Citi Report: AI Models Improve Rapidly, But Infrastructure Struggles to Keep Up

AI large models are rapidly improving in intelligence; Kimi K3 has ranked third among open-source models globally with a score of 57, while the median score of the top 20 models has risen from 34 to 43. Bottlenecks in AI infrastructure are becoming evident: Blackwell GPU rental prices have increased by 27% year-to-date, pricing for China’s cutting-edge models jumped 45% in a single week, and labs such as SpaceX have invested $1 billion in power generation equipment. Citigroup notes that the next competitive moat will shift from acquiring compute power to efficiency and proprietary data, while security risks are escalating simultaneously—autonomous AI agents have progressed from “interrupting lunch” to “infiltrating Hugging Face infrastructure.”

Author and source: Hard AI

Citibank analysis indicates that ROI in the AI industry is accelerating toward the infrastructure layer, and future competitive moats will shift from access to computing power to efficiency and proprietary data.

AI large models are becoming increasingly intelligent, but the physical world supporting them is being pushed to its limits.

In its latest report released on July 24, Citibank wrote that Moonshot's Kimi K3 ranked third globally with a score of 57, just three points behind Claude Haiku, the leader among proprietary models. Meanwhile, pricing for China's cutting-edge models, represented by Kimi K3, surged 45% in one week; Blackwell GPU rental rates have risen 27% year-to-date, and some labs have spent $1 billion to directly purchase power generation units.

Citi analysis indicates that ROI in the AI industry is accelerating toward the infrastructure layer; the next phase of competitive advantage in model development will shift entirely from mere "access to computing power" to "efficient output and proprietary data."

Citigroup's research report noted that Kimi K3 is currently the largest publicly released open-source model. Just a few weeks ago, the highest open-source score was only 51 (Z.ai GLM-5.2); Kimi K3 has now jumped to 57, trailing only Claude Fable 5 (60) and OpenAI GPT-5.6 Sol (59).

The entire industry is accelerating. The median intelligence score of the top 20 model providers has risen from 34 six weeks ago to 43. In the open-source camp, DeepSeek V4 Pro scores 44 with a price as low as $0.03 per million tokens—delivering near-cutting-edge intelligence at two orders of magnitude lower cost.

The closed-source camp has also not slowed down. Google just released Gemini 3.6 Flash (on July 21) and revealed that Gemini 3.5 Pro is still in testing, while pre-training for Gemini 4 has already begun—three generations of models advancing in parallel. However, Citibank believes that the release rhythm of "Flash first, Pro delayed" signals that advancing frontier models is becoming more difficult—consistent with delays other companies have faced in infrastructure development and reflecting the industry’s struggle to compress delivery cycles despite its aspirations.

The larger and more powerful the model, the more rapidly its resource requirements grow.

Trillion-parameter models are redefining the meaning of "computing power."

Citibank noted that, in practice, these ultra-large models are spending an increasing amount of time moving weights and KV-cache data across HBM memory and GPU interconnects, rather than performing matrix computations themselves. The bottleneck has shifted from "how fast can we compute" (FLOPs) to "can we move the data at all"—memory bandwidth, GPU interconnects, and power supply.

Demand for GPUs remains strong, and rental prices for Blackwell architecture have risen 27% year-to-date. However, simply stacking GPUs is no longer enough.

Some labs are now directly entering upstream power generation: SpaceX invested $1 billion in 1 GW of mobile turbine units (July 15), and Georgia Power signed a service agreement the same week (July 22). AI labs purchasing power generation equipment—something almost unimaginable a year ago—is now happening.

Model pricing directly reflects how tight capacity is. The blended price for frontier models in China surged to $0.87 per million tokens, rising 45% week-over-week and month-over-month—the first significant fluctuation in two months. The global average price for frontier models increased 6.8% week-over-week and 11.3% month-over-month. Prices in the U.S. and Europe remained relatively stable (week-over-week: -0.5%), but still rose 4.1% month-over-month.

Citibank believes that as incremental infrastructure comes online, pricing pressure will eventually ease. At that point, valuable data and task-specific performance will form a more durable moat than access to computing power.

The models are getting stronger. But the agents running on these models are also becoming more dangerous.

Citibank's report used a striking phrase: the real-world risk of autonomous AI agents escaping sandboxes has escalated from "interrupting lunch" in April to "compromising Hugging Face's infrastructure" in July.

The stronger and more widespread open-source models become, the broader the surface area for security risks. The debate over AI regulation is ongoing—recent statements were made by NVIDIA (July 24) and U.S. Treasury Secretary Bessent (July 22)—but no clear resolution is expected in the short term. Even as regulatory frameworks remain unimplemented, companies maintaining their own model architectures are already facing higher compliance barriers.

More challenging is that even the providers training these models still cannot fully examine why the models behave as they do. The issue of interpretability remains unresolved.

However, security concerns have not slowed the commercialization of agents. Data from METR (July 21) shows that the economic gap between AI agents and human labor is narrowing. As agents become more autonomous and closer to economic viability, token consumption will continue to accelerate—further driving up infrastructure demand.

The stronger the capability, the tighter the constraints. The tighter the constraints, the greater the infrastructure requirements. This cycle shows no signs of slowing down.