Boost Run Secures $525.6M AI Cloud Deal with Sovereign Company on NVIDIA GPUs

Key Takeaways

• Boost Run signed a $525.6 million AI cloud deal with a sovereign AI company using NVIDIA GB300 NVL72 GPUs
• Alpha Compute's Sweden ALPHA-02 cluster deal is worth $71.9 million and powered by hydropower
• Trump AI Accord is voluntary and relies on self-regulation across Anthropic, Google, Meta, OpenAI, Nvidia and xAI
• GoodVision AI partners with TDTC on a 2.4 MW modular AI Factory in Texas
• SpaceX expects nearly $53 billion in AI spending in H2 2026 with potential $83 billion revenue in 2027
• AI evaluators face conflicts of interest since they are paid by the companies they audit
• MemTrace improves AI coding agent performance by 21.2 points on DeepSWE benchmarks
• Westinghouse launches PWROG AI Assistant built on bertha technology for nuclear industry
• ARUP Laboratories selected for Utah's Redtail AI Factory using EventHorizon foundation model
• Researchers introduce Causally Fair Generation to reduce demographic disparities in LLM outputs

GoodVision AI partners on Texas AI factory and inference pilot

GoodVision AI announced two collaboration agreements in Texas with TDTC and HOP Global. The TDTC deal involves a modular AI Factory with 2.4 MW of confirmed power capacity. The HOP Global project will test RTX PRO 6000 AI inference clusters with 400G high-speed networking. Both projects aim to expand GoodVision's AI inference infrastructure across different deployment conditions.

Boost Run signs $525.6M AI cloud deal, total commitments top $2.6B

Boost Run signed a five-year AI cloud services agreement worth $525.6 million with a leading sovereign AI company. The deal includes compute capacity powered by NVIDIA GB300 NVL72 GPUs, network storage and CPU node services. Initial infrastructure deployment is expected to begin in early Q2 2027. This contract brings Boost Run's total customer revenue commitments to over $2.6 billion.

Alpha Compute signs $71.9M Sweden AI cluster deal

Alpha Compute signed a $71.9 million agreement for its ALPHA-02 cluster in Sweden on September 30, 2026. The cluster uses NVIDIA B300 GPUs for AI training and inference workloads. The data center is powered entirely by hydropower. Alpha Compute expects the deal to add $24 million in new annual contracted revenue, bringing projected annual revenue to $47 million.

AI evaluators face challenges auditing powerful systems

California and dozens of countries are pushing for mandatory AI audits conducted by AI evaluators. Governor Gavin Newsom signed legislation including testing requirements for AI systems. Over 200 AI researchers and evaluators signed a letter calling for industry standards for auditors. Concerns include conflicts of interest since evaluators are paid by the companies they evaluate, similar to issues in the bond-rating industry before the 2008 financial crisis.

Trump AI Accord relies on voluntary audits and self-regulation

President Trump met with executives from Anthropic, Google, Meta, OpenAI, Nvidia and xAI on September 29, resulting in the Trump AI Accord. The agreement is voluntary and not legally binding, but establishes a four-layer system of controls and audits for frontier AI systems. The layers include internal controls, internal review teams, independent external auditors and board oversight committees. Companies also agreed to meet regularly to develop safety standards and best practices.

MemTrace helps AI coding agents track long tasks

MemTrace is a memory system that helps AI coding agents handle long software projects. It stores execution history and checks if old evidence is still valid before using it. The system uses a Memory Trace Graph to organize dependencies. It improved DeepSWE pass@1 by 21.2 points and SWE-Milestone Score by 17.8 points in benchmarks.

Guide to building AI-ready facilities technology

This article explains how to prepare facility management for AI use. It says organizations should first define clear business problems before adopting AI. The guide lists five key technology areas including BAS/BMS for equipment behavior, CMMS for maintenance history, IoT sensors, energy platforms, and IWMS for business context. It stresses that data quality and proper integration matter more than collecting every possible data point.

US debates whether AI companies should pay taxes

A growing debate in the United States asks whether AI companies should pay special taxes. Bill Gates suggested robot-using companies should contribute to Social Security. Lawmakers like Elizabeth Warren and Mark Kelly have proposed taxing AI companies to fund employment programs. Other ideas include taxes on AI data centers, tokens, or usage. Experts say there is no single approach and options range from company-specific taxes to broader tax code reforms.

SpaceX AI spending could bring $83 billion payoff

Evercore expects SpaceX to spend nearly $53 billion on AI in the second half of 2026. The firm sees potential for $83 billion in 2027 AI revenue, higher than the $47 billion consensus. SpaceX has Colossus 1 and 2 with 780,000 GPUs online. Evercore gave the stock an Outperform rating with a $230 price target, saying heavy spending now could create a capacity advantage later.

ARUP AI project chosen for Redtail AI Factory

ARUP Laboratories was selected for Utah's Redtail AI Factory challenge program. Their project uses a foundation model called EventHorizon to improve clinical flow cytometry analysis. The award gives researchers access to Redtail, a statewide AI supercomputer managed by the University of Utah. Redtail ranks 159th on the Top 500 supercomputer list and was built through a $50 million public-private partnership with Hewlett Packard Enterprise and NVIDIA.

Westinghouse launches PWROG AI Assistant powered by bertha technology

Westinghouse launched the PWROG AI Assistant, built on its bertha AI technology. The platform lets nuclear industry personnel ask questions in plain language and get answers from decades of approved technical knowledge. It includes references so users can verify information before acting. The tool is available to domestic and international PWROG members and helps engineers work faster while keeping human oversight central.

New CFG method removes causal disparities from large language model outputs

Researchers introduced Causally Fair Generation, called CFG, to reduce demographic disparities in large language model outputs. CFG uses causal inference to identify and remove selected unfair causal effects from generated text. It also allows justifiable pathways to remain when they serve the task. The method was tested with four LLMs across real-world and synthetic datasets.

Bidirectional Preference Synthesis improves model learning from boundary failures

Bidirectional Preference Synthesis, called BPS, is a new data construction method for Direct Preference Optimization. It handles boundary failures by adding reverse prompt pairs so the same response can be rejected in one context and chosen in another. On Qwen3-4B-Instruct-2507, achieved-side ranking accuracy rose from 6.8% to 62.3%. Human audits and downstream evaluations supported the method's effectiveness.

Sources

• A Texas AI factory could bring in about $210M a year at full capacity, GoodVision AI estimates.
• Customer commitments to Boost Run top $2.6 billion after a new AI computing deal
• A $71.9 million AI computing deal will use a Sweden data center powered entirely by hydropower.
• ‘Evaluators’ are supposed to keep AI from killing us all. No pressure
• Trump’s AI Accord: Can Voluntary Self-Regulation and Independent Audits Protect Against Frontier AI Risks?
• MemTrace: State-Consistent Memory for Long-Horizon Coding Agents
• Tech & FM: Building An AI-Ready Facilities Tech Stack
• Should AI pay taxes? The debate is gaining traction in the US
• SpaceX Is About to Spend $53 Billion on AI. Evercore Sees an $83 Billion Payoff
• ARUP Artificial Intelligence Project Selected for Redtail AI Factory Challenge Project
• Introducing the PWROG AI Assistant: Westinghouse's bertha™ Technology Is Transforming How the Nuclear Industry Accesses Knowledge
• Causally Fair Generation with Large Language Models
• Bidirectional Preference Synthesis: Learning Prompt-Conditioned Preferences from Boundary Failures