Goldman Sachs is making a quiet but significant bet: the future of enterprise AI isn’t exclusively proprietary. The bank’s AI leadership has been vocal about keeping open-weight and open-source models in the mix, pushing back against the instinct to default to closed systems from the usual suspects like OpenAI and Anthropic.
The case for keeping options open
Goldman’s approach centers on a multi-model strategy. The bank’s GS AI Assistant, initially piloted in early 2025, doesn’t lock itself into a single provider. Instead, it incorporates open-source options alongside proprietary arrangements, routing queries to various providers including open-source models.
Chinese AI firms like Moonshot, which developed the Kimi model, and Zhipu, behind the GLM series, have been releasing increasingly sophisticated open-weight models that compete credibly with their closed-source Western counterparts. Examples include the Moonshot Kimi K3 and Zhipu GLM-5.2.
These Chinese developers have also pioneered an interesting middle path: “paid weights.” The core model is accessible, but certain commercial applications require licensing fees, creating a monetization pathway that doesn’t require locking everything behind an API paywall.
Goldman’s analysts have made an even more counterintuitive argument: cheaper open-source models might actually increase demand for compute resources. When AI becomes cheaper to deploy, more companies deploy it, generating more inference workloads and more demand for cloud infrastructure. Analysts specifically noted AT&T cutting model costs by 56% as an example of potential reductions in computing costs for AI applications.
The cognitive atrophy problem
Chris Churchman, who heads the bank’s AI working group, has raised a concern that most tech executives would rather not discuss publicly: the risk of cognitive atrophy. His argument is that when financial professionals lean too heavily on AI for analysis, research, and decision-making, their own analytical muscles start to weaken. Churchman has not suggested avoiding open models altogether — it’s a broader caution about AI dependency across all model types.
What this means for the AI landscape
Goldman hasn’t issued any regulatory warnings against using open models in its public statements over the past six months. In an industry where compliance concerns routinely kill ideas, the absence of a red flag is itself significant.
If Goldman is right that open models will drive broader AI adoption across businesses, cloud providers benefit from increased compute demand, and companies building tools to make open models enterprise-ready — guardrails, fine-tuning platforms, and monitoring systems — become more valuable. Goldman has not issued any regulatory warnings against using open models in its recent public communications, suggesting the bank views open-model deployment as operationally viable within its compliance framework.