Google Cloud Doubled While Amazon Undercuts AI Compute Costs

Google Cloud revenue doubled to $30.71 billion in the third quarter, growing 103% year over year — and a big part of the story is a chip business that now sits inside the cloud line. Alphabet (NASDAQ: GOOGL) books TPU system sales as Cloud product revenue, which means its custom silicon is no longer an internal tool. It's a product.

Amazon (NASDAQ: AMZN) posted cloud numbers of its own in the same earnings season. AWS grew 39% to $45.98 billion, its fastest pace in nearly five years. The two reports show two very different bets on custom silicon, and the gap between those bets is what makes this comparison worth watching.

Sundar Pichai said Google is "seeing strong demand for our TPU systems across a wide range of customers." The demand shows up in the profit line. Cloud operating income jumped to $11.12 billion from $3.59 billion a year earlier, proof that the business makes money at scale rather than just growing the top line.

What did the quarter reveal about each company's stack?

Alphabet's numbers point to a flywheel. Search query growth hit its highest rate in five years. The Gemini app reached one billion users. Each layer of the stack — chips, models, Search — sends demand to the others. That's the mechanism behind the full-stack play: TPU systems run Gemini, Gemini pulls users into Search, and Search revenue funds more TPU capacity.

Amazon's chip story runs through a different door. Andy Jassy said "most of Bedrock today is being run on top of Trainium." Bedrock is Amazon's multi-model marketplace, where AWS customers rent access to AI models from many labs. Trainium and Graviton exist to cut the cost of running other companies' models inside AWS, not to power a single first-party frontier system.

The savings are concrete. Customers using Amazon's chips saved more than $2 billion in Q3 alone. That's a strong reason for them to stay on AWS even as rival clouds pitch similar services.

A side-by-side look at the quarter makes the contrast sharp:

• Cloud growth: Alphabet 103%, Amazon 39%
• Cloud operating income: Alphabet $11.12B, Amazon $17.77B
• Chip showcase: TPU systems sold to outside customers vs. Trainium running most of Bedrock
• Model strategy: first-party Gemini vs. multi-model marketplace

Note the asymmetry. Google Cloud grew far faster but still earns less operating income than AWS. Amazon's slower-growth cloud remains the bigger profit machine — for now.

Why do the two chip strategies differ so much?

Google spent a decade building TPUs into its own platform for frontier training and native JAX scaling. The goal was vertical integration: own the best model and run it on the cheapest chip you control. That's the Alphabet thesis in one sentence.

Amazon designed its chips with a different target: breaking NVIDIA's (NASDAQ: NVDA) pricing power on inference. Inference — running trained models for customers — is where most AI compute dollars go, and Nvidia's grip on that market has let it charge premium prices. Trainium is Amazon's answer.

Jassy's view of the model market explains the strategy. "There is not one tool to rule the world," he said. He expects six to eight models to reach roughly the same quality. If he's right, AWS makes money whichever lab comes out ahead, because every serious model ends up on Bedrock running on Trainium. If Alphabet is right — that owning the best model on the cheapest chip wins — then Google's full stack takes the prize.

The honest answer is that nobody knows yet which thesis holds. Both are internally consistent, and both are expensive to pursue.

What is all this spending doing to the balance sheets?

The cost of these bets is now visible in hard numbers. Alphabet's capital expenditure hit $51.36 billion in the quarter. Its free cash flow fell to negative $9.07 billion. To help fund the buildout, the company sold new shares through its at-the-market (ATM) program — a move that dilutes existing holders.

Amazon is spending even more. Its capex reached $77.26 billion, with about $220 billion expected for 2026. Both companies are essentially converting cash into compute capacity on the assumption that AI demand keeps outrunning supply.

That's the market stake here. Two of the world's largest companies are pouring hundreds of billions into custom silicon, and the shape of the AI compute market — who owns the chips, who sets the prices — depends on which approach pays off.

What happens when Trainium leaves the building?

The most consequential development may still be ahead. Amazon plans to sell Trainium outside AWS late in 2027. That would put it in direct competition with Google's TPU sales for the first time, turning a quiet division of the market into an open fight.

Until then, margins are the number to watch on Amazon's side. Brian Olsavsky noted that AI margins run lower than core AWS margins, and today's 38.7% AWS operating margin could narrow as Bedrock grows. Faster revenue with thinner profit is a real trade-off, not a hypothetical one.

On Alphabet's side, the question is whether TPU customers keep spreading beyond Google's own products. Right now the flywheel mostly feeds itself; outside demand would prove the chip can stand alone.

Which stock looks more compelling?

The valuation gap is wide. Alphabet trades at about 13 times earnings; Amazon trades at about 37. Alphabet's multiple looks cheaper partly because big equity gains earlier in the year inflated its trailing earnings, so the figure deserves some caution rather than pure celebration.

The share prices tell a divergent story. Alphabet shares fell 18.06% over the past month, while Amazon is up 15.64% year to date.

The analyst behind the comparison finds Alphabet more compelling, because it has every layer — from hardware to a billion Gemini users. Amazon's appeal is different: it offers AI exposure without depending on a single winning model. That view could shift, the analyst notes, if Alphabet's share dilution speeds up or TPU demand stalls.

The bottom line: Alphabet is betting that owning the whole AI stack wins; Amazon is betting that cheap, flexible compute for everyone else's models wins. Trainium's external debut in late 2027 will be the first real test of both theses at once.