After a first call in May 2025, PP&A scoped a hands-on review in three phases: setup and planning, including platform accounts and an evaluation framework; platform research at three to six days per platform; and analysis and documentation. The framework set sixteen criteria in four groups. User experience and interface covered navigation, visual design, error handling and accessibility. Onboarding and learning covered time to a first trained model, tutorials, documentation and sample datasets. Development and training covered experiment tracking, the development environment, hyperparameter tuning and collaboration. Deployment and operations covered deployment speed, monitoring and drift detection, scaling, and integration with machine learning operations (MLOps) pipelines.

The method paired quantitative measures such as task completion time, click counts and error rates with qualitative ratings of pain points and workflow coherence, plus a feature parity comparison. PP&A also designed a common test case, an image classifier trained on a public set of labeled product images, to take each platform through the same stages from data preparation to a monitored, autoscaling endpoint.

The take-aways deck of July 2025 compared the platforms of two leading cloud providers, including both generations of one provider's platform. It mapped the model lifecycle into five stages and eighteen steps, rated how far the providers differ at each stage, and set their platform organization and resource deployment models side by side.

The model lifecycle is conceptually the same across providers. The differences lie in implementation. Differentiation was highest at resource deployment and model serving, moderate in training, low in optimization, and high to moderate in operations and management. The sharpest contrasts therefore included model serving, the inference step the client had named as its first priority.

The two providers organize their platforms on fundamentally different principles. One treats each service as an independent tool: developers choose and configure exactly what they need, and cross-service integration requires explicit configuration. The other places everything under a single project that connects storage, access control, monitoring and networking automatically, an approach the deck called platform as a product. The first optimizes for flexibility and specialization, the second for integration and simplicity.

Resource deployment followed the same split. The service-centric platform demands multi-step domain creation, explicit identity and access configuration, quota management and cross-service permissions, in exchange for granular control. The project-centric platform needs a project, enabled interfaces and automatically generated service accounts, with permissions inherited at the project level. Discovery was harder on the service-centric platform, but once inside, its presentation was more focused. That provider had also added integrated generative AI capabilities and a unified development studio to its platform.

The deck gave the client a stage-by-stage view of where the established platforms differ and a clear statement of the design trade-off in front of it: granular control at the cost of setup effort, or simpler setup through an integrated project model. In July, PP&A also prepared a standalone one-page version of the organizing-model comparison.

• --