Nvidia in Talks to Insure AI Chip Loans as DeepSeek and Honor Announce AI Updates

On September 29 in China, multiple developments surrounding AI infrastructure and generative AI emerged simultaneously. They include talks on using insurance to share risks for semiconductor financing, an update to a desktop AI tool, certification for a smartphone marketed as an AI agent, and financial and risk disclosures by a major AI company. Each represents a different layer supporting the social implementation of AI — funding, tools, devices, and governance.

According to reporting by TMTPost’s official account, these developments centered on September 29 were reported together. Notably, the cluster of AI-related announcements came alongside announcements in other sectors, such as adjustments to financial policy and pharmaceutical licensing deals.

Nvidia in Talks with Insurers Over Loan Risks

It has emerged that Nvidia is in talks with insurance companies to explore risk-sharing arrangements for its AI chip financing business, according to sources familiar with the discussions.

The focus of the talks is a framework to transfer part of the risk from capital-intensive semiconductor loans to insurers and other investors. One proposal under discussion is loan insurance for so-called Neoclouds — emerging cloud providers. Under the proposed structure, if a borrower defaults and creditors cannot recoup the loan by reselling the Nvidia chips held as collateral at a sufficiently high price, the insurance mechanism would cover part of the lenders’ losses.

Nvidia has presented insurers with multiple cooperation proposals, aiming to transfer part of the risk from capital-intensive semiconductor financing to insurers and other investors

If such guarantees materialize, they could encourage capital flows to customer segments that lack the financial footing of major technology companies. As demand for AI computing resources grows, loans collateralized by chip procurement are increasing. If insurance can mitigate lenders’ recovery risk, credit hurdles would be lowered and the cycle of capital investment would be easier to sustain.

At the same time, the negotiations are said to be at an early stage and may not lead to any concrete deals. Many elements remain to be designed, including setting premium rates, valuing collateral chips on the secondary market, and the assessment process in the event of default. Semiconductors are assets with rapid generational turnover and highly volatile collateral value. How much risk insurers will assume will depend on modeling price fluctuations and arranging reinsurance.

The significance of these talks lies in the attempt to incorporate insurance as a third channel for AI infrastructure financing, beyond traditional venture investment and corporate bonds. Capital expenditures by cloud providers have reached hundreds of billions of yen, straining the capacity of any single company’s balance sheet. Further diversification and redistribution of risk could affect the pace at which AI computing infrastructure expands.

New Features in DeepSeek Harness v0.2

On September 29, DeepSeek officially released the preview version of DeepSeek Harness v0.2. Desktop installation packages for macOS and Windows are available and distributed ready for immediate use.

The new version comes pre-equipped with frequently used features for everyday office work and development, allowing key functions to be used immediately. A new page for installing and managing plugins has been added, enhancing configurational flexibility. The design is described as allowing users to extend functionality as needed and build smart products tailored to each user.

Harness is positioned not as a model itself, but as a harness — a framework — that makes models easier to use in business and development contexts. The release of a desktop version reflects a trend that emphasizes on-device operation and extensibility, not just API use in the cloud. The separation of plugin management is seen as an effort to decouple the distribution and updating of additional features, allowing users to select only the functions they need.

For developers, the presets of frequently used features reduce friction in initial adoption. For office work, use in routine tasks such as document creation and information organization is envisioned. Resident desktop AI tools enable operations that are difficult to achieve with cloud-only solutions, such as integration with browsers and editors and access to local files. This update is positioned as laying the groundwork for such capabilities.

Honor Magic9 Secures AI Agent Certification

On September 29, Honor announced that the entire Honor Magic9 series has obtained network access certification as part of the first batch of AI agent smartphones. The three models are the Honor Magic9 Pro Max, Honor Magic9 Ultra Edition, and Honor Magic9. The company says it is among the first manufacturers in the industry to obtain the certification.

The term AI agent smartphone refers to a capability beyond conventional voice assistants or single-function generative AI apps, where the device autonomously performs tasks. Network access certification is believed to indicate not only compliance with communications standards but also fulfillment of requirements for agent functions that operate with connectivity. While detailed certification criteria have not been disclosed, the phrase first batch suggests that similar certifications will expand to other models going forward.

Agent functions on smartphones are characterized by the sequential execution of multiple steps, such as scheduling, cross-app operations, and context-aware information gathering and summarization. They combine on-device inference with cloud-based model calls, requiring stable connectivity and fast response times. Obtaining network access certification effectively guarantees such a hybrid execution environment.

Honor’s move indicates that hardware makers have entered a phase of foregrounding AI agents as a differentiator. By integrating the chip, OS, model, and connectivity layers and visualizing the maturity as an agent through certification, the approach could also appeal to enterprise demand and developers.

Anthropic Warns of Catastrophic Risks in

IPO Filing

It was reported that in documents filed as part of its IPO preparations, Anthropic warned that its technology could pose “catastrophic or existential risks” to humanity. A substantial portion of the documents is said to be devoted to potential harms from AI, citing examples such as AI seeking to preserve its own operation, resisting shutdown, and concealing or manipulating information.

Financial information was also disclosed. Revenue last year was about $4.6 billion, an elevenfold increase year over year. However, nearly a quarter of that came from just two customers. Adjusted operating losses exceeded $8 billion, and the company plans to invest $518 billion in cloud, computing resources, and infrastructure over the next several years.

The coexistence of surging revenue and widening losses illustrates how capital-intensive the development and operation of large-scale models is. Spending to secure computing resources and infrastructure expands ahead of returns, with monetization dependent on growth in enterprise business. For OpenAI as well, it was reported on September 29 that annual recurring revenue is approaching $70 billion and that sales revenue from enterprise business has more than doubled since July. Across the industry, enterprise demand is increasingly driving revenue.

Explicitly addressing risks in the prospectus serves to fulfill accountability to regulators and institutional investors. Concerns about AI autonomy and controllability are directly linked to safety evaluations and the establishment of governance structures. For investors, it provides material to factor in not only growth potential but also the long-term costs of risk management.

Structural Shift in AI Infrastructure Investment

The flurry of announcements on September 29 shows that frameworks for funding and implementation are being updated on both the supply and demand sides of AI. Nvidia’s insurance talks concern financing on the supply side, DeepSeek Harness concerns the user environment on the demand side, Honor’s certification concerns devices at the edge, and Anthropic’s disclosure concerns discipline in capital markets — each a different touchpoint.

What they have in common is a shift in AI from competition based solely on model performance to comprehensive competition that includes surrounding systems and infrastructure. Financing that supports chip procurement, tools integrated into everyday work, devices that act as agents, and disclosures that articulate risks are all domains that cannot be addressed by model performance alone.

Adjustments to multiple monetary policy tools announced the same day by the People’s Bank of China also indirectly overlap with this context. They include lowering the pledged supplementary lending rate by 0.25 percentage points from 1.75% to 1.5%, adding six network construction projects to the scope of support, increasing the relending quota for scientific and technological innovation and technological transformation by 200 billion yuan from 1.2 trillion yuan to 1.4 trillion yuan and raising the support ratio from 60% to 100%, among other measures. The quotas for relending and rediscounting to support agriculture and small enterprises were also increased by 500 billion yuan from 4.35 trillion yuan to 4.85 trillion yuan. These are positioned as improvements to the environment underpinning investment, including in AI and computing networks.

Reports that AMD will acquire World Lab for $8.2 billion also point to a reorganization of the computing base. Realignments are proceeding simultaneously across the chip, cloud, device, finance, and capital market layers, indicating that AI implementation is entering a phase defined not only by individual product launches but also by capital flows and institutional design.

Editorial Opinion

In the short term, we see the success or failure of the insurance scheme Nvidia is exploring as determining the pace of credit expansion for Neoclouds. If agreement can be reached on premium rates and collateral valuation, procurement by mid-tier cloud providers could proceed smoothly, potentially boosting GPU utilization and data center demand. Conversely, if conditions are stringent, capital could remain concentrated among large players, raising concerns about oligopolization of computing resources.

In the long term, we assess that the key question is whether on-device AI agent certification will become established as an industry standard. If first-batch moves like Honor’s spread to other companies, a yardstick for comparing agent performance and safety will emerge, and developers will design apps with certification in mind. Vertical integration of devices, operating systems, and models will advance, and user selection criteria may shift from processing performance to autonomy.

The remaining question is how capital markets will price in the catastrophic risks warned of by Anthropic. Explicit disclosure in the prospectus enhances transparency, but without concrete control measures and auditing frameworks, both investors and users will lack material for judgment. Risk disclosure and risk transfer through insurance are two sides of the same coin. The question is whether institutional design can connect the two.

References

• “【钛晨报】重大利好！央行“四箭齐发”，房贷贴息政策落地；中行回应万事达信用卡遭盗刷；英伟达欲拉保险公司入场，为AI芯片融资风险兜底”, by 钛媒体官方账号 — 钛媒体, 2026-09-29T23:20:00.000Z (ARR)
• Source URL: https://www.tmtpost.com/8156446.html