# Amazon Chip Business Hits $25B Run Rate _Amazon custom silicon business passes $25B annual run rate as AWS Trainium3 and Graviton5 capture massive commitments from Anthropic and OpenAI._ **Published:** 2026-07-31 **Source:** https://www.startuphub.ai/hardware/amazon-chip-business-hits-25b-run-rate --- [Amazon (NASDAQ:AMZN)](https://www.google.com/finance/quote/AMZN:NASDAQ) reached a $25 billion annual revenue run rate for its custom silicon business, logging triple-digit percentage growth year over year. The milestone validates a decade-long infrastructure bet that started when AWS bought Annapurna Labs in 2015, according to [Amazon News](https://www.aboutamazon.com/news/aws/amazon-ai-chips-business-history?utm_source=rss). AWS Custom SiliconCore Amazon's decade-long infrastructure bet started with Annapurna Labs acquisition in 2015From the article 3 mentionsAmazon (NASDAQ:AMZN) reached a $25 billion annual revenue run rate for its custom silicon business, logging triple-digit percentage growth year over year.Unified LeadershipContextunified leadership drives next-generation chip development and market adoption$25B Run RateOutcomeFrom the articleAmazon (NASDAQ:AMZN) reached a $25 billion annual revenue run rate for its custom silicon business, logging triple-digit percentage growth year over year.Graviton CPUsCoreFrom the article 5 mentionsGraviton CPU cores now serve 98% of the top 1,000 EC2 customers.Trainium AI ChipsCoreTrainium3 and Graviton5 capture massive commitments from Anthropic and OpenAIFrom the article 9+ mentionsAnthropic committed to using up to five gigawatts of Trainium capacity, powering its Claude models on over one million Trainium2 chips while using tens of millions of Graviton cores.Cost AdvantageEffectFrom the articleDesigning custom processors gives AWS a direct cost advantage over cloud rivals relying on third-party hardware.Performance GainsEffectFrom the article 2 mentionsGraviton5 delivers up to 25% better performance than Graviton4, with quarterly revenue commitments jumping nearly three times quarter over quarter.Big Tech CommitmentsOutcomeAnthropic and OpenAI commit gigawatts, validating Amazon's custom silicon strategy Designing custom processors gives AWS a direct cost advantage over cloud rivals relying on third-party hardware. StartupHub.ai data rates Amazon at 81/100, noting verified raised capital of $10 billion in 2025. Among platform competitors tracked by StartupHub.ai, Wayfair holds a score of 85/100, while Coupang sits at 50/100. ## Silicon Stack from Graviton to Trainium Amazon's chip portfolio spans three primary hardware lines tied together by the Nitro System for networking and security. Graviton CPU cores now serve 98% of the top 1,000 EC2 customers. Graviton5 delivers up to 25% better [performance](/ai-news/technology/2026/amazon-s-chip-business-surges-past-20b) than Graviton4, with quarterly revenue commitments jumping nearly three times quarter over quarter. AI workloads run on Trainium accelerators. The AWS Trainium3 chip delivers up to 40% better price-performance than Trainium2, along with five times higher output tokens per megawatt. Trn3 UltraServers pack up to 144 AWS Trainium3 chips into a single integrated system, achieving 4.4 times more compute performance than previous server racks. Amazon Bedrock runs most of its managed model inference on Trainium hardware. Over half of all new processing power added across AWS runs on Graviton chips. ## Big Tech and AI Unicorns Commit Gigawatts Major AI research labs and tech giants are locking in long-term capacity on AWS hardware. Anthropic committed to using up to five gigawatts of Trainium capacity, powering its Claude models on over one million Trainium2 chips while using tens of millions of Graviton cores. Anthropic's models run on Project Rainier, one of the world's largest AI compute clusters. [Meta (NASDAQ:META)](https://www.google.com/finance/quote/META:NASDAQ) signed an agreement for tens of millions of Graviton cores to run CPU tasks for agentic AI. OpenAI committed to two gigawatts of Trainium capacity starting in 2027. Physical AI and world-building startups are also joining the platform. Odyssey gets nearly twice the useful compute per dollar on Trainium for physical simulation models. Neura Robotics chose Trainium for physical AI models, joining startups like TwelveLabs, DeCart, Poolside, Karakuri, Matagenomi, NetoAI, and Splash music. Enterprise users are migrating core operations as well. [Uber (NYSE:UBER)](https://www.google.com/finance/quote/UBER:NYSE) uses Graviton to match riders with drivers and is piloting AWS Trainium3 for model training. [Pinterest (NYSE:PINS)](https://www.google.com/finance/quote/PINS:NYSE) runs workloads across the custom hardware portfolio. ## Unified Leadership and the Next Generation Amazon CEO Andy Jassy highlighted the momentum during the Q2 2026 earnings call, citing strong adoption across frontier model builders and early-stage startups. Senior Vice President Peter DeSantis now leads an integrated organization combining AI models, custom silicon, and quantum computing. Work on Trainium4 is already underway. Controlling the entire hardware stack allows AWS to adjust chip architecture directly to model demands rather than waiting for general-purpose chip releases. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.