UBS: The low-cost model will not disrupt the AI investment cycle, and chip and cloud giants will be the ultimate winners

The Zhitong Finance App learned that the rapid spread of low-cost open source artificial intelligence models is fundamentally changing the way companies deploy AI, but UBS clearly stated in a recent report that this transformation will not only disrupt the AI investment cycle, but will also make chip makers and cloud infrastructure providers the ultimate winners, while bringing new profit challenges to software companies.

UBS analyst Karl Keirstead pointed out in a report released on July 21 that investors' attention has turned to two closely related trends: enterprises are systematically reducing token costs, and workloads are moving faster to cheaper open models. Based on in-depth discussions with more than 15 private AI companies — including Perplexity, Harvey, Glean, Cursor, ElevenLabs, and Fireworks — the report arrived at a core judgment: the AI industry is moving from an extensive stage of “using the strongest models without discrimination” to a new stage of prioritizing efficiency “detailed configuration according to task difficulty.”

From “Token Maximization” to “Value Maximization”: A Fundamental Shift in Enterprise AI Spending Logic

According to a report released by UBS analyst Karl Keirstead's team on July 21, investors' attention has turned to two closely related trends: enterprises are systematically reducing AI token costs, and increasing willingness to use open source models (including models developed in China) to complete tasks that do not require the most advanced reasoning capabilities.

This shift is not due to a decline in AI demand. Quite the opposite — based on feedback from more than 75 private companies at its annual private AI and software conference, UBS found that enterprise AI adoption is still growing rapidly. The revenue of many AI-native companies has grown extremely fast in the past six months: some companies' revenue rose from 400 million US dollars in January this year to 1 billion US dollars, some companies tripled their revenue during the year and exceeded 500 million US dollars, and others grew from zero to 500 million US dollars within 14 months. The revenue acceleration is mainly due to improvements in model capabilities, the expansion of smart applications that consume more tokens, and some companies are shifting from charging per seat to charging per token usage.

However, as adoption scaled up, cost issues quickly surfaced. UBS previously estimated that about 60% of institutions have viewed AI computation and token spending as real issues; after this survey, it is believed that this ratio may be even higher. Token costs are rising at an astonishing rate: a financial institution initially promised to spend $10 million on Claude, but three months later, it has increased to 60 million US dollars; an AI company's Anthropic spending increased 50 times in 7 months from 20,000 US dollars in December last year to nearly 1 million US dollars in 7 months; some companies' token bills have increased 10 times or even 100 times in a short period of time.

Businesses aren't stopping using AI; they're reducing waste. Many mundane tasks were previously handed over to the most expensive cutting edge model — a bank spent hundreds of thousands of dollars a month calling expensive models, while employees only used it to check the weather, restaurants, and meeting locations. The focus of enterprise AI management is shifting from “token maximization” to “value maximization,” that is, measuring how much business return each unit of token can generate. UBS quoted an interviewee company as saying, “We originally launched five in-house AI tools, and most of our annual token budget has already been spent ahead of schedule. Currently, only two models have been retained, and usage is strictly controlled.”

“Multiple models” become the new normal: China's open source model accelerates penetration

Companies are no longer viewing the AI market as zero-sum competition among a few cutting-edge model developers, but are rapidly shifting to a “multi-model” strategy — mixing cutting-edge models from OpenAI and Anthropic with open source and customized models. Avoiding dependency on a single AI vendor has become an important consideration for large enterprises.

China's open source model is becoming the biggest beneficiary of this trend. UBS pointed out that Z.ai, Dark Side of the Moon, Alibaba, and DeepSeek models are being tested or even imported by more and more companies. Although some highly regulated industries remain cautious, the acceptance of these models by enterprises is steadily increasing in the secure environment provided by cloud platforms such as AWS and Microsoft Azure. According to UBS statistics, the training cost for the Chinese head model is about one-tenth of that of overseas leaders. The inference API price is only 10% to 20% of the overseas benchmark model, and it can still maintain a healthy gross profit level of 20% to 40%. According to data from the OpenRouter platform, the share of tokens used by US companies using the Chinese AI model through this platform has remained above 30% since February 2026, and even reached 46% during some periods.

Nvidia's open source model, Nemotron, is one of the most frequently mentioned US open models at UBS meetings.

Winners and Losers: Restructuring Industrial Chain Value Allocation

Chipmakers: Nvidia is the biggest beneficiary

Despite increasingly intense competition in the model layer, UBS emphasized that the demand for AI hardware has not weakened; on the contrary, it may be further amplified. UBS believes that the open model will not reduce GPU demand, but will shift computing power to more cost-effective inference tasks. Most open models still use Nvidia hardware for training, fine-tuning, and inference. Low-cost models can expand the scope of AI applications and increase infrastructure requirements for inference computing, storage, networking, and deployment.

Nvidia itself also launched the Nemotron 3 Ultra open source model, which has 550 billion parameters, increased inference speed by up to 5 times, and reduced usage costs by up to 30%. This model has become the most mentioned US open source model in UBS reports.

Cloud Infrastructure Providers: “Pipeline” Beneficiaries in the Multi-Model Era

Cloud service providers such as Amazon, Microsoft, and Google are also in a good position because their platforms already support a variety of AI models. Cloud service providers such as Amazon, Microsoft, and Google are also in a good position because their platforms already support a variety of AI models. Cloud platforms such as AWS and Azure already have multi-model capabilities, and no matter what model they choose, enterprises still need cloud inference and computing power. Hyperscale cloud service providers are still facing tight supply and strong customer demand for AI computing resources.

Revenue data from AI-native companies confirms this judgment. In the case cited by UBS, the revenue of some companies rose from 400 million US dollars in January of this year to 1 billion US dollars, while others grew from zero to 500 million US dollars within 14 months. The revenue acceleration is mainly due to improved model capabilities and the expansion of intelligent applications with higher term consumption, indicating that enterprise AI requirements have expanded from the field of programming to a wider range of business processes.

Frontier model developers: Growth pressure is evident

Frontier model developers such as OpenAI and Anthropic are the most vulnerable to cost cuts in the short term. Although UBS believes that the AI market is still in its early stages, and that open and closed source models can grow at the same time, the revenue growth rate of cutting-edge model laboratories may be under pressure.

Software industry: the squeezed “middle layer”

UBS is pessimistic about the future of software vendors. UBS pointed out that in order to control AI costs, enterprises may eventually shift expenses away from the software application field; however, the model routing function was once seen as a competitive advantage, but now it has quickly become an industry standard, making it difficult to achieve profits. In a context where “multiple models” have become the norm, business models that rely solely on models to call the middle tier are facing serious challenges.