AWS G7 Instances Bring Blackwell to the Cloud — and Trainium May Be Going External
The June 20 launch of EC2 G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs is AWS's clearest move yet into the Blackwell GPU generation, arriving the same day reports emerged that Amazon may begin selling Trainium chips to customers outside AWS — a pairing that captures the dual-track tension at the center of its AI infrastructure strategy.
On June 20, 2026, AWS announced Amazon EC2 G7 instances powered by NVIDIA's RTX PRO 4500 Blackwell Server Edition GPUs, marking its commercial entry into the Blackwell GPU generation for a category of deeply optimized AI and professional-compute workloads. The G7 family targets inference, visualization, and AI applications suited to the RTX PRO 4500 architecture, occupying a different market position from the large-scale distributed training addressed by AWS's P-series instances and its proprietary Trainium chips. The launch arrives at a moment of unusual strategic complexity for the world's largest cloud provider: the company is simultaneously extending its NVIDIA partnership into a new GPU generation while accelerating what may become the most ambitious custom-silicon program in the hyperscaler industry.
The Blackwell architecture itself had a troubled gestation. In August 2024, reports emerged that NVIDIA's Blackwell GPUs faced design flaws requiring at least a three-month delay — a setback that cascaded through every major cloud provider's hardware roadmap. AWS responded by doubling down on parallel tracks: by December 2024 it had disclosed plans for Trainium3, and by December 2025 had formally unveiled that chip alongside a commitment to deploy NVIDIA GB300 chips through its AI Factories managed on-premises service. Announcing both on the same day captured the dual-track logic AWS has refined over the past two years — custom silicon for cost-efficient training at scale, NVIDIA partnerships where customer ecosystem familiarity still commands a premium. The December 2024 period also saw AWS disclose that Apple intended to use its custom chips, and Marvell Technology's shares surged on a separately announced agreement covering custom AI silicon and high-speed networking, two data points suggesting AWS was building a diversified silicon supply chain rather than simply deepening a single-vendor dependency.
The economics of that strategy have been visibly pressured by supply constraints. In January 2026, AWS raised GPU EC2 instance prices by 15%, citing AI accelerator shortages — a move that improved near-term margins but signaled real limits on availability heading into the G7 launch. AWS had earlier introduced single-GPU P5 instances with NVIDIA H100s in August 2025, framed explicitly as a cost-saving option for inference workloads, suggesting the company was managing allocation by creating more granular product tiers from its existing NVIDIA inventory. The physical infrastructure layer has seen parallel vertical integration: in July 2025, AWS deployed its own proprietary IRHX liquid-cooling systems for NVIDIA GPU clusters, bypassing third-party cooling vendors in a move consistent with an operator seeking tighter control over thermal budgets as GPU power densities climb.
Customer momentum has continued to accumulate despite the supply friction. By April 2026, Meta was reportedly deploying millions of AWS Graviton ARM cores for AI workloads, a notable endorsement of AWS's custom ARM silicon from one of the industry's heaviest AI compute consumers. The trajectory has not been without interruption, however. In April 2025, AWS and Microsoft both paused or slowed new data center lease agreements, a brief episode in which the two largest hyperscalers appeared to recalibrate capacity commitments against uncertain near-term AI demand — a moment of caution that drew scrutiny given the scale of prior build-out plans. That pause has since given way to resumed expansion: in June 2026, AWS signed a recycled-water supply deal for a planned Melbourne data center, one of several international infrastructure commitments still in active execution. Talent pressure has accompanied the competitive intensity throughout: in January 2026, a senior AWS data center network lead departed for xAI to run its machine-learning infrastructure, an episode illustrating that AWS competes not only with cloud rivals for customers but with well-capitalized newcomers for the specialized engineers who build and operate this infrastructure.
The most strategically consequential item running parallel to the G7 launch may be reporting from June 20, 2026, that Amazon has begun talks to sell Trainium chips to customers outside of AWS. If those discussions yield commercial agreements, they would mark a qualitative shift in Amazon's semiconductor posture — from captive internal supplier to merchant competitor, placing the company more directly against NVIDIA in the chip market rather than solely as a large buyer of NVIDIA silicon. The initiative would represent a departure from Amazon's historical practice of using custom hardware primarily to reduce its own cost structure, raising questions about what commercial concessions may be required to win external customers against an NVIDIA ecosystem backed by a mature CUDA software stack and a partner network built over more than a decade. The initiative would also parallel Google's external TPU offerings, though at a scale and commercial framing that remains to be demonstrated.
Three signals are worth monitoring as this narrative develops. First, the competitive pricing of G7 instances against Blackwell-based GPU offerings from Azure and Google Cloud will reveal whether AWS prioritizes margin or market share in a product category where it is not yet the established incumbent — a repeat of the January 2026 dynamic, or a departure from it. Second, any publicly announced external Trainium customer — with details on volume and pricing terms — would represent a material inflection point for Amazon's semiconductor ambitions and warrant close scrutiny of what was conceded to win business against NVIDIA's entrenched position. Third, Amazon's capital expenditure guidance in forthcoming earnings will clarify whether the April 2025 lease pause was a brief recalibration or the beginning of a more cautious stance on infrastructure commitment. The risks are concrete: continued dependence on NVIDIA for the bulk of high-performance compute demand, exposure to export-control dynamics that have already required Intel to obtain government licenses before selling advanced chips to China, and a talent market where resource-rich rivals are actively competing for the same specialists. The opportunity, if enterprise AI adoption curves track current hyperscaler planning assumptions, is that AWS enters the Blackwell era with a more diversified hardware matrix than it possessed entering the Hopper cycle — one that combines proprietary Trainium silicon, co-developed chips from partners including Marvell, and now the RTX PRO 4500-based G7, giving it more levers to pull on both cost and capability as the AI infrastructure buildout continues.