Everyone expected Amazon to keep its chip game internal. Graviton processors humming away in AWS data centers, Trainium chips training AI models—just for their own empire. No leaks to outsiders. But yesterday, CEO Andy Jassy shattered that in his shareholder letter. If Amazon sold those chips on the open market? $50 billion annual run rate. Fourth biggest semiconductor player worldwide. Behind Nvidia, Samsung, SK Hynix. Ahead of Intel, Qualcomm, everybody else. That’s the flex.
“Our annual revenue run rate for our chips business (inclusive of Graviton, Trainium, and Nitro—our EC2 NIC) is now over $20 billion, and growing triple digit percentages YoY… If our chips business was a stand-alone business, and sold chips produced this year to AWS and other third parties… our annual run rate would be ~$50 billion.”
Jassy’s numbers sting. Especially for Intel.
Amazon’s Chip Arsenal: Better, Faster, Sold Out?
Graviton? Up to 40% better price-performance than x86 rivals. Powers 98% of the top 1,000 EC2 customers. Then Trainium2: 30% edge over comparable GPUs. Largely sold out. Trainium3—shipping now in 2026—30-40% leap again. Nearly fully subscribed. Trainium4? Eighteen months out, chunks already reserved. Bedrock, AWS’s inference darling, leans hard on Trainium. Demand’s exploding.
It’s not hype—it’s happening inside AWS. But Jassy’s dangling the external sales carrot. Racks to third parties? Possible, he says. So much demand.
Here’s the thing. Amazon’s not new to semis. They’ve designed these since 2018. Graviton1 was meh. But iterations? Crushing it. Unique insight: this mirrors IBM’s mainframe days—custom iron locked customers in, then sold outward. Amazon could pull the same. Except today’s fabs are a nightmare. TSMC queues. Geopolitical snarls. Will they scale without tripping?
Short answer: probably. They’ve got the cash. The brains. The AWS moat.
Will Amazon Actually Sell Chips to Outsiders?
Expectations were zero. Chips as cost-savers, not revenue bombs. This flips the script. $20B internal run rate now—triple-digit growth. Hypothetical $50B if external? That’s no joke. Bigger than Intel’s entire chip revenue last year. (Intel’s scrambling with foundry dreams; Amazon’s already printing silicon.)
But call the bluff. Jassy’s letter screams PR polish. “Quite possible we’ll sell racks.” Possible. Not committed. They’re monetizing via EC2 today—understating the true value, sure. Yet why not spin off now? Fear of cannibalizing AWS margins? Or just testing shareholder waters?
Dry humor alert: Nvidia’s printing money on GPUs. Amazon’s Trainium undercuts on price-perf. If they flood the market, Jensen Huang sweats. Samsung, Hynix? Foundry kings, but Amazon’s designs could snag their capacity.
Is Trainium Poised to Dethrone Nvidia GPUs?
Jassy boasts: Trainium2 beats GPUs by 30%. Trainium3, more. But benchmarks? AWS-controlled. Independent tests scarce. (Remember Apple’s M1 hype? Real-world mixed.) Trainium shines in inference—Bedrock’s secret sauce. Training? Still catching Nvidia’s H100s.
Bold prediction: by 2027, Amazon sells Trainium racks to hyperscalers sick of Nvidia premiums. Microsoft? Google? They’ll bite. Internal use proves it works at scale. External? Unlocks billions. Critique the spin: Jassy bundles Nitro (network card) in run rate. Cheeky—pads the number. Pure silicon? Closer to $15-18B maybe.
And fabs. Trainium on TSMC 5nm, eyeing 3nm. Supply chain’s their Achilles. China tensions. US CHIPS Act cash? Amazon qualifies.
Look, Amazon disrupts quietly. AWS ate enterprise IT. Now chips. If they enter semis proper—design house like Broadcom, but cloud-tied—they reshape rankings. Intel’s foundry folly looks dumber daily.
Skepticism reigns. $50B’s a what-if. No product roadmap announced. No sales team hired. But demand’s real. Reservations for unshipped silicon? That’s traction.
Parenthetical: Imagine Qualcomm begging for Graviton arms licenses. Hilarious.
The Broader Shakeup: Semis Oligopoly Cracks?
Nvidia owns AI training. Amazon nibbles inference, general compute. Together? Dual poles. Samsung fabs for all. Hynix stacks HBM memory.
Amazon entering changes math. Reduces reliance on one vendor. Price wars incoming. Developers rejoice—cheaper AI.
But risks. IP theft if racks ship. Support nightmares. AWS lock-in persists.
Wander a sec: back to 2012, AWS launches Graviton precursor. Skeptics laughed. Now? Industry standard inside cloud. History rhymes.
🧬 Related Insights
- Read more: Silicon Motion’s SM8008: The Unsung Hero of AI Server Boot Drives
- Read more: Sharp’s Poketomo: Edge AI That Remembers Without Spilling to the Cloud
Frequently Asked Questions
Is Amazon entering the semiconductor market?
Not yet officially, but CEO Jassy’s $50B run-rate hypothetical signals strong intent to sell chips externally, beyond AWS.
What is Amazon Trainium and how does it compare to Nvidia?
Trainium is Amazon’s custom AI chip; Trainium2 offers 30% better price-performance than comparable GPUs, with Trainium3 improving further.
How big is Amazon’s chip business compared to Intel?
At $50B projected run rate if standalone, it’d surpass Intel, Qualcomm, and others—fourth globally behind Nvidia, Samsung, Hynix.
