Alphabet and Amazon are turning custom silicon into a more direct challenge to Nvidia’s AI infrastructure business, but their Q2 2026 results show two very different routes to market: Google is selling efficiency, while AWS is selling capacity and choice.
As reported by 24/7 Wall St., Google Cloud revenue rose 82% year over year to $24.77 billion, while AWS grew 37% to $42.23 billion. The figures matter beyond earnings season because both companies are increasingly pairing their AI services with in-house processors rather than treating Nvidia GPUs as the only viable foundation for large-scale model training and inference.
Alphabet’s Tensor Processing Units have long powered Google’s own AI workloads, including Gemini. The important shift in Q2 was commercial: Alphabet began recognizing revenue from TPU systems delivered directly into customer data centers.
That creates a new sales motion alongside Google Cloud’s hosted TPU offerings. Organizations with residency, latency, or operational-control requirements can potentially use Google-designed AI infrastructure on premises rather than move every sensitive workload into a public-cloud environment.
Google Cloud’s reported 35.6% operating margin, up from 20.7% a year earlier, is the financial case for that vertical integration. Owning the accelerator, networking, systems software, model stack, and cloud service can lower the cost of serving AI workloads—at least when utilization stays high.
Alphabet has also raised its 2026 capital-expenditure outlook to $195 billion to $205 billion. The company is therefore not claiming that custom silicon eliminates the infrastructure spending problem; it is betting that its TPU stack makes each dollar of infrastructure more productive over time.
AWS has positioned Trainium as a capacity and cost alternative for customers that want something other than GPU-heavy instances. Its commercial advantage is distribution: major AI labs can reserve large blocks of AWS capacity while enterprise customers can consume managed models through Amazon Bedrock.
Amazon has said Graviton5 provides up to 25% better compute performance than its predecessor, and the processor is now available through EC2 M9g instances. For Windows and infrastructure teams, Graviton remains a more qualified proposition than a transparent x86 replacement: application architecture, Windows-on-Arm support, third-party agents, drivers, and licensing all need validation. But for Linux-based services, containers, build farms, and cloud-native workloads, Graviton can be a practical lever for reducing compute cost.
The competitive pressure is nevertheless real. Google now sells access to both TPUs and Nvidia hardware, while AWS offers Nvidia GPUs alongside Trainium and Inferentia. For customers, that means the immediate outcome is less likely to be a wholesale GPU migration than more negotiating power, more workload-specific choices, and a stronger reason to avoid binding every AI project to one accelerator architecture.
The unresolved test is cash flow. Alphabet spent $44.92 billion on capital expenditures in the quarter, according to 24/7 Wall St., while Amazon spent $54.21 billion. Google’s TPU system revenue is expected to become more meaningful in 2027; AWS must show that its large Trainium commitments translate into durable margins rather than simply more expensive capacity build-outs.
Google turns TPUs into a product, not just an internal advantage
Alphabet’s Tensor Processing Units have long powered Google’s own AI workloads, including Gemini. The important shift in Q2 was commercial: Alphabet began recognizing revenue from TPU systems delivered directly into customer data centers.That creates a new sales motion alongside Google Cloud’s hosted TPU offerings. Organizations with residency, latency, or operational-control requirements can potentially use Google-designed AI infrastructure on premises rather than move every sensitive workload into a public-cloud environment.
Google Cloud’s reported 35.6% operating margin, up from 20.7% a year earlier, is the financial case for that vertical integration. Owning the accelerator, networking, systems software, model stack, and cloud service can lower the cost of serving AI workloads—at least when utilization stays high.
Alphabet has also raised its 2026 capital-expenditure outlook to $195 billion to $205 billion. The company is therefore not claiming that custom silicon eliminates the infrastructure spending problem; it is betting that its TPU stack makes each dollar of infrastructure more productive over time.
AWS bets on a wider commercial funnel
Amazon’s chip portfolio is broader in purpose. Trainium targets AI training and inference, Graviton provides Arm-based general-purpose compute, and Nitro underpins AWS’s virtualization and security architecture.AWS has positioned Trainium as a capacity and cost alternative for customers that want something other than GPU-heavy instances. Its commercial advantage is distribution: major AI labs can reserve large blocks of AWS capacity while enterprise customers can consume managed models through Amazon Bedrock.
Amazon has said Graviton5 provides up to 25% better compute performance than its predecessor, and the processor is now available through EC2 M9g instances. For Windows and infrastructure teams, Graviton remains a more qualified proposition than a transparent x86 replacement: application architecture, Windows-on-Arm support, third-party agents, drivers, and licensing all need validation. But for Linux-based services, containers, build farms, and cloud-native workloads, Graviton can be a practical lever for reducing compute cost.
Nvidia is still the compatibility baseline
Neither TPU nor Trainium is a drop-in Nvidia replacement. Nvidia’s CUDA ecosystem, mature enterprise tooling, broad OEM support, and installed software base remain powerful reasons to standardize on its GPUs.The competitive pressure is nevertheless real. Google now sells access to both TPUs and Nvidia hardware, while AWS offers Nvidia GPUs alongside Trainium and Inferentia. For customers, that means the immediate outcome is less likely to be a wholesale GPU migration than more negotiating power, more workload-specific choices, and a stronger reason to avoid binding every AI project to one accelerator architecture.
The unresolved test is cash flow. Alphabet spent $44.92 billion on capital expenditures in the quarter, according to 24/7 Wall St., while Amazon spent $54.21 billion. Google’s TPU system revenue is expected to become more meaningful in 2027; AWS must show that its large Trainium commitments translate into durable margins rather than simply more expensive capacity build-outs.