Citigroup AI tracking: the model is getting stronger and stronger, but chips and power are about to fall behind

Author: Gao Zhimou, Wall Street Insights

AI large models are becoming smarter and smarter, but the physical world that powers them is being pushed to the limit.

In its latest report released on July 24, Citigroup wrote that Moonshot Dark Side’s Kimi K3 scored 57 points to rank third globally, just three points behind the top spot on the closed-source leaderboard, Claude Fable 5. At the same time, the pricing of China’s cutting-edge models represented by Kimi K3 jumped 45% week over week; Blackwell GPU rental prices rose 27% year-to-date; and even one laboratory spent $1 billion to directly buy power generator sets.

Citigroup’s analysis says that investment returns (ROI) in the AI industry are accelerating as they shift toward the infrastructure layer; and the next phase of model-competition moats will move completely from simply “getting compute” to “efficient output and proprietary data.”

The gap is only 3 points

Citigroup’s research report mentions that Kimi K3 is currently the largest publicly released open-source model. A few weeks ago, the highest open-source score was only 51 (Z.ai GLM-5.2). Kimi K3 then leaped to 57, trailing only Claude Fable 5 (60) and OpenAI GPT-5.6 Sol (59).

The whole industry is accelerating. The median intelligence score among the top 20 model providers rose from 34 points six weeks ago to 43 points now. In the open-source camp, DeepSeek V4 Pro (44) is priced at 0.03 per million tokens for Xiaomi (discounted), which is close to frontier-level intelligence—yet the price is lower by two orders of magnitude.

The closed-source camp is also not slowing down. Google just released Gemini 3.6 Flash (July 21), and at the same time disclosed that Gemini 3.5 Pro is still in testing, while Gemini 4’s pretraining has already started—three generations advancing in parallel. However, Citigroup believes the “Flash first, Pro later” release rhythm sends a signal: pushing forward frontier models is becoming harder—consistent with the delays other companies have encountered in building infrastructure, and reflecting the industry’s current reality that it hopes to compress delivery timelines but cannot do so as hoped.

The bigger and stronger the model, the more resources required to run it are rapidly inflating.

The bottleneck has moved

Trillion-parameter models are rewriting what “compute” means.

Citigroup points out that when these ultra-large models actually run, more and more time is spent moving weights and KV-cache data across HBM memory and GPU interconnect networks, rather than on matrix computation itself. The bottleneck has shifted from “whether it can compute fast” (FLOPs) to “whether it can move it”—memory bandwidth, GPU interconnect, and power supply.

GPU demand remains strong, and Blackwell architecture rental prices have risen 27% since the beginning of the year. But simply piling up GPUs is no longer enough.

Some labs have started to move directly into upstream power generation. SpaceX invested $1 billion to purchase 1GW mobile turbine generator sets (July 15), and Georgia Power signed a services agreement the same week (July 22). AI labs buying power-generation equipment—something that was almost unthinkable a year ago—is happening now.

Model pricing directly reflects how tight capacity is. For China’s cutting-edge models, mixed pricing jumped from 0.03 per million tokens to 0.87, with both week-over-week and month-over-month increases of 45%, the first major fluctuation in two months. The global average price of frontier models rose 6.8% week over week and 11.3% month over month. The U.S. and Europe are relatively stable (down 0.5% week over week), but also rose 4.1% month over month.

Citigroup judges that as incremental infrastructure comes online one after another, pricing pressure will ultimately ease. At that time, valuable data and task-specific performance will form a more durable moat than simply acquiring compute.

Agent escapes the sandbox

Models are getting stronger. But the agents running on top of them are also becoming more dangerous.

Citigroup’s report used a thought-provoking phrasing: the real risk of autonomous AI agents escaping the sandbox has escalated from April’s “interrupting lunch” to July’s “penetrating Hugging Face infrastructure.”

As open-source models become stronger and more widely adopted, the scope for the spread of security risks grows larger. The debate on AI regulation is still ongoing—Nvidia (July 24) and U.S. Treasury Secretary Bessent (July 22) have both made statements recently—but no conclusion is likely in the short term. Even if regulatory frameworks have not yet been rolled out, companies that maintain their own model-running architectures are already facing higher compliance requirements.

Even more troublesome is that the providers that train these models themselves still cannot fully inspect why the models act the way they do. The explainability problem has not been solved to date.

But safety concerns have not slowed down the commercialization of agents. METR’s data (July 21) shows that the economic gap between AI agents and humans is narrowing. As agents become more autonomous and closer to economic feasibility, token consumption will continue to accelerate—once again driving up infrastructure demand.

The stronger the capability, the tighter the constraints. The tighter the constraints, the greater the infrastructure demand. There are no signs that this cycle is slowing down.