Bold claim: Amazon is building a gated AI empire for enterprises, presenting itself as the easy button while locking customers into a contained ecosystem. At AWS re:Invent, CEO Matt Garman outlined a vision to make enterprise AI more practical—from infrastructure to custom models and ready-made agents—yet the underlying pattern echoes a familiar move: start with hardware, layer on abstractions, and tighten control as capability grows.
Garman argued that many customers haven’t yet realized AI’s promised returns, a sentiment echoed by an MIT study cited in August. The study reported that enterprises invested roughly $35–$40 billion in generative AI efforts but have yet to demonstrate meaningful results. The implication is clear: there remains significant room for value, provided vendors can clearly prove AI’s business impact.
Amazon’s strategy mirrors its early cloud playbook: reduce friction by building from the bottom up, then monetize scarcity through specialized, high-level services. The trade-off is reduced portability: easy entry comes with stronger lock-in.
Nova Forge: a bridge to domain-specific models
A centerpiece of Amazon’s approach is Nova Forge, a new platform meant to simplify the creation of custom generative AI models. Garman suggested that frontier models today often struggle to align with a company’s unique data and domain, and that integrating data at the right stage of training could yield a proprietary model tailored to the organization. The promise is to empower customers to develop a model that understands their business while preserving the model’s foundational reasoning abilities.
Forge offers access to partially trained Nova checkpoints. Clients can finish training with a blend of their own data and AWS-curated datasets, aiming to produce a domain-specific model without starting from scratch. The expectation is a proprietary end product, branded as Novellas, deployed on AWS Bedrock, AWS’s AI-as-a-service layer. Bedrock runs on a mix of Nvidia GPUs and AWS’s own accelerators, removing the burden of hardware management and low-level software tuning.
However, the resulting models remain confined to AWS. They’re proprietary to the customer, but portability beyond AWS is limited. The same confinement applies to Nova LLMs. Nova 2 comes in four variants—Lite, Pro, Sonic, and Omni—covering reasoning, speech-to-speech capabilities, and multimodal inputs. While Bedrock also hosts open-weight models like Mistral Large and Mistral 3, those Forge-ready capabilities aren’t available for Forge models.
This approach addresses the “stickiness” problem common with API-only services: if a customer builds significant value inside a provider’s stack, it becomes harder to switch away. While beneficial for AWS’s business, it raises questions about long-term portability for enterprises.
Agent orchestration and governance
Beyond models, AWS is pushing tools to simplify building AI agents that perform multi-step tasks with minimal supervision. At re:Invent, two notable additions to Bedrock Agent Core were unveiled to bolster trust and reliability.
First, a new policy extension lets organizations define which tools and data an agent can access and how those resources are used. For example, a policy might prevent a returns authorization for items over a certain value, triggering human oversight instead. Garman argued that clear policies enable deeper trust in agents by ensuring they operate within defined boundaries.
Second, an evaluation suite is designed to monitor agent behavior in real time. Continuous evaluation helps teams detect undesirable actions promptly and adjust as needed. The goal is to prevent subtle regressions in performance when the base model is updated.
In addition to bespoke agents, AWS highlighted a growing catalog of pre-baked agents in the cloud marketplace, including options aimed at accelerating development and strengthening cybersecurity. The overarching message is that agents don’t have to be universal; they should connect to the right mix of tools, services, and models—some of which are AWS offerings and others that aren’t.
A premium on choice, with caveats
Garman emphasized flexibility: builders can select the building blocks that fit their needs rather than follow a single, fixed path. Yet behind the scenes, AWS presents a tightly integrated stack that can be difficult to migrate away from. The alliance between Forge, Nova models, and Bedrock creates a cohesive, end-to-end experience that—while convenient—also locks customers into AWS’s ecosystem.
In short, Amazon is crafting a comprehensive enterprise AI experience that blends custom models, managed agents, and pre-built tools under a single roof. The company argues this simplifies adoption and accelerates value creation. Critics, however, point to limited portability and the risk of vendor lock-in as major concerns for organizations investing heavily in AWS-native solutions. The question for leaders is whether the ease and speed of AWS’s integrated approach justify the potential costs of future migrations and the constraints of sustaining value outside the provider’s ecosystem.
What do you think about this strategy? Is the potential for faster time-to-value worth the trade-off of reduced portability and increased lock-in, or should enterprises push harder for open, interoperable architectures that can roam across clouds? Share your thoughts in the comments.
