The Register

The UK government is hiring someone to help lead its AI strategy for civil servants, and actual AI experience is apparently optional. The Cabinet Office is advertising for a Head of Strategy and Engagement in its Digital, Data, Innovation and AI team, or DDI & AI, with a salary of up to £83,355 ($113,000). The successful applicant will help shape how roughly 550,000 civil servants acquire the skills needed to survive Whitehall's increasingly enthusiastic embrace of AI. According to the job ad, the new hire will "lead on the design and delivery of a long-term, high-impact and adaptive vision" for the DDI & AI team, define its strategy, track developments in the sector, and work with experts across government, industry and academia. They'll also lead on policy positions and ministerial submissions. You might reasonably assume that some previous experience with AI, digital technology, data or innovation would come in handy, but Whitehall does not insist. The essential requirements instead include leadership, strategic thinking, communication skills, and the ability to build "productive collaborative and trusting relationships." Experience in "digital, data, innovation or AI" is tucked away under desirable criteria, alongside experience in learning design and delivery. To be fair to the Cabinet Office, this isn't a job for someone to train a foundation model between meetings with ministers. The DDI & AI team focuses on improving the skills of non-digital specialists across government, and the vacancy is heavy on strategy, planning, and engagement rather than technical work. The Cabinet Office did not respond to The Register’s questions, but a spokesperson told The Telegraph that it was "not a technical AI specialist role," adding that the team already includes people with AI expertise, and works with specialists across government, academia and the tech industry. A couple of months ago, the Cabinet Office was looking for a rather different sort of AI hire. Its £163,000-a-year AI and Innovation Director was expected to help create an "AI-first culture" across the Civil Service – and, for that job at least, a "deep understanding of the AI landscape" was part of the deal. Meanwhile, ministers have been pushing AI tools into government in the hope of squeezing more productivity from the public sector, including the brilliantly named Humphrey suite of AI assistants. The government has also set itself the rather sizeable task of improving AI and digital skills across hundreds of thousands of civil servants. Which brings us back to the new Head of Strategy and Engagement: the person helping decide how Whitehall builds those skills could, according to the government's own criteria, arrive without having previously worked in AI, digital, data, or innovation. One place where AI gets more explicit treatment is the application. Applicants are warned that while they can use the technology to help with their application, their examples must come from their own experience. Passing off AI-generated material as their own could result in the application being withdrawn. So AI experience isn't essential for the job. Just make sure AI doesn't apply on your behalf. ®

AROS is an all-FOSS re-implementation of AmigaOS, and after years of effort, it now supports Arm kit – both 32-bit and 64-bit Raspberry Pi hardware, especially the Pi 3. The AROS project's nightly builds page has some interesting new offerings. Notably, there's raspi-aarch64-system for 64-bit capable Raspberry Pi hardware, including the Pi 3 and Pi Zero, and raspi-armhf-system for older ARMv6 based devices. We first mentioned AROS 13 years ago, and last took a proper look in May 2025. Building a version of AROS that runs natively on Arm processors has been a long term goal, as the AROS Arm support page details. It has made good progress, though, and now the nightly images include a version which runs well on the Raspberry Pi 3. (Unfortunately, the Reg FOSS desk’s Pi 3 is currently running PiHole full-time, and Pi 4 support is slightly lagging, with no sound or networking.) There's no specific mention of what has changed on the project's Github repo, but we suspect that some judicious use of LLM has enabled this sudden burst of progress. Since May 2026, there have been 13 commits by Claude Code. Even so, it works. This is an existing, human-written OS which can now boot and run on a Raspberry Pi 3. The install images are small – only slightly over 100 MB in size – but if you want to give them a try, you should also copy the contents of the additional raspi-aarch64-contrib or raspi-armhf-contrib files, which contain an additional gigabyte of compressed files with apps, demos, accessories, and supplementary tools. Whether or not it has been made with LLM bot assistance, we think this is an impressive step forwards. AROS has been able to run on Arm CPUs for years, but until this, it ran hosted on top of another Arm OS – usually Linux. Now it can boot and run on bare metal. We feel this is a good step for several reasons. Although the native platform for AROS was for many years x86-32, on that platform, AROS is a small fish in a big pond: there's a vast array of off-the-shelf hardware for x86 PCs, and it’s impossible for a tiny experimental OS to support more than a fraction of it. We have tried and failed to get AROS running on bare metal more than once. For many years, the best way to try it was to run it under VMware thanks to a specialized distro called Icaros Desktop. The Raspberry Pi represents a much narrower platform: there are only a handful of models of Pi, and as a result, a very small number of GPUs and network interfaces that AROS needs to support. That makes it much easier to offer complete coverage. The Pi is also a fairly severely constrained platform: except for the most recent - and most expensive - models, most of the older and cheaper Pi kit only has either 512 MB or 1 GB of RAM. That isn’t enough to do full justice to Linux, Windows, or any BSD with a graphical desktop – but it’s plenty for AROS. This initial build boots to the Workbench desktop in about five seconds. And, of course, those lower-end Pis are still cheap, which is always welcome. For now, the Pi version of AROS is a prototype: it’s incomplete, may not be very stable, and lots of things don’t work yet. An important missing part is that it doesn’t have any 680x0 emulation yet, so it can’t run any classic Amiga apps. Even so, we suspect that quite a few old Amiga fans might enjoy playing with a brand-new £30 ($40) Amiga. If you just want to play your old games, there are much easier ways to do that on a PC. Last year, we also looked at Commodore OS 3 – and in there, we gave PiMiga a mention as well. Both products are free Linux distros bundled with emulators and tons of Amiga apps, games, and demos. Easier still, if you’re a Windows user, just buy a copy of Amiga Forever. Either way, knock yourself out. AROS is re-invigorating other bits of the Amiga world, as well. Later in 2025, we also reported on a new Amiga web browser, which runs on the A600 GS and A1200 NG computers. These are Arm-based machines which run Amiga software via emulation on an OS called AmiBench – which uses parts of AROS. TheA1200 is now officially an Amiga Alongside the A600 GS and A1200 NG Arm-based machines, last year we also mentioned THEA1200 from Retro Games Ltd. This is another Arm-based machine, based around a tweaked emulator, like RGL’s existing Sinclair reproduction The Spectrum. However, THEA1200 just recently moved one step closer to Amiga legitimacy. On its Facebook page, RGL announced that THEA1200 is officially an Amiga product: “The Amiga name has a new chapter. With the recent announcement from Mike Battilana and the Amiga Corporate [sic], we’re delighted to see the Amiga brand move forward with a clear and unified future. We’re proud to be part of that future with THEA1200, bringing the Amiga experience to a new generation while celebrating everything that made the original so special.” As we said last year, THEA1200 is an Arm computer running Linux, which runs a bundled emulator derived from Amiberry. The same is true of the AmigaKit machines we mentioned earlier, the A600 GS and A1200 NG: they too run an emulator based on Amiberry on top of Linux, and inside that, they run AmigaKit’s AmiBench OS. When RGL says it has come to an agreement with Mike Battilana and Amiga Corp, that means Cloanto. Cloanto is the company that sells AmigaForever, and it owns the Amiga trademarks, the Kickstart ROM, and the rights to the original 680x0 AmigaOS. Inside AmigaForever, along with various retro models for games compatibility, there's even a higher-spec VM running Cloanto’s updated version of classic AmigaOS, which it calls 3.X – it also has a page about the improvements 3.X contains. A deal with Cloanto means RGL can bundle the genuine original AmigaOS with TheA1200. This could give TheA1200 an edge over the AmigaKit models in terms of its compatibility with older games and original Amiga apps. (As one sign of the complexity of Amiga ownership, German developer Hyperion also has an AmigaOS license, for the purposes of developing a PowerPC version. The Register covered this back in 2004, and Hyperion used it to develop AmigaOS 4 – used in, among other things, the Amiga X1000 which The Reg covered in 2010. Controversially, Hyperion is using this license to flog updates to classic AmigaOS: in 2022, we reported on AmigaOS 3.2.2 and in 2025 on AmigaOS 3.2.3.) Christian Simpson – known on YouTube as Peri Fractic – recently revived Commodore International, and following its FPGA-powered Commodore 64 Ultimate computer, the company also launched a Sailfish-powered smart flip-phone, as The Register covered back in June. The Commodore brand is back, and it's in the news. However, the revived Commodore doesn't own the Amiga name, or IP, or code. Feel free to call us excessively cynical, but we suspect that this could be a motivation behind licensing some Amiga rights to another vendor in 2026. This exposure will increase the value of Cloanto's IP. Hypothetically, if Cloanto were to sell its rights to the revived Commodore, that could mean it could ask for more money. ®

BORK!BORK!BORK! Wide, super wide, ultra wide, and monitor-munga screens might be all the rage these days. But check out the girth of this Debian display in Scotland. Spotted by eagle-eyed Register reader Kenny Millar, this digital billboard in the west central Lowlands town of Greenock has been showing a Debian login prompt for a few weeks now. We know the Linux fraternity are less than keen on advertisements in their operating system, but this is getting ridiculous. Did you know that Debian had its own tartan? No, we didn't either. Our reader provided a link to a Google Street View shot of the board in happier times, in 2024. However, the base of the sign in 2024 hinted at the fate that has since befallen it. It was a 75Media billboard, something now removed. 75Media, a major outdoor advertising company in the UK, ceased trading at the end of 2025, and we fear that the Greenock signage whoopsie is a real-world example of the lights being on, but nobody's in. Or, in this case, Debian prompting for a login, but no humans to handle the request. So the board waits… It looks like Debian 11 – Bullseye – is the operating system in use. Released in 2021, the operating system reaches the end of long-term support in 2026, although fixes can continue until 2031 (assuming someone is paying the bills). With 75Media out of the picture, responsibility for the board has likely shifted to a different advertising agency. Despite its imposing physical dimensions, the display resolution comes in at just 1,200 x 320 pixels which is plenty for running a couple of Doom sessions side by side, but tight for actual desktop use. It's fitting, then, that the terminal driving the board is left on display. No true IT professional, after all, likes to stray far from the command line. There is something strangely plaintive about this board with its login prompt bared to the world, but with no techie behind the scenes to either put it out of its misery with a sudo shutdown -h or set off the next slideshow of ads to entice passing motorists. A case of perhaps IT is aware, but nobody cares. ®

Around 80 percent of UK teachers now regularly use AI in their work, although the minority say it has cut their hours. Research from YouGov shows that AI use among teachers has doubled in the past year. Among those who use it, just over half (51 percent) said it had reduced their workload, but only around a third (35 percent) said they were working fewer hours as a result, according to the poll of 1,033 teachers by YouGov. Although some time was saved with AI, that was more than made up for with demand to complete other tasks, resulting in 55 percent of teachers who say they use AI working the same hours as before, and 4 percent saying they work more. In April, the UK government launched a £23 million trial program for AI and educational technology it hoped would boost results and cut teachers' workloads. Many teachers work more than 50 hours a week, which is higher than contractual limits. But the YouGov survey — funded by education technology exhibition Bett — found reduced working hours do not necessarily follow from increased use of AI. Teachers said they were using AI across a spread of tasks: 76 percent to build lesson plans and worksheets, and around 39 percent to draft letters and emails to parents or write pupil reports. However, only 8 percent said they use AI to mark or grade pupils’ work. AI is not just for teachers, though. Pupils are using it too, and 71 percent of secondary school teachers (pupil-age 11-16 years) said AI had made it easier for children to cheat and get away with it. Fifty-seven percent of secondary teachers suspected at least one pupil of handing in AI-assisted work without permission in the prior four weeks, the report found. The research comes amid broader concerns about the effectiveness of AI across the economy. Despite the layoffs associated with AI in some tech companies, a study from the Atlanta Federal Reserve study found that about 90 percent of executives believe AI is yet to improve productivity in their organizations. ®

ICANN is poring through over 1,600 applications for new global top-level domains (GTLDs) and will reveal which ones it thinks are valid in a few weeks – but hundreds of names applicants hope to secure have already appeared online. GTLDs are the last part of an internet address. The most common and well-known are .com, .net, and .org. In the early 2000s, ICANN allowed a few more GTLDs, and in 2012 ran an open application process that allowed anyone to pay for the privilege of operating their own GTLD. That program saw over 1,900 applications and led to the creation of over 1,200 new GTLDs. Tech companies, including Microsoft, Cisco, and Oracle, scored their own GTLDs. One registry applied for, and won, .sucks – one of many novelty domains that ICANN approved. As The Register reported in May, ICANN this year opened a round of applications for GTLDs and recently announced that it received over 1,600 submissions. The internet governance organization plans a “Reveal Day” in September or October, on which it will publish a list of all applications that have passed administrative checks. Some applicants have jumped the gun and published the list of GTLDs they applied for. The Link Freedom Group has applied for 316 GTLDs, listed here. Among the strings it hopes to turn into GTLDs are .slop, .therapy, and .hype. The outfit also hopes to own .con, .token, and .moon, plus AI-related strings including “agentic” and “AGI”. An unofficial ICANN community Wiki, which The Register understands some applicants have used to share the GTLDs they desire, offers the chance to filter GTLDs in the 2026 application and lists the names sought by Link Freedom Group plus hundreds more. Predictably enough, many are strings related to cryptocurrencies, datacenters, and AI, or locations in China. ICANN charged an evaluation fee of $227,000 for each submission, and the cost can be higher if a desired string needs extra review due to being considered a highly risky GTLD. It can take months or years for new GTLDs to go live after Reveal Day, because ICANN performs additional reviews and operates a 104-day window in which it considers objections. If you want a domain using any of the GTLDs mentioned above, it’s unlikely registries will be ready to sell you one before late 2027 – and likely rather later than that. ®

Nutanix is closer to porting its stack to the Arm architecture, spurred by the high price of hardware and a desire to ensure customers can run its wares wherever they want to. CEO Rajiv Ramaswami today told The Register Nutanix customers are currently acutely aware of hardware costs and availability, issues that are slowing some purchases of its software. Nutanix is trying to avoid such delays by expanding its hardware compatibility list and supporting external storage devices – moves Ramaswami said mean users can migrate away from VMware without needing to replace hardware. He said Nutanix also continues to optimize the footprint of its stack and pointed to its decision to allow bare metal installations of its Kubernetes Platform as another hardware-minimization effort. The CEO today told The Register Nutanix now sees Arm support as another way to ensure its software platform can run wherever users need it, and believes that adopting the architecture will mean it can run on lower-cost hardware that users won’t balk at buying in the current climate. That’s an advance on Nutanix’s 2024 position on an Arm port as a worthy future consideration, but not an item on its development to-do list. AI support is, understandably, higher on that list and yesterday Nutanix delivered an update to its Enterprise AI suite that added a Model Context Protocol (MCP) gateway – a tool that sits between agents and MCP servers to provide identity management and security services to control the data and other services agents can access. MCP Gateways are quickly becoming standard issue in packaged AI infrastructure stacks, so Nutanix has made sure it’s on par with competitors. Another cost Ramaswami thinks his customers want to control is spending on tokens used in and produced by AI apps, and the company has dogfooded in this field by spending $20 million on its own AI infrastructure – a sum that he expects to recoup in a year “Our software teams have been using AI for coding across the lifecycle,” he told The Register, and initially used tools including Copilot and Claude. “Usage exploded and so did costs,” the CEO admitted. The company has therefore moved to using open weight models and an on-prem cluster. “We are no longer paying on a per-token basis,” Ramaswami said. Some users occasionally use frontier models or external clusters, when necessary, an arrangement the CEO suggested is increasingly common as organizations match models and infrastructure providers to workloads rather than using LLMs for all AI apps. Ramaswami made his remarks about Arm and Nutanix’s on-prem cluster during a media briefing about the company’s Q4 results, which saw it win $757 million revenue, a 16 percent year-over-year jump. Full year revenue grew 12 percent to $2.85 billion. Net income remains modest – just $1.5 million for the full year. “Our fiscal 2026 results demonstrated a good balance of top and bottom line performance with,” said CFO Rukmini Sivaraman. “We remain focused on delivering sustainable growth and improving profitability.” Ramaswami pointed to the fact that Nutanix won 3,000 new customers in the financial year as evidence of growth and said many chose the company’s platform as an alternative to VMware. Nutanix will continue to win customers from VMware for another five years, he predicted. ®

The FBI on Wednesday said it disrupted a botnet and seized two platforms that Chinese-government cyberoperatives used to hack NASA, the US Senate, the Department of Energy, and several other government agencies and critical networks. The Federal Reserve, Department of Justice, Department of Health and Human Services, and the National Institutes of Health were also among those victimized by the two now-seized hacking tools: a vulnerability scanning and exploitation malware named QScan, and an obfuscation network named QTRouter. The FBI says a People’s Republic (PRC) of China-backed group called QTFY created and operated the two platforms, plus botnets of compromised IoT devices. The Bureau says QTFY’s hackers work for a private PRC company called Nanjing Xinjiuwei. “Payments from the PRC's Ministry of State Security (MSS) to Nanjing Xinjiuwei, for example, indicate that the company conducts malicious cyber activities on behalf of the PRC Government,” according to court documents. “QTFY actors include former members of the PRC's People's Liberation Army (PLA), and they use their PLA relationships to obtain contracts and subcontracts supporting offensive cyber operations,” the documents state. How to build a botnet QScan scans and automatically infects thousands of IoT devices worldwide, and then adds them to the QTRouter network of QTFY-controlled devices. The QTRouter botnet – consisting of these compromised IoT devices, plus commercial proxy service devices, and leased virtual private servers – then serves as an obfuscation network, allowing QTFY and other criminals who pay for the service to conceal the origin of their digital intrusion activities, making these communications appear to originate from local computers. On Monday, a US federal court granted seizure warrants for three domains linked to QTFY: qtproxy.xyz, qt-proxy.org, and qt-team.com. All three domains were hardcoded into both the QScan and QTRouter malware, and the court-authorized seizures made both hacking services inoperable, the Justice Department said. Hacking critical networks since 2018 (at least) The hacking services and malware have been in use since at least 2018, and as recently as this year when QTFY infrastructure compromised the US Senate, according to court documents. The FBI investigated an attempted computer intrusion at NASA in August 2019, during which the Chinese government snoops tried to exploit CVE-2019-11510, a critical vulnerability in Ivanti’s Pulse Secure VPN that allowed attackers to learn legitimate users' usernames and passwords, effectively granting them unauthorized access to protected networks. Ivanti patched the flaw in April 2019. As The Register previously reported, China also abused this bug as a zero-day to break into dozens of defense companies, government agencies, and financial organizations in America and abroad. QTFY later exploited this same CVE in 2020 during the COVID-19 pandemic to attack a medical center in Ohio, according to court documents. Other victims in 2019 and 2020 include unnamed financial groups in Michigan and South Korea, and a Missouri insurance agency. The insurance agency attack abused a different vulnerability: CVE-2019-19781, a critical flaw in Citrix VPN products that allowed arbitrary code execution with no account credentials. Years later, in 2024, QTFY hackers also broke into computers at three DOE National Laboratories, NIH, and a US security device manufacturer. “These entities were victims of a zero-day attack against Ivanti Cloud Services Appliance,” the court documents say. The FBI did not respond to The Register’s inquiries, including how many computers QTFY compromised, and whether the crew has ties to any of China’s "Typhoon" groups. PRC goon squad whac-a-mole This latest disruption follows a series of court-ordered seizures intended to hamstring China’s hacking activities over the last few years. In 2025, the FBI removed PlugX surveillance malware from more than 4,000 US computers that had been infected by the PRC-sponsored group Mustang Panda. A year earlier, in 2024, China’s Flax Typhoon burned down their own botnet consisting of hundreds of thousands of infected internet-of-things devices when confronted by the feds. And in late 2023, the FBI disrupted a botnet used by yet another Chinese government attack crew, Volt Typhoon, to attack US and foreign critical infrastructure. In June, however, Lumen’s Black Lotus Labs reported a “significant resurgence” of a botnet linked to Volt Typhoon, with this cluster of injected machines surging to 1,500 compromised routers and IoT devices.®

OpenAI has published its technical report detailing "the Hugging Face incident," the compromise of the eponymous LLM repository by unreleased, ill-supervised AI models. The incident, widely reported, has prompted concern among technical types, the public, and lawmakers about how automated software was able to escape containment and hack an external organization, and about what can be done to prevent similar incidents. OpenAI's explanation addresses what happened, but its call for keeping a closer watch on AI activities won't elicit much enthusiasm. "The incident occurred during cybersecurity evaluations of several OpenAI models, and was primarily driven by a highly capable, internal-only research model comparable in scale to GPT‑5.6 Sol," the company said in a blog post. "The models, operating under reduced safeguards, took actions that were misaligned with the goals of their assigned tasks – they communicated through unauthorized channels, exploited vulnerabilities in shared infrastructure, gained internet access, and accessed third-party systems." The incident became the subject of discussion at the Black Hat security conference and elsewhere. The basic storyline begins with a notionally sandboxed AI agent trying to solve an impossible task in ExploitGym, a security benchmark test. Unable to complete the challenge, the agent explored its environment to find ways to make progress toward its goal. It figured out that it could communicate with other AI agents by using Artifactory, an internal package management system, as a message board. The agents then collaborated with each other to cheat on various ExploitGym tasks. The machine learning models eventually identified a server-side request forgery (SSRF) zero-day vulnerability in Artifactory's code and used that to gain internet access. They subsequently found exposed Hugging Face credentials. The agents then used those credentials to find and chain several security exploits that gave them full code execution on several Hugging Face servers, all in the hope of finding information to solve the ExploitGym task. "Ultimately, agents powered either by the internal-only research model, or by GPT-5.6, executed code on 41 Hugging Face production dataset server workers, obtained root access on at least one production node, accessed Hugging Face production credentials and limited internal data, and downloaded four private Hugging Face code repositories," OpenAI's technical report [PDF] explains. The details are fascinating and troubling, more so because Anthropic's and Meta's models have also acted in ways that would constitute a crime if a human took the same actions. The ChatGPT-maker identified four misalignment patterns that made things go awry: "reward hacking, persistence on seemingly impossible tasks, unauthorized communication, and agents adopting goals from one another." OpenAI has already taken steps to shore up its security, which was relaxed for its model testing. The company says it's trying to improve security and monitoring to mitigate alignment issues like how models cheat, behave when given impossible tasks, and how alignment can be maintained while multiple agents work, including over long-duration tasks. But the central issue is not that machine learning software armed with tools and a task will try every possible option, even ones its programmers failed to consider. The problem, as OpenAI acknowledges, is that people don't watch over their AI agents at all times. "We are taking this incident as a 'warning shot' that today’s model capabilities present the possibility of loss-of-control incidents," the AI biz said. "Companies that build AI systems will need to ensure that their systems always remain under meaningful human control, and that meaningful safeguards constrain their ability to cause harm." Throughout the tech industry, companies like Anthropic, AWS, Google, OpenAI, Microsoft, and Salesforce talk about "autonomous agents." But agents are no longer autonomous under persistent, meaningful human control. ®

If you’re going to build a custom AI accelerator, inference is a good place to start. The clusters are smaller, the chips can be simpler, and at the end of the day they only need to do one thing: serve tokens. OpenAI’s spicy new Jalapeño inference chips are just the latest example. Social media magnate Meta is bucking this trend. Its first proper generative AI accelerator, the MTIA 400 — short for Meta Training and Inference Accelerator — is aimed squarely at LLM training. The Facebook parent is no stranger to custom silicon. But, much like Amazon and Google, its first AI accelerators weren’t built to run AI chatbots or train generative AI models. They were built to serve ads. The MTIA 400, teased earlier this year and detailed at the annual Hot Chips semiconductor development conference this week, will perform some inference duties. Just not of the GenAI variety. Instead, it’ll be saddled with running the ad recommender systems that actually pay Meta’s bills. The combination of LLM training and the deep learning recommender model (DLRM) inference used for serving ads is unusual, as the two have wildly different performance demands. LLM training is enormously compute-intensive, often requiring tens or even hundreds of thousands of accelerators to train models in a reasonable amount of time. DLRM inference, on the other hand, is a predominantly memory-bound job, which means most of the FLOPS that make the chip good at training are going to be left sitting idle. But given how much cash Meta is burning on compute these days, it’s probably not a bad thing for its chips to pull double duty – even better if one of the two use cases is highly profitable. While MTIA may play a bigger role in how Meta serves its generative AI workloads, its in-house silicon probably won’t replace AMD or Nvidia’s GPUs any time soon. The two chip designers' GPUs are currently much better suited to LLM inference and are almost certainly what Meta Superintelligence Labs is using to train frontier models like Muse Spark. Application-specific hardware has always been better suited to well-understood workloads, and Meta’s accelerators are no exception. But enough about market positioning and workloads. Let’s dig into the chip itself. Dissecting the MTIA 400 At first blush, the MTIA looks a lot like either Nvidia’s Rubin GPUs or AMD’s MI355X accelerators. The MTIA 400 is based on a heterogeneous multi-die architecture, which means it is built from a collection of different chips each with their own job. It features two compute dies, two I/O dies, and an SoC die that handles host connectivity and workload orchestration — all pretty standard stuff. Broadcom’s influence here is hard to miss. The chip almost certainly was built using the IP house’s XPU tech, which we explored in detail last year. Building a competent AI accelerator isn’t a trivial endeavor. So, if you’re trying to bring one to market quickly, it makes sense to focus on what makes your chip special and let someone like Broadcom handle the rest. Digging deeper into the MTIA 400’s heart, the compute chiplets are built on a 3nm process tech, presumably from TSMC, and feature a 6x8 grid of processing elements responsible for the bulk of the accelerator’s AI performance. Combined, the two chiplets are capable of outputting 12 petaFLOPS of MXFP4 compute at 1.7 GHz. To put that in perspective, the MTIA 400 is about 20 percent faster than Nvidia’s top-specced Blackwell accelerators at the higher precisions more commonly used to train models, while sucking back roughly the same power. But the parts don’t hold up nearly as well when compared to Nvidia and AMD’s latest chips – the MTIA 400 is between 3x and 3.3x slower than Rubin and the Instinct MI455X, respectively. Meta’s latest accelerator is fed by eight 36 GB HBM3e stacks that deliver 288 GB of memory and about 9.2 TB/s of bandwidth. Again, that’s about 15 percent faster than Nvidia and AMD’s last-gen parts, but with less than half the bandwidth of their new GPUs. This is probably why Meta is positioning the part as a training chip. It has much better and faster options for LLM inference. The MTIA 400 features a pair of I/O chiplets that provide 1.2 TB/s of chip-to-chip bandwidth over RDMA. We don’t know what transport tech Meta is using at this point, but given Broadcom’s involvement, Ethernet would be the obvious choice. Scaling up MTIA 400’s physical resemblance to Nvidia Rubin and AMD MI350-series parts also extends to the system design. Each compute blade features four MTIA 400 accelerators connected via a PCIe switch to an x86 CPU and a scale-out NIC. A single rack is equipped with 18 compute blades and eight switch blades for a total of 72 accelerators in a single unified domain. Meta shared a pic of the rack earlier this year, so while the compute blade is highly reminiscent of AMD's Helios blades, it's not using the same double wide OCP racks. Meta hasn't shown how large a cluster its chips can support just yet — the slides simply state "multi-thousand accelerator scaling" — but even at rack scale the systems should be quite competent relative to Nvidia’s GB200 and GB300 rack systems for training. More on the way While the MTIA 400 is aimed primarily at LLM training, Meta is already working on an inference-optimized version of the chip. Disclosed back in March, the MTIA 450 will double the chip’s memory bandwidth, presumably by swapping HBM3e for much faster HBM4. That part is expected to enter production next year, and should be extremely well suited to running those LLM-based recommender models Zuck and crew have been talking about the past few quarters. Meanwhile, the MTIA 500, which is also slated for 2027 release (probably in H2 if we had to guess), will increase memory bandwidth by another 50 percent — likely using 4 additional HBM4 stacks — and double the number of compute chiplets. ®

Wednesday's release of Kubernetes v1.37 shows how the team behind the popular open source resource orchestrator is prioritizing production readiness, pruning outdated features even while it adds new ones. Overall, the release, nicknamed Garhwal after India’s lush Himalayan region of Uttarakhand, includes 67 changes: 16 features that have graduated to stable, production-ready status, along with 23 features being tested as beta, 27 new alpha features, and 1 deprecation. The Cloud Native Computing Foundation (CNCF), which stewards (but does not manage) Kubernetes among more than 220 cloud native projects, has to be mindful of its user base, which may be growing out of its infatuation with the bleeding edge. Approximately 82 percent of container users run Kubernetes in production, up from 66 percent in 2023, according to a CNCF January survey. Kubernetes has also become a big driver of AI workloads, with CNCF reporting that 66 percent of organizations hosting generative AI models use Kubernetes to manage some or all of their inference workloads. Metrics ratified One of the first rules of maintaining production software is to keep the betas moving along. This release finally bumps the Metrics API to general availability, after being in beta status for the last nine years of K8s’ 12 years of public existence. In 2020, the maintainers of the Kubernetes resource orchestration platform set down the law: Any new API-based feature must not linger in beta status for more than three Kubernetes minor releases (about 12 months in the current release cadence). After three releases, the beta must move to general availability, or be replaced by another attempt at offering the same functionality. The thinking was that forcing a technology off beta status would inspire the maintainers to finish the work. If beta APIs stick around too long, temporary designations (“v1beta1”) might become embedded in automated systems and require changes later. Their temporary status will also trip up security software, which frowns on the production use of “pre-production” software. The CNCF declined to comment on why this particular API beta got lodged in the slipstream of new releases. A CNCF spokesperson noted that the organization does not interfere in the decisions of the projects under its purview. "Despite the Beta label, the metrics.k8s.io API has been widely used in production for years," commented Kubernetes 1.37 Release Lead Dipesh Rawat in a statement to The Register. Like the millionaires who found a loophole in New York City’s pied-à-terre tax, the Metrics API skirted out of these mandates perhaps due to a technicality, in that it isn’t managed by the API server itself, but rather is an aggregated API served out-of-tree by a separate metrics server. Also, the Instrumentation Special Interest Group was busy refining the underlying performance of the metrics serving engine after replacing the previous version, Heapster. The API provides CPU and memory usage for pods and nodes. These numbers can be used not only for compiling reports but also as input for other features, such as pod autoscaling. It’s stable, everybody uses it, and now it’s official. "No functional changes [are] expected," the release notes state. "Its move into v1 in this release formally recognizes that the API has already proven itself in production, and this is simply a mature project addressing a long-standing permanent Beta status." Unloading the baggage Other technical debt has been addressed in this release as well. The software’s original built-in DNS server, kube-dns, is being retired in favor of CoreDNS, the default cluster DNS add-on since Kubernetes v1.13. Those running kube-dns must update their clusters before Kubernetes 1.40, or they will stop working. Also, word has been handed down that kube-proxy, which handles the routing amongst the pods, will no longer support the Linux IPVS (IP Virtual Servers), beginning with Kubernetes 1.43. This load balancer software was introduced to solve the bottlenecks caused by iptables-based routing. Well, IPVS didn’t work out, and nftables has now been decreed as the way forward. Also on the way out is the first version of Linux control groups (cgroup v1), which Kubernetes uses to group server processes into hierarchies for easier management. The updated cgroup v2 offers more granular oversight of container resources. Since v1.35, Kubernetes won’t initialize nodes that still rely on cgroup v1 (though there is a temporary work-around that will be supported for a limited time). KYAML flow The v1.37 development cycle ran for 15 weeks, with contributions from over 1,700 individuals and 212 different companies. Of the new features, many ease deployment, which is always appreciated in production environments. Brand new is pod-level checkpoint and restore (Kubernetes Enhancement Proposal 5823), introduced as Alpha. Entire pods can be checkpointed, capturing the runtime state of their containers so they can later be restored from that checkpoint. It’s up to the container runtimes to manage this. Autoscaling is getting a boost with HPAScaleToZero (KEP #2021), a beta feature now enabled by default that allows the Horizontal Pod Autoscaler to scale workloads down to zero pods when idle, using object or external metrics as indicators that they can then be restored upon demand. The resilient watch cache initialization (KEP #4568) manages the thundering herd problem during API server startup and recovery. Previously, an uninitialized watch cache caused the kube-apiserver to either hang on long-running watch requests or delegate heavy LIST queries directly to etcd, overwhelming the backend database. This fix essentially acts as a bouncer, rejecting watchcache-dependent read requests with retryable errors until the cache is fully initialized. Lastly, there's a bit of news that will please anyone fed up with working with Kubernetes’ occasionally ambiguous YAML-based configuration. KYAML (KEP 5295) is a new subset of YAML that restricts formatting to Kubernetes norms, making manifests parseable by other tools. Mostly, it’s super-strict about string quoting and indifferent to whitespace. It’s stable, you don’t have to make any changes to existing manifests, it doesn’t replace YAML, and is already understood by the kubectl command line. “It is less of a migration and more of a better habit,” explained computer science student Kashish Verma in a recent Kubernetes blog item. It’s just better to be consistent across large repositories and team projects. ®

A mischief-maker in a Darth Vader suit recently endorsed Flock at a city council meeting, and it has not helped the company's fortunes. Around the US, citizens have been protesting and sometimes vandalizing the company's network of surveillance cameras. And the advocacy has caused some governments to change course. Government officials, commonly hesitant to second-guess law enforcement practices, have become concerned enough to question these mass surveillance practices. US Senator Josh Hawley (R-MO) on Wednesday wrote to Flock CEO Garrett Langley to announce an investigation of "Flock Safety's collection, retention, and dissemination of data gathered by its artificial intelligence cameras." Despite "Darth Vader's" comments last week at a San Diego City Council meeting about the need to "continue utilizing Flock technologies so that we can follow and surveil the rebel scum," a growing number of local governments in the US have been cancelling or choosing not to renew their contracts with Flock and other makers of automated license plate readers (ALPRs). The Institute for Justice (IJ), a Virginia-based legal advocacy non-profit, on Wednesday published a new database that tracks the termination of local government ALPR contracts. As of Wednesday, the database documents 93 local governments that canceled ALPR contracts. The IJ, which is challenging ALPR use on Fourth Amendment grounds in Norfolk, Virginia and San Jose, California, previously published a database that documents police abuse of ALPRs, a topic addressed in Hawley's letter. Incidents include romantic stalking, wrongful traffic stops and detentions, and other forms of misconduct. "I think what's happened is that the American public has woken up to the mass surveillance that quietly arose around them," said Robert Frommer, senior attorney at IJ, in an interview with The Register. "And they don't like it. "It does seem to me that the rate of cancellations, the rate of public protest has increased over the past few months. I think that's both due to more news attention being paid to the real life abuses of this technology, as well as the Supreme Court's recent decision in the geofence case," Frommer said. In that case, SCOTUS ruled that cops need a warrant to grab location data. Frommer said that he believes the IJ's efforts to document ALPRs and associated abuses have helped focus the public's concern and have illustrated the problems with warrantless mass surveillance. ALPRs, from vendors like Flock, Motorola Solutions, and PlateSmart, among others, capture license plates whenever cars drive by, along with geodata and potentially images of people. This data can be combined with other information, effectively putting all passersby under constant surveillance without any warrant. "I think what we're seeing now is the reaction to that," said Frommer, who explained that the problem with Flock cameras is that they're accessible to anyone with login credentials, without any authorization from a supervising authority. "You don't need to go to a judge and get a warrant," Frommer explained. "You don't need anything. And when you basically give that search power to the bored officer in his cop car in the middle of a shift, it inevitably happens that they start using it and start misusing it, whether to look people up personally or whether to investigate people at political protests." These sorts of issues, Frommer said, have made people question whether we should live in a panopticon. The IJ's litigation and its Plate Privacy Project, said Frommer, follow from the belief that there are solutions, whether they come from the courts or lawmakers. Pointing to Hawley's letter, he said that this is not just a state or local issue, but has become a federal concern. "I think what the din that you've heard over the past couple of months shows that there absolutely needs to be a change, a reckoning," he said. "And that's what we're pushing for in our litigation. We're pushing for a warrant requirement. You want to go look up somebody's past travels and see where they have been? Go to a judge, explain it, and get a warrant." Flock did not respond to a request for comment. ®

The US government disclosed that crims targeted more than 100 internet-exposed water systems during July cyberattacks. That's the first time the feds have put a number on the digital intrusions, but they have yet to attribute the campaign, widely suspected to be linked to Iran, to a particular group. “In July 2026, CISA observed malicious cyber activity targeting over 100 internet-exposed systems in the Water and Wastewater Systems (WWS) Sector, commonly via programmable logic controllers (PLCs) connected directly to a cellular modem,” America’s lead cyber-defense agency said, adding that connecting PLCs directly to the internet “can create significant security risks.” Suspected Iranian attackers targeted water and wastewater facilities across at least a dozen states in July, including internet-exposed PLCs. While neither federal nor state officials have identified all 12, we know that the cyberattacks occurred at mostly small, rural utilities in Minnesota, Michigan, Georgia, South Dakota, and New Jersey. “This is very serious. What stands out isn't any single incident. It's the scale,” Matt Hartman, chief strategy officer at the Merlin Group and CISA’s former acting head of cyber, told The Register. “More than 100 water systems with internet-exposed assets were hit in a single month, which points to a systemic vulnerability across the sector, not a run of isolated, unlucky targets,” Hartman said. “Much of this infrastructure runs on operational technology that was built for closed, physical environments. It was never designed with the assumption that it would be reachable from the open internet.” John Gallagher, VP at Viakoo, an OT and IoT cybersecurity provider, told us that while 100 systems represent a small fraction - only about 0.5 percent - of water utilities in the US, the “real threat is that these are test runs for a larger-scale attack.” While the 100-plus water incidents occurred in July, just last week five US federal agencies warned that attackers are using AI-generated exploitation scripts to break into internet-exposed Siemens S7 Series PLCs at water, manufacturing, energy, and other critical facilities. “This appears to be a continuation of the same suite of activity we suspect is affiliated with Iran targeting PLCs,” Halcyon Ransomware Research Center SVP Cynthia Kaiser told The Register a week ago. “Iran-affiliated actors and adversaries are actively targeting a wide swath of operational technology because these PLCs underpin essential health, safety, and critical infrastructure across society,” Kaiser, a former FBI cyber division deputy assistant director, added. While third-party analysts have largely blamed Iran for the intrusions, the federal government has not attributed the attacks to anyone. “Attribution in cyber incidents is inherently difficult and often takes time. Adversaries deliberately obscure their infrastructure, reuse tools and techniques, and route activity through compromised systems, so the government needs to be diligent before publicly assigning responsibility,” Hartman said. “In this case, CISA has done the most important thing: quickly getting actionable information into the hands of water-sector operators so they can defend their systems,” he added. “From a defender’s perspective, the ‘who’ matters less in the immediate term than understanding how the attacks are occurring and taking steps to stop them.” In its advisory, CISA recommended organizations disconnect PLCs from the internet and ensure any remote access goes through a VPN or gateway device rather than connecting directly to the PLC. The cyber-defense agency also advised owner-operators to enable password protection (we suggest multi-factor authentication) and change any default passwords. Also: ensure that allowlist IPs only allow remote access from known engineering laptops or other critical OT assets.®

GitHub Actions stumbled again on Wednesday, days after the code host renewed its promises to improve reliability. Wednesday’s problem, as has so often been the case, hit Actions, GitHub’s CI/CD platform for automating software builds, tests, and deployments. According to the incident report GitHub put out for the disruption, things started going south at 1511 UTC. The source code host identified an issue with a database primary and failed over to a replica, but said the move "did not fully mitigate" the degradation. GitHub then throttled inbound traffic while investigating upstream Vitess issues before gradually restoring traffic. By 1800 UTC, it said Actions was operating as expected and inbound queues had recovered. The latest disruption isn’t particularly reassuring given that GitHub claimed last week that it’s now serving double the commits it was dealing with in April, which wasn’t exactly a good month for GitHub either. No month this year has been great at the ‘Hub, really. The history archive on GitHub’s status website indicates there were 26 issues with the platform in April. There were 23 in May and June, 26 in July, and there’ve been 23 so far in August with just under a week left to go. Whether this month can top March, with 32 incidents, or February’s 37, remains to be seen. There were 25 in January, too, meaning GitHub has suffered at least 23 reliability issues every month this year. GitHub Actions is arguably a central part of the platform for many developers using CI/CD workflows and other forms of automation, and it has been among the services hardest hit by GitHub’s ongoing reliability problems. As everyone who uses a software-as-a-service product knows, uptime is a key element in measuring reliability, and Actions isn’t exactly at triple nines right now - as of Wednesday, GitHub’s uptime page for Actions shows it at just 98.13 percent for August - nearly in danger of slipping into 97 percent reliability territory. That’s a bad place to be when you’re supposedly dealing with 2.9 billion commits, 24 million new repos, and 130 million merged pull requests a month, as GitHub claims it is. While GitHub’s issues this year have been many and frequently reported on here at The Register, August has been a particularly bad month for the operation. August 17 saw GitHub suffer from a nearly eight-hour outage that hit multiple services, with Issues, Pull Requests, APIs, Actions, and Copilot all producing elevated errors and hamstringing customers’ ability to do work. GitHub has pointed the finger at AI for many of its issues, blaming bots and agents for skyrocketing usage it hasn’t been able to cope with. The same went for that August 17 outage, with GitHub CTO Vladimir Fedorov issuing a mea culpa for the incident, saying that his operation had let users down and promising, just like he did back in April when GitHub admitted it was having issues, to scale enough to support its growing user base, human or otherwise. “We'll earn your trust through the scaling and reliability of the platform,” Fedorov wrote in last week's postmortem of the August 17 outage. Six days after he promised to fix things, here we are with reliability slipping and the issue count growing. GitHub didn’t respond to questions for this story. ®

MNT is launching the Station, a miniature desktop case for motherboards that come from its Reform open-source laptop. If you upgraded the laptop, you can now give your old motherboard a new life. The MNT Station is very pricey for a case at $299. That’s not a complete computer – it is just the case for an existing motherboard from its Reform laptop line, which itself goes for $429. Then again, MNT’s Reform laptops are very much not ordinary laptops. The company describes the original MNT Reform as “the ultimate open hardware laptop”. It’s a modular, repairable, replaceable and upgradable machine, complete with extensive documentation. The first model was launched in 2019 and The Register mentioned it when discussing the optional OLKB ortholinear keyboard in 2021. Today, MNT offers several Arm processor options across the Reform family, including a Quasar module based on Qualcomm’s QCS6490 via a Thundercomm C6490P, with 8 GB of RAM and 128 GB of flash, and an RCORE module based on Rockchip’s RK3588, with 16 GB of RAM and 256 GB of flash. In 2023, the full-size laptop was joined by the much smaller MNT Pocket Reform with a seven-inch screen. The company is currently working on the Reform Next, a slimmer, lighter, faster model that it successfully crowdfunded early last year and which should ship very soon. The Reform starts at $1,299, the Pocket Reform costs from $1,200 depending on spec, and the Reform Next is currently planned to be $1,249. These are not bargain-basement laptops. If you want fully open, documented, and maintainable Arm-based hardware, it’s going to cost you. That may help explain the appeal of MNT’s new desktop case: if you upgrade your Reform laptop with a newer, faster motherboard, you can install the old board in the MNT Station and continue using it. This seems to be very much in the model of the Framework Desktop Case, whose announcement The Reg covered early last year. The new Station case is CNC-milled from solid aluminum. It accepts Reform motherboards version 2.0, 2.5, and 3.0. You can add various optional extras, including a daughterboard to provide two more USB-C ports, a Wi-Fi module with external antennas, battery packs, and more. It’s not just a simple plastic enclosure, but even so, it’s quite a lot for just a case. You could buy a used or surplus mini PC for that kind of money, and sites such as LowCostMiniPCs will help you find them – complete with options to look in the US, UK, or Germany (for the Eurozone). At the time of writing, the site has units from $65, so you could build a redundant cluster of the things. Sadly, fully open development costs. ®

Meta announced on Wednesday that it has offered to pay about $18 billion to settle the lawsuit against it brought by 52 attorneys general from US states, territories, and the District of Columbia. But part of that payout and some of its concessions are being conditioned on whether YouTube and TikTok will adopt its new teen-targeted usage restrictions, too. The case, led by the California Attorney General, was first filed in 2023, and accused Zuckercorp of not only illegally collecting data on its underage users, but also designing its platforms in such a way that it put young users at risk and drove excessive use. All the while, the suit alleged, Meta lied about those risks to users, their families, and the public in violation of state and federal laws. Meta admitted no wrongdoing in the proposed settlement, naturally, though it’s still going to make considerable changes to ensure the not-risks it didn’t inflict on children won’t continue to cause them any nonexistent mental health harms on its pair of totally innocent social media platforms. And nothing happens unless a judge approves. What Meta is offering If the settlement is approved by the judge in the case, Meta will implement changes to accounts held by kids under 18, including a two-hour daily time limit that’s shared across Facebook and Instagram, default blocking of its apps between midnight and 6AM, muted notifications during school hours (8AM - 3PM), prompts every 15 minutes encouraging them to stop mindlessly scrolling, an option to disable autoplay on videos, hiding likes and reactions by default, and a block on “extreme makeup filters.” Most interestingly, Meta said it’ll also give teens the option to choose a non-algorithmic feed that “isn’t personalized by our recommendation systems … as their default” in order to weed out the garbage that Meta algorithms tend to surface to users. No word on whether that option will be made available to adult users too – we asked, but didn’t hear back. Direct messages on Facebook and Instagram, we note, are exempt from night mode, time limits, and school notification pauses, as Meta wants “to allow teens to stay connected with friends and family.” Meta also said it’ll dole out the roughly $18 billion settlement across the states and territories party to the deal, with funds ultimately used at those governments’ discretion, though with the intention it go toward addressing mental health harms to kids due to social media. The payment will be distributed in annual installments over 10 years, making this a drop in the bucket for Zuck and friends. For reference, Meta’s Q2 2026 revenue, reported at the end of July, was $60.8 billion. “Meta has agreed to make massive transformations that will reduce the risk of harm from its platforms — and will do it within months,” California AG Rob Bonta said of the deal. “I am proud to deliver this settlement that addresses the concerns at the core of our lawsuit and institutes real change, real transparency, and real enforceable protections for children on Facebook and Instagram — right now, no more waiting.” Meta wants its peers to share the blame Meta did its damndest in its settlement announcement to deflect blame and shift responsibility, tossing YouTube and TikTok under the bus and conditioning a good deal of its concessions on whether the pair willingly adopt the moves it was forced into in court. “Teens move fluidly between dozens of apps a day,” Meta said. “For meaningful progress to happen, we urge TikTok and YouTube to join us and state attorneys general in adopting this new standard, to ensure teens use social media in a healthy and responsible way.” To put the pressure on the pair, Meta said that it’s going to withhold 30 percent of its settlement payment (around $5.3 billion, per the company) until both YouTube and TikTok agree to implement similar daily limits, night mode, and age assurance measures to the ones Meta has agreed to. Meta also expects the pair to pay an amount matching that 30 percent withholding. If those two conditions aren’t met, Meta’s bill for the settlement will decrease to around $12.7 billion. Along with that demand, Meta also said that the time commitment for its daily time limit and night mode measures would be strengthened if YouTube and TikTok sign on, too. While most of the terms in the settlement are required to be kept in place for 10 years, Meta’s only on the hook to retain its night mode and time limits for teens for five years. If YouTube and TikTok agree to its terms, Meta will not only agree to keep those measures in place for 10 years, but also strengthen them, reducing the daily limit for minors to just one hour per app and expanding night mode to 10PM - 7AM. “This framework will only work if all our peers join us,” Meta legal chief C.J. Mahoney said. “We therefore call on our industry peers, TikTok and YouTube, to implement this new framework, right away.” We reached out to YouTube and TikTok for comment, but didn’t hear back from either. Speaking to Meta’s agreement, Forrester principal analyst Kate Winick described the settlement as the largest penalty ever seen for a social media platform, while qualifying Meta’s concessions as largely in line with what it’s agreed to in other markets that’ve ruled against it. Winick agreed that the social media problem facing teens is bigger than Meta and said its fate is irrevocably linked to those of TikTok and YouTube, but called the company out for deflecting responsibility for much of its concessions to others. “The most pernicious problems are still being displaced onto others: preventing young users from signing up for standard accounts is a problem Meta says it needs the app stores to solve, and ensuring they use the opt-in features is a burden placed on parents,” Winick said in an email. “While limits like these are helpful in [reducing] the mental burden on teenagers, the challenge of fundamentally changing what content they are exposed to remains as big as ever.” ®

The rising cost of contract prices for memory chips could see DRAM and NAND flash account for more than two-thirds of the capital expenditure of cloud service providers (CSPs) by next year. Taiwan-based market watcher TrendForce forecasts that total capex by cloud operators will almost double this year - up 98 percent - and grow a further 50 percent in 2027. That rapid rise is driven partly by growing investment in new infrastructure, but also by soaring memory bills. The upshot is that TrendForce estimates DRAM and NAND flash will account for 47 percent of the total spend on hardware this year, jumping to 68 percent of it in 2027. The analyst doesn’t put revenue figures to its predictions, but reckons server DRAM components will have grown by 270 percent year-on-year by the end of 2026, with enterprise SSD prices up 235 percent over the same period. As Reg readers know, growing demand for high-performance memory from big customers is prompting DRAM and NAND chip suppliers to prioritize production capacity for server applications. TrendForce estimates that HBM - used in GPUs - and RDIMMs used in servers will together account for more than half of DRAM bit supply this year. The effects are already visible in PC prices climbing by double-digit percentages due to a shortage of standard memory parts, and overall PC shipments falling by 5 percent in response. The smartphone market has been hit in a similar fashion, with shipments expected to fall by 15 percent this year. For enterprise customers, the knock-on effects are becoming hard to ignore. As memory eats an ever-larger share of CSP capex, that cost is unlikely to stay with the cloud providers and will be passed down. Earlier this month, Euro operator OVHcloud warned it was planning to inflate its charges by up to 87 percent to cover rising memory costs in a preview of what other providers may follow with. As if that isn’t bad enough, TrendForce says that elevated memory costs provide server and AI chip suppliers such as Nvidia with greater justification for raising their own prices. Sure enough, recent news reports claim Nvidia plans to charge 15 percent more for products containing its Vera Rubin and Grace Blackwell chips. To counter this cloud operators may need to increase capital expenditure even further to maintain their targeted level of infrastructure expansion - which means more price increases for users. ®

Medical device maker Boston Scientific’s global operations have been disrupted by an ongoing cyberattack, the company disclosed on Wednesday. According to an SEC filing, the “cybersecurity incident” affecting its IT systems started on Tuesday, and resulted in a “global disruption to the company’s operations.” Upon detecting the digital intruders, the company began an investigation with third-party infosec experts who are working to contain the threat, the filing says. The medtech firm did not immediately respond to The Register’s inquiries, including if this was a ransomware infection and what data, if any, the criminals stole after breaking into the IT systems. “The incident has caused, and is expected to continue to cause, disruptions and limitations of access to certain of the company’s information systems and business applications that support aspects of the Company’s operations, including the ability to process and ship customer orders,” Boston Scientific reported to federal regulators. The company doesn’t have a timeline for full restoration, and noted the “full scope, nature and impacts, including operational and financial impacts” of the cyberattack remain unknown. The news did, however, send Boston Scientific shares spiraling down more than 4% on Wednesday morning. At press time, none of the usual suspects had claimed responsibility for the attack. In recent months, both ransomware/extortion gangs and government-backed hackers have disrupted other medtech firms’ operations and stolen sensitive data. In March, Stryker was hit by a cyber crew with ties to Iran's intelligence agency, causing a global network outage at the company. A month later, medical-device maker Medtronic disclosed a cyberattack in a filing with federal regulators. Notorious data-theft-and-extortion group ShinyHunters claimed to be behind this intrusion, and the company in July warned patients that their names, contact details, dates of birth, Social Security numbers, and health information was stolen in the breach. ®

Workwear and fashion retailer Carhartt’s data breach was about half as bad as ShinyHunters claimed, according to Troy Hunt, who reviewed the data dump before uploading it to his Have I Been Pwned website. Hunt’s HIBP service reported 12.9 million individuals affected by the alleged breach, around half of what ShinyHunters purported when it claimed to leak the company’s data earlier this month. ShinyHunters dumped what it said was 50GB worth of Carhartt’s data on August 13 after the retailer hired what the criminals called “a very unskilled and incompetent negotiator” to haggle the crooks down from their $3.3 million extortion demand. According to Hunt, ShinyHunters data was injected with millions of lines of synthetic data, substantially padding out the number of affected individuals. “You're not going to believe this, but turns out you can't always take criminals at their word,” said Hunt, before detailing the investigative process that goes into calculating the scale of breaches that make it to the HIBP site. Hunt usually starts with HIBP's open-source email address extractor, which spat out nearly 25 million addresses, before running it through OpenClaw to analyze the contents further and sift through the mass of information, looking for anomalies. For a retailer, the AI thought that the millions of .edu and .org email domains looked off - like signs of TPC-DS synthetic data injection. Examples included michael.ware@c.edu and michelle.larue@lkvb06fkzsjv.org. The first and last names look real, but a common finding among TPC-DS-generated data is that it will use completely random strings for domains. A manual look at the data points with which these email addresses were associated further revealed these "individuals" were located in countries, such as Benin, which don’t represent dominant Carhartt markets. Further, the AI found more customers registered in Montenegro than in the US, where Carhartt is headquartered. It also found a suspiciously large proportion of customers with birth dates set in the early 1900s – unlikely given the company’s clientele is more hipster-oriented than than blue-collar nowadays. Hacking away at all the clearly bogus data, OpenClaw dropped the estimated number of genuine individuals among the data trove from 24.8 million to 13.6 million. Hunt carried on eliminating suspicious findings he noticed, such as Microsoft 365 duplicate email addresses and addresses marked for deactivation, as well as prompting OpenClaw to continue doing the same. He finally arrived at 12,933,413 accounts believed to be genuine among the ShinyHunters-leaked dataset. That’s the number that made it to the HIBP platform, which states that 83 percent of these were already gathered up in previous breaches. The moral of the story is, as Hunt says, to “take headline numbers with a grain of salt unless you're confident in the processes of those making the claims." And until trustworthy sources do the arduous work, stop treating the word of cybercriminals as gospel. The real data contains names, email addresses, phone numbers, and physical addresses. Carhartt did not respond to our request for comment on Hunt’s findings. The company is yet to comment on the breach anywhere publicly. ®

For users of vSphere 8, October 11, 2027 looms large on the calendar. That is the day general support for vSphere 8 ends, along with bug fixes and security patches, and the day traditional perpetual licensing for the technology also expires. From then on, customers must upgrade to VMware Cloud Foundation (VCF) 9 or VMware vSphere Foundation (VVF) 9, and shift to a bundled subscription model widely seen as much more expensive and geared toward larger installations. VMware launched vSphere 8 in October 2022, ahead of its takeover by Broadcom in 2023. Broadcom made no secret of its intention to reorganize the VMware product lineup and licensing regime, and train its sights on the biggest, most lucrative users. vSphere 8 offered customers breathing space in which to watch Broadcom's plans play out and plan their own next move. Aside from the higher licensing costs, VCF 9 offers a radically different proposition that includes much more onerous audit requirements. Technically, VCF 9 is a take-it-or-leave-it bundle of products that gives customers less flexibility than its predecessor, which came in Standard and Enterprise Plus editions. According to Harsha Kotikela, senior director for product and solutions marketing at Nutanix, calling this an upgrade doesn't do justice to what shifting to VCF 9 entails. VCF is bundled much more tightly with both vSAN and NSX, the latter of which was an add-on option in vSphere 8. Broadcom's whole vision, he says, is heavily geared toward private clouds. That might work for some businesses but it will go against the grain for many enterprise organizations pursuing hybrid cloud or more distributed infrastructure strategies. Add all this together, argues Kotikela, and what sounds like a routine upgrade starts to look much more like a major architectural migration to a more expensive, less flexible platform by an immovable end date. Customers who want to stay within the VMware ecosystem have an alternative in the shape of VMware vSphere Foundation, which Broadcom pitches as delivering "enterprise-class hyperconverged infrastructure." The platform is widely perceived as a cut-down alternative to VCF, with a narrower range of features and reduced storage per core, along with many of the same drawbacks around unwanted bundled features and costs. It does not offer the broader Kubernetes and AI support found in VCF. That means most customers will still face a decision at some point about how to support these two trends. As Kotikela says: "If an organization decides, 'We can't afford VCF 9 right now, so we're gonna go with this lower cost bundle,' then two years down the road, they're going to be saying, 'Oh, I do want to use AI', then they're going to have to migrate again. That's a hidden danger." The vSphere 8 cut-off may be over a year away, but many businesses are already considering their options, and plenty have identified alternatives and voted with their feet. Gartner research shows that 35 percent of respondents have already migrated or are in the process of migrating their entire portfolios away from VMware. Alternatives to the VMware stack are developing at an accelerated pace, the analyst house added, "with a diverse ecosystem of enterprise-ready solutions projected to reach full feature-parity and operational maturity within the next 24 months." Gartner pointed out that "Nutanix is expanding its ecosystem support, including integration with external third-party storage vendors." Defining your future platform The October 2027 cut-off should be about more than simply finding a replacement for VMware infrastructure assembled over years, even decades. Organizations need to think well beyond a straight lift-and-shift of their existing architectures, Gartner added, and consider broader modernization strategies such as hybrid architectures, open source, and container management. The future of virtualization, Gartner said, is one where "VMs and containers are managed through a single control plane. Hypervisors integrate natively with Kubernetes." That single control plane would allow infrastructure to be provisioned and governed through "declarative automation." That is the Nutanix Cloud Platform architecture, Kotikela says, with workflows, lifecycles, and operations managed in one place. "It's one platform, so, you'll be able to manage both VMs and containers. You have one platform to manage your different clouds, plus on prem." More recently, Nutanix Agent Gateway has added AI agent observability and token management to the platform. The alternative, Kotikela said, is a migration strategy that potentially leaves tech teams juggling separate silos and separate lifecycles for private cloud and public cloud. That is challenging enough during the migration itself, but from a day-two perspective, he says, "It becomes increasingly difficult." Who switched early? Nutanix is keen to highlight businesses that have already shed aging VMware infrastructure to modernize their technology stacks and harvest tangible benefits. Western Union migrated at least 900 applications across a 3,900-core server fleet, while South Korea’s Everland theme park also moved from Broadcom to Nutanix, citing the ease of the migration path. Looming over all of this is the rush toward implementing AI, particularly the promise of agentic AI. As strategic leaders in organizations, Kotikela says, "CIOs are also thinking about where I should drive my organization three years from now, right?" It's increasingly clear that hybrid cloud is going to be the key underlying model for AI, Kotikela argues, and "You have to have a seamless way of managing this distributed environment." Kotikela adds, "The foundation for AI, obviously, is Kubernetes and containers. So, you have to modernize from an application perspective. But it also doesn't make sense that you modernize every legacy application." For that reason, Kotikela said, it's vital that IT should be able to "manage simultaneously both your VM-based applications as well as container-based applications or new applications like AI." That includes the ability to accommodate and manage both CPU- and GPU-based architectures, whether on-premises, in colos, or in the cloud. "That's something that Nutanix provides at the top level," Kotikela says. In the coming years, as businesses move beyond experiments and pilots, which are often on-premises, they will begin deploying inference in production. "It should be able to run anywhere in this distributed environment." Given the scale of the challenge around transitioning to AI, imposing a major migration effort that locks organizations into a single stack, both in AI and other domains, is hard to justify. That is especially true with technology on such a rapid innovation trajectory. CIOs cannot predict or control the innovations in software or silicon they might want to adopt in two or three years' time, Kotikela says. "So, you want to pick a platform that allows you that flexibility to mix and match stuff. So that's what's in your hands, that's what you can control." Sponsored by Nutanix.

Microsoft is giving customers who are still clinging to its doomed legacy Whiteboard technology a lifeline, and has pushed the deletion date to October 16, 2026. The migration tooling for "legacy enterprise whiteboards" will be retired on September 25, 2026, and the standalone Microsoft Whiteboard app will be deprecated on November 30, 2026. To be clear, Microsoft Whiteboard as a concept is not going away - this is more about migration for Azure-based files. The original dates were August 22 for the migration tooling, September 5 for the permanent deletion, and September 14 for the app deprecation. Therefore, customers have been granted a little more time to make the transition to OneDrive-backed whiteboards. Microsoft did not respond to a query about why it had moved the milestones. The lateness of the announcement suggests it either ran into technical problems or that enough customers were still on the legacy platform that Redmond uttered the computing equivalent of FINE! Have another month! New whiteboard files for commercial customers have been in OneDrive rather than Azure since 2022. However, legacy whiteboards could still be stored in Azure and, if not migrated to OneDrive, will be deleted on October 16. Migration happens automatically when a whiteboard is opened in Whiteboard on the web, Teams desktop, Teams web, or using the Whiteboard Windows app (through to September 22, 2026). If migration doesn't happen, then users have until October 16 before somebody at Microsoft hits the delete key. However, only owners of Azure Whiteboards can trigger migration. Users on vacation, or who have left the organization, could also present a headache for administrators, who might need to transfer ownership to active users. In Microsoft's original announcement, it promised a set of PowerShell cmdlets to ease the process. The Microsoft Whiteboard app became generally available in 2018. It was a useful collaborative tool for Microsoft 365 subscribers (or could be used standalone with a Microsoft account, although these days organizations are the target). It is very much a relic of a past era at Redmond, when touch-based experiences were all the rage. Today, the company would likely prefer users to bark instructions at an AI assistant. According to Microsoft, "Moving these files to OneDrive improves security, accessibility, revision history, eDiscovery, and access to current and future Whiteboard features." ®