Last week handed us two Amazon announcements that, read together, tell a story the industry has been watching build for three years. On June 10, AWS made Graviton5 generally available — a 192-core ARM processor purpose-built for agentic AI, the most ambitious CPU Amazon has ever shipped. Then, on June 18, Bloomberg reported that Amazon is in active talks to sell its Trainium AI accelerators directly to outside data centers, bypassing AWS entirely. Taken alone, each development is interesting. Together, they are a signal that Amazon has decided it is a chip company now, not just a cloud company that makes chips to keep its Nvidia bill down.
What Graviton5 Actually Is
I want to spend a minute on the specs because they matter operationally. Graviton5 packs 192 ARM Neoverse V3 cores into a single package using a four-chiplet design on TSMC's 3nm process. That's not a small step — the previous generation, Graviton4, topped out at 96 cores. Amazon also gave Graviton5 192 MB of L3 cache, five times larger than its predecessor, along with DDR5-8800 memory and PCIe Gen 6 connectivity. Inter-core latency dropped 33%.
The official performance claims: 25% higher overall compute throughput, 35% faster web application workloads, 35% faster ML inference, and 30% faster database performance versus Graviton4. Tom's Hardware found the chip competitive with high-end AMD EPYC and Intel Xeon parts in cloud configurations. For a processor designed in-house rather than sourced from established silicon vendors, that is a remarkable result.
Amazon is calling Graviton5 "purpose-built for the agentic AI era." That phrase is doing real work in the positioning. The enlarged L3 cache and lower inter-core latency directly address the memory-access patterns of transformer-based inference — lots of KV cache shuffling, attention computations that thrash on memory bandwidth, long context windows that punish high latency. The 192-core design also means you can run more concurrent agent sessions on a single instance before threads start waiting on each other. Graviton5 is available today in M9g and M9gd instances, with compute-optimized C9g and memory-optimized R9g variants coming later this year.
The Trainium Move Is the Bigger Story
Graviton5 is AWS infrastructure getting better, which is normal and expected. The Trainium external-sales conversation is something else entirely.
According to Bloomberg's June 18 report, Amazon is in talks to sell Trainium AI accelerators directly to external data center operators. Peter DeSantis, Amazon's AI chief, told Bloomberg: "We view AI infrastructure as rapidly evolving. And we're constantly looking at ways to get to more customers." CEO Andy Jassy had telegraphed this direction on the Q1 2026 earnings call in April, saying there was "a good chance" Amazon would offer Trainium beyond AWS "in the next couple of years." The Bloomberg reporting suggests those discussions are moving faster than the cautious earnings-call language implied.
Consider what this means practically. Right now, if you want to use Trainium, you use AWS. You consume it as a cloud service. Trainium is not something you can buy and rack yourself. If that changes — if Amazon starts selling Trainium to data center operators, hyperscalers in other regions, sovereign cloud builds — the competitive dynamics of the AI accelerator market shift meaningfully. Not immediately, not decisively, but the direction of travel becomes clear.
DeSantis argued that external sales will not hurt AWS revenue because "there's so much underconsumption in AI" that demand will absorb both channels. That framing is probably right in the short term. But the longer-term implication is that Amazon is willing to compete in the market for AI silicon directly, not just offer it wrapped in cloud services. That puts them in the same conversation as Nvidia, which has been selling H100s and H200s to anyone with a purchase order and a data center.
How We Got Here
I've been watching this progression since the original Graviton launched in 2018. At the time, the story was about cost savings — AWS making its own chips to run commodity workloads cheaper, reducing exposure to Intel's pricing power. Graviton2 in 2020 started to actually perform, not just save money. By Graviton4, the chip was genuinely competitive with the x86 parts it replaced, and AWS could credibly market it on performance rather than price alone.
Trainium followed the same arc. Trainium1 was Amazon's answer to Nvidia's dominance in ML training — get some cost leverage, reduce dependency. Trainium2 started to compete on actual throughput. And now Trainium3 is reportedly nearly sold out. Amazon has signed commitments totaling over $225 billion in Trainium revenue. OpenAI has agreed to roughly two gigawatts of Trainium capacity through AWS. Anthropic has committed to up to five gigawatts of current and future Trainium chips. Those are not hedge purchases. Those are primary workloads.
The custom silicon business — Trainium, Graviton, and the Nitro security chip — crossed $20 billion in annual revenue run rate in Q1 2026, growing at over 100% year-over-year. Jassy said on the earnings call that if this business were standalone, selling chips to outside customers as traditional chip companies do, the annual revenue run rate would be $50 billion. Amazon deployed over 2.1 million AI chips in the past 12 months alone.
What This Means If You Run Infrastructure
A few things worth paying attention to:
- Graviton5 deserves a serious look for agentic workloads. If you're running LLM inference — especially multi-agent pipelines with lots of concurrent sessions — the combination of 192 cores, the huge L3 cache, and DDR5-8800 memory bandwidth makes M9g instances worth benchmarking against your current GPU-backed inference setup. GPU inference wins at high request rates with good batching. CPU inference at the right price-latency tradeoff is often better for lower-concurrency, latency-sensitive endpoints. Graviton5 sharpens that tradeoff considerably.
- Trainium commitments signal a maturing alternative to Nvidia for training. If you're designing a training pipeline and evaluating H100/H200 versus Trainium, the OpenAI and Anthropic commitments are directionally meaningful. The ecosystem is maturing. The software compatibility story is still behind CUDA, but the gap is closing faster than it was 18 months ago.
- If external Trainium sales materialize, on-premise AI gets more interesting. Financial services, government, healthcare — organizations that can't or won't run sensitive training workloads in public cloud — would gain a meaningful new option if they can buy Trainium hardware and operate it themselves. Watch this space over the next 12 to 18 months.
The Nvidia Question
I want to be direct: Nvidia is not in trouble. H100 and H200 demand remains backlogged. Blackwell is shipping. The CUDA software ecosystem is still the dominant surface area for AI compute, and that moat does not dissolve quickly. The threat from Amazon's silicon — and from Google's TPUs, Microsoft's Maia, Meta's MTIA — is real but long-cycle.
What Amazon is building changes the ceiling on alternatives, not the current market-share numbers. The hyperscalers collectively have the R&D budgets, the manufacturing relationships, and the captive deployment scale to build chips that are genuinely competitive. The $20 billion run rate from Amazon's silicon group is proof that internal deployment creates enough scale to justify the fixed costs. External sales are the natural next step — they spread those fixed costs further and potentially accelerate the roadmap.
I've watched what happens when large cloud providers decide to compete in markets that used to be vendor lock-in: networking silicon, storage controllers, database engines. The pattern is consistent. They start by reducing external dependency, achieve competitive performance, and then the economics of external sales become too good to ignore. Amazon is at step three with Graviton. They're at step two with Trainium. That progression has been reliable, and I see no reason it stops here.
The Bottom Line
This week's two announcements — Graviton5 GA and the Trainium external-sales talks — are individually notable. Together they represent something more: Amazon has made a strategic decision to become a serious silicon company. For infrastructure operators, that means a new competitive option in AI compute is materializing in real time. It will take years to fully develop, and Nvidia will not yield market share easily. But the direction is set. The question now is pace, not destination.
If you're planning AI infrastructure investments for the 2026–2028 window — training hardware, inference clusters, on-premise AI for regulated workloads — Amazon's silicon roadmap should be a variable in your planning now, not an afterthought.