BriefTechNews
Intel's EMIB deal, Palo Alto's $4.5B AI bet, and the day's other big moves

Intel's EMIB deal, Palo Alto's $4.5B AI bet, and the day's other big moves

7 min read · 11 sources

TL;DR
  • Microsoft and Intel expanded their partnership, with Microsoft as the first customer for Intel Foundry's EMIB packaging and securing multi-year Xeon supply including Diamond Rapids CPUs.
  • Palo Alto Networks agreed to acquire Applied Compute for $4.5 billion in stock to add tools for training and serving customized open-weight AI models.
  • Workday will establish an Ireland-based operating company with an EU-citizen board to govern its EU Sovereign Cloud, operational in early 2027.
  • Google's Sashiko, an open-source agentic code review system for Linux, completes about 1,200 reviews daily and correlates with a 35 percent drop in 60-day follow-up fixes.
  • Amazon CloudFront Functions now supports network calls via cf.fetch() and cf.beacon() at $0.10 per million calls.

Microsoft and Intel just made the kind of announcement that quietly reshapes what you can rent in Azure for the next five years. Microsoft is the first customer for Intel Foundry’s EMIB advanced packaging, has locked in multi-year Xeon supply including the upcoming Diamond Rapids CPUs, and is collaborating on CXL memory, QuickAssist, and TDX confidential computing. The claimed numbers are the headline: up to 2.3x performance for Azure M-series v4 VMs with CXL and Flat Memory Mode, and up to 2.5x faster SQL backups with QAT-enabled VMs.

This isn’t just a press release about a nice partnership. It’s a signal that Microsoft is betting its first-party silicon roadmap on Intel’s foundry, and that the next wave of Azure VM generations will be tuned around Intel’s packaging and accelerators rather than just raw core counts. For anyone sizing workloads, the M-series v4 numbers mean the memory-bandwidth bottleneck just got a lot cheaper to address.

The rest of the day is about who’s buying what to win the AI infrastructure game, and how Europe is trying to keep its data on its own soil.

Microsoft will be the first customer for Intel Foundry’s EMIB packaging, with claimed gains of up to 2.5x faster SQL backups on QAT-enabled VMs.

Palo Alto buys its way into the AI training rack

Source: investors.paloaltonetworks.com ↗

Palo Alto Networks agreed to acquire Applied Compute for $4.5 billion in stock, adding tools for training and serving customized open-weight AI models. The deal is expected to close in Palo Alto’s fiscal third quarter of 2027 and will support deployments across public and sovereign clouds.

The strategic logic is clear: security vendors are realizing that the next attack surface is the model itself, and the next product is the platform that trains and serves it securely. Applied Compute gives Palo Alto a way to sell the infrastructure and the guardrails together. For engineers, the interesting bit is the sovereign cloud angle - this is another sign that the “open-weight model on your own rack” market is becoming a real procurement category, not just a research hobby.

Workday moves its sovereign cloud governance to Ireland

Source: investor.workday.com ↗

Workday announced it will establish a dedicated EU-based operating company in Ireland with a board composed of EU citizens to govern its EU Sovereign Cloud. The service runs on the AWS European Sovereign Cloud, and the new entity is expected to become operational in early 2027.

This is a structural answer to a legal problem. Simply running the servers in Frankfurt isn’t enough anymore - regulators and customers want to know that the people making decisions about data access are also subject to EU law. For engineers evaluating HR and finance workloads, this changes the compliance calculus: the governance layer is now a separate legal entity, not just a checkbox on a shared services agreement.

The AI decision problem: what gets to act on its own

Source: federalnewsnetwork.com ↗

IT leaders are wrestling with which decisions machines can make without a human in the loop, according to a piece featuring Mark Holterman, CIO of the Idaho National Laboratory. The tension is between rules-based automation and LLM-driven synthesis in high-security environments, where the failure modes are not just bad output but unauthorized action.

The practical hurdles are data readiness, over-trust in AI results, and classifying sensitive information. The emerging pattern is localized LLMs, strict data governance, and mandatory human-in-the-loop checkpoints for anything that can trigger a consequential action. This is the operational reality behind all the agent hype: the bottleneck isn’t the model, it’s the control plane around it.

Legacy software is the real AI bottleneck

Source: federalnewsnetwork.com ↗

Google Cloud’s Jeanette Manfra makes the case that maintenance burdens and vendor lock-in are strangling AI ambitions. The numbers are stark: 59% of the 1,250 surveyed U.S. enterprise technology decisionmakers report excessive time spent maintaining legacy systems, and 78% say it diverts capital and talent from AI.

The scale of the patching burden is the real story. Microsoft addressed 3,338 vulnerabilities across July-October 2026, including 1,407 in October alone, versus 1,130 in all of 2025. The “release-then-patch” model is becoming unsustainable as AI accelerates both the discovery and the exploitation of flaws. For engineers, this is the argument for vendor consolidation and for pushing security requirements into procurement - the cost of maintaining a sprawling estate is now measured in lost AI capacity.

Nvidia and Microsoft go deeper on Azure and on-prem

Source: blogs.nvidia.com ↗

Nvidia and Microsoft expanded their collaboration at Ignite, adding Azure VMs powered by Vera Rubin NVL72 systems and extending confidential computing to bare-metal and rack-scale AI infrastructure. Microsoft is also using Nvidia Dynamo for post-training and RTX PRO 6000 Blackwell GPUs for Bing search models.

The MLPerf results are the concrete takeaway: Azure scaled to 10,240 GPUs with GB200 NVL72 on DeepSeek-V3, and GB300 NVL72 led cloud submissions on DeepSeek-R1 interactive throughput by over 10%. The on-prem angle matters too - planned Azure Local support for additional Nvidia GPUs and Jetson platforms means the same stack can run in a factory or a data center, not just in a Microsoft region.

OVHcloud bets big on European sovereignty

Source: cloudcomputing-news.net ↗

OVHcloud announced 55 new cloud and AI products at its AI Summit, covering computing, AI development, networking, and enterprise applications. The announcement includes plans to add Nvidia Blackwell B300 and Rubin R200 GPUs.

The context is the European Commission’s push for digital resilience, with contracts worth up to €180 million over six years and OVHcloud’s consortium achieving SEAL-3 classification. The number that should worry every European cloud engineer: the ECB estimates AWS, Microsoft, and Google hold 82% of Europe’s broader cloud market. That concentration is the operational risk that sovereignty mandates are trying to address, and OVHcloud is positioning itself as the alternative.

Microsoft's agent controls: a control plane for behavior

Source: commandline.microsoft.com ↗

Microsoft introduced agent controls, a managed Azure service for centrally defining and evaluating policies governing AI agent behavior across frameworks and clouds. It’s built on the open-source Agent Control Specification (ACS) and Agent Hooks, which define intervention points across the agent lifecycle: startup, input, model calls, tool calls, output, shutdown.

The evaluation stack uses content safety classifiers, LLM judges, and deterministic Rego rules. For engineers, this addresses the gap between safe answers and safe actions - a model can give a good response, but the thing that takes an action needs a policy layer that can stop it. The managed service angle means you don’t have to build that policy infrastructure yourself.

Sashiko: the code reviewer that scales

Source: cloud.google.com ↗

Google built and open-sourced Sashiko, an agentic code review system powered by Gemini, donated to the Linux Foundation. The Linux kernel spans over 35 million lines of code with over 85,000 changes per year, and maintainers manually review patches to catch memory leaks, deadlocks, race conditions, and security bugs.

The system separates bug finding from verification using parallel specialist sessions and per-bug deep checks. The production data is the headline: seven months of use on kernel mailing lists shows Sashiko completes about 1,200 reviews daily and correlates with a 35 percent drop in 60-day follow-up fixes for merged commits. This is the proof point that code review, not code generation, is the real bottleneck in large open-source projects.

CloudFront Functions grows some legs

Source: aws.amazon.com ↗

Amazon CloudFront Functions now supports network calls via new cf.fetch() and cf.beacon() methods, allowing lightweight JavaScript at edge locations to call APIs and act on responses in real time. This enables edge-based personalization, authentication, routing, and security checks without needing Lambda@Edge or origin-side logic.

Network calls cost $0.10 per million calls on top of invocation charges, are available at all edge locations, and are included in Pro and above flat-rate plans. This closes a significant gap - previously, CloudFront Functions were for simple transforms, and anything requiring an upstream call meant a more expensive Lambda@Edge or a round trip to the origin.

The GPU is not the product

Source: siliconangle.com ↗

Vultr CMO Kevin Cochrane made the case at KubeCon that enterprise AI adoption is pushing infrastructure beyond GPUs, requiring full cloud platforms that combine compute, networking, storage, security, and application lifecycle management. Vultr positions itself as “the tortoise in AI infrastructure,” prioritizing enterprise workload and platform-engineering requirements over just GPU availability.

The shift is real: training a model is a batch job, but running AI agents in production is a distributed systems problem. Agents need CPUs, networking, storage, and isolated environments to execute code and interact with enterprise systems. The vendors that win the next phase will be the ones that can run the whole stack, not just the ones with the biggest GPU allocation.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like