Custom Silicon Can't Break Alphabet and Amazon From Nvidia

Google Cloud revenue doubled to $30.71 billion last quarter, yet Alphabet still fills its data centers with Nvidia systems — and Amazon, which saved over $2 billion on custom silicon in Q3 alone, has committed to deploy another 2 million Nvidia GPUs through the second quarter of Nvidia's fiscal 2029. That is the uncomfortable arithmetic at the center of this earnings season: the two hyperscalers building the most aggressive in-house AI chip programs remain Nvidia's most dependable buyers.

Nvidia's own stock tells you the market is testing that dependence. Shares fell 18.29% over the past week and 26.34% over the past month, leaving the stock near 34x earnings — still richer than Alphabet's 13x, cheaper than Amazon's 38x. All three companies have now reported, and their numbers describe one connected business rather than three separate ones.

What do the earnings actually show?

The headline figures stack up fast:

• Google Cloud: revenue rose 103% to $30.71 billion, making it Alphabet's main growth engine. Alphabet's capex hit $51.36 billion.
• AWS: revenue rose 39% to $45.98 billion, with a 38.7% operating margin. Amazon's capex reached $77.26 billion.
• Nvidia: hyperscale revenue came in at $49 billion, with Q3 revenue guided to $108.0 billion plus or minus 2%.

Sundar Pichai pointed to "strong demand for our TPU systems across a wide range of customers." Nvidia, for its part, lists Google Cloud among the sites where its next-generation Vera Rubin platform is already in full production. Google sells its own silicon and runs Nvidia's hardware in the same buildings — simultaneously.

Amazon's position looks similar. Most of its Bedrock AI service runs on Trainium, its in-house accelerator, and CEO Andy Jassy said custom-silicon customers saved more than $2 billion in Q3 alone. Even so, the 2-million-GPU deployment commitment stands.

Why does custom silicon coexist with CUDA?

Jassy himself flagged the constraint: all advanced chips come from one foundry — Taiwan Semiconductor — and "everybody gets an allocation." That cap limits how fast Trainium can scale, no matter how much Amazon wants it to grow.

Jensen Huang framed the rivalry narrowly: "Many of these XPUs are inference-specific chips for one cloud or one service." His argument is that frontier model training demands the densest compute and fastest interconnects available, while most of the machine learning world still writes code for CUDA. Cloud providers, in his telling, have to stock the hardware developers actually ask for.

The financials show why each company tolerates the arrangement differently. For Amazon, Trainium mostly protects margins — and with AWS posting a 38.7% operating margin, that strategy is working. Alphabet's spending is harder to absorb: free cash flow came in at negative $9.07 billion, and the company issued $28.7 billion in new notes. Add an $8.70 billion loss on equity securities, and the strain is visible.

When could the dependence actually break?

Nvidia expects Vera Rubin to make up roughly 20% of data center revenue in the coming ramp, and Huang described his fiscal 2028 growth outlook of about 70% as limited by supply, not demand.

The test arrives in late 2027, when Amazon plans to start selling Trainium chips externally to other companies. If those sales — or Google's TPU sales — begin pulling real workloads away from CUDA, the calculus changes. Amazon's premium valuation also leans on Jassy's forecast of AWS growth above 40% in both 2027 and 2028, a bar that assumes the AI buildout keeps accelerating.

Until external Trainium sales gain genuine traction, the picture stays stubbornly the same: Alphabet and Amazon will keep putting Nvidia hardware in their data centers, and Nvidia will keep shipping to the two companies ostensibly trying to replace it.