Google and Amazon to Sell AI Chips Directly, Challenging Nvidia

Google and Amazon — Nvidia's two largest customers — are both preparing to sell their custom AI chips directly to enterprises. This marks a fundamental shift in the AI chip market that could reshape Nvidia's $4.9 trillion empire.

The critical detail isn't that these tech giants make chips. Google's TPU and Amazon's Trainium have existed for years. The game-changer is distribution: previously, customers could only access these chips by renting compute through Google Cloud or AWS. Now, both companies plan to sell hardware directly, letting enterprises install chips in their own data centers.

Google Confirms Direct TPU Sales This Year

Google CEO Sundar Pichai confirmed during the Q1 earnings call that Google will deliver TPU chips to 'select customers' this year for deployment in their own facilities. He added a caveat — 'the majority of revenue won't materialize until 2028' — but the timeline is now officially on the table.

Morgan Stanley has already begun modeling the potential impact. The signal is clear: Google isn't treating this as an experiment. It's a strategic business line with a multi-year roadmap.

Why This Changes Everything for Nvidia

Nvidia's extraordinary run over the past 3 years rests on a single premise: the entire AI industry depends on its chips. That premise is now under pressure from multiple directions simultaneously.

Here's what makes this moment different from previous 'Nvidia killer' narratives:

• Direct sales, not just cloud rental — Enterprises can now deploy Google and Amazon chips in their own data centers, directly competing with Nvidia's core market
• Proven silicon — TPU and Trainium aren't paper launches; they've been battle-tested at massive scale inside Google and Amazon's own infrastructure
• Ecosystem leverage — Both companies control major cloud platforms, AI frameworks, and developer communities that can accelerate adoption
• Cost incentives — Custom chips designed for specific AI workloads can offer better price-performance ratios than general-purpose GPUs
• Supply chain independence — Enterprises desperate for AI compute now have alternatives when Nvidia supply is constrained

From Monopoly to Multi-Polar Competition

The AI chip market is transitioning from what was essentially a single-vendor monopoly to a multi-polar competitive landscape. This shift has been building quietly, but the simultaneous moves by Google and Amazon signal an inflection point.

Nvidia still holds overwhelming market share in AI training and inference chips. Its CUDA software ecosystem remains a powerful moat — thousands of AI models and applications are optimized specifically for Nvidia hardware. Switching costs are real and significant.

But the economics are shifting. As AI infrastructure spending accelerates into hundreds of billions of dollars annually, even Nvidia's biggest customers are motivated to reduce their dependency. Building and selling custom chips isn't just a technical exercise — it's a strategic imperative for companies spending $50 billion or more per year on AI compute.

The Software Moat Is Eroding

Nvidia's most durable competitive advantage has always been software, not silicon. CUDA, built over nearly 2 decades, creates deep lock-in across the AI development stack. But this moat is showing cracks.

Google's JAX framework and its tight integration with TPUs offer a credible alternative for large-scale AI training. Amazon has invested heavily in making Trainium compatible with popular frameworks like PyTorch. The open-source Triton compiler, originally developed by OpenAI, is reducing CUDA dependency across the industry.

What Comes Next

This competition is just beginning. Pichai's 2028 revenue timeline suggests we're in the early innings of a multi-year transition. Several factors will determine how this plays out:

Nvidia's response will likely include more aggressive pricing, tighter integration with cloud partners, and accelerated product cycles. The company has navigated competitive threats before — AMD, Intel, and countless startups have tried and failed to dislodge it.

But this time is different. The challengers aren't scrappy startups. They're $2 trillion companies with captive demand, world-class engineering teams, and direct relationships with every major AI customer on the planet.

For enterprise buyers, the message is straightforward: the era of a single-source AI chip supply chain is ending. Planning for a multi-vendor future isn't just prudent — it's becoming inevitable.