Cloudflare Open-Sources Decision Model Clef; Amazon Moves $8B in Nvidia Chips Off-Balance-Sheet

Over the past 24 hours, the AI industry has been focused on two parallel storylines: open-source models and compute capital. On one side, Cloudflare open-sourced a decision model that topped benchmark leaderboards. On the other, Wall Street and tech giants are engaged in multi-billion-dollar financing rounds and asset restructuring centered on Nvidia chips. Open-source models and chip-backed debt — both happening on the same day — are a sign that the compute ledger keeps growing longer.

Open-Source Models: Cloudflare Launches the Clef Decision Model Family

Cloudflare has open-sourced two decision models — Clef and Clef Flash — under the Apache 2.0 license, both deployable locally via Hugging Face. This means developers can embed decision-making capabilities into their own workflows without depending on closed-source APIs.

According to published benchmark data, Clef tops the Decision Index and leads all 43 benchmark tests on latency. The model includes a built-in vision encoder and supports a 64K context window — twice that of comparable models. Median latency clocks in at 209ms, while the lightweight Flash variant compresses that down to just 38.8ms. On the training side, Clef uses Coin as a frozen base, and Cloudflare has also released an accompanying RL fine-tuning platform so users can perform reinforcement learning tuning for their own use cases.

Open licensing, local deployment, low latency, and an integrated fine-tuning platform — the intent behind this combination is clear: lower the barrier to deploying decision-oriented AI and extend Cloudflare's reach into a broader developer ecosystem.

Decision Models are AI systems specifically designed to make optimal choices among multiple options. Unlike general-purpose conversational models, their core function is to output an action policy given a particular state — commonly used in scenarios like routing, tool-call selection, and automated workflow orchestration. Cloudflare positions this model type at the "workflow decision" layer, with the goal of letting developers plug Clef into multi-step AI Agent pipelines to replace human judgment on questions like "which tool should be called next, or which branch should be taken."

The Apache 2.0 license is one of the most commercially permissive open-source licenses available. It allows users to commercialize, modify, and redistribute software without being required to open-source their own code — a sharp contrast to GPL-family licenses, which require modifications to be contributed back to the community. This license choice means companies can integrate Clef directly into closed-source products without exposing internal code, significantly reducing legal friction for commercial deployment. It's a strong signal that the model is designed to attract enterprise adoption.

Compute Capital: $60B Financing Round and Amazon's Asset-Light Pivot

The activity on the compute capital side carries a distinctly financial character. A Wall Street banking consortium has launched a $60 billion AI chip financing deal, structured as $42 billion in Class A senior secured debt and $18 billion in Class B subordinated debt. Blackstone is leading the Class B tranche, with its funds committing $9 billion. AI companies like Anthropic will use the proceeds to purchase chips and servers, while Broadcom is seeking to leverage the opportunity to challenge Nvidia and expand its data center equipment footprint. For context, back in August, Nvidia partnered with six institutions to raise roughly $500 billion in a similar initiative.

Amazon's move is equally noteworthy. Through an investor transfer, Amazon has moved $8 billion worth of Nvidia chips off its balance sheet — thousands of Grace Blackwell units will be placed into a Special Purpose Vehicle (SPV), which Amazon will then lease back while the SPV raises capital through debt issuance. Amazon retains up to a 10% equity stake, representing a broader shift toward asset-light operations. The chips will be deployed across more than ten data centers in five states, including Nevada. According to disclosures, Amazon's capital expenditures are expected to remain elevated this year; the company also issued $50 billion in bonds back in March.

This "off-balance-sheet hold, lease-back" model is essentially a way to continue expanding compute capacity without inflating the company's own balance sheet. When chip procurement runs into the tens of billions of dollars, managing that spend has become a central challenge of financial engineering for tech giants.

Special Purpose Vehicles (SPVs) are isolated legal entities commonly used in asset securitization and project finance. The core logic is to "carve out" specific assets from a parent company's balance sheet into an independent legal entity, which then issues debt or raises financing, while the parent continues to use the assets via lease or service contracts. For Amazon, this means Nvidia chips no longer appear as fixed assets on its own books; capital expenditure pressure is converted into operating lease expenses, which can improve balance sheet metrics like the debt-to-asset ratio and free cash flow. This structure has long precedent in aviation (aircraft leasing) and telecom (tower spin-offs), and is now being transplanted by tech giants into large-scale AI compute procurement — reflecting a reality where compute assets have grown large enough to require dedicated financial engineering.

Embodied AI: Humanoid Robot Shipments Surge 432%

The humanoid robotics sector has delivered an impressive first-half scorecard. According to IDC data, nearly 25,000 units were shipped globally in the first half of the year, a year-over-year increase of 432.1%, with the market exceeding $740 million in value — up 320% year-over-year.

China dominated the market, shipping over 19,000 units and capturing a 77.9% global share. In terms of competitive rankings, Zhiyuan (AgiBot) led with over 8,600 units shipped, overtaking Unitree for the top spot for the first time. Unitree shipped approximately 5,900 units in the first half, holding a 31% share and ranking second globally. IDC has also revised its 2030 forecast upward, projecting shipments to exceed 750,000 units with a compound annual growth rate of approximately 50%. The more-than-fourfold year-over-year growth is being driven by a combination of maturing supply chains and robust capital enthusiasm.

Embodied AI refers to AI systems where perception, decision-making, and planning capabilities are embedded in robots with physical bodies, enabling them to perceive and manipulate objects in real-world environments — rather than simply processing information in a digital space. Humanoid robots are the most representative form of embodied AI. Because they are highly compatible with human workspaces (they can use human tools and facilities directly), they are seen as the ultimate form of general-purpose labor substitution. The explosive growth in shipments is being driven partly by large language models significantly improving robots' perceptual and language-understanding capabilities, and partly by China's supply chain achieving cost reductions in core components like motors and sensors. Zhiyuan's rise to the top spot over Unitree also signals that market competition has moved from a technology validation phase into a mass production phase, where shipment volume has become the new primary competitive dimension.

Products & Compliance: ChatGPT Try-On, Chip Smuggling Case, and Zhipu AI Compensation

On the product front, ChatGPT's shopping feature now includes a virtual try-on capability. A "Try On" button appears on product pages, allowing users to upload a selfie or take a photo in the moment to generate a rendered image of themselves wearing the selected item. The feature is powered by the ChatGPT Images 2.5 image generation model. A new Favorites folder has also been added, where products can be saved to the ChatGPT Library and organized into folders, with try-on images and shopping history stored together. JD.com and Google are also advancing similar AI virtual try-on features, and this space is heating up quickly.

On the compliance and regulatory front, the U.S. Department of Justice has indicted a California man for allegedly smuggling AI servers valued at $300 million, containing U.S.-made GPUs — components commonly used for AI training. He faces three charges including conspiracy to violate the Export Control Reform Act. The 38-year-old suspect founded a company called Earthmade Computer, and the shipments in question were allegedly transited through Malaysia and Singapore before their final destination. He made his initial court appearance in the U.S. District Court for the Central District of California. The case underscores the tight constraints export controls place on the circulation of AI hardware.

On the data security front, Zhipu AI has announced a compensation plan in response to its recent Z-code repository upload controversy: paid users will receive 8 quota reset cards and additional daily token allocations. The company has also removed the warehouse snapshot upload pipeline and deleted the associated cloud-stored data. The incident originated from a repository sync setting that was enabled by default. Zhipu AI open-sourced Z-code on September 21 and has brought in CAICT and a third-party auditor for review. Financially, the company posted total revenue of 954 million yuan in the first half of the year, with models and API services accounting for 86.5% of that figure.

Wrap-Up

Today's headlines reveal a clear bifurcation: on the technology side, open-source models and embodied AI continue to break new ground; on the capital side, the financing and financial maneuvering around compute infrastructure is growing increasingly complex. As chips become both a means of production and a financial asset, whoever can more efficiently orchestrate compute resources will hold the initiative in the next phase of the race.