<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>ihos</title><link>https://ihos.eu/</link><description>Infrastructure news, explained.</description><language>en</language><atom:link href="https://ihos.eu/rss.xml" rel="self" type="application/rss+xml"/><item><title>US hyperscale data centers used up to 2.2% of the grid in a year</title><link>https://ihos.eu/hyperscale-data-centers-us-power-use/</link><guid isPermaLink="true">https://ihos.eu/hyperscale-data-centers-us-power-use/</guid><pubDate>Tue, 17 Nov 2026 09:40:00 GMT</pubDate><category>Pricing &amp; Market</category><description>A new study counts 403 large US data centers: 68–99 TWh of electricity, more than half of it from fossil fuels. The real total is probably higher.</description><content:encoded><![CDATA[<p>A new study published in Nature Sustainability puts numbers on something the industry has found hard to measure: how much electricity the largest US data centers actually use. The researchers started from a commercial database of 675 facilities and kept the 403 for which they could verify a reported electrical capacity and a building large enough for the expected cooling and power equipment. The average was about 40 megawatts.</p>
<p>Because they could not see real load, they used a range of utilization, about 48% of capacity for conventional sites and about 70% for AI training. That gives 68 to 99 terawatt-hours between May 2024 and April 2025, or 1.5 to 2.2% of total US electricity use. Ars Technica notes this is about what Wisconsin or Indiana use. Ars Technica puts the associated emissions at 21 to 31 million metric tons of CO2.</p>
<p>The sites are not in unusually clean grid regions: their mix was only slightly cleaner than the US average. About 54% of the power came from fossil fuel plants, mainly natural gas, with 25% from renewables and 21% from nuclear.</p>
<p>The study does not count behind-the-meter generation. Cleanview, a market intelligence company, tracks 116 current or planned data centers with on-site generation, and the generators used are much less efficient than utility power plants. If the buildout continues, the authors say, repeating the analysis in later years could show how much of the growth is AI.</p>
]]></content:encoded></item><item><title>New US rule targets cloud rental as a way around AI chip limits</title><link>https://ihos.eu/us-rule-remote-access-ai-servers/</link><guid isPermaLink="true">https://ihos.eu/us-rule-remote-access-ai-servers/</guid><pubDate>Mon, 16 Nov 2026 21:50:00 GMT</pubDate><category>AI &amp; GPU</category><description>A BIS interim rule limits remote access to AI servers for firms in higher-risk countries and adds licences for five more. It applies in 90 days.</description><content:encoded><![CDATA[<p>The US Commerce Department’s Bureau of Industry and Security, working with the Treasury Department and the National Security Council, has issued an interim final rule on remote computing. Tom’s Hardware quotes its aim: existing restrictions on sensitive American technology should not be circumvented through remote computing arrangements.</p>
<p>Nvidia can ship its high-performance AI chips only to approved buyers, but reports of black-market GPUs and of Chinese companies renting overseas GPUs suggested the controls had holes. Tom’s Hardware notes that Blackwell Ultra B300 GPUs were recently found in a data center used by China Telecom.</p>
<p>Under the rule, overseas data centers that use US technology will have to do more to limit who can reach their services, and operators that give remote access to companies headquartered in higher-risk countries such as China, Russia and Iran will face restrictions. Exporting advanced chips to companies headquartered in Indonesia, Malaysia, Singapore, Thailand or Brazil will need a licence. The requirements take effect in 90 days.</p>
]]></content:encoded></item><item><title>Google Project Zero will cut disclosure deadlines to 30 days</title><link>https://ihos.eu/google-project-zero-30-day-disclosure/</link><guid isPermaLink="true">https://ihos.eu/google-project-zero-30-day-disclosure/</guid><pubDate>Mon, 16 Nov 2026 21:30:00 GMT</pubDate><category>Security</category><description>Project Zero is dropping its 90+30 policy: 90 days now, 60 from 2 February 2027 and 30 from 3 January 2028. AI speeds up exploits, it says.</description><content:encoded><![CDATA[<p>Google’s Project Zero is replacing its “90+30” disclosure policy with shorter fixed deadlines, BleepingComputer reports. The change is gradual. It began on 13 November, when the 30-day extension was removed. On 2 February 2027 the deadline drops to 60 days, and on 3 January 2028 to 30 days. After it, Project Zero will publish vulnerability details even if there is no fix.</p>
<p>Project Zero says AI has accelerated vulnerability discovery and exploitation and made it easier for attackers to reverse-engineer fixes as soon as patches ship. Vendors had warned Google that moving straight to 30 days could expose users to unfixed vulnerabilities, and could make maintainers focus on Project Zero findings over other, more serious reports, which is why the cut is staged.</p>
<p>There are exceptions: early disclosure when a flaw is already exploited in the wild, when others have published details, or when the vendor asks. Google can also hold details back for longer, as it did with Spectre and Meltdown (more than 120 days).</p>
]]></content:encoded></item><item><title>Lambda runs the first production co-packaged optics GPU cluster</title><link>https://ihos.eu/nvidia-lambda-co-packaged-optics-cluster/</link><guid isPermaLink="true">https://ihos.eu/nvidia-lambda-co-packaged-optics-cluster/</guid><pubDate>Mon, 16 Nov 2026 21:25:00 GMT</pubDate><category>Network &amp; IP</category><description>NVIDIA and Lambda say a 10,000-GPU cluster already uses co-packaged optics switches, years before many expected the technology in production.</description><content:encoded><![CDATA[<p>Co-packaged optics put the optical modules on the same package as the switch chip instead of in pluggable modules on the front panel. The industry has tried to bring that from the lab to production for decades. At SC26, according to ServeTheHome, NVIDIA and Lambda described a production cluster with more than 10,000 GPUs that already runs it.</p>
<p>NVIDIA’s first CPO version is in its Quantum-X switch, with Spectrum-X to follow in a second generation. The Quantum-X CPO switch has 144 MPO/MTP ports at 800 Gbps each. Instead of 144 optical modules, it uses 18 pluggable light sources on the front of the chassis. Fewer parts can fail, and they are easier to cool. Because the optics sit at the switch package, the high-speed copper links that would otherwise carry signals to the front cages are not needed, which cuts power use.</p>
]]></content:encoded></item><item><title>NoisyTCP: a remote Linux crash bug that sat in the kernel since 2015</title><link>https://ihos.eu/noisytcp-linux-remote-crash/</link><guid isPermaLink="true">https://ihos.eu/noisytcp-linux-remote-crash/</guid><pubDate>Mon, 16 Nov 2026 18:30:00 GMT</pubDate><category>Security</category><description>CVE-2026-99288 lets anyone who can reach an open TCP port crash a Linux host, even from inside a container. A fix landed upstream on 3 November.</description><content:encoded><![CDATA[<p>The researcher Hyunwoo Kim, who reported the Dirty Frag kernel bug earlier this year, has published another Linux flaw. Tracked as CVE-2026-99288 and named NoisyTCP, it sits in the kernel’s TCP code and dates from October 2015. Two parts of the kernel can race while handling incoming traffic, and one can end up using memory that has already been freed. The result can be a kernel panic.</p>
<p>According to The Register, an attacker needs no credentials, only access to an open TCP port on a system that spreads incoming traffic across several receive queues or CPUs. No special configuration is required: Nginx, Apache and SSH are affected in their defaults. In tests on QEMU/KVM virtual machines, Kim crashed them in between five seconds and two minutes. Bare metal was harder: after tailoring the exploit to the target, a crash of a server running Nginx took about 14 hours.</p>
<p>Kim describes a second path through the use-after-free that could in principle go further, and says he has a limited exploit against virtual machines in a lab. He has not shown general remote code execution, so the confirmed impact is remote denial of service. The details were published on 15 November after the embargo ended, and the fix was committed upstream on 3 November.</p>
]]></content:encoded></item><item><title>Microsoft 365 outage: a July bug, a Teams failover and lost capacity</title><link>https://ihos.eu/microsoft-365-outage-latent-bug/</link><guid isPermaLink="true">https://ihos.eu/microsoft-365-outage-latent-bug/</guid><pubDate>Mon, 16 Nov 2026 16:45:00 GMT</pubDate><category>Cloud</category><description>Three small problems combined into a 24-hour Exchange Online outage, Microsoft says. Its load tests used weaker hardware than production.</description><content:encoded><![CDATA[<p>Microsoft’s preliminary post-incident report in the Microsoft 365 Admin Center blames three problems for the outage that began at 11:20 AM ET on 12 November and lasted until 11:10 AM ET the next day, according to BleepingComputer.</p>
<p>The first was a previously unknown bug in an update for the networking gateway servers, shipped in July. It ran for more than three months without trouble, but could make servers stop responding under certain traffic conditions. Microsoft says its tests missed it because the lab traffic mix and server types were weaker than production: the contention problem is worse on machines with more cores and newer hardware.</p>
<p>The second was extra load. About 15 hours earlier, Microsoft Teams automatically failed over to the networking infrastructure used by Exchange Online. The third was capacity: nearly 14% of the gateway’s servers were already offline for an unrelated reason. With that combination, the bug triggered, and as servers stopped responding their traffic moved to the remaining ones, which then failed too. In under two hours the defect was hit across the fleet.</p>
<p>Teams, OneDrive and SharePoint could switch to other infrastructure and recovered earlier. Exchange Online could not, so mailboxes stayed unreachable. Restarting servers did not help at first because backed-up traffic overwhelmed them again. Microsoft has added capacity, improved monitoring and added safeguards against redirected traffic causing cascades, and has promised a final report.</p>
]]></content:encoded></item><item><title>AMD EPYC Venice supports MRDIMM-12800, but you can’t buy it yet</title><link>https://ihos.eu/amd-epyc-venice-mrdimm-12800/</link><guid isPermaLink="true">https://ihos.eu/amd-epyc-venice-mrdimm-12800/</guid><pubDate>Mon, 16 Nov 2026 15:10:00 GMT</pubDate><category>Dedicated</category><description>Phoronix re-tested EPYC 9686F at 8800 MT/s to show what the new memory adds. MRDIMM-8800 already costs $3–4k per 64GB module.</description><content:encoded><![CDATA[<p>Phoronix’s launch coverage of the AMD EPYC 9006 (SP7) processors prompted a common question: how much of the performance comes from the new MRDIMM-12800 memory? The site’s answer is a re-test of an EPYC 9686F two-socket system with the modules limited to 8800 MT/s, the speed of first-generation MRDIMM that Intel’s Xeon 6 Granite Rapids supports.</p>
<p>The review server arrived with 32 SK Hynix 64GB MRDIMM-12800 modules. The author could not compare different modules or standard RDIMMs, only change the target speed in the BIOS. Settings below 8800 MT/s in the BIOS fell back to 12800 in the operating system, which suggests 8800 is the supported baseline for first-generation parts.</p>
<p>Phoronix notes that MRDIMM-12800 is not for sale at any retailer yet, and that MRDIMM-8800 listings show about $3–4k per 64GB module.</p>
]]></content:encoded></item></channel></rss>