亚马逊官方重申AWS自研芯片业务年化收入超250亿美元

Key takeaways
Amazon's custom chips business exceeded a $25 billion annual revenue run rate with triple-digit growth year over year.
Trainium3 delivers up to 40% better price-performance than Trainium2.
Graviton serves 98% of the top 1,000 EC2 customers with up to 40% better price-performance.
When Amazon acquired chip design company
Annapurna Labs
in 2015, the AI boom was still years away. The bet was straightforward: if you design chips specifically for cloud workloads rather than relying on general-purpose hardware, you can deliver better performance at lower cost.
A different bet on AI infrastructure
A decade later, that bet has become one of the fastest-growing businesses in Amazon’s history.
Amazon’s chip business
—spanning
Trainium
for AI training and inference and
Graviton
for general cloud computing (and increasingly, agentic AI)—all based on the Nitro System for security and networking—recently exceeded a $25 billion annual revenue run rate, growing triple-digit percentages year-over-year.
We're unusually well-positioned for this AI inflection given our leading price-performance chips in both AI (with Trainium) and CPUs (with Graviton).
Andy Jassy
Amazon CEO
Chronology of Amazon Designed Chips
The below chronology details the multiple generations of Amazon's purpose-built chips.
Filter by family
All
Nitro
Inferentia
Why Amazon builds its own AI chips
The
AI chip
market hasn’t historically been in a hurry to lower costs. Market leaders rarely disrupt themselves. Amazon chose a different path.
Rather than relying solely on off-the-shelf processors,
Amazon designs chips from the ground up
for specific workloads. The approach is vertically integrated: hardware and software engineers collaborate from chip design through server deployment, working backwards from the system to create silicon tailored for the workloads customers actually run.
AWS Trainium3 chip
The result is three chip families that serve distinct but complementary roles:
is purpose-built for AI training and inference—the computationally intensive work of teaching AI models and then running them at scale.
handles general cloud computing—the websites, applications, databases, and increasingly, the agentic AI workloads that power modern software. Graviton is used by 98% of our top 1,000
EC2
customers, and revenue commitments have increased nearly three times quarter over quarter.
Graviton5
, the latest generation, delivers up to 25% better performance than its predecessor, and is growing nearly two times faster than Graviton4 did.
powers the networking, storage, and security behind AWS cloud infrastructure.
As CEO Andy Jassy noted on the company’s
Q2 2026 earnings
call: "Our Chips revenue run rate is now over $25 billion. We're unusually well-positioned for this AI inflection given our leading price-performance chips in both AI (with Trainium) and CPUs (with Graviton). In addition to the two leading AI labs in the world in Anthropic and OpenAI making multi-year, multi-gigawatt commitments to Trainium, an increasing number of AI start-ups are also adopting Trainium.”
What Amazon’s chips mean for AI customers
For companies building with AI, the economics of chips matter enormously. Training a frontier AI model can require hundreds of thousands of chips running for weeks or months. Running that model at scale—answering millions of queries per day—demands efficient inference infrastructure. Even small improvements in price-performance translate into significant savings at scale.
The AI startups building beyond chatbots are choosing Amazon’s custom chips. Here's why.
Beyond Anthropic and OpenAI, a new wave of AI startups is choosing Trainium to train models that simulate the physical world.
That’s why the world’s leading AI labs and tech companies are making major commitments to Amazon’s silicon:
Anthropic
has committed
to using up to five gigawatts of current and future generations of Trainium chips to train and power its Claude models, which run on more than one million Trainium2 chips. Anthropic is also using tens of millions of Graviton cores to deliver scalable performance and cost efficiency across a broad range of generative AI workloads.
OpenAI
to consuming two gigawatts of Trainium capacity through AWS infrastructure to power its frontier models, beginning in 2027.
Meta
has signed
an agreement to deploy tens of millions of Graviton cores to power the CPU-intensive workloads behind its agentic AI efforts.
Uber
is using
Graviton to match riders with drivers in fractions of a second and piloting Trainium3 to train the AI models that make every ride smarter.
AI chips built to meet customer demands at unprecedented scale
Amazon Bedrock
, the company’s managed AI service used by hundreds of thousands of customers, runs most of its inference on Trainium. An increasing number of AI start-ups are also adopting Trainium as well, including unicorns like Neura Robotics (which chose Trainium for physical AI) and Odyssey, which chose Trainium to build world models that simulate physics, getting nearly twice the useful compute per dollar compared to alternatives. They join other startups like TwelveLabs, DeCart, Poolside, Karakuri, Matagenomi, NetoAI, and Splash music, as well as larger companies Uber, and
Pinterest
.
On the Graviton side, more than 130,000 customers use Graviton-based servers today. Over half of all new processing power added to
AWS
runs on Graviton chips.
AWS Graviton5 chip
The demand reflects a shift in how AI infrastructure gets built. As AI systems move from answering questions to taking actions—real-time reasoning, code generation, and multi-step task orchestration—the compute required pulls heavily on both AI accelerators and CPUs. Amazon is positioned across both.
But individual chips are only part of the story. Amazon connects them into increasingly powerful systems:
Trn3 UltraServers
pack up to 144 Trainium3 chips into a single integrated system, delivering up to 4.4 times more compute performance than Trainium2 UltraServers. This allows customers to train models in weeks instead of months.
Project Rainier
is the world’s largest AI computing cluster, purpose-built for training frontier models at scale. Project Rainier is running Anthropic's Claude models.
Energy efficiency:
Trainium3 delivers over five times higher output tokens per megawatt than Trainium2—efficiency that matters at scale for both cost and environmental impact.
Inside Project Rainier: One of the world’s largest AI compute clusters.
AI, silicon, and quantum: The future of Amazon’s chips business
Amazon’s chip roadmap continues to accelerate. Trainium4 is already in development. Graviton continues to evolve for the agentic AI era. And under Peter DeSantis, who now
leads a new organization spanning AI models
, custom silicon, and quantum computing, the company is integrating these technologies so they reinforce each other.
“One of the ways that we can give ourselves an advantage in how we build these models is by using our deep investments in chips to deliver both performance and cost efficiency that will allow us to differentiate our model development,” said Peter DeSantis, SVP, Amazon AI Models, Chips, & Quantum Computing.
It’s a strategy that bets on a simple premise: the companies that control their own silicon will set the pace for AI. With a chip portfolio spanning training, inference, and general computing, and a roadmap that’s already demanding customer interest years into the future—Amazon is building the infrastructure layer that
the next era of AI
will run on.
Learn more about
AWS silicon innovation
Trending news and stories
Amazon.com announces second quarter results
Q2 Earnings: Andy Jassy on why AWS is booming and what it means for customers
How Amazon Leo plans to connect mobile devices from space
Q2 Earnings: Andy Jassy on what's driving Amazon Ads growth
Related Tags
Artificial Intelligence
Amazon Custom Chips
Innovation
Share