AWS Says AI Inference Now Drives Over Half of AI Revenue as Amazon Pushes Trainium and Nova Ahead of re:Invent
With AWS running near a $182 billion annualised pace, leadership says inference now exceeds training revenue while capacity constraints fuel external Trainium chip sales and Amazon Nova adoption ahead of re:Invent 2026.

By Clara Tan, Artificial Intelligence
, updated 6 min read
Artificial intelligence inference has overtaken model training to generate more than half of Amazon Web Services' AI revenue, cloud leadership disclosed in mid-November 2026, underscoring a structural shift in cloud economics as AWS prepares to showcase its Trainium silicon and Amazon Nova models at re:Invent later this month.
Operating at an annualised revenue run rate near $182 billion with strong double-digit expansion, AWS executives say enterprise demand for running production models has outpaced available data centre power and silicon, driving both cloud migrations of legacy databases and a push to sell custom Amazon chips directly to external buyers.
Inference Economics Reshape the Cloud Stack
For the past three years, hyperscaler capital expenditure narratives revolved around massive pre-training clusters built to train frontier weights. According to AWS leadership comments in early November, the revenue centre of gravity has now flipped: inference workloads account for over 50% of AWS AI revenue, and executives estimate enterprise inference deployment still represents less than 1% of its eventual market scale.
Unlike sporadic training jobs, inference scales linearly with daily enterprise usage across customer support bots, code review agents, document pipelines, and voice interfaces. Crucially, AWS notes that enterprises have realised they cannot deploy reliable AI agents on top of fragmented on-premises databases, accelerating core lift-and-shift migrations of ERP, transactional, and analytics data into AWS. Even when a customer only wants to automate customer billing disputes or internal IT tickets, connecting models to clean, low-latency operational records requires consolidating legacy silos onto managed cloud databases first.
Expanding External Sales of Trainium Silicon
Tight data centre and GPU supply has turned AWS's in-house silicon programme from a defensive cost-control measure into a primary commercial differentiator. AWS CEO Matt Garman emphasised a strategic push to expand external sales and adoption of Trainium processors as an alternative to Nvidia GPUs for both training and high-throughput inference.
Software friction has historically slowed adoption of non-Nvidia accelerators, since most open-weight inference stacks optimise for CUDA first. AWS moved to narrow that gap in November by releasing SGLang-Neuron in a public AWS Deep Learning Container, allowing developers to serve models and execute reinforcement learning rollouts on Trainium using standard SGLang and PyTorch interfaces without rewriting model serving code. Paired with AWS's claim that its data centres now add a custom AI chip roughly every 14 seconds, Amazon is betting that software compatibility plus raw cluster availability will persuade AI startups and regulated enterprises to run production inference on Trainium3 rather than waiting in Nvidia allocation queues.
What to Watch at AWS re:Invent 2026
Scheduled for 30 November to 4 December in Las Vegas, AWS re:Invent 2026 features more than 140 dedicated AI and machine learning sessions. Pre-conference tracks focus on four engineering priorities: moving beyond single-prompt chatbots to stateful, multi-agent systems built on Amazon Bedrock AgentCore; routing workloads dynamically across Nvidia GPUs and AWS Trainium for lower token costs; extending Amazon's first-party Nova foundation models into multimodal enterprise deployments; and binding vector stores, S3 data lakes, and IAM identity boundaries directly into agent runtimes.
Enterprise buyers are also watching how AWS packages first-party and partner models together. OpenAI's 9 November decision to list ChatGPT Enterprise in the AWS Marketplace underscores how cloud marketplaces have become neutral procurement hubs, even as AWS competes aggressively with Bedrock hosting for Anthropic's Claude family, Meta's open weights, and its own Nova models.
Competitive Stakes Across APAC and Global Cloud Buyers
In markets such as Singapore, Tokyo, and Sydney, cloud buyers face acute sensitivity to token unit economics and data residency rules. If AWS can demonstrate at re:Invent that Trainium clusters running SGLang-Neuron match Nvidia latency on 70B-to-400B parameter open models at substantially lower hourly rates, cost-conscious SaaS startups and banks gain genuine leverage in multi-cloud negotiations with Microsoft Azure and Google Cloud.
At the same time, capacity constraints remain AWS's primary governor on growth. Until newly contracted power interconnects and custom chip racks come online across 2027, balancing GPU allocation between frontier partners such as Anthropic and mainstream enterprise Bedrock customers will remain AWS's sharpest operational test.
Sources
- 1AWS re:Invent 2026 Official Event Hubaws.amazon.com
- 2AWS What's New: SGLang-Neuron and AgentCore Updatesaws.amazon.com