• AI KATANA
  • Posts
  • Microsoft is now hosting xAI’s Grok 3 models

Microsoft is now hosting xAI’s Grok 3 models

Also: Congress Moves to Override State AI Laws in Draft Tax Bill

Author

AI KATANA
May 19, 2025

Hello!

Today’s edition captures major power plays in the AI and semiconductor world, with Microsoft making a bold move by hosting xAI’s Grok 3 models on Azure—an expansion that could redefine its relationship with OpenAI while signaling Azure’s ambition to become the premier platform for diverse AI capabilities. This strategic shift is matched by rapid innovation from Microsoft researchers who unveiled “PipelineFlow,” a fine-tuning wizard that slashes training time and costs for enterprises. Nvidia, not to be outdone, showcased its next-gen NVLink Fusion and a major expansion in Taiwan, underscoring its continued lead in AI infrastructure. Meanwhile, a controversial federal tax bill aims to override state AI laws, igniting a fierce debate over governance. On the compliance front, Azure AI now offers built-in guardrails that streamline model certification and risk reporting. Globally, U.S. chipmakers are eyeing Saudi Arabia as the next frontier for AI fabs, driven by geopolitics and energy incentives. Finally, the developer and venture worlds see fresh momentum, from AWS open-sourcing a robust agent SDK to VC backing new AI tools in law and pharma.

Sliced just for you:

  • 🦾 Microsoft is now hosting xAI’s Grok 3 models

  • 🧑‍💻 Microsoft’s “PipelineFlow” Training Cuts Fine-Tuning to Minutes

  • 🔌 Nvidia Debuts NVLink Fusion and a Taiwan AI Hub

  • 🏛️ Congress Moves to Override State AI Laws in Draft Tax Bill

  • 🔒 Azure AI Adds Built-In Impact-Assessment & Compliance Guardrails

  • 🏜️ U.S. Chipmakers Court Saudi Arabia for Next-Gen AI Fabs

Microsoft has officially added xAI’s Grok 3 and Grok 3 mini models to its Azure AI Foundry service, a move that expands its growing portfolio of third-party AI offerings and positions Azure as a go-to platform for businesses and developers seeking access to cutting-edge models. The decision, reportedly driven by CEO Satya Nadella’s aggressive push to host prominent models, highlights Microsoft’s strategy to diversify beyond its existing partnership with OpenAI—a relationship that may face further strain due to this development. The timing follows recent controversies involving Grok, including unauthorized code modifications that led to politically charged outputs, and continues the backdrop of legal tensions between Elon Musk and OpenAI. Despite potential friction, Microsoft is providing full service-level agreements for Grok on Azure, reinforcing its ambition to dominate the AI model hosting market.

Microsoft researchers have wrapped data ingestion, synthetic augmentation and LoRA-style weight updates into a single “PipelineFlow” wizard inside Azure Machine Learning. In demos, a 13-billion-parameter model hit domain-specific accuracy targets using just 500 labelled examples and 120 GB of GPU memory—reducing fine-tuning time from hours to 18 minutes. The workflow auto-evaluates bias, injects red-team prompts and stores checkpoints in Fabric by default, letting small teams spin up bespoke chatbots without MLOps overhead. Microsoft claims enterprises can cut training costs by 65 percent and CO₂ emissions by 40 percent versus previous pipelines, though critics say results may vary outside benchmark datasets. The feature ships in public preview this week with pay-as-you-go GPU billing.

🔌 Nvidia Debuts NVLink Fusion and a Taiwan AI Hub

Nvidia’s keynote revealed a fibre-optic NVLink capable of lashing 256 Blackwell GPUs into one logical processor, quadrupling inter-GPU bandwidth while trimming rack cabling. The firm also announced a 600-engineer campus in Hsinchu to co-design liquid-cooled AI servers with Taiwanese partners, reinforcing the island’s role in Nvidia’s supply chain. NVLink Fusion, slated for Q4 sampling, already has design wins with MediaTek and Marvell, signalling a new push into custom accelerators. Observers note the upgrade keeps Nvidia a step ahead of chiplet-based rivals even as U.S. export rules tighten.

A surprise clause in a sweeping tax package would pre-empt state-level AI regulations, centralising rule-making under federal agencies. Proponents argue a single framework will prevent a compliance maze that stifles innovation, while attorneys-general from 18 states warn the “one-size-fits-all” approach strips local consumers of vital protections against deepfakes and biometric abuse. Policy analysts expect a fierce floor fight: civil-rights groups back state autonomy, but tech lobbyists say fragmented laws could freeze out small vendors. The clock is ticking—House leaders plan a vote before the July recess, aiming to bundle the measure with semiconductor tax credits.

Microsoft expanded Azure AI Foundry with an “AI Governance Kit” that automates algorithmic-impact assessments, generates risk reports and pipes evidence directly into Purview Compliance Manager. Hooks to third-party auditors Credo AI and Saidot let developers submit models for external review without leaving the console. The update also introduces a policy engine that blocks deployment if declared use cases diverge from documented intents, aiming to head off regulatory fines. Early enterprise testers say the kit shaves weeks off certification cycles, though some fear lock-in to Microsoft’s governance schema.

With Washington’s China export curbs still biting, U.S. semiconductor firms are looking to Riyadh’s $100 billion tech initiative for fresh growth. Industry executives told a Gulf summit they’re eyeing joint ventures that pair Saudi capital with American know-how to build wafer-scale AI accelerators inside the kingdom’s planned Neom megacity. Officials pitched low-cost solar power and streamlined permits, while suppliers stressed the need for U.S. export-license clarity. Analysts see the overture as a hedge that could diversify supply chains away from East Asia and unlock new compute hubs nearer to African and European markets.

🛠️ AI tools updates

Amazon Web Services released Strands, a Python toolkit that lets developers spin up multi-step AI agents with built-in memory, tool-calling and autonomous error recovery. Strands scaffolds vector stores, schedules retries and logs provenance metadata to CloudWatch, aiming to make agentic apps as easy to deploy as a Lambda function. Early adopters report slashing prototype time from days to hours, and the Apache-2.0 license signals AWS’s bid to compete with LangChain-style frameworks on openness.

NLWeb, an open-source project unveiled at Build, wraps a JavaScript widget and serverless back-end that inject natural-language search into existing websites. It indexes structured and unstructured content, then routes queries to a choice of LLMs—local or hosted—returning citations in line. Early testers plugged NLWeb into static docs sites in under 30 minutes, giving users GPT-style Q&A without handing data to third-party SaaS. The code is on GitHub under MIT license.

💵 Venture Capital updates

California-based Theo Ai secured a $4.2 million seed round co-led by NextView Ventures and Collide Capital. Its platform ingests docket data and judge histories to forecast case trajectories, offering law firms probabilistic settlement ranges and strategy suggestions. Funds will expand the training corpus beyond U.S. federal courts and add multilingual support for cross-border disputes.

Persist AI closed a $12 million Series A to scale its lab-automation robots that use reinforcement learning to fine-tune pill coatings and injectable suspensions. The Y Combinator-backed firm says its platform cuts formulation cycles from months to weeks, improving drug stability and patient adherence. The cash will fund a second automated line and expand partnerships with biopharma CDMOs.

🫡 Meme of the day

⭐️ Generative AI image of the day