What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Stability AI selected Amazon Web Services (AWS) as its preferred cloud provider at AWS re:Invent in late November 2022. The plan was to use Amazon SageMaker, large GPU and AWS Trainium clusters, and distributed-training software to build foundation models for image, language, audio, video, and 3D generation. AWS later added a distribution role: Amazon Bedrock lets customers call supported Stability AI models through managed APIs.
The announcement was therefore both an infrastructure deal and a route to enterprise adoption. It does not establish that AWS is Stability AI’s exclusive cloud provider, that every Stability AI workload runs on AWS, or that every historical Stable Diffusion model remains available through Bedrock.
What was announced at AWS re:Invent 2022?
AWS said Stability AI had chosen it as a preferred cloud provider for developing and scaling foundation models. Stability AI described a broad target spanning image, language, audio, video, and 3D systems. The company used SageMaker-managed infrastructure and large accelerator clusters rather than building every training component itself.
AWS’s announcement is the primary account of the arrangement: Stability AI builds foundation models on Amazon SageMaker. Contemporary coverage published by VentureBeat on December 2, 2022 reported that Stable Diffusion 2.0 had been built on AWS and that Stability AI was training GPT-NeoX across approximately 1,000 Nvidia A100 GPUs: VentureBeat’s December 2022 report.
Recommended Free Tools
#1 Best Overall
“Preferred cloud provider” is narrower than “exclusive provider.” The public announcement does not prove that all models, products, customer requests, or inference workloads used AWS.
Why foundation-model training needs hyperscale infrastructure
Training a large model is a distributed-computing problem. Data must be stored and streamed to many accelerators; the accelerators must exchange updates over high-bandwidth networks; and the system must recover from hardware or software failures without losing days of work.
- Accelerators: Thousands of GPUs or purpose-built chips provide the parallel arithmetic required for model training.
- Networking: Fast interconnects reduce the time machines spend waiting for one another.
- Storage: Datasets, intermediate checkpoints, logs, and model versions can occupy substantial capacity.
- Fault tolerance: Checkpointing and restart mechanisms protect long-running jobs from failures.
- Inference capacity: Once a model is released, serving unpredictable demand requires autoscaling, monitoring, and reliable endpoints.
Cloud is not the only option. A company can build its own data center, lease dedicated capacity, or spread workloads across providers. Hyperscale cloud mainly lowers the operational barrier to obtaining capacity and supplies integrated networking, storage, identity, and monitoring.
Rank #2
What SageMaker contributed
SageMaker is AWS’s model-development and machine-learning platform, not merely a collection of hosted chat or image APIs. In the 2022 arrangement, it provided managed infrastructure for distributed training, model-parallel libraries, and access to GPU and Trainium clusters.
AWS said Stability AI reduced training time and cost by 58% on a GPT-NeoX-related workload using SageMaker and its model-parallel library. That is an AWS-reported result, not an independently audited benchmark, and it should not be treated as a universal saving for every model or customer.
The same caution applies to an image-generation figure reported from the 2022 re:Invent presentation: image-generation time was said to fall from about 5.6 seconds to 0.9 seconds during Stable Diffusion 2.0 development. The result depends on hardware, resolution, sampler, batch size, software version, and whether the measurement covers inference alone or the full request path. The historical account is available at VentureBeat.
Rank #3
How the relationship expanded through Bedrock
Training infrastructure served Stability AI itself. Bedrock changed the customer-facing side of the relationship. In April 2023, Stability AI announced that Stable Diffusion and future Stable models would be accessible through Amazon Bedrock: Stability AI’s Bedrock announcement.
Bedrock is a managed foundation-model service. An AWS customer sends an API request instead of provisioning GPU servers, downloading weights, and operating an inference stack. Bedrock can fit into existing AWS identity, networking, logging, and application systems, and it lets teams evaluate models from multiple providers through one service family.
Stability AI’s announcement also described private customization and integration with AWS tools such as SageMaker Experiments and Pipelines. Those capabilities do not mean the customer is training the underlying Stability AI foundation model; they support application-level customization, evaluation, or deployment around a hosted model.
What is available through Bedrock now?
AWS’s current Stability AI documentation lists a narrower, updated catalog than historical articles imply. Availability, model identifiers, quotas, and Regions can change, so verify the live documentation before committing an architecture: AWS Stability AI model parameters.
| Current category | What AWS documentation lists | Qualification |
|---|---|---|
| Image generation | Stable Image Ultra; Stable Image Core; Stable Diffusion 3.5 Large | Supported models and Regions are subject to change. |
| Image editing and control | Inpainting, outpainting, background removal, search and replace, search and recolor, sketch-to-image, structure control, style guide, and style transfer | These are specialized services, not a promise that every historical model is available. |
| Older models | Some other Stability AI models | AWS documentation warns that support for certain models is being deprecated. |
AWS’s product overview is at AWS Bedrock Stability AI. Regional endpoint information is maintained separately at AWS Bedrock model and endpoint availability.
What Stable Diffusion 3.5 Large added
Stable Diffusion 3.5 Large became available in Bedrock on December 19, 2024, initially in the US West (Oregon) Region according to AWS: AWS availability announcement. AWS describes the model as having 8.1 billion parameters and producing high-quality, one-megapixel images with broad style support and improved prompt adherence. AWS also says it was trained on Amazon SageMaker HyperPod.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minuteBest Value
Those are vendor and platform claims, not a neutral ranking against every competing image model. AWS and Stability AI position the model for advertising, gaming, media, retail, and product imagery. Technical details appear in AWS’s SD3.5 Large coverage and Stability AI’s announcement.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Bedrock versus SageMaker
AWS treats the services as complementary. Bedrock is primarily for consuming managed foundation models and building applications; SageMaker supplies broader tools for training, customizing, evaluating, and deploying models. AWS’s decision guide sets out the distinction: Bedrock or SageMaker decision guide.
| Requirement | More relevant service |
|---|---|
| Call a hosted Stability AI model through an API | Amazon Bedrock |
| Experiment with image generation without operating GPUs | Amazon Bedrock |
| Use several managed model providers in one application | Amazon Bedrock |
| Train or fine-tune a large model | Amazon SageMaker AI and HyperPod |
| Deploy a custom model with control over instances and serving | Amazon SageMaker AI |
| Control inference architecture, kernels, or hardware selection | SageMaker AI or self-hosting |
Costs and operational trade-offs
Bedrock’s managed API reduces infrastructure work but does not make the workload free. Model invocation, provisioned capacity where applicable, storage, networking, logging, monitoring, orchestration, and surrounding application services can all contribute to the bill. Current rates should be checked on AWS Bedrock pricing.
SageMaker can provide more control, but provisioned endpoints and accelerator instances can continue generating charges when utilization is low. Training at scale can also make cloud GPU costs substantial. Review SageMaker pricing and model-specific licensing before deployment.
- Bedrock advantages: fast integration, managed scaling, AWS identity and governance, and access to multiple model vendors.
- Bedrock limitations: less low-level control, regional restrictions, service-specific quotas, and dependence on the managed catalog.
- SageMaker or self-hosting advantages: control over weights, fine-tuning, hardware, network isolation, and inference optimization.
- SageMaker or self-hosting limitations: greater MLOps, security, capacity-planning, and maintenance responsibility.
Open model weights do not make every deployment route identical. Self-hosting, Stability AI’s direct services, and Bedrock can carry different licenses, pricing, rate limits, moderation rules, and commercial-use terms.
When another route may be better
AWS is most compelling when an organization already relies on AWS networking, IAM, data services, and procurement, or wants a managed multi-model platform. Other choices can be rational:
Quick Recap
- Direct Stability AI services: a narrower integration for users focused on Stability AI’s own tools. See Stability AI and its developer platform.
- Microsoft Azure: a natural fit for Microsoft identity, security, and enterprise purchasing. See Azure AI services.
- Google Vertex AI: useful for organizations invested in Google Cloud data, analytics, TPU capacity, or Google’s model ecosystem. See Vertex AI.
- Self-hosting and model platforms: Hugging Face, Replicate, or RunPod may offer broader experimentation or lower-level GPU control, but with different enterprise, privacy, and operational trade-offs. See Hugging Face, Replicate, and RunPod.
What the headline does—and does not—mean
- It refers primarily to a late-2022 preferred-provider decision, not a new 2026 announcement.
- It describes Stability AI’s announced model-building workloads, not proof that every Stability AI request runs on AWS.
- It does not mean AWS owns Stability AI’s models.
- It does not guarantee that every Stable Diffusion model mentioned in older coverage remains in Bedrock.
- It separates training from inference: Stability AI used SageMaker infrastructure, while AWS customers can consume selected models through Bedrock.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
