Models can be "rolled", computing power must be "burned"! UBS: AI giants intensively promote new models, computing power investment will continue to increase

UBS Group AG pointed out in its latest Artificial Intelligence (AI) research report that recent launches of new generation large language models (LLMs) by companies such as Alphabet Inc. Class C (GOOGL.US), Anthropic, and DeepSeek are driving continuous advancements at the forefront of intelligence, leading to increasing industry competition. The bank believes that the "scaling laws" of model performance are still effective, and computational power investment will continue to be a key factor determining the competitive landscape of AI. The report highlights a wave of intensive releases in the AI model field in the past few weeks. Alphabet Inc. Class C launched the Gemini 3 Pro multimodal model on November 18, which achieved top rankings in comprehensive performance, text understanding, and visual tasks on the Hugging Face leaderboard, as well as second place in web development tasks. Following closely, Anthropic released the Claude Opus 4.5 model on November 24, excelling in software engineering tasks and significantly improving model security and abuse prevention, ranking at the top in web development tasks. On December 1, DeepSeek introduced the open-source model DeepSeek V3.2, which, through innovations in reinforcement learning algorithms and sparse attention architectures, significantly reduced inference costs while maintaining performance close to cutting-edge closed-source models, with the high computational power version "Speciale" matching Gemini 3 Pro's performance. The new generation of models has shown breakthrough progress in core performance metrics. Benchmark tests such as ARC-AGI-2 showed that Gemini 3 Deep Think and Claude Opus 4.5 achieved scores of 45% and 38% respectively in multi-step reasoning tasks, far surpassing the 10%-20% levels of many previous state-of-the-art models and approaching the human average of 60%. UBS Group AG analysts pointed out that this achievement confirms the effectiveness of the scaling laws of AI model pre-training, in which increasing computational power continuously brings nonlinear improvements to model capabilities. Competition in the chip sector is also under close scrutiny. Alphabet Inc. Class C revealed that Gemini 3 Pro was entirely trained on its in-house TPU chips, sparking discussions in the market about the efficiency of GPU and AI-specific ASIC chips. The report analyzed that ASIC chips are more efficient for specific AI tasks, but GPUs, with their flexible architecture and extensive software ecosystem support, still hold 90% of the data center chip market share. With the advancement of collaborations such as OpenAI with Broadcom Inc. (AVGO.US) and Anthropic with Alphabet Inc. Class C, the focus on ASIC chips is continuously increasing, and it is expected that both types of chips will develop in parallel in the future. NVIDIA Corporation's latest financial report shows that its next-generation GPU products have visibility of $500 billion in revenue, demonstrating a strong growth trend in computational demand. In general, with the introduction of new generation chips such as those from NVIDIA Corporation Blackwell and Rubin, UBS Group AG expects the competition for computational expansion to continue, supporting the recent upward revision of the institution's AI capital expenditure forecast. In addition to the significant progress brought by Alphabet Inc. Class C, the new models from Anthropic and DeepSeek are also adding competitive pressure to the industry, primarily impacting companies like OpenAI, driving the evolution of the AI industry towards a multi-model, multi-vendor landscape. This trend is expected to continue at least until 2026.