Kimi posted a picture of an “AI-native credit card.”

Author: Cai Pengcheng; Source: Barron’s

On July 10, Kimi, under the Mantle of the Moon (Moonshot AI), officially launched the “world’s first AI-native credit card” in partnership with China Agricultural Bank and American Express. The card had been in the works since April this year, opened for pre-orders on June 12, and was rolled out in just three short months. Against the backdrop of a credit card industry that has shrunk in issuance for three straight years, this is a “lightning marriage” between an LLM startup and a state-owned commercial bank plus an international card network.

As defined on the Kimi website, an “AI-native credit card” means embedding Kimi membership, credit limits, and other key benefits into the credit card’s account opening and points-spending system. Through activities such as prompt-course programs and AI sharing sessions, users can gain an AI usage experience while using the card.

Unlike the co-branded card model previously common in the market—“banks lead, with AI benefits as add-on perks”—Kimi’s co-branded credit card directly links Kimi membership tiers to card tiers. The product comes in two tiers: the standard version corresponds to Kimi’s entry tier (Andante) membership benefits. The primary card annual fee is RMB 580, with the annual fee waived in the first year, and next year’s fee waived when spending is done 10 times. The premium Platinum card corresponds to Kimi’s advanced tier (Allegretto) membership benefits. The primary card annual fee is RMB 880, and 200k points can be used to offset the annual fee.

In the design of core benefits, cardholders can exchange points for AI productivity resources such as Agent quota, Kimi Code quota, and access to Kimi new model private testing. The points redemption rule is 1000:1. Daily spending points can be exchanged for Kimi membership benefits at an 80% rate, with up to 50% supported for exchanging benefits using points. Premium Platinum cardholders can also enjoy customized benefits such as “AI priority dispatch during peak periods,” premium private events, and prompt-course programs.

Five “Token cards” launched in concentrated fashion within a month

From the partnership structure, the commercial interests behind the three parties are not the same.

For Kimi, this is an attempt at a “second growth curve” beyond the subscription model. Long-term monetization of large-models-to-consumers relies heavily on monthly subscriptions, with a clearly limited ceiling. By linking everyday spending—points—computing power conversion, the credit card effectively embeds customer acquisition and retention into a high-frequency payment scenario.

Choosing American Express is especially notable. American Express has strong overlap with target demographics such as high-net-worth professionals doing business both domestically and internationally, and cross-border workers. These also match the profile of people willing to pay for productivity AI. And according to disclosures from a B-end leader at Moonshot AI at the end of June this year, Kimi’s overseas paying users and API revenue both grew by 400%, and the product has entered more than 200 countries and regions. For a company that treats globalization as a growth engine, American Express’s international network provides a ready-made low-cost route to reach overseas users.

For Agricultural Bank of China (ABC), this is a way to seize a technology-oriented customer segment in the “battle over existing credit card stock.” By the end of 2025, the number of credit cards issued nationwide fell to 696 million. Compared with the peak at the end of Q3 2022, it had cumulatively shrunk by 111 million, declining for three consecutive years. Traditional perks such as dining discounts and travel memberships have diminishing marginal appeal. Token, described as “electricity fees for the AI era,” offers real utility for high-value groups such as developers, tech professionals, and high-frequency AI users. It can simultaneously support operational goals such as new-customer acquisition, usage activation, and value/assets retention.

For American Express, at a time when Visa is loudly partnering with OpenAI and Mastercard is pushing Agent Pay, co-branding with a leading Chinese large-model vendor is a step to ensure it doesn’t fall behind in the AI payments narrative.

Since June, many banks—including CMB, Ping An, SPDB, and internet-based business banks—have rolled out card products equipped with AI computing-power benefits, creating a wave of industry innovation with considerable momentum. In mid-June, CMB launched an engineer credit card with exclusive AI benefits tied to the MiniMax function. On June 30, Ping An Bank jointly with China UnionPay and Tencent Cloud launched the “AI Smart Computing Card,” integrating a TokenPlan computing-power package with Tencent WorkBuddy. In early July, SPDB Bank, together with China UnionPay and Alibaba Cloud, released the “SPDB Technology Elite Credit Card (Cloud Intelligence Edition),” offering total computing-power subsidies with a maximum as high as 3 billion Qianwen model tokens. In late June, Webank jointly with Alibaba Cloud and Sesame Enterprise Credit upgraded its business card, offering small-and-micro business operators up to 10 million Token trials and AI capabilities of an operational nature.

With at least five “Token cards” appearing in concentrated fashion within the past month, it clearly shows how much banks value this differentiated track.

Token credit cards still have to clear several hurdles

Amid the excitement, scaling these products still faces practical constraints.

First is the data compliance red line. Under the current regulatory framework, financial asset transaction data is core sensitive information, while large-model call logs are classified as personal behavior information. According to Wang Pengbo, Chief Analyst at Bain & Company (Botoon Consulting), when two categories of data are transferred across institutions, separate and explicit user authorization must be obtained. Physical isolation must be established between banks and computing-power platforms, and financial core data must not be directly synchronized to third parties—nor used for model training.

Second is cost and standards. Dynamic fluctuations in computing-power costs may make the cost of redeeming benefits uncontrollable. Large-model inference prices have continued to trend down over the past two years. Token benefits promised today may lose value or even become “inverted” in real worth tomorrow. Token measurement standards are not unified across vendors, creating barriers to cross-use of benefits. Unclear allocation of responsibilities and rights in computing-power services may also lead to consumer disputes.

Third is customer-segment fit. Token benefits naturally skew toward developers and heavy AI users, while ordinary cardholders without AI usage habits have limited perceived value. Zeng Gang, Deputy Director of the National Financial Research Lab, pointed out that in the short term this is a differentiated customer-acquisition tool; in the long term, it may help build a value closed loop of “consumption—computing power—production.” “But at this stage the audience is still relatively niche, and benefits lack general applicability. Whether large-scale replication is possible still needs to be tested by the market.”

For Kimi, the significance of this credit card may be more than just customer acquisition. In a global competitive landscape where OpenAI and Anthropic valuations are both nearing the trillion-dollar mark and have kicked off IPO processes, Chinese large-model firms are rushing to prove the viability of C-end monetization.

In fact, this wave of “AI + finance” has already seen more aggressive exploration overseas.

In June this year, Visa announced a strategic partnership with OpenAI. AI agents can use tokenized payment credentials (not raw card numbers) to directly initiate transactions within “guardrails” set by users, such as preset spending limits and merchant ranges. Mastercard, starting in 2025, has pushed Agent Pay. This year it also teamed up with more than 30 institutions including Stripe to roll out Agent Pay for Machines for machine-to-machine transactions. Compared with the overseas approach to “reconstruct infrastructure around agent payments,” the domestic “put computing power into cards”玩法 is still more cautious—at essence, it is an upgrade to the benefits system.