人工智能：花旗前瞻——前沿的阻碍

研报英文原文证据摘录

Vi ewp oint |

14 Aug 2026 16:55:49 ET │ 25 pages

Artificial Intelligence

Citi’s Inference Ahead - Holding Back the Frontier

CITI'S TAKE

With leading frontier labs holding back out of security concerns and related

government reprisals, publicly available progress at the frontier has stalled

(Figure 1), giving open models the opportunity to catch up (Z.ai, 8/14), at

least as measured by public benchmarks. We believe significant progress is

being made in unreleased models, with behind the scenes work to be done

to find a path to release for more advanced proprietary models. Volatility in

financing costs and questions around the ability of markets to scale to the

degree needed is being addressed as Nvidia and partners target over $500B

of capital (8/10), potentially creating a vehicle to secure market share, lower

customer borrowing costs, and accelerate data center construction.

Models. Grok 4.6 gained 5 intelligence points over Grok 4.5, offering a model within

2 points of the leader at a 64% lower blended price (Figure 4), with expectations

higher for 4.6’s successors (Figure 5). Grok 4.7 should be incrementally better when

released in the next 3-4 weeks (X, 8/12), with Grok 5 expected to benefit from

Cursor’s integration into training. Gemini 3.7 Flash’s 4-point intelligence gain and

29% lower blended price versus 3.6 Flash show model design,

introductory/discounted pricing, and capacity coming online lowering cost per task

(Figure 11). Conversely, DeepSeek’s 50%-1,100% API price increases and new

peak/off-peak rates as of August 17 indicate providers could leverage pricing to

ration demand by time of use (D8/13). Consistent releases and gains should support

frontier staying power, while delays give competitors room to pull ahead (Figure 6).

n

Heath Terry AC

Shelby Spencer

Ashley Kim

Janna Withrow

Infrastructure. CoreWeave contracting A100s through 2029 (9 years after the chip’s

debut) suggests prior-generation GPUs can migrate toward inference and retain

value longer than expected, with A100 rental rates rising 0.6% WoW to $1.65 and

Nvidia finding new ways to finance the frontier (Figure 15). Given architecture

…