation
