Amazon plans to begin selling its custom Trainium AI accelerator chips outside of AWS by late 2027, shifting from a cloud-only model to a merchant hardware business. This strategic pivot allows Amazon to capture upfront hardware revenue, though it risks cannibalizing the long-term service fees earned when those chips operate within AWS data centers. Success depends on more than just chip availability; buyers will require comprehensive support, software compatibility, and hardware integration packages to effectively deploy these processors in private data centers or alternative environments, rather than just renting them as cloud services.
Amazon plans to make its Trainium AI chips available outside AWS beginning in late 2027, CEO Andy Jassy said on the company’s October 29, 2026, earnings call. The move would expand Amazon’s role from supplying AI computing through its cloud to selling the processors that power it. (Earnings call transcript)
Jassy linked the decision to customer requests. As CNBC reported, he said:
“With Trainium demand so strong, customers continue encouraging us to make it available outside of AWS. We intend to do so late in 2027.”
This is a strategic plan, not a product launch. Amazon has not announced external pricing, configurations, distribution partners, supported markets or a firm on-sale date beyond that late-2027 window.
The consequential detail is the business tradeoff Jassy described: Amazon could collect hardware revenue when it sells a chip, or potentially earn more total revenue by operating that chip inside AWS for five to six years. That choice explains why external demand alone has not made direct sales an obvious decision.
Trainium Could Become a Business Beyond AWS
Trainium is Amazon’s custom AI accelerator family, used to provide computing capacity for workloads such as training AI models. Its established route to customers has been through AWS services. Customers buy access to computing capacity while AWS operates the underlying infrastructure.
External availability would create another route: customers could obtain Trainium-based computing outside Amazon’s cloud. That would move Amazon toward a merchant AI-chip business, meaning it would supply processors for use beyond its own cloud operation.
The distinction matters because the two models sell different things. A cloud customer pays for access to an operating service. A hardware customer buys infrastructure and needs a way to deploy, support and maintain it, whether directly or through another provider.
Amazon has not specified the commercial package, so it would be premature to picture a particular accelerator card, server or rack. “Outside AWS” describes where the technology could be used; it does not yet describe everything a buyer would receive.
My view is that the important expansion is in customer choice. Amazon could compete for an AI workload even when the organization running it has decided against putting that workload in AWS. That opens a sales opportunity the cloud-only model cannot fully reach.
The Tradeoff: Sell a Chip Once or Operate It for Years
Jassy described external sales as a way to recognize hardware revenue immediately, while keeping chips in AWS could generate more total revenue across five to six years of operation. That is Amazon’s economic assessment, not an independently measured comparison of the two businesses. (Earnings call transcript)
The basic difference is straightforward:
Business model
What the customer buys
Revenue pattern Amazon described
Main operating responsibility
Trainium inside AWS
Access to AI computing through cloud services
Service revenue over years of operation
AWS runs the infrastructure
Trainium outside AWS
Hardware supplied for use beyond AWS
Revenue associated with the hardware sale
Depends on the eventual product and support arrangement
A chip deployed inside AWS can support repeated periods of paid computing. Amazon also supplies the surrounding infrastructure needed to keep that computing available. An external sale brings forward the hardware transaction, but the chip’s subsequent use does not automatically become AWS compute revenue.
That makes the relevant question bigger than how many processors Amazon could sell. The company must consider what it gives up when a processor is sold instead of installed in its own fleet.
AWS Also Values the Services Around the Chip
Jassy’s broader explanation of AWS economics helps clarify that calculation. On the call, he argued that customers prefer production AI inference near their existing applications and data. He also described AI work creating demand for other AWS computing services, including those used in post-training and tool use. (Earnings call transcript)
Inference is the work a model performs when responding to a request after training. If the application, its data and the model all run within the same cloud environment, the customer can have practical reasons to keep buying services there.
The resulting implication is that retaining a chip may help Amazon earn more than the revenue from that accelerator alone. It can also support a broader customer relationship involving storage, applications and other computing.
An external sale does not necessarily eliminate those opportunities. A buyer could purchase hardware for one deployment and continue using AWS elsewhere. The point is that Amazon cannot assume a chip sale and an equivalent cloud deployment have the same lifetime value.
More Revenue Does Not Automatically Mean More Profit
Jassy’s five-to-six-year comparison concerns total revenue. It is not enough, by itself, to establish which option produces a better margin or return on invested capital.
Running hardware for years also requires operating expenditure, infrastructure and continued investment. Hardware sales would bring their own costs, including whatever warranty, support and distribution arrangements Amazon ultimately chooses.
A useful comparison would therefore need information Amazon has not supplied for the external offering: selling prices, costs, utilization assumptions, service obligations and the timing of payments and expenditure.
I would read the announcement as a decision to pursue two revenue models, not proof that Amazon has found one universally better model. Cloud operation may be more valuable for some capacity and customers; external sales may reach demand that AWS would otherwise miss.
Customer Demand Is the Reason, but Deployment Is the Test
Amazon said demand for prior-generation Trainium remained strong and that customers had encouraged it to offer the chips externally. That is management’s account of demand, without a disclosed external order book or named group of buyers. (Earnings call transcript)
Interest in an earlier generation is a useful detail. It suggests Amazon sees demand extending beyond the newest hardware cycle. It does not identify which generation will be offered outside AWS, or establish that Amazon plans to sell used equipment.
External Trainium could appeal to organizations with reasons to operate infrastructure outside a public cloud. Those reasons might include existing data-center investment, requirements governing where data is processed, or a preference for controlling equipment and capacity planning. These are possible uses of the planned offering, not customer motivations Amazon disclosed in the announcement.
The practical benefit would depend on the complete system. A processor needs compatible software, memory, networking, power, cooling and support to become useful AI infrastructure. Buyers would also need to know how their models and applications move onto the platform.
That is why I would treat software and support details as being as important as chip pricing. A lower hardware price can lose its appeal if adapting a workload or operating the system adds substantial cost. Amazon’s announcement establishes an intended sales channel; the deployment burden remains an essential part of the eventual product.
Google Provides a Precedent, While NVIDIA Remains a Partner
CNBC placed Amazon’s plan alongside Google’s strategy of making its tensor processing units, or TPUs, available outside Google’s own facilities. On that evidence, Amazon is following a direction another major cloud provider has already taken: extending custom AI processors beyond the cloud infrastructure that originally gave them a market. (CNBC)
The comparison has limits. The available reporting does not provide matching external configurations, prices, software terms or support arrangements for Trainium and Google’s TPUs. It supports a comparison of business direction, not a conclusion that the offerings will be interchangeable or that one will be cheaper.
For NVIDIA, external Trainium could introduce another option into hardware purchasing decisions. A customer that wants to run infrastructure outside AWS could eventually consider Amazon’s processors alongside other available AI platforms.
That would broaden the scope of competition. Today, the choice can involve which cloud service to use. An external Trainium business could bring Amazon into decisions about which hardware platform an organization deploys itself or obtains through another operator.
But Amazon’s custom-chip ambitions do not imply an immediate break with NVIDIA. CNBC also reported that AWS announced plans during the quarter to expand its portfolio of NVIDIA GPUs for AI workloads. Amazon can develop a competing processor business while continuing to buy and offer NVIDIA technology.
That coexistence is more informative than a simple replacement narrative. Different workloads and customers can favor different hardware and software combinations. External Trainium would give Amazon another way to compete for those choices.
The Details Buyers Will Need Before Making Plans
AWS’s existing growth gives the announcement context. Jassy reported 39.3% year-over-year AWS revenue growth for the third quarter of 2026 and described strong demand for newly installed capacity. Those are Amazon’s reported figures and observations; they do not isolate Trainium’s contribution or forecast sales outside AWS. (Earnings call transcript)
Amazon is therefore proposing an external business while still expanding the cloud business that uses its chips. That makes capacity allocation worth watching: supplying external buyers could create a new revenue stream, but AWS also has customers waiting for computing capacity.
For prospective buyers, the next useful disclosures will be operational:
Software access: Which tools are available outside AWS and how customers maintain compatibility over time.
Infrastructure requirements: The networking, power and cooling needed for supported deployments.
Support responsibility: Who handles failures, replacements, software updates and system troubleshooting.
Commercial terms: Pricing, availability commitments and the conditions attached to purchases.
These are evaluation questions, not announced features.
The most useful test of the strategy will be whether Amazon can make Trainium practical outside the environment it controls. AWS currently combines the chip with infrastructure and services. An external offering must define which parts Amazon supplies and which parts the customer must assemble or operate.
For now, I would put Trainium on a late-2027 evaluation roadmap, not a committed procurement schedule. Jassy has identified the direction and explained the revenue tradeoff. Buyers still need the product and operating terms that turn that direction into a usable alternative.
Frequently Asked Questions
3 questions
1
When Will Trainium Be Available Outside AWS?
Amazon intends to begin external availability in late 2027. Jassy announced that plan on October 29, 2026, but did not provide a firm on-sale date.
2
Which Trainium Generation Will Amazon Sell?
The announcement did not specify a generation or configuration. Amazon’s comments about strong demand for prior-generation chips do not identify the external product.
3
Will External Sales Be More Profitable Than AWS Rentals?
Amazon has not supplied enough information to determine that. Jassy said operating chips inside AWS for five to six years could produce more total revenue, but revenue alone does not establish profitability.
