Alphabet and Amazon’s Custom Silicon Bets Face a Margin Test

By Tom Whitfield

Google Cloud revenue rose 103% to $30.71 billion in the third quarter, its fastest growth in years, while AWS revenue increased 39% to $45.98 billion. Those results place Alphabet’s TPUs and Amazon’s Trainium and Graviton processors at the center of two different cloud strategies.

Alphabet now records external TPU system sales as Google Cloud product revenue, so the chip business contributes directly to the cloud line. Sundar Pichai said Google was “seeing strong demand for our TPU systems across a wide range of customers.” Google Cloud operating income rose to $11.12 billion from $3.59 billion, giving the business substantial scale.

What changed in the third quarter?

Search query growth reached its highest rate in five years, and the Gemini app reached one billion users. Those services support demand for Google’s own TPUs, while Google Cloud can sell TPU systems to outside customers. In short, Alphabet’s internal workloads and external chip sales reinforce each other.

AWS delivered its fastest growth pace in nearly five years, reaching $45.98 billion. Andy Jassy said: “most of Bedrock today is being run on top of Trainium.”

Customers using Amazon’s chips saved more than $2 billion in Q3, giving AWS a direct reason to retain those workloads. AWS operating income reached $17.77 billion. Both businesses grew quickly, but Alphabet now books outside TPU system sales inside its cloud revenue.

How do the chip strategies differ?

Alphabet spent a decade building TPUs into a platform for frontier training and native JAX scaling. The chips run behind Gemini and Search, linking model development, software and cloud infrastructure. Its approach joins the model and the silicon that serves it.

Amazon designed Trainium to weaken Nvidia’s pricing power in inference, while Trainium and Graviton cut the cost of running other companies’ models in AWS. Bedrock offers a multi-model marketplace rather than centering the service on one first-party model. Andy Jassy said: “most of Bedrock today is being run on top of Trainium.”

Customers using Amazon’s chips saved more than $2 billion in Q3, giving AWS a direct reason to retain those workloads. Jassy expects six to eight models to reach roughly equal quality and argues that no single tool will dominate.

If that forecast holds, AWS can earn whichever model leads; Alphabet benefits most when the best model runs on the cheapest proprietary chip. The strategic split is plain: Alphabet controls one integrated stack, while Amazon competes across a multi-model marketplace.

Where does capital spending create pressure?

Both strategies require heavy investment, and the reported cash numbers show why. Alphabet spent $51.36 billion on capital expenditures, while free cash flow fell to negative $9.07 billion, and its at-the-market share sales diluted existing holders. Amazon spent $77.26 billion on capital expenditure and expects roughly $220 billion for 2026.

That spending supports different businesses: Alphabet ties hardware to Gemini, Search and external TPU sales, while Amazon builds AWS services and prepares a broader Trainium channel. Neither model is capital-light, and both carry execution risk as spending rises. The financial takeaway is simple: Alphabet adds cash and dilution pressure, while Amazon carries the larger planned budget.

Can Trainium become a product beyond AWS?

Amazon plans to sell Trainium outside AWS in late 2027, placing it in direct competition with Google’s TPU system sales. The announced target gives no shipment volume or unit pricing, so the scale of that business remains unconfirmed. Its commercial appeal rests on the inference economics supporting Bedrock: lower compute costs can make model services more valuable to customers.

Brian Olsavsky noted that AI margins run below core AWS margins, while AWS’s current operating margin stands at 38.7%. That margin could narrow as Bedrock grows, and Alphabet’s corresponding test is whether TPU customers keep buying beyond Search and Gemini. The practical test is whether outside demand can turn an internal cost tool into a product without eroding margins.

Why does the analysis favor Alphabet?

The analysis places Alphabet at about 13 times earnings, compared with 37 times for Amazon. It cautions that earlier-year equity gains inflated Alphabet’s trailing earnings, making the lower multiple less decisive. Alphabet shares fell 18.06% over the past month, while Amazon gained 15.64% year to date.

Despite that caution, the analysis favors Alphabet because it owns every layer from hardware to a Gemini app with one billion users. Amazon offers AI exposure without depending on one model to win. The preference could change if Alphabet’s dilution accelerates or TPU demand stalls, so the conclusion remains an analyst judgment rather than a settled result.

What comes next?

The next checkpoints are concrete: late-2027 Trainium sales outside AWS, continued TPU adoption beyond Alphabet’s own services, and AWS margin pressure as Bedrock grows. Amazon’s roughly $220 billion of expected 2026 capital spending raises the stakes on execution. Those tests will show whether proprietary silicon can convert lower inference costs into durable pricing power without eroding cloud profit.