Open-source artificial intelligence models are drawing greater enterprise interest as restrictions on some leading, proprietary systems tighten and certain vendors slow how they release new models, Citi analysts said in a research note.
According to Citi, demand for open-weight and open-source models has climbed sharply as companies confront rising costs and reduced access to top proprietary models. The bank's analysis said the performance differential between closed and open models has come down in recent weeks after Z.ai released its GLM-5.2 model, a development Citi said strengthens the argument for enterprises to adopt open alternatives for targeted use cases.
Citi's internal model rankings put several open models from developers including Z.ai, DeepSeek, MiniMax (HK:0100) and Moonshot AI increasingly close to proprietary systems when evaluated on intelligence and cost metrics. The bank flagged notable expansion among firms building around open architectures: Cohere reportedly raised its internal annual recurring revenue projection for 2027 by threefold, while OpenRouter's share of open-source tokens processed rose to 65% in June from 34% in January, according to Citi's note.
The bank also cited fast growth in open-token processing volumes for other platforms. Citing industry data, Citi noted that Fireworks' open-source token volume roughly doubled between April and June, reaching about 30 trillion tokens in that period.
Investment and corporate activity aimed at cutting AI deployment costs was highlighted as additional evidence of momentum behind open systems. Citi pointed to Baseten's $1.5 billion funding round, Upscale AI's $190 million raise, and Qualcomm's planned acquisition of Modular as signs that investors and technology companies are backing more open and flexible software and hardware architectures.
Despite these developments, Citi cautioned that significant bottlenecks persist. The bank identified shortages of skilled AI researchers and engineers as a structural constraint, and flagged infrastructure limits tied to data-center expansion as another key challenge. Citi said opposition to data-center construction has emerged as a potential drag on capacity growth.
On the policy front, Citi reported that more than 300 data-center bans or moratoriums have been enacted by U.S. localities since 2023, with over 275 such measures introduced since the start of the current year. The bank suggested that these constraints are likely to advantage established AI leaders that possess deeper talent pools and greater access to compute resources, even as open-source models continue to capture market share.
In sum, Citi's research frames a shifting landscape in which enterprises weigh the trade-offs between proprietary and open models amid changing access, cost and infrastructure dynamics. The bank's findings point to accelerating adoption of open architectures in specific applications, tempered by persistent human and physical resource limitations.
Summary
Citi analysts say enterprises are increasingly adopting open-weight and open-source AI models as regulators restrict access to some frontier systems and developers slow releases. The performance gap narrowed after Z.ai's GLM-5.2 launch, and Citi highlighted rising token volumes, stronger revenue projections from open-focused firms, and sizable funding rounds as evidence of growing momentum. However, talent shortages and data-center constraints, including numerous local bans and moratoriums, remain key bottlenecks that could favor established AI leaders with deep resources.