- Amazon CEO Andy Jassy projected a $50 billion annual revenue run rate for a standalone chip business.
- Amazon’s custom silicon business crossed a $20 billion annual revenue run rate in Q1 2026, with Trainium3 nearly sold out and Trainium4 largely reserved.
- Direct sales of Trainium chips would reposition Amazon as a chip vendor, competing with Nvidia and offering enterprises more procurement options.
Amazon is in talks to sell its Trainium AI chips directly to third-party data centers, a move that AWS AI chief Peter DeSantis confirmed to Bloomberg in mid-June 2026. If it goes ahead, it repositions Amazon from cloud-infrastructure operator to chip vendor competing head-on with Nvidia. CEO Andy Jassy flagged the ambition in his April 2026 shareholder letter, putting a $50 billion annual revenue run rate on a standalone chip business.
A Direct Challenge to Nvidia’s Market Position
Nvidia’s GPUs have underpinned the generative AI boom, holding a commanding position in high-performance AI training hardware. Amazon has spent years building around that dominance rather than confronting it directly. That changes if Trainium ships to data centers that have no AWS relationship at all.
Trainium is designed specifically for AI training workloads; Inferentia handles inference. Both have already attracted serious customers: OpenAI and Anthropic have committed to large capacity blocks through AWS. A direct-sales model would let Amazon reach data center operators and enterprises that want the silicon without the cloud layer, putting pressure on Nvidia’s pricing in markets where AWS currently has no footprint.
The $50 Billion Revenue Case
Amazon’s custom silicon business already crossed a $20 billion annual revenue run rate in Q1 2026, growing at a triple-digit pace, according to DeSantis. Trainium3 is nearly sold out. Trainium4 is largely reserved ahead of launch. Jassy’s $50 billion projection for a standalone chip business assumes that demand extends well beyond AWS’s own infrastructure needs, and current supply signals suggest he may be right about the appetite, if not yet about the capacity to fill it.
Selling chips rather than compute cycles also changes the revenue structure. AWS currently monetises Trainium by renting access to it. Direct sales would capture the hardware margin upfront, diversifying Amazon’s income beyond cloud services and into the semiconductor supply chain itself.
What This Means for Cloud and Data Center Procurement
Cloud providers have traditionally been chip consumers, not chip vendors. Amazon moving into direct sales blurs that boundary. For enterprises, the practical upside is real: Trainium chips in an on-premise rack or a third-party colocation facility, without an AWS contract required. That is a meaningful option for organisations with data sovereignty requirements or those already running hybrid infrastructure.
AWS executives have reportedly pushed back on concerns about cannibalising cloud revenue, arguing that demand for AI compute is large enough to support both channels. How widely that holds in practice is harder to verify from public material, but the framing is consistent with Amazon’s general posture on AI infrastructure capacity. The G7’s push to diversify chip supply chains also creates an opening for non-Nvidia silicon at the infrastructure-planning level.
Supply Chain Is the Real Constraint
Demand is not the problem. Supply is. Trainium3 is already nearly exhausted and Trainium4 is heavily reserved before it has formally launched. Amazon manufactures through TSMC, which is running at high utilisation serving Nvidia, Apple and others. Securing additional wafer capacity in that environment is genuinely difficult, and Nvidia recently became TSMC’s largest customer, which does not make Amazon’s position easier.
The harder operational question is allocation. Diverting existing Trainium supply to external buyers risks shorting AWS’s own cloud customers. Amazon will need to scale production substantially before a direct-sales business can run at meaningful volume without creating internal friction. How it manages that allocation will determine whether this is a real hardware business or a well-publicised pilot.
Cost Economics for External Buyers
AWS Inferentia chips are positioned as offering substantial cost reductions for inference tasks compared to Nvidia GPUs, though independent benchmarking of those figures remains limited.
The constraint for external buyers will be software. Nvidia’s CUDA ecosystem has a decade of tooling, optimised libraries and developer familiarity behind it. Trainium runs on AWS Neuron SDK, which is capable but narrower in third-party support. Enterprises buying Trainium hardware outside AWS would need to invest in porting workloads, a real but not insurmountable cost for large-scale training shops. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
