Open-source AI models benefit from restrictions on frontier systems, Citi says

Citi says open-source AI models are gaining momentum as access restrictions and rising costs push enterprises toward lower-cost alternatives, with open models from Z.ai, DeepSeek, MiniMax, and Moonshot AI narrowing the gap with proprietary systems. The bank also cited strong activity in the open AI ecosystem, including Cohere tripling its 2027 ARR projection, OpenRouter's open-source token share rising to 65% in June from 34% in January, and Fireworks' volume roughly doubling to 30 trillion tokens. Funding and deal activity remain supportive, highlighted by Baseten's $1.5 billion round, Upscale AI's $190 million raise, and Qualcomm's planned acquisition of Modular, though talent and data-center constraints could still favor established AI leaders.

Analysis

The key second-order implication is not that open-source AI wins outright, but that model economics are moving from a scarcity regime to a distribution regime. That compresses monetization for pure-play frontier model vendors while broadening demand for the picks-and-shovels layer: inference hosting, enterprise integration, networking, power, and edge silicon. In that setup, the most durable winners are the companies that can sell lower-latency, lower-power deployment rather than the ones trying to defend model exclusivity.

QCOM is better positioned than the market usually gives it credit for if open models accelerate edge inference. A more modular AI stack increases the value of on-device compute, especially when enterprises want to avoid recurring cloud inference costs and data-exposure issues. That creates a path for higher content per device and a longer replacement cycle for premium AI-capable endpoints, but the upside is gradual rather than immediate because enterprise deployment decisions tend to lag model breakthroughs by 2-4 quarters.

The real constraint here is capacity, not model quality. Talent scarcity and permitting friction for data centers should keep hyperscaler capex elevated, which is supportive for semiconductor demand even if software pricing gets more competitive. But the same constraint also raises the odds of a bifurcated outcome: winners with access to power and skilled labor keep compounding, while smaller AI infrastructure players face margin pressure as customers arbitrage toward cheaper open stacks.

The contrarian read is that the market may be overestimating how fast open-source adoption translates into broad enterprise revenue. In practice, open models often increase deployment complexity, support burden, and customization spend, which can benefit integrators more than model developers. That argues for owning the enablers of adoption rather than the theme itself, while fading names whose premium valuation depends on model lock-in surviving a 12-18 month normalization cycle.

Key Decisions for Investors

• Long QCOM on a 3-6 month horizon: open-model diffusion should incrementally support edge-AI content and handset/PC upgrade demand; risk/reward is favorable if the market keeps underpricing on-device inference versus cloud-only AI.
• Pair trade: long QCOM / short a basket of high-multiple frontier-model beneficiaries with weak distribution moats; thesis is that model exclusivity compresses before hardware demand does.
• Buy semis-on-dips versus software AI names over the next 1-2 quarters: if open-source adoption accelerates, monetization migrates from model access to compute, networking, and deployment layers.
• Consider a call spread on QCOM into the next earnings cycle: upside catalyst is management commentary on AI-capable edge content; downside is limited if the market is already skeptical, making skew attractive.
• Avoid chasing pure-play open-model hype after the initial rerating; use rallies to trim into strength, as enterprise adoption should unfold over quarters, not weeks.