Amazon’s AI Chips: Nvidia Competition by 2026

Table of Contents

• 1. Amazon aims to compete with Nvidia in AI chips
• 2. What is Amazon Web Services planning with its AI chip Trainium?
• 3. Why is Amazon considering selling its homegrown AI chips?
• 4. What is the estimated annual run rate for Amazon’s chips business?
• 5. Why has AWS resisted selling its AI chips until now?
• 6. What is the current demand for Amazon’s Trainium chips?
• 7. Conclusion: The Future of AI Chips and Market Dynamics
• 7.1 Amazon’s Strategic Positioning in AI Chip Market
• 7.2 Implications for Nvidia and the Competitive Landscape

Amazon aims to compete with Nvidia in AI chips

• AWS is in early-stage talks to sell its Trainium AI chips to other companies for use in data centers.
• CEO Andy Jassy has floated a chips business of about $50 billion if sold to AWS and third parties.
• AWS has historically kept Trainium in-house because chip profits compound through cloud services like storage and networking.
• Demand is tight: Trainium capacity has been selling out quickly, including future Trainium4 capacity.

Hyperscalers Challenge Nvidia Dominance
Nvidia remains the default accelerator supplier for much of the AI industry largely because of its hardware scale and the CUDA software ecosystem. At the same time, hyperscalers have been building custom silicon (e.g., AWS Trainium, Google TPU, Microsoft Maia) to reduce cost and dependency on a single vendor—an industry shift that multiple market analyses describe as growing, though market-share figures vary by methodology and time window.
For readers, that’s the backdrop for why “AWS selling Trainium” is a bigger story than a new product SKU: it would move Amazon from using custom chips mainly to optimize its own cloud economics to competing for accelerator budgets in third-party data centers—where Nvidia has historically been strongest. (Market-share estimates for Nvidia’s AI accelerator position are often cited in the ~85–92% range in 2026 commentary, but they are estimates rather than audited figures.)

What is Amazon Web Services planning with its AI chip Trainium?

Amazon Web Services is exploring a shift that would move its custom AI silicon from an internal advantage to an external product line. AWS AI chief Peter DeSantis said the company is in talks to sell its Trainium chips to other companies, according to Bloomberg. An AWS spokesperson confirmed the conversations are real.

If it happens, the move would put AWS more directly into Nvidia’s core territory: supplying accelerators that power AI training and inference. Beyond raw compute, the competitive gap is also shaped by software: Nvidia’s CUDA ecosystem remains a major switching-cost advantage, while AWS’s Neuron stack is positioned as the path to run common frameworks like PyTorch on Trainium. Until now, Trainium has primarily been a lever to make AWS’s own AI compute cheaper and more available—an alternative to buying as many Nvidia GPUs as possible. Selling Trainium externally would change the nature of the competition: not just “AWS vs. other clouds,” but “Amazon silicon vs. Nvidia silicon” in third-party facilities.

The plan, as described publicly, is not framed as selling individual chips to hobbyists or small labs. The language from Amazon points toward selling “racks” of chips—data-center-scale systems rather than a retail component business. That matters because the competitive battlefield in AI is increasingly about integrated infrastructure: chips, networking, and the ability to deploy at scale.

Still, AWS is not naming potential buyers. DeSantis declined to specify which companies might purchase Trainium, leaving open whether the first customers would be enterprises building private AI capacity, other infrastructure providers, or specialized data-center operators.

Staged Trainium Rack Rollout
A practical way to interpret “selling Trainium” (based on Amazon’s own “racks” wording) is a staged rollout:
1) Define the sellable unit: rack/system configurations (not loose chips), including networking, power/thermal envelopes, and management tooling.
2) Choose initial buyer profiles: a small set of data-center operators or enterprises that can integrate rack-scale systems and commit to volume.
3) Validate software portability: confirm Neuron + common frameworks (e.g., PyTorch) meet the buyer’s model/training/inference needs, including profiling and debugging workflows.
4) Secure supply allocation: carve out manufacturing capacity that doesn’t materially extend AWS customer waitlists.
5) Support and lifecycle: establish firmware updates, failure/replace processes, and a roadmap cadence (e.g., how Trainium generations map to system refresh cycles).
Checkpoint to watch: if Amazon announces “racks” with delivery timelines and support terms (not just talks), that’s a strong signal the effort has moved beyond exploratory conversations.

Why is Amazon considering selling its homegrown AI chips?

The most direct catalyst is Amazon CEO Andy Jassy’s own framing: demand is high enough that the company is thinking about selling what it has built for itself. In his annual shareholder letter in early April, Jassy wrote that Amazon’s homegrown AI chips were “so coveted” he was considering selling them—an unusually explicit signal that AWS sees Trainium as more than a cost-saving internal tool.

Jassy also sketched the ambition in financial terms. If the chips business were standalone and sold chips produced this year to AWS and third parties, he said, the annual run rate would be about $50 billion. That kind of number is not incremental; it implies Amazon believes it can create a major new revenue stream by productizing what has been an internal capability.

Strategically, selling Trainium would also let Amazon challenge Nvidia’s dominance more directly. Nvidia’s founder and CEO Jensen Huang has been publicly expanding Nvidia’s narrative beyond GPUs—talking about a large market in CPUs for AI as well. In that context, Amazon’s move reads like a counterpunch: if Nvidia is broadening its reach across the AI data center, Amazon is signaling it may broaden its own reach beyond AWS.

There’s also a broader industry trend behind the decision. In practice, many large buyers are moving toward multi-vendor procurement—mixing Nvidia hardware with custom silicon—rather than betting on a single accelerator supplier. Hyperscalers have been designing custom silicon—Amazon with Trainium, Google with TPU, Microsoft with Maia—to optimize for their workloads and reduce dependence on a single supplier. Selling those chips externally would be the next step: turning internal scale into a product that can compete for budgets outside their own walls.

Balancing Trainium Sales and AWS
What Amazon potentially gains by selling Trainium:

• New revenue stream: monetizes silicon beyond AWS consumption.
• Market leverage: gives buyers (and Amazon) more negotiating power in a GPU-constrained world.
• Ecosystem pull: if Trainium becomes “standard enough,” it can attract more tooling, partners, and workloads.

What Amazon risks or must manage:

• AWS flywheel dilution: external deployments could reduce the “waterfall effect” of attached AWS services.
• Allocation tension: every rack sold externally is capacity not used to serve AWS customers—unless manufacturing expands.
• Support burden: selling to third-party data centers implies enterprise-grade lifecycle support, spares, and roadmap commitments.

The core decision is less “can we sell chips?” and more “can we sell them without slowing AWS growth or degrading customer experience?”

What is the estimated annual run rate for Amazon’s chips business?

Amazon has put a headline number on the opportunity: roughly $50 billion, under a specific assumption. Jassy wrote that if the chips business were standalone and sold chips produced this year to AWS and to third parties—“as other leading chips companies do”—it would be on a ~$50 billion run rate.

That estimate is important for two reasons. First, it frames Trainium not as a niche accelerator but as a business that could sit in the same conversation as the largest semiconductor franchises. Second, it signals Amazon is thinking in “merchant silicon” terms: selling to itself and to external customers, rather than treating chip design purely as an internal optimization.

Even at $50 billion, Amazon would not automatically dethrone Nvidia. Nvidia is described as being on a $326 billion revenue run rate, so a $50 billion competitor would not “tank” Nvidia by itself. But it would be large enough to change pricing dynamics and procurement strategies—especially if it gives big buyers credible leverage in negotiations.

Jassy’s comparison point is telling: he suggested a $50 billion chips business is akin to Intel’s annual revenues. That doesn’t mean Amazon would instantly replicate Intel’s breadth or manufacturing footprint. But it does underline the scale of what Amazon believes it could build if it can produce enough chips and find enough buyers beyond AWS.

The caveat is embedded in the premise: the run rate assumes chips produced “this year” could be sold both internally and externally. That makes supply—how many Trainium chips Amazon can actually get manufactured—central to whether the number is aspirational or achievable.

Interpreting the ~$50B Run Rate
How to read the “~$50B annual run rate” claim (based on the wording in the shareholder letter):

• Unit of measure: revenue run rate, not profit, and not a guarantee of realized external sales.
• Included demand: chips “produced this year” sold to (a) AWS itself and (b) third parties.
• Key variables that drive the number:
• Volume: how many Trainium chips/racks are actually produced and delivered in the year.
• Mix: how much goes to AWS internal consumption vs. third-party buyers.
• Pricing: internal transfer pricing vs. external “merchant” pricing (these can differ materially).
• Packaging: whether revenue is counted as chips alone or as rack-scale systems (Amazon’s “racks” language implies systems could matter).

Practical takeaway: the ~$50B figure is best understood as an “if we sold what we already plan to produce, at chip-company-style economics” framing—highly sensitive to supply and pricing assumptions.

Why has AWS resisted selling its AI chips until now?

AWS’s historical reluctance is rooted in how cloud economics work. The biggest reason, as described, is that the money AWS makes on its chips is a “waterfall effect.” AWS can charge customers for the AI tokens processed on its cloud, but the real compounding value comes from everything attached to those workloads: storage, security, networking, and monitoring services.

In other words, keeping Trainium inside AWS doesn’t just monetize silicon; it pulls customers deeper into the AWS platform. Selling chips to a third-party data center could weaken that flywheel if customers run workloads off-AWS while still benefiting from Amazon-designed hardware.

There’s also a practical constraint: capacity. Amazon has repeatedly emphasized that Trainium supply has been tight relative to demand. Jassy said current Trainium chip capacity sold out almost instantly. He also said capacity for the next generation, Trainium4—despite being more than a year away—was already fully reserved. That kind of forward sell-through makes it hard to justify diverting supply to external buyers without leaving existing AWS customers waiting.

Manufacturing realities reinforce the constraint. If Amazon wants to sell externally at scale, it likely needs surplus production through manufacturing partners such as TSMC. But TSMC capacity is fiercely contested, and Nvidia is a major customer. The brief notes that TSMC has recently supplanted Apple to become the foundry’s largest customer—an indicator of how intense the AI-driven demand cycle has become.

Finally, AWS has also been pursuing a hybrid approach rather than an all-in break from Nvidia. The brief notes a multi-year agreement between AWS and Nvidia for the supply of one million GPUs and related AI infrastructure through 2027. That suggests AWS has been balancing: build Trainium, but keep Nvidia supply flowing to meet customer needs.

Signals Behind AWS’s Holdout
Concrete signals behind AWS’s “why not sell (yet)?” stance:

• “Waterfall effect” economics (as stated in the article): AWS monetizes not only token processing but also the attached stack—storage, security, networking, monitoring—which is harder to capture if workloads run off-AWS.
• Capacity sell-outs (as stated in the article): Jassy said Trainium capacity sold out “almost instantly,” and Trainium4 capacity was “already fully reserved” despite being more than a year away.
• Foundry constraint (as stated in the article): scaling external sales likely requires surplus output via partners such as TSMC, where capacity is heavily contested.
• Ongoing Nvidia reliance (corroborating context): reporting cited in the dossier (Techzine, 2026) describes a multi-year AWS–Nvidia agreement for up to one million GPUs and related AI infrastructure through 2027—consistent with a hybrid strategy rather than an immediate break.

What is the current demand for Amazon’s Trainium chips?

Demand is described as exceptionally strong—strong enough to shape Amazon’s entire decision-making about whether it can even afford to sell externally. In Jassy’s April shareholder letter, he said the current Trainium chip capacity had sold out “almost instantly.” He added that capacity for Trainium4, which won’t be available for more than a year, was already fully reserved.

That matters because it implies demand is not just reactive to what AWS can offer today; customers are reserving future capacity well ahead of availability. In practical terms, it suggests Trainium is already part of long-range infrastructure planning for some AWS customers, not just a spot-market alternative when Nvidia GPUs are scarce.

The timing is also notable. The sell-out comments came before AWS formally added OpenAI to the models it was serving up, according to the brief. If AWS is expanding the set of major models and partners it supports, that could add further pressure on compute capacity—making the question of external sales even more complicated.

This is the central tension in Amazon’s chip ambitions: selling Trainium to third parties could open a new market, but it could also force AWS to choose between external revenue and serving existing cloud customers. Unless Amazon can materially expand manufacturing output, selling chips externally could mean longer waiting lists for current customers—an outcome AWS has strong incentives to avoid.

At the same time, the fact that AWS is even discussing external sales, despite sell-outs, signals how large Amazon believes the opportunity is—and how confident it may be that it can expand supply over time.

Conclusion: The Future of AI Chips and Market Dynamics

Amazon’s Strategic Positioning in AI Chip Market

Amazon’s Trainium discussions mark a potential pivot from “custom silicon as an internal advantage” to “custom silicon as a product.” The company is signaling that its AI chips are not only cost-effective inside AWS but potentially valuable enough—and in demand enough—to sell to other data-center operators.

Yet the move is constrained by the same factor that makes Trainium attractive: scarcity. With current and even future capacity selling out quickly, Amazon’s ability to compete with Nvidia as a supplier will depend on whether it can secure enough manufacturing capacity without undermining its own cloud growth. The company’s public language—selling “racks” in the future—suggests it is thinking in systems and scale, not small-batch experimentation.

Implications for Nvidia and the Competitive Landscape

For Nvidia, Amazon’s ambitions represent a more direct challenge than typical cloud competition. A credible alternative supplier with a stated ~$50 billion run-rate ambition could influence pricing, availability, and negotiating power across the AI infrastructure market.

But the brief also makes clear this is not a simple displacement story. Nvidia remains enormous by revenue run rate, and AWS itself continues to buy Nvidia hardware under a multi-year supply agreement. The likely near-term outcome is a more hybrid market: Nvidia plus custom silicon, with buyers increasingly expecting choice.

If Amazon can translate internal demand into external supply—without starving AWS customers—it could become one of the most consequential new competitors Nvidia has faced in AI accelerators. The next phase will be defined less by announcements and more by execution: manufacturing, delivery, and whether Trainium can become a standard option beyond AWS’s walls.

Signals to Watch Next
What to watch next (signals that matter more than the headline):

• First disclosed buyers: whether early customers are enterprises, colo/data-center operators, or infrastructure providers.
• What exactly is sold: “racks/systems” details (networking, management plane, support terms) vs. vague chip talk.
• Supply allocation: any indication Amazon has secured incremental foundry/packaging capacity rather than reallocating from AWS.
• Software maturity: clearer proof points that Neuron workflows meet production needs beyond AWS-native teams.
• Pricing posture: whether external pricing undercuts Nvidia materially or is positioned as a “second source” option.
• Timeline specificity: delivery windows and volume commitments (the difference between exploration and a real product line).

Perspective: This analysis is written from the lens of building and scaling technology businesses in regulated, multi-stakeholder environments—where platform economics, procurement constraints, and software ecosystems often matter as much as headline chip performance (Martin Weidemann, weidemann.tech).

This piece reflects what Amazon and AWS leaders have said publicly about Trainium and what that may imply for competition with Nvidia. Some figures cited in public commentary—such as market-share estimates and revenue run-rate framing—are directional and may vary by source or definition. Availability, pricing, and buyer details can change over time as new information emerges.