$50 billion. That’s the annual run rate AWS’s homegrown chips would clock if they operated as a standalone business, selling to outsiders just like Nvidia or Intel do.
Andy Jassy dropped that bombshell in his latest shareholder letter, and it’s no idle boast—it’s backed by two massive customers already pleading for all of AWS’s 2026 Graviton capacity. They got turned down, sure, but Amazon’s raking in $20 billion yearly from services powered by these custom silicon beasts.
Chips that pay their way—and then some
Look, Jassy’s not just puffing his chest. Trainium AI chips are the real story here. The Trainium3, fresh out early this year, has capacity “nearly fully-subscribed.” And get this—a “significant chunk” of Trainium4 reservations are locked in, even though it’ll be 18 months before broad availability. That’s demand so hot it’s melting the supply chain.
Jassy spells it out bluntly:
“There’s so much demand for our chips that it’s quite possible we’ll sell racks of them to third parties in the future.”
Racks. Not loose chips—whole racks. Imagine AWS shipping pre-loaded server farms to rivals or startups, undercutting the usual cloud lock-in. It’s a page from Nvidia’s playbook, but with Amazon’s hyperscale muscle behind it.
And the economics? Trainium alone promises “tens of billions” in yearly capex savings, plus hundreds of basis points in operating margin over buying from others for inference workloads. That’s not hype; it’s math that keeps Wall Street happy.
Will AWS Actually Sell Graviton Racks to Outsiders?
Here’s the thing—demand isn’t theoretical. Those two big customers wanted every Graviton instance in 2026. Graviton, Amazon’s ARM-based CPUs, power a chunk of AWS’s empire, delivering better price-performance than x86 rivals. But denying them? Smart move. Keeps AWS’s fleet humming for its own AI services, which just hit a $15 billion revenue book in three years—faster than the entire AWS ramp-up back in the day.
Yet Jassy’s dangling the carrot: sell racks if demand overflows. Bold? Absolutely. Risky? You bet—handing rivals your secret sauce could backfire spectacularly.
My take: it’ll happen selectively. Think enterprise deals with NDAs thicker than a data center cable bundle. AWS won’t flood the market like a fabless startup; they’ll cherry-pick, much like how Apple doles out M-series silicon to select partners.
Short para for emphasis: Power’s the bottleneck.
AWS added 3.9 gigawatts in 2025, plans to double total capacity by 2027—and still can’t serve all comers. “Capacity constraints yield unserved demand,” Jassy laments. Electricity, not silicon, is the real scarce resource in this AI gold rush.
Why Does Sold-Out Trainium Capacity Matter for AI Buyers?
Because it flips the script on Big Tech dependency. AWS isn’t just a cloud landlord anymore; it’s a chip powerhouse eyeing $50 billion in standalone revenue. Compare that to three years post-launch AWS revenue run rate: $58 billion total. AI chips alone are closing the gap fast.
Jassy notes 85% of global IT spend lingers on-premises. “This will change,” he promises. And with Trainium4 reservations pouring in, AWS is betting big on inference efficiency to lure workloads off legacy setups—and Nvidia’s GPUs.
But let’s cut through the spin. Amazon’s not reinventing the wheel; they’re iterating ruthlessly, just like in e-commerce. Robotics tell the tale: over a million bots in fulfillment centers, handling stowing, picking, sorting. And now? Hints at selling those too, leveraging “scale and real-time feedback” from their robot army.
Drones next—Prime Air scaling to 30 million customers by year-end, half a billion packages this decade. Satellite broadband mid-2026. It’s all connected: chips power the brains, robots and drones the brawn.
The hidden edge: A parallel to ARM’s quiet conquest
Here’s my unique angle—AWS chips echo ARM’s 90s rise. Back then, ARM didn’t battle Intel head-on; they powered mobile quietly, then dominated. Graviton and Trainium could do the same for cloud and edge AI. Prediction: by 2030, AWS custom silicon snags 20% of inference market share, slashing Nvidia’s margins in that segment. Not because they’re faster (yet), but because racks at hyperscaler prices crush per-unit costs.
Critique time: Jassy’s letter reeks of controlled megalomania—“one of the largest job creators” while deploying a million robots? Please. It’s effective, though. Amazon creates jobs because of efficiency, not despite it.
And AI adoption? Jassy admits it’ll stumble: new interfaces may flop before sticking, just like past pivots. Retail via AI chat? Voice? It’ll evolve messily.
Yet the data screams bullish. AWS revenue at $142 billion run rate, AI accelerating it. Chips as a business? Logical next step—why hoard when you can monetize overflow?
One caveat: electricity. Without grids doubling down, this party’s over before racks ship.
Beyond chips: Drones, robots, and the sprawl
Prime Air’s new design scales from Same Day centers—90,000 top products, 30-minute delivery. Amazon Now? Micro-fulfillment with thousands of SKUs, 20 minutes flat. Over 360 already in India.
Robots: exploring new forms, smarter grasping. Sell to factories? Why not—Amazon’s got the data firehose.
Satellites: 200 birds by mid-2026. Kuiper’s no side hustle; it’s broadband for the unconnected masses.
🧬 Related Insights
- Read more: AWS’s Trainium Lifeline: Will Amazon’s Custom Chips Rescue Its AI Laggard Status?
- Read more: Intel’s Terafab Pact with Tesla and SpaceX: The Foundry That Could Power Mars Rovers and Robotaxis
Frequently Asked Questions
What’s the AWS Trainium chip used for?
Trainium powers AI training and inference on AWS, optimized for large models—cheaper and more efficient than GPU alternatives for certain workloads.
Can customers buy AWS Graviton servers directly?
Not yet, but Jassy hints at future rack sales to third parties as demand overflows AWS’s internal capacity.
Why is AWS AI chip capacity sold out?
Explosive demand for Trainium3 and pre-bookings for Trainium4, coupled with power constraints limiting expansion.
