The AI race is shifting from bigger models to cheaper, smarter systems

The landscape of artificial intelligence is undergoing a significant transformation. For the past couple of years, the competition in AI was straightforward: larger models delivered superior benchmarks, and companies vied for leadership with each new release. However, this paradigm is evolving. As organizations transition from AI experimentation to real-world applications, the focus is shifting from merely identifying the best model to finding the most suitable one for specific tasks, while also considering cost, available data, and the operational environment. This change heralds a new era of AI competition that prioritizes efficiency over sheer size. Aravind Srinivas, CEO of Perplexity, emphasized this point, stating, "The model alone is no longer the product; it’s about the orchestration system that effectively utilizes the model alongside various tools." This approach means that AI products are becoming more sophisticated systems capable of determining the optimal model for different tasks, depending on the context. For instance, customer service functions may not require the most advanced and expensive models, while intricate coding challenges might benefit from them. Conversely, simpler workflows could efficiently operate on less costly open-source models, with the potential to escalate to more powerful options as needed. As Srinivas articulates, "The answer is always to use whatever is best suited for the task." This shift to alternative models coincides with corporate America tightening its AI budgets, posing new challenges for leading firms such as OpenAI and Anthropic, which have thrived by offering top-tier technology. Recently, Perplexity showcased a new system designed around the GLM 5.2 model from China's Z.ai, which aims to maximize efficiency by relying on a more affordable model for most tasks and only engaging a stronger model when necessary. This trend reflects a broader market evolution. Open-weight models—those that can be downloaded, customized, and operated by companies—are gaining capabilities and proving to be more cost-effective compared to expensive proprietary alternatives from major AI laboratories. Peter Fenton, a general partner at Benchmark, forecasts a dramatic transition, positing that over the next 18 to 24 months, more than 90% of tokens processed will originate from open-weight models, potentially even sooner. Tokens, which are the units of data processed by AI models, are expected to shift the balance of power in the industry. Fenton noted that the profit margins of leading model companies may face pressure as these open models gain traction, allowing businesses to operate without the added costs associated with high-end proprietary models. He also highlighted that smaller, task-specific models often outperform larger, general-purpose models in terms of speed and efficiency. Companies like Ollama, which facilitate the downloading and management of open models, are gaining momentum, with their solutions adopted by over 85% of Fortune 500 companies across various sectors, including aviation and healthcare. Many organizations initially deploy smaller models close to their data before gradually incorporating larger models as they become more comfortable. Moreover, the rise of open models presents strategic implications for the U.S., especially as many of the leading models are emerging from Chinese laboratories. This reality transforms open-source AI into a pivotal business, policy, and national competitiveness issue. Srinivas advocates for U.S. support of open models, arguing that they increase accessibility and affordability, making AI benefits available to small businesses and allied nations. Furthermore, this shift could influence the ongoing expansion of data centers in the tech industry. As demand for AI continues to grow, there’s potential for some AI tasks to be executed locally on consumer or business-owned devices, leading to a hybrid AI system where routine operations are handled locally while more complex tasks are processed in the cloud. For investors, the key question remains whether the top AI laboratories can sustain their pricing power amidst the rise of open models and a more discerning approach to model utilization by companies.