.png){kind=logo}
Google has turned TPUs into a major cloud asset, with TPU 8t for large models and TPU 8i for inference.
Amazon has scaled Trainium across huge AI clusters, with Anthropic and OpenAI as major future customers.
Meta uses MTIA chips for specific AI workloads, with newer versions focused on higher memory bandwidth and better inference performance.
Google, Amazon and Meta now treat custom processors as a core part of AI infrastructure. NVIDIA and AMD still hold roles, yet hyperscalers want tighter control over cost, power and memory. Small efficiency gains can save billions.
Google has the most mature custom chip effort. Its latest TPU family has two roles. TPU 8t targets large model work, while TPU 8i targets inference, reasoning tasks and agentic AI. Google says TPU 8t can scale to 9,600 chips in one superpod and deliver about 121 exaflops. TPU 8i targets low latency and claims 80% better performance per dollar than the prior generation.
Google has also moved beyond internal use. Google Cloud offers TPU access to outside customers. Google Cloud CEO Thomas Kurian said the TPU accelerator business has grown to more than twice the size of its nearest hyperscaler rival. Google also counts third-party TPU sales in Cloud revenue reports. The same chip can support Gemini, Google infrastructure and Cloud customers.
Also Read - AI Chips Explained: GPU vs NPU vs TPU
Amazon has built its own path through Trainium and Inferentia. Trainium targets AI model work, while Inferentia targets inference. Amazon says Trainium and Graviton have passed a USD 25 billion annualized revenue run rate. Trainium3 now sits in production, with up to 4.4 times the performance and four times the performance per watt of Trainium2 in the relevant UltraServer test.
Project Rainier has more than 500,000 Trainium2 chips, and Anthropic uses the system for Claude model work. Amazon says its total Trainium2 deployment has reached about 1.4 million chips. Anthropic has committed to as much as 5 GW of Trainium capacity, while OpenAI has committed to 2 GW through AWS from 2027.
On September 8, 2026, Amazon announced a multi-generation custom AI chip deal with Qualcomm, with an initial focus on inference. The potential purchase commitment could reach USD 60 billion over a decade, although that figure is a maximum rather than guaranteed revenue. Qualcomm aims for USD 15 billion in data center chip revenue by 2029.
Meta has developed MTIA, a custom AI accelerator for recommendation, rank and generative AI workloads. The MTIA 300, 400, 450 and 500 chips cover recommendation models, generative AI and inference. Meta says hundreds of thousands of MTIA chips are in production.
Memory has a major role. MTIA 450 doubles HBM bandwidth versus MTIA 400, while MTIA 500 adds another 50% increase. Meta has expanded its Broadcom partnership for several MTIA generations. Reuters reported that Meta planned to put a new AI chip, codenamed Iris, into production in September 2026.
General-purpose GPUs support many workloads. A hyperscaler can create an ASIC around a smaller set of known tasks. That approach can remove unused features and tune compute, memory access, precision and data movement for specific models.
Power consumption adds another major concern. AI data centers need vast amounts of electricity, so performance per watt has direct financial value. Google says its Ironwood TPU offers about 3.7 times better compute carbon intensity than TPU v5p. Modern AI systems also need high-bandwidth memory, large capacity and fast chip links.
Also Read - Why is Elon Musk Spending $16.8 Billion on AI Chips
Custom silicon does not mean the end of NVIDIA. GPUs still offer broad model support, mature software, CUDA and fast access to new AI systems. Custom ASICs fit best for huge, stable workloads with strict cost or power targets.
A hybrid AI data center is the likely result. Meta can use NVIDIA, AMD and MTIA. AWS can combine NVIDIA, Trainium, Inferentia, Graviton and partner chips. Google can offer NVIDIA GPUs beside TPUs and Arm CPUs.
The next test will come from inference. Google must prove that TPU sales can grow into a major external chip business. Amazon must expand Trainium use across more customers. Meta must prove that MTIA 450 and 500 can deliver strong economics at scale.
The AI chip market now points toward choice. NVIDIA can remain the flexible workhorse, while custom processors can handle high-volume workloads with tighter cost and power targets. That split may reshape AI infrastructure economics more than any single chip launch.
1. Why are Google, Amazon and Meta building AI chips?
Custom chips can lower AI costs, reduce power use and give each company greater control over its infrastructure.
2. What is Google’s custom AI chip called?
Google develops Tensor Processing Units, or TPUs. TPU 8t targets large model work, while TPU 8i focuses on inference and agentic AI.
3. What are Amazon’s Trainium chips used for?
Trainium chips support AI model work at large scale. Amazon has deployed Trainium2 across major clusters, including Project Rainier.
4. What is Meta’s MTIA chip?
Meta Training and Inference Accelerator, or MTIA, is Meta’s custom processor for recommendation, ranking and generative AI workloads.
5. Will custom AI chips replace NVIDIA GPUs?
Not entirely. NVIDIA remains important for flexible AI workloads, while custom processors can offer stronger economics for large, predictable workloads.
