UBS has released a report stating that the rapid adoption of low-cost, open-source artificial intelligence models is fundamentally changing how enterprises deploy AI. However, the report clarifies that this shift will not disrupt the AI investment cycle. Instead, it will make chip manufacturers and cloud infrastructure providers the ultimate winners, while presenting fresh profitability challenges for software companies.
UBS analyst Karl Keirstead's report, published on July 21, notes that investor focus has shifted to two closely related trends: the systematic reduction of token costs by enterprises and the accelerated migration of workloads toward cheaper open-source models. The report is based on in-depth discussions with over 15 private AI companies, including Perplexity, Harvey, Glean, Cursor, ElevenLabs, and Fireworks. The core conclusion is that the AI industry is moving from a broad phase of "using the most powerful model for everything" to a new efficiency-first phase of "fine-tuning deployment based on task difficulty."
From "Token Maximization" to "Value Maximization": A Fundamental Shift in Enterprise AI Spending
The report, released by the team of UBS analyst Karl Keirstead on July 21, points out that investors are now focusing on two closely related trends: enterprises are systematically reducing the cost of AI tokens, and they are increasingly willing to use open-source models, including those developed in China, for tasks that do not require the most advanced reasoning capabilities. This shift is not due to a decline in AI demand. On the contrary, based on feedback from over 75 private companies at UBS's annual Private AI and Software Conference, UBS found that enterprise AI adoption continues to grow at a rapid pace. Several AI-native companies have seen extremely fast revenue growth in the past six months. For example, one company's revenue jumped from $400 million in January to $1 billion, another's revenue tripled this year to exceed $500 million, and a third grew from zero to $500 million in just 14 months. This revenue acceleration is primarily driven by improvements in model capabilities, the expansion of agent-based applications that consume more tokens, and a shift by some companies from per-seat pricing to charging based on tokens or usage.
However, as adoption scales, cost issues have quickly come to the forefront. UBS previously estimated that around 60% of institutions view AI computing and token spending as a real problem. After this survey, the firm believes the percentage may be even higher. The rate of token cost increases is staggering. One financial institution initially committed $10 million for Claude, but that figure grew to $60 million after three months. An AI company's spending on Anthropic rose from $20,000 in December last year to nearly $1 million in July this year, a 50-fold increase in seven months. Some companies have seen their token bills increase tenfold or even a hundredfold in a short period. Enterprises are not stopping their use of AI; they are reducing waste. Many routine tasks were previously assigned to the most expensive frontier models. One bank was spending hundreds of thousands of dollars per month on high-cost models, only for employees to use them for queries about the weather, restaurants, and meeting locations. The focus of enterprise AI management is shifting from "token maximization" to "value maximization," which means measuring the business return generated by each unit of token. UBS quoted one surveyed company as saying, "We initially launched five AI tools internally, but most of the annual token budget was already consumed ahead of schedule. We are now retaining only two, while strictly controlling usage."
"Multi-Model" Becomes the New Normal: Chinese Open-Source Models Gain Traction
Enterprises no longer view the AI market as a zero-sum competition between a few frontier model developers. Instead, they are rapidly adopting a "multi-model" strategy, mixing and matching frontier models from OpenAI and Anthropic with open-source and custom models. Avoiding dependence on a single AI supplier has become a key consideration for large enterprises. Chinese open-source models are among the biggest beneficiaries of this trend. UBS specifically named models from Z.ai, Moonshot AI, Alibaba, and DeepSeek, noting that they are increasingly being trialed and even adopted by enterprises. While some highly regulated industries remain cautious, acceptance of these models is steadily improving within the secure environments provided by cloud platforms like AWS and Microsoft Azure. According to UBS statistics, the training cost of leading Chinese models is about one-tenth that of their overseas counterparts, and their inference API pricing is only 10% to 20% of that of foreign peers, while still maintaining healthy gross margins of 20% to 40%. Data from the OpenRouter platform shows that the token share of Chinese AI models used by American companies on the platform has remained above 30% since February 2026, reaching as high as 46% in some periods. NVIDIA's open-source Nemotron model was one of the most frequently mentioned US open-source models at the UBS conference.
Winners and Losers: A Restructuring of Value in the Industry Chain
Chip Manufacturers: NVIDIA is the Biggest Beneficiary
Despite the increasingly fierce competition in the model layer, UBS emphasizes that demand for AI hardware is not weakening; it may actually be amplified. UBS believes that open-source models will not reduce the demand for GPUs, but rather shift computing power toward more cost-effective inference tasks. Most open-source models still use NVIDIA hardware for training, fine-tuning, and inference. Low-cost models can expand the scope of AI applications, increasing demand for inference computing, storage, networking, and deployment infrastructure. NVIDIA itself has launched the Nemotron 3 Ultra open-source model, which boasts 550 billion parameters, up to 5 times faster inference speed, and up to 30% lower usage costs. This model has become the most frequently mentioned US open-source model in the UBS report.
Cloud Infrastructure Providers: Beneficiaries of the "Pipeline" in the Multi-Model Era
Cloud service providers such as Amazon.com (AMZN), Microsoft (MSFT), and Google are also well-positioned because their platforms already support a variety of AI models. Cloud platforms like AWS and Azure already have multi-model capabilities, meaning that regardless of which model an enterprise chooses, it still needs cloud-based inference computing power. Hyperscalers continue to face supply constraints and strong customer demand for AI computing resources. The revenue data from AI-native companies supports this assessment. In the cases cited by UBS, one company's revenue rose from $400 million in January to $1 billion, and another grew from zero to $500 million in 14 months. The revenue acceleration is primarily driven by improvements in model capabilities and the expansion of agent-based applications that consume more tokens, suggesting that enterprise AI demand has expanded from the programming domain to a broader range of business processes.
Frontier Model Developers: Growth Pressure is Emerging
Frontier model developers like OpenAI and Anthropic will be the most vulnerable to cost-cutting in the short term. While UBS believes the AI market is still in its early stages and both open and closed-source models can grow simultaneously, the revenue growth rate of frontier model labs may face pressure.
Software Industry: The Squeezed "Middle Layer"
UBS holds a more pessimistic view of the prospects for software vendors. The report notes that, in order to control AI costs, enterprises may ultimately shift spending away from software applications. Furthermore, model routing functionality, once considered a competitive advantage, is rapidly becoming a standard industry feature, making it difficult to achieve profitability. In a "multi-model" reality, the business model of purely relying on the middleware layer for model invocation is facing severe challenges.
Comments