Citigroup: AI inference demand remains tight, bottlenecks are spreading from chips to power and data centers BlockBeats News: On June 16th, Citigroup announced that the intensity of artificial intelligence inference demand is still continuing, and the scarcity of computing power is overflowing from the latest generation chips to the previous generation GPUs, driving model manufacturers to accelerate monetization through price, quota, and routing mechanisms. In a report released on June 14th, Citigroup analysts such as Heath Terry wrote that the rental price of A100 GPU has risen by 0.6% in the past week and 11% in six weeks, indicating that the demand for AI computing power is not only concentrated on the most advanced hardware. At the same time, some cutting-edge models have significantly increased their prices after improving their intelligence scores, with Citigroup stating that "scarcity is being monetized faster than it is being resolved. The report points out that currently no model supplier can simultaneously occupy the advantages of intelligence, speed, and price. The latest cutting-edge model's intelligence score has increased by about 4 points, but the overall price is close to doubling; The mid-range model has made progress in speed, with the median output speed of the Top 20 models increasing from 64 tokens/s to 105 tokens/s within six weeks. Citigroup also stated that the capability gap between closed source models and open source models is widening, with proprietary models leading the way in intelligence relative to open source models increasing from about 6 points to about 10 points. This means that top model manufacturers still tend to hold onto the high-end market with stronger capabilities, rather than directly competing with open source models in terms of price. Beyond computing power, the location of power and data centers is becoming a new constraint for AI expansion. The report mentioned that a private Neocloud has signed up for a 4.9GW demand, but plans to have a pipeline of over 40GW, highlighting the gap between demand explosion and supply implementation. Citigroup stated that data centers tend to be located in areas with electricity prices of about 9-12 cents/kWh, while the proportion of renewable energy and long-term power purchase agreements are also affecting site selection. The report suggests that the cost of AI infrastructure will continue to rise in the future. With the increase in component prices, power access, and early-stage infrastructure investment, capital expenditures calculated based on H100 equivalent computing power are rising, and power costs are shifting from the operational phase to pre construction capital investment. Citigroup stated that the next phase of value may flow towards the 'inference routing layer', which is a platform capable of determining which model, quantization method, and hardware should be used for different tasks. This layer can reduce inference costs and improve output efficiency, but enterprise data, intellectual property, and privacy protection will be the implementation difficulties. From the perspective of the industry chain, the report not only points to GPUs, but also includes data centers, power, optical communications, cloud infrastructure, and model applications. Citigroup listed relevant coverage targets such as Ciena, Lumentum, and MiniMax in the appendix, indicating that the AI inference cycle is spreading from chips to a wider range of infrastructure and application layers.