Amazon’s first-quarter 2026 earnings and the launch of Amazon Bedrock Managed Agents powered by OpenAI together paint the clearest picture yet of how AWS is positioning itself as the primary infrastructure layer for enterprise AI — with both of the leading frontier AI labs, Anthropic and OpenAI, now running production agent workloads through Bedrock. The earnings and the Bedrock launch happened the same week, making the overall strategic picture unusually readable.
The Capital Expenditure Scale
Amazon confirmed it is on track to spend $200 billion in capital expenditures in 2026 — nearly double the company’s capex in 2025 and the largest single-year infrastructure investment in the company’s history. CEO Andy Jassy has described AI as “an extraordinarily unusual opportunity to forever change the size of AWS and Amazon as a whole.” The $200 billion figure is the financial expression of that conviction.
The spending covers several categories:
- Data center construction across US and international regions, specifically optimized for AI training and inference workloads
- AWS Trainium and Graviton custom silicon — Amazon’s in-house chip program now generating over $20 billion annually in revenue, roughly double the figure cited earlier in 2026
- Amazon Bedrock infrastructure — the managed AI platform that now hosts both Anthropic’s Claude and OpenAI’s GPT and Codex products
- Networking and cooling infrastructure for GPU-dense data center configurations that AI training requires
Bedrock Managed Agents: What the OpenAI Launch Revealed
The launch of Amazon Bedrock Managed Agents powered by OpenAI in San Francisco this week provided the clearest description yet of what the product actually is. AWS CEO Matt Garman’s framing: enterprise customers who have been using AWS as their primary infrastructure for years have been forced, if they want OpenAI’s most capable models via API, to route those calls to Microsoft Azure. That friction has been one of the most common enterprise complaints about the OpenAI-Microsoft arrangement.
Bedrock Managed Agents resolves it. The product allows enterprise customers to build agents using OpenAI’s models — including Codex for coding workflows — that run entirely within AWS’s security perimeter. Customer data stays within the enterprise’s Virtual Private Cloud (VPC). The agent infrastructure handles memory persistence across sessions, context management for long-running tasks, and tool-call orchestration — the same capabilities that Anthropic’s Claude Managed Agents provides for enterprises using Claude.
OpenAI’s chief revenue officer Denise Dresser, who joined the company four months ago, described the demand dynamic at the San Francisco launch: “The hundreds of enterprise customers I’ve met since joining are past the experimentation phase with AI. They understand that to do that, they need powerful models — but even more importantly, they want those models in a trusted environment that they know and a trusted infrastructure.”
The Dual-Lab Bedrock Strategy
Amazon is now running both major frontier AI labs on Bedrock simultaneously — Anthropic through its deep investment partnership and multi-year deployment agreement, and OpenAI through the new post-exclusivity arrangement. This positions Bedrock not as a model provider but as a model-agnostic enterprise AI platform: the infrastructure through which enterprises access whichever frontier AI capabilities they need, regardless of which lab produced them.
The strategic logic is strong. Enterprise buyers increasingly want flexibility — the ability to use Claude for tasks where its instruction-following and writing quality lead, and GPT for tasks where OpenAI’s tool ecosystem or Codex coding capabilities are the priority. Running both through a single platform eliminates the multi-cloud management overhead while retaining model flexibility.
Amazon has invested up to $25 billion in Anthropic and committed $50 billion to OpenAI — a total AI commitment from the equity side alone that dwarfs any single AI infrastructure investment made by a competitor. The Bedrock platform is where that capital is being operationalized into enterprise revenue.
The Trainium Signal
AWS Trainium’s $20 billion annual revenue figure — disclosed in Amazon’s Q1 earnings and CEO Andy Jassy’s shareholder letter — is particularly significant for the competitive dynamics of the AI hardware market. Trainium is Amazon’s in-house alternative to NVIDIA GPUs for AI training workloads. A $20 billion annual run rate for chips that barely existed as a commercial product two years ago is a meaningful challenge to NVIDIA’s dominance in the AI training market.
Both Anthropic and OpenAI have committed to running workloads on Trainium as part of their AWS agreements. As Trainium’s training performance and availability improve, the effective cost of frontier model training on AWS decreases — benefiting both labs’ cost structures and potentially enabling lower API pricing for enterprise customers over time.
Conclusion
Amazon’s $200 billion capex commitment and the Bedrock dual-lab launch together represent the clearest statement yet that AWS is betting on infrastructure dominance, not model development, as its AI strategy. Whoever builds the best models, Amazon will host them — and take a share of the resulting enterprise revenue. It is, in some ways, the same strategy that made AWS dominant in cloud computing, now applied to AI. Browse our directory to explore Claude, ChatGPT, and every AI tool running on the infrastructure Amazon is building at this scale.
