Amazon tops $3 trillion as AWS puts OpenAI, Anthropic, Meta and xAI models in its government cloud

What the Q2 results do — and do not — show

When Amazon reported Q2 2026 results, the stock moved sharply and the company's valuation again became part of the AI infrastructure story. The linked TradingView report describes the market reaction and analyst target changes, including Citizens Bank's $315 target and Evercore ISI's $355 target. Those targets are analyst opinions, not company guidance.

The underlying figures should be taken from Amazon's Q2 2026 earnings materials. This article does not repeat previously stated share-price, market-capitalisation, AWS revenue, AWS operating-income, total-sales or AWS-profit-share figures because those figures require matching them precisely to the release and its accounting presentation. The same caution applies to previously stated dollar-to-euro conversions and the annualised AWS run-rate calculation.

Amazon's capital-spending outlook should likewise be read only in the wording and period used in the official earnings materials. It should not be presented as a quarterly commitment, as an amount already spent or as AWS-only expenditure. Capital expenditure can cover AI servers, networking and data centres across the group.

A fair-value gain tied to an investment such as Amazon's stake in Anthropic would also need to be distinguished from operating performance. Such a gain is a non-cash accounting adjustment rather than revenue from selling a product, and it is not the same as the value of Amazon's entire stake. It may be included in reported net income, subject to applicable tax effects, while remaining separate from AWS's recurring revenue and operating income. It does not represent a cash distribution from Anthropic or a second revenue stream. The exact amount should be taken from the official earnings disclosure rather than repeated without verification.

The government cloud becomes a model bazaar

The stock story overshadowed a product story worth your attention. AWS's August 31, 2026 announcement described expanded Amazon Bedrock availability in AWS GovCloud (US), covering the two GovCloud regions, US-East and US-West. The source material supplied for this article does not include a direct URL to that announcement, so the official Amazon Bedrock documentation should be used to verify the current service page and model catalogue.

No specific OpenAI model ID is verified by the supplied evidence, and the material does not establish a complete list of confirmed models from Anthropic, Meta, xAI or NVIDIA. Those provider names should therefore not be treated as proof of availability. The announcement concerns selected third-party models; it does not mean that every model from any named provider is available in both GovCloud regions. Exact model IDs, region support and inference options must be checked in current AWS documentation and in the customer's AWS account.

In practice, the announcement gives US government and other eligible users a broader route to multiple model families through one cloud service. It is better understood as evidence of a multi-model platform strategy than as proof that all named providers' models can be called interchangeably. Model availability, security authorisations, quotas and routing rules remain model- and region-specific.

That distinction also matters when assessing the Anthropic investment. Amazon can benefit from a higher valuation of that investment through a disclosed fair-value accounting adjustment, while AWS can earn infrastructure and service revenue from workloads using a range of models. That is a platform hedge, but it does not turn an investment gain into operating revenue or guarantee that every competing model will be offered in every AWS environment.

What this changes for European companies

GovCloud itself is reserved for US government use and eligible US entities; European agencies should not expect access to it. European companies can access a comparable multi-model menu through Amazon Bedrock in EU regions, though the EU-region model list is not identical to GovCloud's. Availability and inference routing vary by model, region and account. Data residency is not automatic: Bedrock regional or cross-region inference settings and the individual model's terms determine where data is processed. The practical question for most European organisations is not only which model to use, but also which contractual and regulatory framework applies.

Under the EU AI Act, obligations apply according to the relevant role, system category and implementation timeline. Providers of general-purpose AI models, including companies such as Anthropic, OpenAI, Meta and xAI, have obligations under the GPAI provisions where the Act applies. European companies integrating those models may be providers or deployers depending on what they place on the market and how they use it. Article 50 covers transparency obligations for certain AI systems, including informing people when they interact with an AI system in the relevant circumstances and disclosing or marking certain AI-generated or manipulated content. It does not automatically impose the same duty on every model deployment. High-risk obligations apply where an AI system falls within the Act's high-risk categories; they are not triggered merely because a company uses Bedrock. GDPR controller-processor agreements must meet Article 28 requirements, and a processor may use a subprocessor only with the controller's prior specific or general written authorisation, subject to the required conditions. Bedrock in an EU region can simplify the contract surface, but it does not erase those obligations.

The pricing comparison in the earlier version used model names and rates that are not supported by the linked official pricing pages. Claude Opus 5 and GPT-5.6 Luna should therefore not be presented as verified products or prices here. The official pages should be checked for the exact model ID, region, modality and billing tier before a routing calculation is made. List prices can also differ by input and output tokens, context length, batch processing, provisioned capacity and the way Bedrock charges for a managed model. The general economic point remains valid: a workload that meets its quality and compliance requirements on a less expensive model should not automatically be routed to the most expensive one.

What we would do with this

From our own production experience — article pipelines, transcription, TTS — model choice is an operational decision, not a brand decision. We keep an abstraction layer between our services and the models, benchmark new releases on our AI Arena rig, and route each task to the cheapest model that meets the quality bar. The AWS announcement makes that approach easier in principle: the platform supports mixing vendors per workload, subject to actual model and region availability.

For European companies, the takeaway is practical: you do not have to pick a single AI vendor. The infrastructure supports mixing providers per workload, but model availability, pricing and data routing must be checked rather than assumed from a launch headline. Just make sure your compliance layer keeps pace with your routing layer — the GDPR and AI Act obligations follow the processing arrangement and the use case, not simply the console.

Amazon's broader investment in cloud infrastructure may improve model availability, but it does not by itself guarantee European access to US-only GovCloud services or identical EU Bedrock functionality. Do not wire your workflows so deeply into one platform that switching becomes impossible.