June 30, 2026, (Inside AI) — Two frontier AI models landed today on Kiro, the integrated development environment and command-line tool, inside the AWS GovCloud (US-West) Region. OpenAI’s GPT-5.4 and NVIDIA’s Nemotron 3 Super 120B are now selectable for developers building sensitive, regulated workloads.
The launch puts both proprietary and open-weight reasoning engines directly into the hands of government and enterprise teams that require isolated infrastructure. Kiro users must update to the latest IDE or CLI version and restart to see the new options.
This dual release signals a deliberate push by AWS to offer choice and control inside its air-gapped cloud. GPT-5.4 targets complex multi-step agentic flows, while Nemotron 3 Super 120B delivers an open-weight, compute-efficient alternative.
The Models: Agentic Power and Open Efficiency
OpenAI’s GPT-5.4 is positioned for advanced reasoning, coding, document analysis, and multi-step agentic workflows. It runs on Amazon Bedrock’s next-generation inference engine, which provides isolated queues and durable execution designed for resilient, long-running tasks.
The model ships with a 272,000-token context window and carries a 1.2x credit multiplier. That pricing signal suggests AWS expects heavier consumption for sustained agentic loops compared to simpler chat models.
NVIDIA’s Nemotron 3 Super 120B enters as a hybrid mixture-of-experts architecture. It activates only 12 billion of its 120 billion parameters during inference, a design choice that aims to slash compute cost while preserving performance on agentic tasks.
Its 256,000-token context window supports up to 32,000 tokens of output. The 0.25x credit multiplier makes it a cost-conscious option for high-volume deployments where budget and sovereignty matter equally.
Why GovCloud and Why Now
Placing these models inside AWS GovCloud (US-West) addresses a persistent tension: how to use frontier AI without data leaving controlled environments. Government agencies, defense contractors, and regulated industries often cannot send prompts to public endpoints.
By routing through Bedrock’s isolated inference engine, GPT-5.4 inherits the compliance posture of the GovCloud fabric. Meanwhile, Nemotron 3 Super 120B’s open-weight nature allows security teams to inspect, fine-tune, or even self-host the model later.
The move follows a broader industry pattern where model providers and cloud vendors compete on deployment sovereignty. Microsoft’s Azure Government and Google’s air-gapped offerings have similarly courted public-sector AI workloads, but the combination of Kiro’s developer surface and Bedrock’s managed inference gives AWS a distinct toolchain play.
What remains unstated is the latency profile inside GovCloud. Neither AWS nor the model builders disclosed inference speed benchmarks. For agentic workflows that chain dozens of tool calls, tail latency can make or break a production system.
Also absent is any mention of fine-tuning availability. GPT-5.4’s Bedrock deployment may eventually support customization, but no timeline was given. Nemotron’s open weights theoretically allow it, yet AWS did not confirm whether GovCloud offers the necessary compute for training runs.
Pricing beyond the credit multipliers remains opaque. Kiro’s credit system abstracts underlying costs, but heavy users will need to map consumption carefully to avoid surprises. The 1.2x multiplier on GPT-5.4 suggests a premium for agentic reasoning, while the 0.25x on Nemotron hints at a volume play.
For developers already inside the AWS GovCloud ecosystem, the update removes friction. They can now experiment with frontier reasoning models without leaving their compliant toolchain. For everyone else, the announcement is a reminder that the most powerful AI models are increasingly gated behind infrastructure choices.
