Choosing US-only inference for Kimi K3 costs about 11% more than Bedrock's global profile

Two strings carry most of the cost decision. For workloads with no regional restriction, AWS recommends the global profile `global.moonshotai.kimi-k3`, which sends each request to any supported commercial AWS Region worldwide ^[7]. The US geographic profile, `us.moonshotai.kimi-k3`, keeps processing inside the US geography for data residency requirements ^[8]. Read the 10 percent discount from the other end and the residency choice costs roughly 11 percent more per request: if global is 0.90 of geographic, geographic is 1 divided by 0.90 ^[10]. That premium applies to every token the agent spends, including the ones it spends re-reading a file it read two turns ago.

The token multiplier is the other half. AWS cites Gartner's 2026 analysis putting agentic workflows at 5 to 30 times the token consumption of ordinary use ^[15]. Consumption pricing removes the per-seat floor and replaces it with a bill that tracks how often the agent retries.

What you run locally is a Go binary ^[1]. It edits files, runs shell commands, and reads project structure through Language Server Protocol diagnostics ^[2], and it speaks to more than 75 model providers, Bedrock among them ^[3]. Inference happens inside your AWS account, with no GPU provisioning and no inference infrastructure to manage ^[4][20]. Access rules are IAM policies, the audit trail is CloudTrail, the network path can be PrivateLink, and encryption controls are the same ones proprietary models on Bedrock inherit ^[14]. Bedrock is in scope for HIPAA, SOC 2, ISO 27001, FedRAMP and GDPR ^[13].

The performance argument leans on one number from someone else's workload. AWS cites CrowdStrike's fine-tuned NVIDIA Nemotron at 96 percent valid query accuracy, against GPT-4o at 61 percent and Claude Sonnet 4.5 at 94 percent ^[17]. That is not one experiment: the winning model was fine-tuned on CrowdStrike's domain, and the metric is whether a generated query is valid, not whether a patch compiles. The margin over Sonnet 4.5 is two points ^[21]. For the figure to transfer to your repository you would need labelled data in your own domain, a fine-tuning run, and a validity metric that predicts working code; the walkthrough's Nemotron 3 Super 120B is served stock from Bedrock ^[5].

AWS scopes its cost claim narrowly: at the order of multimillion conversations per month, switching to open weight models on Bedrock can reduce annualized costs ^[16]. The post does not publish a per-seat comparison, and the thing it argues against is per-seat subscriptions charged regardless of how much you use them ^[19]. McKinsey's 2025 open-source report, also cited, says 76 percent of organizations expect to increase open source AI usage and that leading AI adopters are 40 percent more likely to use open weight models ^[18]. That report measures expectation, not deployments. A team with a residency clause gets what it needs from the profile ID and its existing IAM and logging ^[12]. A team hoping to cut a seat bill has to measure its own token volume first.