BriefTechNews
Microsoft's email crackdown bounces legit mail as Google, Broadcom, and Red Hat circle Nvidia

Microsoft's email crackdown bounces legit mail as Google, Broadcom, and Red Hat circle Nvidia

8 min read · 15 sources

TL;DR
  • Microsoft's Outlook.com is bouncing legitimate email after authentication changes, with SPF and DMARC failures on apparently compliant domains
  • Google unveiled two TPU variants, the 8tx with AMD processors and hardware encryption and the 8i Max with more memory for reasoning workloads
  • A forensic report claims Chinese labs including DeepSeek and Moonshot AI trained models on Nvidia Blackwell despite export controls
  • Broadcom's VM API project was accepted into the CNCF Sandbox for Kubernetes-based VM management
  • Linux Foundation research found VMs account for 53% of production applications on average across 492 cloud native practitioners

Microsoft’s consumer email services started bouncing legitimate messages this week, and the culprit is the company’s own authentication crackdown. Users report SPF and DMARC validation failures on domains that appear fully compliant, and a Microsoft support moderator has acknowledged a service degradation. The result: administrators are chasing intermittent delivery failures on mail that should sail through. If you run a mail server, today is a good day to check your bounce logs.

The timing is brutal, because this is the week Google decided to take a direct swing at Nvidia’s data center monopoly. Two new TPU variants are out, and one of them pairs Google silicon with AMD processors and hardware-level encryption. That is not a niche product. That is Google telling hyperscale customers they can run sensitive AI workloads without touching an Nvidia GPU. The rest of the day is a pile-on: Broadcom is pushing a vendor-neutral VM API into the CNCF, Red Hat shipped OpenShift 5 with AI-powered troubleshooting, and a forensic report claims Chinese labs trained on Blackwell despite export controls. Nvidia’s week is not going well.

Google’s 8tx TPU pairs its silicon with AMD processors and hardware-level encryption for sensitive workloads, a direct shot at Nvidia’s data center dominance.

Google's TPU 8tx and 8i Max take aim at Nvidia's crown

Source: reuters.com ↗

Google introduced two specialized TPU variants as cloud AI demand surges. The 8tx is the interesting one: it combines Google’s Tensor Processing Unit with AMD processors and hardware-level encryption for sensitive workloads. That is a direct answer to enterprises that cannot put regulated data on shared infrastructure. The 8i Max is the brute-force option, adding memory and compute capacity for reasoning, coding, and scientific applications that need to hold larger models in memory.

For engineers, the 8tx matters because it changes the threat model. Hardware-level encryption means data is protected even if the hypervisor is compromised, which is a real selling point for financial services and government customers. Google is not just offering a cheaper chip; it is offering a chip that solves a compliance problem Nvidia’s current lineup does not. Whether it beats Nvidia on raw price-performance is an open question, but the encryption angle is a wedge that does not exist on the other side.

Marvell and Penguin Solutions take on the RAMpocalypse with photonic memory

Source: theregister.com ↗

Memory prices are still climbing, and Marvell and Penguin Solutions have a proposal: a photonic memory fabric that lets appliances share 16 TB of DRAM and half a terabyte of HBM across up to 16 systems. The pitch is that you stop buying memory per machine and instead pool it, using photonics to keep latency low enough that remote memory behaves like local memory.

This is a big deal for anyone running AI inference clusters, where HBM is the scarcest and most expensive resource. If the fabric works as advertised, you can over-provision compute and share a smaller pool of expensive memory across it. The catch is the usual one: photonic interconnects add complexity, and the failure domain grows from one box to sixteen. But with DRAM prices where they are, refurbished kit is looking less secondhand by the day, and pooling is starting to look rational.

The campaign to put Jensen Huang back in charge at Nvidia

Source: theregister.com ↗

Elon Musk is leading a Silicon Valley campaign to put Jensen Huang back in charge at Nvidia, and it has collected more than 173,000 signatures plus a few famous billionaires. The catch, per the reporting: there is absolutely no guarantee those signatures are genuine. The petition is a pressure campaign, not a governance mechanism, and Nvidia’s board is under no obligation to act on it.

The context is a company in what the analysis calls a midlife crisis. Jensen Huang’s self-sabotage has the competition circling, and Google’s TPU announcement today is exactly the kind of opportunistic strike that happens when a leader is distracted. For engineers, the practical takeaway is that Nvidia’s roadmap could shift if leadership changes, and that uncertainty is already affecting procurement decisions. Nobody wants to standardize on a platform whose CEO might be ousted by a petition.

Chinese labs trained on Blackwell despite export controls, forensic report claims

Source: theregister.com ↗

A forensic report from a former OpenAI engineer claims to find Blackwell’s fingerprint in models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. The export controls were supposed to keep Nvidia’s most advanced silicon out of Chinese labs, and this report says they failed. The implication is that the hardware got in through a third country or a re-export loophole.

For engineers running models from these labs, the report matters because it affects trust. If a model was trained on hardware that was not supposed to be there, the provenance and compliance posture of that model is questionable. For everyone else, it is a reminder that export controls are only as good as the enforcement, and that AI capability is flowing regardless of policy.

A third of US AI chatters are insulting their bots

Source: theregister.com ↗

Anthropic has set behavior standards to protect Claude from rough treatment, and the reason is that a third of US AI chatters are hurling insults at LLMs. That is a lot of people yelling at software. Anthropic’s move is about defining what the model should tolerate and what it should push back on.

The engineering angle is the data. If a third of traffic is abusive, that is a significant fraction of the training and fine-tuning pipeline, and it shapes how models respond to edge cases. Setting behavior standards is not just PR; it changes the reward model and the safety tuning. If you are building on Claude, expect it to be less deferential to hostile input going forward.

The Cyberpocalypse hits China harder

Source: chinatalk.media ↗

Researchers used an OpenAI model to find a critical zero-day in PHP, the language underpinning roughly 70% of the web including Wikipedia and WordPress. Exploitation attempts began before a patch was released, and Google’s incident response team reports the average time to exploit a new vulnerability is -7 days. Attacks routinely precede patches. One week after public disclosure, the US patched over 25% of 14 million exposed services. China patched only 4%.

The takeaway is not about China specifically. It is that AI is collapsing the time and cost of finding and exploiting flaws, and the patching ecosystem cannot keep up. If you run PHP anywhere, assume it is being probed right now and that the window between disclosure and exploitation is negative. Patch aggressively and assume your perimeter is already compromised.

VMs are still 53% of production, and Broadcom wants to manage them with Kubernetes

Source: cncf.io ↗

A Linux Foundation Research study sponsored by CNCF and Broadcom surveyed 492 cloud native practitioners and found VMs account for an average of 53% of production applications. 71% expect that to stay stable over the next year. Larger organizations run even more VMs: 64% for companies with 10,000+ employees versus 45% for smaller ones. The container-everything narrative is wrong for most of production.

That is why Broadcom’s VM API project landing in the CNCF Sandbox matters. It is a vendor-neutral way to manage VMs through Kubernetes APIs, which means operators can stop juggling two control planes. Broadcom also announced a workload conformance initiative and more than 20 validated VMware vSphere Kubernetes Service partner solutions. If you run a mixed environment, this is the bridge you have been waiting for.

OpenShift 5 ships with AI-powered operations

Source: itpro.com ↗

Red Hat announced general availability of OpenShift 5 at KubeCon, promising built-in operational intelligence. The headline features are faster root-cause analysis, automated update pre-checks for API compatibility and risk assessment, and a five-year predictable lifecycle. OpenShift Lightspeed is evolving from a conversational assistant to an agent with proactive troubleshooting.

For platform teams, the update pre-checks are the sleeper feature. Automating API compatibility checks before an update means fewer surprise breakages in production. The five-year lifecycle is a commitment that matters for enterprises that cannot upgrade every six months. Red Hat is betting that AI-assisted operations is the differentiator, and it is shipping that bet now.

Google Cloud offers self-managed VMware on bare metal

Source: infoq.com ↗

Google Cloud announced general availability of a self-managed VMware Cloud Foundation 9.x solution on dedicated bare metal, giving customers root access across ESXi, vCenter, and NSX in 22 regions. The hardware is serious: z3-highmem-192-highlssd-metal instances with 96 physical cores, 1.5 TB memory, 72 TB local NVMe, and 200 Gbps networking. Customers handle their own upgrades and patching, and can bring existing VCF subscriptions.

This joins similar offerings from AWS and Oracle, and the economics are the pitch: up to 40% lower server TCO and 39% lower storage TCO. For engineers who have been fighting for control over their VMware stack, this is the option that gives it to you on Google Cloud. The trade-off is that you own the patching burden, which is exactly what some teams do not want.

Cisco tripled GPU utilization by building its own AI factory

Source: blogs.cisco.com ↗

Cisco IT built a full-stack, multi-tenant AI factory to support MyAgent for 90,000 employees and Cisco IQ for production workloads. The four barriers were familiar: stranded GPU capacity at 10-15% utilization, unpredictable token costs, and security constraints. Cisco reports tripling GPU utilization and expanding its footprint sixfold without adding data center space.

The blueprint is useful for anyone running enterprise AI at scale. Pooled networking, rack-level cooling, and integrated security turned idle GPUs into production capacity. The 10-15% utilization number is the one to remember; if your cluster is in that range, you are leaving money on the table and Cisco just showed a path out.

The bottleneck is you

Source: frantic.im ↗

A developer’s essay on programming with AI agents describes the exhaustion that comes from context switching between threads, addictive social-app-like interfaces, complexity overload, and decision fatigue. The author names it “idle agent anxiety” and describes a death march feeling. The systems get more complex, the output quality drops, and the human gets burned out.

The engineering lesson is that AI agents are tools with real cognitive costs that do not show up in benchmarks. If your team is adopting agents, the workflow design matters as much as the model choice. Redesign around human attention and recovery, not just code output.

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