Foreword¶

On July 30, 2026, Amazon released its Q2 financial report ending June 30. For developers focused on cloud-native and AI infrastructure, the most impactful signal in this report was not the update of any single product, but that AWS regained its fastest growth rate in 18 quarters — with quarterly net sales of $42.2 billion, a year-over-year increase of 37%, and an annualized run rate of $169 billion. Meanwhile, AWS’s internal AI business and self-developed chip business both exceeded an annualized revenue run rate of $2.5 billion, both achieving double-digit year-over-year growth; the company also raised its 2026 full-year CapEx guidance to approximately $220 billion. CEO Andy Jassy said bluntly on the earnings call: Even at this scale of investment, we will not be able to meet all demand in 2026, and this will likely be the case in 2027, while the order volume for 2028 is already “strikingly large”.

This article is based on Amazon’s official IR announcement and Jassy’s public interpretations, sorting out the technology stack layout behind these figures and what it means for the落地 of cloud-native AI.

Overview of Q2 Core Data¶

First, list the key indicators verified by official announcements (currency unit: USD):

It should be noted that AWS’s quarterly growth rate has two expressions in the announcement: 37% (rounded) and 36.7% (exact value), which are essentially different precisions of the same set of data. The “2.5 billion” for AI and chip business refers to the annualized run rate — the annual level折算d by the current quarterly revenue rhythm, not single-quarter revenue.

Why AWS Returned to an Accelerated Growth Track¶

Wall Street’s previous expectation for AWS’s growth rate was about 31%, and the actual 37% was significantly higher than expected, making it the fastest quarter since 2021. Jassy attributed the growth to two interlocking curves: AI workloads directly drive computing power and model service consumption, while post-training, reinforcement learning, Agent tool calling and other links after AI training largely run on CPUs rather than AI accelerators, thereby driving the synchronized growth of “core cloud” services such as storage, vector databases, and Graviton instances.

For developers, this logic chain is very specific: When you run inference on Bedrock or fine-tune models on SageMaker, you not only need GPU/Trainium, but also S3 for data storage, OpenSearch for retrieval, and Lambda for Agent logic. AWS’s selling point is “inference as close as possible to existing applications and data” — the company claims that more enterprise data is hosted on AWS than on any other cloud platform worldwide. After the Q2 earnings report, if AWS were to be listed independently, it would rank among the top 25 of the Fortune 500 based on its 169 billion run rate, with a scale close to that of a giant independent software company.

AI and Self-Developed Chips: Two $2.5 Billion Runway Tracks¶

AI Service Layer: Bedrock and SageMaker¶

Among the earnings highlights, AWS AI business run rate exceeding 2.5 billion with double-digit growth mainly corresponds to model construction, inference, and enterprise Agent deployment.
- Amazon Bedrock: A model hosting and inference platform. Over 10 fully managed foundation models were added in Q2, including OpenAI GPT-5.6, Anthropic Claude Opus 5, Google DeepMind Gemma 4, etc. Official data shows that Bedrock has hundreds of thousands of customers; the number of new customers in the past six months exceeded the sum of the first two years after the product launched; total customer spending in Q2 exceeded the sum of all previous quarters.
- SageMaker AI: Targeted at enterprises that want to build small and medium-sized proprietary models using their own data. Jassy has repeatedly emphasized that companies with strong technical capabilities will not only rely on frontier large models, but will train dedicated models adapted to their own data, and SageMaker is the main tool for this path.
- Bedrock AgentCore: Solves the “dirty work” of bringing Agents into production — secure runtime, memory, identity, tool connection and observability. Capabilities such as Policies (deterministic permission control), Payments (autonomous transactions), Web Search (domain knowledge grounding) and Harness (accelerated assembly) were added in Q2.

Chip Layer: Trainium, Inferentia and Graviton¶

The self-developed chip business run rate also exceeds 2.5 billion, covering two lines: AI accelerators and general-purpose CPUs:
- Trainium: Targeted at AI training and inference. Anthropic and OpenAI have signed multi-year, multi-GW-level Trainium procurement commitments; Uber, Pinterest, and AI startups such as NEURA Robotics and Odyssey are also adopting it.
- Inferentia: An earlier launched inference-specific chip, which forms AWS’s self-developed AI computing power matrix together with Trainium; delivered on top of the Nitro security and network architecture.
- Graviton: A general-purpose cloud CPU, which the company claims has a 30%-40% cost-performance advantage over similar instances; Graviton5 has been GA, with computing power increased by about 25% compared to Graviton4, and is used by 98% of the top 1000 EC2 customers. With the increase in Agent-based workloads, the demand for “non-accelerator” computing power carried by Graviton has also risen synchronously.

The significance of the chip run rate is that a considerable proportion of the computing power revenue brought by AI demand remains in AWS’s own silicon gross profit structure, rather than all flowing to external GPU suppliers. This is crucial for understanding AWS’s long-term AI infrastructure ROI.

$220 Billion CapEx: The Game Between Investment and Return¶

CapEx was another focus of this earnings report. Single-quarter PPE acquisition in Q2 was approximately $54.2 billion, a year-over-year surge; Jassy raised the 2026 full-year cash CapEx guidance from approximately $200 billion to approximately $220 billion, partly due to rising memory costs.

More noteworthy is the management’s judgment on supply and demand:

Even with a $220 billion investment, we will not be able to meet all demand in 2026; 2027 is expected to be equally tight; the locked-in demand for 2028 is strikingly large.

We can also see the intensity of investment from free cash flow: In the 12 months ending June 30, 2026, operating cash flow was $161.4 billion (+33%), but free cash flow was a net outflow of $7.6 billion, mainly due to a $66.1 billion year-over-year increase in PPE purchases, which the company explicitly stated “mainly reflects investments in artificial intelligence”.

For the industry, these figures show that the verification period for AI infrastructure ROI is shortening — AWS has proven the real existence of demand with its $2.5 billion AI run rate, but bottlenecks on the supply side (data centers, chips, memory) mean that customers may still face capacity queuing and price pressure. When cloud-native teams conduct capacity planning for 2026-2027, they should include “locking in computing power and regions in advance” into architectural decisions, rather than assuming that on-demand resources are always available.

Implications for Developers and Cloud-Native AI¶

Combined with Jassy’s statements in the official blog post, there are several points with direct reference value for engineering practice:
1. Deploy inference and data in the same domain: Enterprises tend to place AI inference close to existing applications and data to reduce cross-cloud egress and compliance frictions. Existing workloads already on AWS are a natural source of incremental demand for Bedrock/SageMaker.
2. The threshold for bringing Agents into production lies in “infrastructure glue”: Q2 releases such as AgentCore, Lambda MicroVMs (supporting up to 8 hours of stateful sessions), and OpenSearch Serverless all point to the same problem — the gap between demo and 24/7 production lies in managed runtime, memory, and observability, not just switching to a larger model.
3. Hybrid computing power model: Training and inference lean toward accelerators (Trainium/GPU), while post-training and Agent toolchains lean toward CPUs (Graviton). When doing cost modeling, you should not only estimate based on GPU hours, as CPU, storage, and vector retrieval will account for a significant proportion.
4. Security and governance have become hard indicators for selection: Continuum (code vulnerability discovery and repair Agent), Bedrock governance controls, and Quick enterprise access controls reflect that enterprise customers’ top concern for AI deployment is security and governance. Technical selection needs to simultaneously evaluate policy layers (Policies) and audit capabilities.

Conclusion¶

Amazon’s Q2 2026 financial report provides a verifiable answer with AWS’s 37% growth rate and dual $2.5 billion run rates (AI + chips): AI infrastructure investment is being translated into substantial revenue, and growth is not limited to “selling GPUs”, but covers the full stack of model services, Agent platforms, and self-developed silicon. The $220 billion CapEx and the statement that production capacity is still in short supply remind developers that in addition to model effectiveness, the next competitive dimension of cloud-native AI also includes computing power availability, regional layout, and full-stack cost.

For teams that are evaluating Bedrock fine-tuning, Trainium training, or AgentCore production deployment, what is worth doing now is not waiting for the “bubble burst”, but breaking down their own workloads: How much computing power is on accelerators, how much is on CPUs and storage, and which managed capabilities are missing for Agent launch. The earnings figures are a weathervane, but engineering落地 still needs to be aligned item by item in the architecture diagram.