Just two weeks later, on the Q1 2026 earnings call (April 29, 2026), Jassy went further, saying external sales could begin within two years and describing it as "a good chance" the company will do so . A Bloomberg report on June 18, 2026 added that Amazon is already in active negotiations to sell Trainium chips directly to companies for deployment in their own data centers, with a particular focus on international clients seeking localized data solutions .

This would mark a fundamental strategic pivot. Historically, Amazon only "leased" chip compute through AWS cloud instances; selling chips directly would transform the company from a chip consumer into a direct merchant of silicon .

Chip Business Financials

The chip division's financial trajectory underscores why Amazon is considering the move:

• $20 billion annual revenue run rate for the custom silicon division (Graviton + Trainium), growing at triple-digit year-over-year rates .
• $50 billion estimated standalone value: Jassy said on the Q1 2026 earnings call that if the chips business were a stand-alone company selling to both AWS and third parties, its annual run rate would be approximately $50 billion .
• $225 billion+ in customer commitments, and the chip division is described as "sold out two generations ahead" .

Demand is so intense that two large AWS customers asked if they could buy all of Amazon's Graviton instance capacity in 2026 — a request Amazon declined in order to preserve service availability across its broader customer base .

Performance Metrics and Competitive Advantages

Amazon has released specific performance data for two generations of Trainium:

• Trainium2: Delivers roughly 30% better price-performance than comparable GPU-based cloud instances and is "largely sold out" .
• Trainium3: Began shipping in early 2026. It offers an additional 30–40% improvement in price-performance over Trainium2 and is "nearly fully-subscribed" . Built on a 3nm TSMC process, it delivers 2.52 PFLOPS FP8 per chip. Amazon says its Trn3 UltraServers cost up to 50% less to run for inference than H100 clusters .
• Software ecosystem catch-up: Amazon is actively investing in its software stack (via AWS Neuron) to close the gap with Nvidia's CUDA, which remains a key adoption barrier .

Customer Traction

Amazon has secured commitments from some of the largest AI companies in the world:

• Anthropic: Amazon's most significant Trainium customer. Anthropic committed to spending over $100 billion on AWS over the next decade and locked in 5 gigawatts of Trainium computing capacity for training and deploying Claude across Trainium2, Trainium3, and Trainium4 generations . Over 500,000 Trainium2 chips are planned for the "Project Rainier" Anthropic cluster alone, with total deployments across all Amazon data centers exceeding one million chips . Three AWS campuses in final stages of construction boast over 1.3 GW of IT capacity for the sole purpose of serving Anthropic's training needs .
• OpenAI and Apple have also been reported as Trainium customers, with OpenAI locking in multi-gigawatt Trainium capacity .
• Uber: Reported as an AWS customer using Trainium-based inference .

Peter DeSantis and the Vertical Integration Vision

In December 2025 through January 2026, Amazon created a unified AI and Silicon organization under 27-year veteran Peter DeSantis, who now oversees AI models (Nova), custom silicon (Trainium/Inferentia), and quantum computing .

DeSantis's philosophy is explicit: he has described controlling "things down to the silicon" as a way to optimize AWS in ways that uniquely serve customers, arguing that tight integration of chip design, model architecture, and infrastructure delivers compounding advantages . In a June 2026 interview, he said AI needs "a couple more orders of magnitude" of improvement and that new model architectures beyond today's transformers will emerge, requiring deep co-design of hardware and software .

The core bet is that vertical integration — designing custom chips alongside Nova AI models and cloud infrastructure — creates a cost-performance moat that rivals like Nvidia (which sells a standard chip to everyone) cannot easily replicate .

Positioning vs. Nvidia

Jassy has directly compared Trainium to Nvidia GPUs. In his shareholder letter, he stated that "almost all AI development to date has utilized Nvidia, but customers are waking up to the fact that there's a better price-performance option" . He also noted that Trainium will save Amazon "tens of billions of capex dollars per year" and provide "several hundred basis points of operating margin advantage versus relying on others' chips for inference" .

Amazon continues to maintain a strong partnership with Nvidia — Nvidia GPUs remain widely available on AWS — but Trainium is increasingly positioned as the company's strategic differentiator for the long term .

The company's $225 billion customer backlog and $20 billion+ chip revenue run rate represent the most credible competitive challenge to Nvidia's near-monopoly in AI accelerators from any single company to date .

Just two weeks later, on the Q1 2026 earnings call (April 29, 2026), Jassy went further, saying external sales could begin within two years and describing it as "a good chance" the company will do so . A Bloomberg report on June 18, 2026 added that Amazon is already in active negotiations to sell Trainium chips directly to companies for deployment in their own data centers, with a particular focus on international clients seeking localized data solutions .

This would mark a fundamental strategic pivot. Historically, Amazon only "leased" chip compute through AWS cloud instances; selling chips directly would transform the company from a chip consumer into a direct merchant of silicon .

Chip Business Financials

The chip division's financial trajectory underscores why Amazon is considering the move:

• $20 billion annual revenue run rate for the custom silicon division (Graviton + Trainium), growing at triple-digit year-over-year rates .
• $50 billion estimated standalone value: Jassy said on the Q1 2026 earnings call that if the chips business were a stand-alone company selling to both AWS and third parties, its annual run rate would be approximately $50 billion .
• $225 billion+ in customer commitments, and the chip division is described as "sold out two generations ahead" .

Demand is so intense that two large AWS customers asked if they could buy all of Amazon's Graviton instance capacity in 2026 — a request Amazon declined in order to preserve service availability across its broader customer base .

Performance Metrics and Competitive Advantages

Amazon has released specific performance data for two generations of Trainium:

• Trainium2: Delivers roughly 30% better price-performance than comparable GPU-based cloud instances and is "largely sold out" .
• Trainium3: Began shipping in early 2026. It offers an additional 30–40% improvement in price-performance over Trainium2 and is "nearly fully-subscribed" . Built on a 3nm TSMC process, it delivers 2.52 PFLOPS FP8 per chip. Amazon says its Trn3 UltraServers cost up to 50% less to run for inference than H100 clusters .
• Software ecosystem catch-up: Amazon is actively investing in its software stack (via AWS Neuron) to close the gap with Nvidia's CUDA, which remains a key adoption barrier .

Customer Traction

Amazon has secured commitments from some of the largest AI companies in the world:

• Anthropic: Amazon's most significant Trainium customer. Anthropic committed to spending over $100 billion on AWS over the next decade and locked in 5 gigawatts of Trainium computing capacity for training and deploying Claude across Trainium2, Trainium3, and Trainium4 generations . Over 500,000 Trainium2 chips are planned for the "Project Rainier" Anthropic cluster alone, with total deployments across all Amazon data centers exceeding one million chips . Three AWS campuses in final stages of construction boast over 1.3 GW of IT capacity for the sole purpose of serving Anthropic's training needs .
• OpenAI and Apple have also been reported as Trainium customers, with OpenAI locking in multi-gigawatt Trainium capacity .
• Uber: Reported as an AWS customer using Trainium-based inference .

Peter DeSantis and the Vertical Integration Vision

In December 2025 through January 2026, Amazon created a unified AI and Silicon organization under 27-year veteran Peter DeSantis, who now oversees AI models (Nova), custom silicon (Trainium/Inferentia), and quantum computing .

DeSantis's philosophy is explicit: he has described controlling "things down to the silicon" as a way to optimize AWS in ways that uniquely serve customers, arguing that tight integration of chip design, model architecture, and infrastructure delivers compounding advantages . In a June 2026 interview, he said AI needs "a couple more orders of magnitude" of improvement and that new model architectures beyond today's transformers will emerge, requiring deep co-design of hardware and software .

The core bet is that vertical integration — designing custom chips alongside Nova AI models and cloud infrastructure — creates a cost-performance moat that rivals like Nvidia (which sells a standard chip to everyone) cannot easily replicate .

Positioning vs. Nvidia

Jassy has directly compared Trainium to Nvidia GPUs. In his shareholder letter, he stated that "almost all AI development to date has utilized Nvidia, but customers are waking up to the fact that there's a better price-performance option" . He also noted that Trainium will save Amazon "tens of billions of capex dollars per year" and provide "several hundred basis points of operating margin advantage versus relying on others' chips for inference" .

Amazon continues to maintain a strong partnership with Nvidia — Nvidia GPUs remain widely available on AWS — but Trainium is increasingly positioned as the company's strategic differentiator for the long term .

The company's $225 billion customer backlog and $20 billion+ chip revenue run rate represent the most credible competitive challenge to Nvidia's near-monopoly in AI accelerators from any single company to date .