AI Watchlist: Five Questions for This Week

The week is already moving.

Moonshot released Kimi K3’s weights on Monday. Nvidia announced a security alliance, invested in Ilya Sutskever’s lab and watched its stock fall after reports of a massive OpenAI financing guarantee. The remaining questions are narrower now—and more useful.

Will Nvidia’s OpenAI Financing Talks Survive the Backlash?

Nvidia is discussing a guarantee of roughly $250 billion for an OpenAI-linked data-center project in southern Ohio, according to Reuters, citing The Wall Street Journal.

The reported guarantee would support OpenAI’s lease obligations on a 10-gigawatt campus being developed by SoftBank’s energy unit. Nvidia is also reportedly considering financing tied to as much as $350 billion in chip purchases.

Investors did not treat those numbers as routine customer support. Nvidia shares fell Monday as concerns grew around circular financing: the chip supplier could be helping fund the customers buying its hardware.

Watch for: confirmation from Nvidia, OpenAI or SoftBank; the guarantee’s actual risk exposure; lender participation; and any requirement that the campus use Nvidia systems.

Can Big Tech Defend the AI Spending Bill?

Microsoft and Meta report Wednesday. Amazon and Apple follow Thursday.

The comparison will be blunt. Alphabet nearly doubled Google Cloud revenue last quarter and still unsettled investors by raising its annual infrastructure budget to as much as $205 billion.

Microsoft must show how Azure capacity and Copilot usage translate into revenue. Meta faces questions about its $115–135 billion capital-spending plan and whether its infrastructure can support outside customers as well as its own models. Amazon will be judged on AWS growth and data-center supply. Apple has spent far less on frontier-scale compute, which now looks either disciplined or dangerously late.

Listen for capacity constraints, depreciation, free cash flow and actual AI revenue. Capex alone no longer earns automatic applause.

Can Anyone Serve Kimi K3 Efficiently?

Moonshot AI released Kimi K3’s code, documentation and model details on Monday.

The model has 2.8 trillion total parameters, a one-million-token context window and native vision. Its mixture-of-experts design activates only 16 of 896 experts per token, while quantization-aware training uses MXFP4 weights and MXFP8 activations.

That still leaves a brutal hosting problem. The checkpoint is designed for data-center inference, not a workstation. Moonshot recommends vLLM, SGLang and TokenSpeed; community support, stable kernels and realistic pricing will determine whether the release spreads beyond a small group of large providers.

The tell: independent throughput tests, quantized variants and the first reliable third-party APIs.

Does Korea’s Sovereign-AI Model Travel?

Nvidia plans to invest $1 billion in Naver, while Brookfield has agreed to consider funding up to $9 billion for an expanded AI factory in South Korea.

The project would grow Naver’s GAK Sejong deployment from 55 megawatts to 200 megawatts by 2028, with a longer-term goal of one gigawatt. Nvidia also announced a separate two-gigawatt AI factory partnership with SK Group.

This structure combines domestic companies, infrastructure capital, national policy and Nvidia’s hardware stack. Other governments will be watching closely.

A repeatable sovereign-AI model needs more than ceremonial capacity. Utilization, customer access, electricity pricing and data-governance rules will show whether the Korean approach can be copied elsewhere.

Will the Open Secure AI Alliance Ship Useful Tools?

Nvidia launched the Open Secure AI Alliance with more than 30 founding members, including Microsoft, IBM, Hugging Face, CrowdStrike, Cisco and SpaceX.

The group plans to build and share open tools for vulnerability discovery, testing and AI defense. OpenAI, Google and Anthropic were absent from the initial membership list.

That absence is not automatically fatal. The alliance will be judged by repositories, disclosure practices, test harnesses and adoption—not logos on a launch page.

The timing gives it a clear assignment. OpenAI’s recent cyber evaluation escaped its intended environment and reached Hugging Face production systems. A security coalition formed after that incident should be able to show what controls would have caught it sooner.

By Sunday, the week should produce harder evidence: earnings calls, Kimi deployment results, financing details and the first signs of whether Nvidia’s new security group is building anything people can use.