{"id":2081,"date":"2026-01-04T14:48:14","date_gmt":"2026-01-04T14:48:14","guid":{"rendered":"https:\/\/blog.thomarite.uk\/?p=2081"},"modified":"2026-01-04T14:48:41","modified_gmt":"2026-01-04T14:48:41","slug":"ai-labs-power-colossus2-aws-reinvent-2025-asml-power-laws-deepmind-hotchips-tax-strategy-groq-polyglot","status":"publish","type":"post","link":"https:\/\/blog.thomarite.uk\/index.php\/2026\/01\/04\/ai-labs-power-colossus2-aws-reinvent-2025-asml-power-laws-deepmind-hotchips-tax-strategy-groq-polyglot\/","title":{"rendered":"AI Labs Power,  Colossus2, AWS Reinvent 2025, ASML, Power Laws, DeepMind, HotChips, Tax strategy, Groq, Polyglot"},"content":{"rendered":"\n<p><a href=\"https:\/\/newsletter.semianalysis.com\/p\/how-ai-labs-are-solving-the-power\">AI Labs Powers<\/a>: Interesting articule for solutions to get power for AI infra<\/p>\n\n\n\n<p><a href=\"https:\/\/newsletter.semianalysis.com\/p\/xais-colossus-2-first-gigawatt-datacenter\">Colossus2<\/a>: The tricks to get power between states&#8230; and cost.<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=45MFTpKfxII&amp;list=PLxF-QlFVz84S34FMpmdK7Ha23WGh0hmiV&amp;index=26\">Amazon Leo Architecture + Internet Edge ARC320<\/a>: I didn&#8217;t know AWS was going to compete with Starlink&#8230; but having JB&#8217;s BlueHorizon, I guess it makes sense.<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=YQOuYK23VEI\">AWS Network Infra 2025 NET402<\/a>: From minute 29 is the most interesting for me. Precast fiber duct banks, TE pre-signalled bypass IP tunnels for each path (without RSVP-TE?) and constant recalculation. UltraCluster3, connector improvements (36% reduce link failures! 76% reduce time for cabling!). UltraSwitch with dynamic LB and adaptive routing (like IB and UltraEthernet)<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=YZUNNzLDWb8&amp;t=1s\">AWS DynamoDB outage lessons learnt<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/aiengineer.net\/tech\/kubernetes\/kubernetes-pod-networking-from-scratch\">Kubernetes pod networking<\/a>: Good refresher<\/p>\n\n\n\n<p><a href=\"https:\/\/blog.apnic.net\/2025\/12\/02\/hot-takes-on-13-talks-from-nanog-95\">Nanog95 summary<\/a>:<\/p>\n\n\n\n<p>Veritasium: Why <a href=\"https:\/\/www.youtube.com\/watch?v=MiUHjLxm3V0\">ASML<\/a> is so critical.<\/p>\n\n\n\n<p>Veritasium: <a href=\"https:\/\/www.youtube.com\/watch?v=HBluLfX2F_k\">Power Laws<\/a>: I think I get it&#8230; but it is scary: forest fires, sanpiles, money, investment&#8230;<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=uakvvweLDRk\">Extreme success, you can&#8217;t be a balanced person<\/a>: I like this format of smaller conversations. It is difficult to find time for 1h+ videos<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=d95J8yzvjbQ\">Google DeepMind<\/a>: I like the history behind DeepMind and his founder. I didnt know they tried to beat the Go champion in China&#8230;. and disconnected.<\/p>\n\n\n\n<p>HotChips2025:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><a href=\"https:\/\/www.youtube.com\/watch?v=xG0WJ4HoMXM&amp;t=5262s\">Datacenter<\/a><a href=\"https:\/\/www.youtube.com\/watch?v=xG0WJ4HoMXM\"> Racks part2<\/a>: \n<ul class=\"wp-block-list\">\n<li>Nice the NVIDIA presentation from a mechanical engineer.<\/li>\n\n\n\n<li>Meta Catalina pod 33:39 &#8211; 4xracks for liquid cooling! for 2xIT racks. 42:39 3 networks: frontend (N-S), backend(E-W) and management\/console. Leaking monitoring<\/li>\n\n\n\n<li>Google TPU rack Ironwood: I need to research the 3D Torus connection. 1h15m16<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><a href=\"https:\/\/www.youtube.com\/watch?v=z1cu-LBeRx4&amp;t=5882s\">Networking<\/a>:\n<ul class=\"wp-block-list\">\n<li>Intel IPU E220: This is a NIC although I wasn&#8217;t 100% sure until I checked in another <a href=\"https:\/\/techovedas.com\/intel-ipu-e2200-the-silent-powerhouse-reshaping-data-centers\/\">site<\/a>. You can use P4<\/li>\n\n\n\n<li>AMD Pensando Pollara 400 (NIC). P4 architecture. UltraEthernet ready. 48:36 95% network utilization (intelligent LB, congestion mgmt (RTT-based), fast failover and loss recovery (select ACK)= the trinity)<\/li>\n\n\n\n<li>NVIDIA ConnectX-8 SuperNIC: 1h13m multiplane &#8211; as I understand, you use more leaf switches, in the example 4 planes = 4 leaf, and you have 64x gpu scale <\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>switch radix 512,  gpu 8x100G, two-level non-blocking<\/p>\n\n\n\n<p>standard nic 2-level non-blocking: 64&#215;32=2k<\/p>\n\n\n\n<p>connect8-x multiplane: 256&#215;512=128k<\/p>\n\n\n\n<p>Spectrum-X:  network round-trip time: 5-10us, packet transmisson 2ns<\/p>\n\n\n\n<p>demo 1h24m (with grafana!)<\/p>\n\n\n\n<p>Tax Strategy: <a href=\"https:\/\/www.youtube.com\/watch?v=WgzyFrIwYi4&amp;t=12s\">ex1<\/a>, <a href=\"https:\/\/www.youtube.com\/watch?v=22CjsRG0sTc&amp;t=12s\">ex2<\/a><\/p>\n\n\n\n<p><a href=\"https:\/\/www.theregister.com\/2025\/12\/31\/groq_nvidia_analysis\/\">NVIDIA buys Groq<\/a>: interesting details between SRAM vs HBM. And it seems the key is &#8220;assembly line architecture&#8221;<\/p>\n\n\n\n<p><a href=\"https:\/\/www.youtube.com\/watch?v=XWMEcUBbDTI\">Polyglot<\/a>: no magic tricks. I need more exposure.<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>AI Labs Powers: Interesting articule for solutions to get power for AI infra Colossus2: The tricks to get power between states&#8230; and cost. Amazon Leo Architecture + Internet Edge ARC320: I didn&#8217;t know AWS was going to compete with Starlink&#8230; but having JB&#8217;s BlueHorizon, I guess it makes sense. AWS Network Infra 2025 NET402: From &hellip; <a href=\"https:\/\/blog.thomarite.uk\/index.php\/2026\/01\/04\/ai-labs-power-colossus2-aws-reinvent-2025-asml-power-laws-deepmind-hotchips-tax-strategy-groq-polyglot\/\" class=\"more-link\">Continue reading<span class=\"screen-reader-text\"> &#8220;AI Labs Power,  Colossus2, AWS Reinvent 2025, ASML, Power Laws, DeepMind, HotChips, Tax strategy, Groq, Polyglot&#8221;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13,32,2],"tags":[],"class_list":["post-2081","post","type-post","status-publish","format-standard","hentry","category-aws","category-cpu","category-networks"],"_links":{"self":[{"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/posts\/2081","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/comments?post=2081"}],"version-history":[{"count":3,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/posts\/2081\/revisions"}],"predecessor-version":[{"id":2088,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/posts\/2081\/revisions\/2088"}],"wp:attachment":[{"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/media?parent=2081"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/categories?post=2081"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.thomarite.uk\/index.php\/wp-json\/wp\/v2\/tags?post=2081"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}