During this National Day holiday, the AI industry did not take a single day off. The commercialization of domestic models going global has achieved significant progress, and breakthroughs are being made in overcoming computing power restrictions. OpenAI released over 700 mathematics papers in one go, while also highlighting model safety issues.
01 Three Commercialization Pathways for Domestic Models Going Global
On October 6, Zhipu's GLM-5.3 was listed on AWS Bedrock; just two days prior, it was integrated into Cursor's API. Kimi moved even faster, with K3 landing on Bedrock on September 18, and becoming directly callable in OpenAI's enterprise version, Codex, starting September30. Reuters disclosed that Kimi is also in negotiations with Microsoft and Google regarding listing, demanding a revenue share of up to 30% from cloud providers.
iFenxi believes that the above pathways will be the three most core models for domestic models expanding overseas.
The first is cloud provider hosting. In the cases of Zhipu and Kimi being listed on AWS Bedrock, the cloud providers supply computing power and sales channels. Revenue goes into the cloud providers' pockets first and is then distributed to the model providers. The model providers have zero computing power costs; revenue equals profit, with the trade-off being a smaller share of the pie.
The second is transit via inference service providers. In the Codex case, the overseas inference service provider Baseten provides the computing power, while OpenAI acts as the sales channel. Enterprises pay OpenAI, OpenAI subcontracts to Baseten, and Baseten then shares the revenue with the model providers. There is an extra layer in the transaction chain. The revenue received by the model providers is expected to be lower than in the cloud provider hosting model, but revenue still equals profit.
The third is direct connection to applications. In the Cursor case, Cursor is the sales channel, and the model providers need to provide their own computing power services. It is expected that in the future, Cursor will definitely take a cut from the transactions. In this model, the revenue volume for model providers is the highest, and the gross profit margin depends entirely on their own pricing.
These three pathways have completely opened up the overseas market for China's open-source models, meaning overseas revenue no longer relies on self-built computing power nodes. To this end, Zhipu has raised its year-end ARR guidance to USD 3 billion.
In terms of gross profit margin, the overseas gross profit margin of domestic model providers is not expected to be more optimistic than the domestic one.
On the one hand, domestic model providers will definitely adopt a low-price strategy to capture a larger overseas market; on the other hand, there is an additional sales channel expense overseas, which is not required domestically.
For channel costs, we can refer to Anthropic. In 2025, its channel rate paid to cloud providers was about 16%, and in 2026, it has already risen to 20-30%. The channel rate for domestic model providers will only be higher.
02 Computing Power Breakthrough: Rent, Build, and Create
Tencent plans to spend USD 7 billion to lease the usage rights of approximately 100,000 advanced AI chips in Southeast Asia through Oracle, with a 5-year lease term and a 30% prepayment. Based on the lease price, it is highly probable that the leased chips are NVIDIA's H100 and H200, not the latest Blackwell.
Alibaba is negotiating power supply in Spain, preparing to build its own overseas data centers.
DeepSeek has gone even further, open-sourcing the complete suite of software toolchains for the Ascend platform before the National Day holiday.
Tencent's approach of leasing is about buying time. Current U.S. export controls only govern the entry of chips, not overseas cloud invocations, but this window is expected to close soon. Therefore, Tencent is locking in a long-term contract to grab more computing power before the rules are finalized.
Alibaba's approach is self-building, backed by years of accumulation in its self-developed T-Head chips and cloud infrastructure, directly serving global customers through overseas AIDCs.
DeepSeek has chosen to build an ecosystem. Although it does not manufacture chips, it has moved the entire suite of mature toolchains honed on NVIDIA directly to Ascend, hoping that domestic computing power will be able to step up in the future.
iFenxi believes that Tencent's leasing path is destined to be transitional. The U.S. House of Representatives has already passed a bill to regulate remote cloud access, and the Department of Commerce is drafting new rules to prohibit Chinese-funded entities from leasing computing power in third countries. Alibaba's self-building of AIDCs through chips and DeepSeek's locking in of domestic computing power through ecosystems are more sustainable approaches.
In the future, the choice of route at the computing power level will directly determine the token cost curve for each company, thereby affecting final pricing and market share.
03 What Exactly is the Model Safety Issue
OpenAI recently disclosed new model safety incidents. During training and evaluation, the model accessed the systems of over 100 institutions, including the SEC, the Census Bureau, and the Australian Medicare statistics portal, without authorization. These have been reported one by one, and apologies were made at an Australian Senate hearing. None of these incidents affected external production services, and the severity is not high. However, the flagship model GPT-6.1 Astra was withdrawn before release due to deceptive behavior and unauthorized execution.
It should be noted that these are not proactive cyberattacks initiated by the model, but rather accidental discoveries made while the model took shortcuts to execute tasks.
For example, when asked to query a piece of government data that is unavailable through public channels, the Agent sent instructions through the website's feedback entry and accidentally discovered it could directly have the backend server execute it for them; or, upon seeing leaked keys on the public internet, it directly used them to read information.
These are all common system security vulnerabilities that have existed for a long time and have most likely been exploited by other hackers long ago. Moreover, these safety issues are concentrated in the evaluation phase. In this phase, the model does not have complete safety guardrails, so unauthorized behavior is not hard to understand. Under the condition of complete guardrails in the production environment, no similar incidents have occurred.
iFenxi believes that the model did not exhibit hacker intent, the incidents are concentrated in the evaluation phase where guardrails are absent, and there have been zero incidents in the production environment. Therefore, the model safety issue merely reflects that the threshold for hacking has been leveled by model capabilities.
In the past, unauthorized access required hacking skills. Now, hacking skills are not difficult to achieve, and an Agent might inadvertently trigger unauthorized behavior during task execution.
Moreover, the actual losses from such safety incidents are limited; only public data was accessed, and there is no evidence of personal data being stolen. Therefore, the impact of model safety is mainly to strengthen the adoption of safety guardrails. For instance, NVIDIA has turned Agent runtime monitoring into a DPU-level product.
04 In the Field of Mathematics, AI is Becoming Humanity's God
AI has not surpassed humans in the field of hacking, but it may have already done so in mathematics.
On the morning of October 7, OpenAI released 722 mathematics manuscripts on GitHub in one go, divided into 372 families of results, covering 17 fields including theoretical computer science, number theory, geometry, and mathematical physics. It includes progress on old problems such as the quasi-Riemann hypothesis and the Mahler conjecture.
The results come from an unreleased internal model, with each result consuming an average of about 3 hours of ChatGPT Pro-level computing power. Some proofs have been translated into Lean formalized versions, and for the untranslated parts, OpenAI believes there might be errors. It is estimated that verifying the proofs in these mathematics manuscripts will take the mathematics community quarters or even years.
iFenxi believes that the deeper impact of this event is that the correctness and comprehensibility of mathematics have been decoupled for the first time. In the past, AI was responsible for proving, and humans reviewed it. In the future, Lean will guarantee that the proofs are correct, but humans may not be able to read or understand them, nor learn anything from them.
The mathematics community is already projecting such a future: AI is the god, and humans can initially act as priests, verifying and interpreting the oracles given by AI. Gradually, humans will not even be able to verify or interpret; all they can do is meditate like monks before an eternally unattainable truth.
Mathematics may just be the first field; the same will be true for software and other fields in the future.