Amazon Web Services posted $45.99 billion in third-quarter 2026 revenue, up 39.3% year over year. Backed by a new AI accelerator installed every 14 seconds, the cloud unit is pushing a record investment budget, and Andy Jassy now targets $1 trillion in annual revenue within five to ten years.
🔑 In brief
- Growth: AWS revenue rose 39.3% year over year, a $184 billion annualized run rate and a sixth straight quarter of growth.
- Capex: Amazon expects about $220 billion of 2026 cash capex, with significantly higher spending planned for 2027.
- Target: Jassy expects growth above 40% in 2027 and 2028, and a path to $1 trillion in revenue within five to ten years.
- Risk: Amazon shares fell about 6% after hours as investors demanded evidence of returns.
A strong quarter, with a smaller revenue share than the source suggests
Amazon generated $180.17 billion in total revenue during the quarter. AWS contributed $45.99 billion, or roughly 25.5% of the total by direct calculation. One source places AWS at about 68% of revenue, a figure that cannot be reconciled with the reported amounts, so we rely on the direct calculation. The $184 billion run rate cited by Jassy is exactly four times the quarterly figure.
The segment still grows fastest in the group. Retail and operations rose only 6%, with thinner margins. Prime Video has become profitable and generates revenue in the tens of billions of dollars, but it remains secondary to AWS’s momentum.
| Metric | Figure | Context |
|---|---|---|
| AWS revenue, Q3 | $45.99B | +39.3% year over year |
| Amazon total revenue, Q3 | $180.17B | All segments combined |
| AWS annualized run rate | $184B | Four times the quarter |
| Capex, Q3 | $75.9B | Mainly for AWS |
| Expected 2026 capex | about $220B | Majority for AWS |
An investment program without precedent
Amazon raised its 2026 cash capex budget to roughly $220 billion, most of it for AWS, and expects to spend meaningfully more in 2027. Capital expenditure reached $75.9 billion in the third quarter alone, mainly for AWS. Jassy said spreading the program over twenty-five years would not match the scale of the opportunity, and he called AI the biggest transformation of his career.
On funding, he was direct.
“We can fund this plan, and I think it will forever change the size of Amazon. It will forever change our revenue.”
Andy Jassy, CEO, Amazon
A second excerpt from his remarks added that the program will change free cash flow as well as the size of the company, and that Amazon intends to remain the leader it is today. Free cash flow is therefore the metric to watch when judging the return on this spending.
Demand reaches well beyond accelerators
Jassy estimated that about 85% of global IT spending still sits on premises, a balance he expects to reverse over the next decade or two. Workloads once considered fixed, including mainframes and storage networks, are already migrating to AWS.
Customers running AI workloads on AWS need inference systems close to their existing data and applications, which lifts demand for conventional compute. AI itself consumes that capacity for reinforcement learning and post-training, creating a cycle that pulls investment toward broader cloud functions.
Bedrock and AI margins
AWS Bedrock is described as the multibillion-dollar service with the fastest growth in Amazon’s history. Core compute and storage revenue also grew at the fastest pace in several years. Jassy said AI services already carry higher margins than core cloud services did at comparable stages of development.
- Bedrock: fastest growth in Amazon’s history for a service now worth several billion dollars
- Compute and storage: acceleration at the fastest pace in several years
- AI margins: already above core cloud margins at a comparable stage
Jassy described enterprise AI adoption as still in its early days, with AWS in what he called “the very early days,” and said the strategy is working. The sources provide no margin figure, so these claims remain qualitative.
Investors want proof, not just spending
Investors remain focused on the cost of the build-out. On the call, Wall Street analysts pressed Jassy to justify the investments and to name the indicators that would show growth and expected margins materializing. Amazon shares fell about 6% in after-hours trading.
The sector as a whole combines heavy spending with strong revenue growth. Microsoft also reported another record revenue increase, driven mainly by its cloud business, while spending heavily. The Register summed up the tension in one line: Amazon could become “a cloud company with a retail side hustle.”
A reader comment in The Register’s forum linked cloud migration to Dell’s share price, which the author said rose from $50 to $400. That is an individual reaction rather than verified market data, so we do not use it as analysis.
Inferentia versus GPUs: what AWS’s test shows
An AWS technical post notes that after training a model, machine learning teams can spend several weeks choosing the instance types, accelerators, containers and compilation settings that best fit their throughput, latency, resource use and cost targets. Amazon SageMaker provides containers for XGBoost, scikit-learn, PyTorch, TensorFlow and Apache MXNet, along with real-time inference endpoints and the AWS Inferentia accelerator. SageMaker Neo automatically compiles models for different target hardware.
The test uses a PyTorch ResNet50 computer vision model across three setups: a c5.xlarge CPU instance, a g4dn.xlarge instance with an NVIDIA T4 GPU, and an inf1.xlarge instance based on AWS Inferentia. Simultaneous client connections were used to reach maximum throughput. The input was a 21 KB JPG image of 3 × 224 × 224 pixels showing a beagle puppy, sent to a SageMaker endpoint.
| Setup | Throughput | Latency | Cost per million inferences |
|---|---|---|---|
| g4dn.xlarge, uncompiled | Baseline (1x) | Baseline | Baseline |
| g4dn.xlarge, compiled with Neo | 1.9x | 50% lower | 50% lower |
| inf1.xlarge, compiled | 4.1x versus uncompiled g4dn | 77% lower | 90% lower |
AWS also states that an Inferentia instance delivers 5.4 times the price-performance of a compiled g4dn instance, and up to 2.3 times the throughput with up to 70% lower cost per inference than comparable GPU-based EC2 instances. ResNet50 has 48 convolution layers, one max pooling layer, one average pooling layer and 3.8 billion floating-point operations. The model was pretrained on ImageNet and sourced from PyTorch Hub, and the architecture dates to the 2015 paper “Deep Residual Learning for Image Recognition.”
These figures come from the vendor, were obtained on a single model and a specific configuration, and should be read as orders of magnitude rather than guarantees that apply to every workload.
Outlook and scenarios
The central case rests on growth staying above 40% in both 2027 and 2028, as Jassy suggests. The sources contain no direct numerical contradiction, but two nuances matter: the $184 billion figure is described at times as an annualized run rate and at other times as expected revenue for the fiscal year, two different metrics; and a secondary summary speaks of growth of “40% or more” in either year, whereas Jassy’s quoted words point to at least 40% in each of the two years.
In a bullish scenario, AI demand and the migration of conventional workloads to the cloud sustain revenue, and AI service margins confirm their early lead. In a bearish scenario, spending that stays above revenue growth, a slower payback or margin pressure would weigh on the valuation, as the after-hours share price move suggested. Between the two, the indicators requested by analysts, particularly AI service margins and the trajectory of free cash flow, will be the key checkpoints in the coming quarters.
Sources
- The Register, AWS article (October 30, 2026)
- Redirect link to The Register
- The Register forum, reader comments (October 30, 2026)
- Lavx.hu, summary of the announcement
- AWS Machine Learning Blog, choosing an accelerator and compiling for inference
Cet article est publié à titre informatif et éducatif. Il ne constitue en aucun cas un conseil en investissement. Faites vos propres recherches (DYOR) avant toute décision.
