August 5 marked the day open autonomous driving models entered the commercialization stage, and we began to see creative approaches to separating massive AI infrastructure investments from companies' financials. In healthcare, too, the shift from single-task AI to general-purpose foundation models is gaining momentum.
NVIDIA Makes Open Reasoning Model "Alpamayo 2 Super" for Robotaxis Available for Commercial Use
NVIDIA has released "Alpamayo 2 Super", an open reasoning model for robotaxis and autonomous vehicles (AVs), on Hugging Face with commercial use now permitted. The model is post-trained with reinforcement learning on top of the company's "Cosmos 3 Super Reasoner" and is the latest version of the Alpamayo family, the most widely adopted autonomous driving reasoning model on Hugging Face. Alpamayo has racked up more than 500,000 downloads in total.
The biggest point is the license. Under the Linux Foundation's permissive open license "OpenMDW-1.1", fine-tuning, derivative models, and even commercial redistribution are allowed. AV developers and automakers can adapt the model to their own driving data and driving policies. Unlike the initial release, which was restricted to research use, this license now applies to the entire Alpamayo family, opening a path for companies to handle everything from adaptation to production deployment in-house.
In terms of performance, it ranks first among the roughly 40 models evaluated on LingoQA, a reasoning benchmark for autonomous driving. In the company's tests (using the Lingo-Judge metric), it reportedly outperformed Qwen2.5-VL 72B by 17.0 points, Gemini 2.5 Pro by 15.1 points, and GPT-4o by 23.2 points. The model integrates camera footage from the front, rear, left, and right of the vehicle to reason about the situation, and can output planned trajectories, traces showing its reasoning process, and meta-actions indicating intent, all together. The design is deliberately tilted toward making it easier for developers to verify "what the model saw and how it judged".
Google Separates Risk of Anthropic Chip Procurement from Its Balance Sheet
It has emerged that Google is structuring a multi-billion-dollar financing arrangement to supply AI chips and data centers to Anthropic in a way that keeps most of the risk off its own balance sheet. Under the structure, which involves Broadcom, Apollo, Blackstone, and Morgan Stanley, contracts worth roughly $200 billion would ultimately depend on Anthropic's growth and its ability to make lease payments.
What matters here is that the move to "separate massive AI infrastructure investments from the parent company's financials" is taking concrete shape. Liabilities that would directly hit margins and credit risk if booked on Google's own books are instead borne by external financial and infrastructure players. As AI demand balloons, we are getting a glimpse of how big tech distributes procurement risk — a part of the business that has rarely been visible.
Separately, Anthropic has secured on the order of $10 billion in compute from Volta Infra Holdings, a cloud startup founded just a few months ago, but that deal was already reported previously, so this article focuses on Google's financing structure.
Microsoft Releases "PRISM2", a Foundation Model for Pathology Diagnosis
Microsoft Research and Paige (now under Tempus) have developed "PRISM2", a foundation model for pathology diagnosis, and published their findings in Nature Medicine. Whereas previous pathology AI had to be rebuilt for each individual task, such as "detecting a specific cancer", PRISM2 is trained on a combination of tissue images and language data from pathology reports. With a single model, it reportedly achieved results matching or exceeding specialized systems across multiple benchmarks, including detection of prostate cancer, breast cancer, and lymph node metastases.
Pathology sits at the heart of many cancer diagnoses, and pathology reports shape treatment decisions. PRISM2 can handle images alone or both images and text, and can be queried via prompts. The researchers say it will make it easier to build and adapt future pathology tools without rebuilding them from scratch for each task. The model weights are available on Hugging Face for research purposes.
NVIDIA Overhauls the AI Storage Layer — Open-Sources the cuFile API
Coinciding with this week's Future of Memory and Storage (FMS) conference, NVIDIA announced a set of moves in the storage layer that underpins AI data access. The largest is open-sourcing the "cuFile API" and its underlying stack, which lets GPUs read from and write to storage directly without going through the CPU. Google, Intel, and Meta have signed on as initial maintainers.
The backdrop: AI agents consume enormous amounts of data, and GPUs now generate thousands of parallel trace requests. NVIDIA's Vera CPU (part of Vera BlueField-4 STX) reportedly delivered up to 3.21x the throughput of x86 CPUs in a two-stage compression-and-encryption pipeline. The company also introduced the "Storage-Next" initiative, involving more than 40 storage vendors, and "SCADA", a framework that lets GPUs pull only the data they need directly into high-speed memory. It is no longer just about AI compute power — how fast and how securely data is supplied is becoming the next bottleneck.
Data Center Regulation: Texas Halts New Construction, US Government Weighs Banning Chinese Components
While infrastructure keeps expanding, regulatory pressure is mounting as well. Data centers have clustered in Texas thanks to its light regulation and abundant power, but the governor has called for a review of new data centers, and new construction has been halted. Even Texas, known for its lax regulation, is starting to show its limits.
In addition, Reuters reported that the Trump administration is drafting a plan to ban the use of Chinese-made components in AI data centers. From a security and supply chain perspective, the provenance of the equipment used in AI infrastructure is once again being called into question.
Apple Asks Federal Judge to Bar OpenAI from Using Trade Secrets
In its lawsuit alleging that OpenAI systematically stole confidential information about future Apple products, Apple has asked a federal judge for an immediate injunction. The preliminary injunction would require OpenAI to return the confidential information it obtained to Apple and to stop acquiring other non-public information, and is expected to remain in effect while the case proceeds. Among big tech companies, the battle over secrets in the AI race continues.