A Mudança no Silício: Como a Amazon e o Google Estão Desafiando a Hegemonia da Nvidia na IA (AI)

For the past several years, the narrative of the artificial intelligence (inteligência artificial) revolution has been inextricably linked to a single hardware provider: Nvidia. Its H100 and upcoming Blackwell GPUs have been the currency of the AI realm—scarce, expensive, and absolutely essential. However, a significant shift is currently reshaping the landscape. At Creati.ai, we are observing a pivotal moment where major Cloud Service Providers (CSPs), specifically Amazon and Google, are transitioning from mere customers to formidable competitors.

By developing custom silicon—Amazon’s Trainium and Google’s Tensor Processing Units (TPUs)—these tech giants are not only reducing their reliance on Nvidia but are also generating billions in revenue and offering viable, high-performance alternatives for industry leaders like Anthropic. This evolution marks the beginning of a heterogeneous hardware era, challenging the "Nvidia tax" (imposto Nvidia) that has long dominated AI infrastructure economics.

AWS and the Rise of Trainium

Amazon Web Services (AWS) has aggressively pursued a strategy of vertical integration with its custom silicon lineup. While the company has long offered its Graviton processors for general-purpose computing, its recent focus has shifted sharply toward AI-specific acceleration through its Trainium (training) and Inferentia (inference) chips.

The Anthropic Alliance

The most significant validation of Amazon’s hardware strategy comes from its deepened partnership with Anthropic. As one of the world's leading AI labs, Anthropic requires immense massive compute power to train its Claude models. Historically, this would have required tens of thousands of Nvidia GPUs. However, AWS has successfully positioned its Trainium chips as a potent alternative.

Anthropic is now utilizing AWS Trainium 2 chips to build its largest foundation models. This is not merely a cost-saving measure; it is a strategic alignment. Trainium 2 is designed to deliver up to four times faster training performance and two times better energy efficiency compared to the first generation. For a company like Anthropic, where training runs can cost hundreds of millions of dollars, the efficiency gains offered by custom silicon translate directly into a competitive advantage.

Revenue Implications

The financial impact of this shift is profound. By moving workloads to its own silicon, Amazon retains margin that would otherwise flow to Nvidia. Furthermore, Amazon is turning its chip development into a revenue generator. Reports indicate that AWS is now generating billions of dollars in revenue from its custom chips de IA. This creates a flywheel effect: revenue from Trainium usage funds further R&D, leading to better chips, which in turn attracts more customers away from standard GPU instances.

Google's TPU Maturity and Ecosystem Lock-in

While Amazon is making waves with recent partnerships, Google has been the pioneer of custom AI silicon. Google introduced its Tensor Processing Units (TPUs) nearly a decade ago, initially for internal use to power Search, Photos, and later, the revolutionary Transformer models that birthed modern IA Generativa (Generative AI).

From Internal Utility to Public Cloud Powerhouse

Today, Google’s TPUs have matured into a robust platform available to Google Cloud customers. The introduction of the TPUs (specifically the sixth generation, Trillium) represents a massive leap in performance. Google has successfully demonstrated that its hardware can handle the most demanding workloads in the world. Notably, heavyweights like Apple have reportedly utilized Google’s TPU infrastructure to train components of their AI models, underscoring the reliability and scale of Google's custom silicon.

The Software Advantage: JAX and XLA

Google’s strength lies not just in the silicon but in the software stack. While Nvidia relies on CUDA, Google has built a deep integration between TPUs and JAX, a Python library used extensively for high-performance numerical computing. This software-hardware synergy allows for optimizations that are difficult to replicate on general-purpose GPUs. For developers deeply entrenched in the Google ecosystem, the switch to TPUs often brings performance-per-dollar benefits that Nvidia’s hardware, with its high markup, cannot match.

The Economic Imperative: Why the Market is Shifting

The dominance of Nvidia has created a bottleneck in the AI supply chain. The "Nvidia tax" (imposto Nvidia)—the premium paid for their market-leading GPUs—pressures the margins of every AI company, from startups to hyperscalers. The move by Amazon and Google to develop proprietary chips is driven by three critical factors:

1. Cost Control: Custom silicon allows CSPs to control their manufacturing costs and offer lower prices to end-users (or higher margins for themselves) compared to renting out Nvidia GPUs.
2. Supply Chain Independence: During the peak of the AI boom, obtaining H100s was nearly impossible. By controlling their own chip design, Amazon and Google reduce their vulnerability to external supply shortages.
3. Power Efficiency: As AI data centers consume an alarming amount of global electricity, chips designed specifically for a single cloud architecture (like Trainium or TPU) can be optimized for cooling and power usage more effectively than off-the-shelf GPUs.

Comparative Analysis: Custom Silicon vs. Nvidia

To understand the competitive landscape, it is essential to compare the current offerings of these tech giants against the industry standard.

Table 1: AI Hardware Landscape Comparison

The Software Barrier: Nvidia's Moat

Despite the impressive hardware specifications of Trainium and TPUs, Nvidia retains a massive defensive moat: CUDA. The Compute Unified Device Architecture (CUDA) is the software layer that allows developers to program GPUs. It has been the industry standard for over 15 years.

Most open-source models, libraries, and research papers are written with CUDA in mind. For Amazon and Google to truly break Nvidia's dominance, they must do more than build fast chips; they must make the software experience seamless.

AWS is investing heavily in its Neuron SDK to ensure that switching from a GPU to a Trainium instance requires minimal code changes. Similarly, Google is pushing XLA (Accelerated Linear Algebra) compilers to make models portable. However, inertia is powerful. For many engineering teams, the risk of migrating away from the battle-tested stability of Nvidia/CUDA to a cloud-specific chip is still a significant hurdle.

Future Outlook: A Fragmented but Efficient Future

The inroads made by Amazon and Google suggest that the future of AI hardware will not be a monopoly, but an oligopoly. Nvidia will likely remain the gold standard for research, development, and cross-cloud compatibility. However, for large-scale production workloads—where improving margins by even 10% translates to millions of dollars—custom silicon from AWS and Google will become the default choice.

At Creati.ai, we anticipate that 2026 will be the year of "Inference Economics." As the focus shifts from training massive models to running them (inference), the cost-per-token will become the most critical metric. In this arena, the specialized, low-power, high-efficiency chips like Inferentia and Google’s latest TPUs may well outpace Nvidia’s power-hungry GPUs.

The chip wars are no longer just about who has the fastest processor; they are about who controls the entire stack—from the energy grid to the silicon, up to the API endpoint. Amazon and Google have proven they are not just renting space in the AI revolution; they are building the foundation of it.