<?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" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Infrastructure - Soursen</title><description>Infrastructure covers how networks and power fail, checked against the release or the incident report that records what failed.</description><link>https://soursen.org/</link><language>en-gb</language><atom:link href="https://soursen.org/beat/infrastructure/rss.xml" rel="self" type="application/rss+xml"/><item><title>How China Built the Thing Everyone Else Has to Buy From</title><link>https://soursen.org/articles/how-china-built-the-thing-everyone-else-has-to-buy-from/</link><guid isPermaLink="true">https://soursen.org/articles/how-china-built-the-thing-everyone-else-has-to-buy-from/</guid><description>China ran a trade surplus of $938 billion in the first 9 months of 2026, installed 59% of the world’s new factory robots, and refines around 90% of the planet’s rare earths. The interesting question is not whether it is ahead in manufacturing, the numbers settle that, but how a few deliberate choices compounded into a lead that tariffs and chip bans have not reversed. I went through the mechanics.</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Walk into a car showroom almost anywhere in Europe this year and you will find brands that barely existed there a few years ago. That is the surface of a much larger shift. In the first 9 months of 2026, China’s auto exports rose about 54% by value over the same stretch of 2025, according to its customs agency, driven by demand for electric and hybrid cars. In June, monthly vehicle exports crossed 1 million units for the first time. A growing share of the new cars on the world’s roads now ship from China.&lt;/p&gt;
&lt;p&gt;I do not find the fact of the lead very interesting on its own, because the data is not really in dispute. What is worth understanding is the machinery underneath, the specific decisions and physical advantages that let China keep pulling ahead while the countries trying to catch up stay stuck. That is what this is about.&lt;/p&gt;
&lt;h2 id=&quot;the-number-that-captures-it&quot;&gt;The number that captures it&lt;/h2&gt;
&lt;p&gt;Start with trade, because it ties everything else together. China ran a goods trade surplus of $938 billion in the first 9 months of 2026, according to &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.ajc.com/news/2026/10/chinas-exports-surge-363-in-september-as-pressure-ramps-up-to-rebalance-trade/&quot;&gt;customs data reported by the Associated Press&lt;/a&gt;. The world bought nearly a trillion dollars more from China than China bought back, and the full year is on track to beat 2025’s record of about $1.2 trillion.&lt;/p&gt;
&lt;p&gt;The mix matters more than the size. This is not cheap clothing and plastic toys anymore. In that same stretch, China’s semiconductor exports rose 108% by value, its data processing equipment rose 55%, and its auto exports rose 54%. The things driving the surplus now are chips, batteries, cars, and machinery, which is to say the high value goods that rich countries built their economies on. For much of the world’s manufacturing, the supply chain now starts in China.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DxyS575c_Zj4AAp.webp&quot; alt=&quot;What China’s export surge is actually made of. The growth is in high value goods, not cheap ones.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;What China’s export surge is actually made of. The growth is in high value goods, not cheap ones. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.ajc.com/news/2026/10/chinas-exports-surge-363-in-september-as-pressure-ramps-up-to-rebalance-trade/&quot;&gt;China customs data via AP, Jan to Sept 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;why-the-factories-move-so-fast&quot;&gt;Why the factories move so fast&lt;/h2&gt;
&lt;p&gt;The clearest way to see China’s manufacturing advantage is to follow one product from idea to shelf. Picture an engineer in Shenzhen, in the Pearl River Delta, designing a new electric car. Within a short drive of the office are the mold makers who shape the plastic, the factories that print the circuit boards, the plants that make the battery cells, and the software teams that write the code. Midway through testing, the engineer decides the battery housing needs a different shape. She drives to the tooling shop that afternoon, the mold gets recut, the new part gets tested, and the change is locked in within days.&lt;/p&gt;
&lt;p&gt;Run the same change at a carmaker whose suppliers are spread across several countries. Drawings go out by email, samples ship by freight, and each round of revisions waits for the last to arrive. A change that takes days in Shenzhen takes months across a scattered supply chain, and a car is thousands of those decisions stacked on top of each other. The same density exists around Shanghai in the Yangtze River Delta. The speed comes from everything a product needs sitting in one region.&lt;/p&gt;
&lt;p&gt;Geography is only half of it. For decades China assembled phones, laptops, and appliances for brands worldwide, and every one of those contracts left knowledge behind. Tooling engineers who know which mold will crack. Line managers who can clear a bottleneck before lunch. Inspectors who spot a flaw by feel. That knowledge moves between companies, passes to younger workers, and compounds with every product cycle, so each new product starts from a higher base than the one before it.&lt;/p&gt;
&lt;h2 id=&quot;involution-the-competition-most-markets-never-see&quot;&gt;Involution, the competition most markets never see&lt;/h2&gt;
&lt;p&gt;There is a Chinese word for the next piece, involution. It describes competition so intense that everyone has to run faster just to hold their position. The scale of it is hard to overstate. Around 2018 China had more than 500 registered electric vehicle startups. By 2025, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://evboosters.com/ev-charging-news/400-chinese-ev-companies-ceased-operations-between-2018-2025-only-a-few-will-dominate-towards-2030&quot;&gt;roughly 400 of them had ceased operations&lt;/a&gt;, leaving about 100 active brands, and McKinsey analysts expect fewer than 50 to survive by 2030. The ones still standing, names like &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/BYD_Auto&quot;&gt;BYD&lt;/a&gt;, Geely, NIO, and Xpeng, learned to cut cost, ship updates, and improve at a brutal pace through years of domestic price wars, because the alternative was running out of cash and disappearing, which is what happened to most of them.&lt;/p&gt;
&lt;p&gt;So by the time a Chinese car reaches a European showroom, it has already survived the hardest car market on earth. The company exporting it is a survivor of a cull that killed about 4 out of every 5 of its rivals, hardened rather than coddled, which is a different kind of competitor than a legacy automaker that has had a comfortable domestic market for 50 years. That history is a large part of why the imports are hard to price against.&lt;/p&gt;
&lt;h2 id=&quot;the-3-bets-that-built-the-lead&quot;&gt;The 3 bets that built the lead&lt;/h2&gt;
&lt;p&gt;China’s lead did not happen by accident, and it comes down to 3 strategies that each took years to pay off.&lt;/p&gt;
&lt;p&gt;The first was picking industries early and staying in them through the lean years. Under &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Made_in_China_2025&quot;&gt;Made in China 2025&lt;/a&gt;, Beijing named batteries, solar, and electric vehicles as sectors it meant to lead, then kept the money flowing when demand was thin and nobody was making a profit. Western firms did the opposite. The American battery maker &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/A123_Systems&quot;&gt;A123 Systems&lt;/a&gt; went bankrupt in 2012 and was bought by China’s Wanxiang Group. Many early US and European solar manufacturers collapsed or left the business. So when global demand finally arrived, the Chinese companies were the ones that already had the factories, the suppliers, and the trained workers in place.&lt;/p&gt;
&lt;p&gt;The second was a bet on battery chemistry. Most Western carmakers built around NMC cells, using nickel, manganese, and cobalt. Chinese makers backed &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Lithium_iron_phosphate_battery&quot;&gt;lithium iron phosphate&lt;/a&gt;, or LFP, which leaves out the expensive nickel and cobalt. CATL and BYD scaled LFP to enormous volumes, and BYD built its Blade battery around it. The battery is a huge share of what an electric car costs, so a cheaper chemistry at that scale showed up directly as a cheaper car.&lt;/p&gt;
&lt;p&gt;The third was a response to being cut off. US export controls blocked Chinese AI labs from the most advanced chips, so instead of waiting for better hardware, labs like &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/DeepSeek&quot;&gt;DeepSeek&lt;/a&gt; and the teams behind Qwen and Kimi squeezed more out of what they had. They leaned on a mixture of experts design, where a model is split into specialized sections and only the ones a given question needs switch on, which saves a large amount of computing power. Then they released the models as open weights, so anyone could download them, run them, and build on top. Chinese AI spread by being efficient and easy to adopt rather than by owning the biggest data centers.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.Dy1-lyZt_UDuCW.webp&quot; alt=&quot;The robot race is the clearest single snapshot. China installed 59% of the world’s new industrial robots in 2025.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;The robot race is the clearest single snapshot. China installed 59% of the world’s new industrial robots in 2025. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ifr.org/&quot;&gt;IFR World Robotics 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-chokepoint-hiding-inside-everything&quot;&gt;The chokepoint hiding inside everything&lt;/h2&gt;
&lt;p&gt;This reaches into products that say they were made somewhere else. Take an electric car assembled in Europe and trace it backward. The motor’s magnets are made from rare earth metals. The battery needs graphite, lithium, and cobalt. The chips rely on gallium and germanium. Those raw materials get dug out of the ground on several continents, but before any factory can use them, they have to be refined, the step where raw ore is purified into something a manufacturer can work with. That step happens overwhelmingly in China.&lt;/p&gt;
&lt;p&gt;China handles around 90% of the world’s rare earth refining, plus the bulk of refining for graphite, cobalt, and gallium.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.BkNzD2GY_Z15Sm23.webp&quot; alt=&quot;Where the control actually sits. The refining step, not the mining, happens mostly in China.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Where the control actually sits. The refining step, not the mining, happens mostly in China. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.iea.org/&quot;&gt;IEA critical minerals data&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;So a made in label tells you where the final assembly happened, not where the product really started. Beijing understands exactly how much power that gives it. In April 2025 it put &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.hklaw.com/en/insights/publications/2025/04/china-imposes-export-controls-on-medium-and-heavy-rare-earth-materials&quot;&gt;licensing requirements on 7 medium and heavy rare earths&lt;/a&gt;, including metals used in high performance magnets, and carmakers and electronics firms outside China suddenly waited on paperwork for supplies they used to just order. A broader wave of controls followed later in the year before being paused as part of a truce with the US.&lt;/p&gt;
&lt;p&gt;The same pattern showed up in a less obvious place. Tunnel boring machines, the giant rotating drills that carve rail and metro tunnels, are another thing many countries buy from China. In late 2026 Beijing announced &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.mlex.com/mlex/articles/2539359&quot;&gt;export controls on those machines&lt;/a&gt; too. India had already seen Chinese made machines held at customs, and its bullet train and metro projects faced fresh delays. A rule written in Beijing can hold up trains on other continents, which is what real control over a supply chain looks like.&lt;/p&gt;
&lt;h2 id=&quot;why-tariffs-and-chip-bans-did-not-reverse-it&quot;&gt;Why tariffs and chip bans did not reverse it&lt;/h2&gt;
&lt;p&gt;The obvious counter to all this was pressure from Washington, tariffs and technology bans. Neither changed the underlying position, and understanding why is the whole point.&lt;/p&gt;
&lt;p&gt;Take tariffs. The US put heavy duties on goods shipped from China, so Chinese manufacturers moved final assembly to Vietnam, Mexico, and elsewhere in Southeast Asia. The components still came from Chinese suppliers, got assembled in the new hubs, and shipped to America under a different country of origin. The shipping routes and the customs forms changed, the label on the box changed, and the dependence did not change at all. China’s trade surplus kept climbing the entire time.&lt;/p&gt;
&lt;p&gt;The chip bans ran into a similar reversal. The US cut Chinese firms off from the most advanced AI chips and the tools to make them, meaning to freeze China’s chip industry in place. What it did instead was remove the option of buying American, so Chinese firms started building their own. Huawei pushed ahead with its Ascend AI chips, SMIC kept driving toward more advanced manufacturing, and memory makers CXMT and YMTC kept expanding, while Beijing steered its own tech companies toward domestic chips and handed its suppliers a guaranteed home market. Restrictions meant to slow China down pushed it to build its own alternatives sooner. I am not claiming the bans did nothing, they raised costs and bought time, but they did not do the thing they were designed to do.&lt;/p&gt;
&lt;h2 id=&quot;the-next-race-is-already-running&quot;&gt;The next race is already running&lt;/h2&gt;
&lt;p&gt;The same clusters that won electric cars are now pointed at what comes next. In 2025 China installed &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/industrial-robots-world-robotics-2026-china-59-eu-falls&quot;&gt;354,000 industrial robots&lt;/a&gt;, about 59% of every new industrial robot installed on earth, according to the International Federation of Robotics. The US installed 38,400 and the EU 60,500, with EU installations actually falling 11%. China put in roughly 9 times as many robots as the US, and more than half of its own came from Chinese suppliers.&lt;/p&gt;
&lt;p&gt;The pattern repeats across the new frontiers. In drones, DJI is the name most people already know. At IFA in Berlin in September, Chinese firms staged the largest humanoid robot showcase in the event’s history. CATL is developing sodium ion cells and solid state battery designs for the generation after lithium. The factories, engineers, and supply chains that won the car race are being aimed at robots, drones, and the batteries to power them, which means the advantage compounds into the next set of industries rather than resetting.&lt;/p&gt;
&lt;h2 id=&quot;the-bill-that-comes-with-the-lead&quot;&gt;The bill that comes with the lead&lt;/h2&gt;
&lt;p&gt;I do not want to write the triumphant version, because the same model that produced the lead is doing real damage inside China. Involution cuts both ways. The price wars that forged world beating companies also pushed the auto industry’s profit margins down to around 1.8%, left most makers losing money on every car, and helped drag the whole economy toward deflation, where falling prices feed on themselves. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bloomberg.com/news/articles/2025-09-17/china-s-involution-problem-how-price-wars-overcapacity-are-fueling-deflation&quot;&gt;Bloomberg has described involution&lt;/a&gt; and the overcapacity behind it as a core problem for China’s economy, not a feature of it.&lt;/p&gt;
&lt;p&gt;Beijing clearly agrees, because it keeps trying to stop it. Regulators have summoned automakers more than once over what they call an irrational price war, and have drafted rules to curb the endless discounting. BYD, the strongest of the survivors, reported a profit slump as the downturn ran on. The honest picture is a country that built a genuine manufacturing lead and is now struggling to control the destructive competition that built it. The lead is real, and so is the hangover.&lt;/p&gt;
&lt;h2 id=&quot;what-it-means-for-the-rest-of-us&quot;&gt;What it means for the rest of us&lt;/h2&gt;
&lt;p&gt;Set the geopolitics aside and look at what lands in ordinary life. Across much of the world outside the US, the list of new cars now includes Chinese models that were not there a few years ago, carrying cheaper LFP batteries and costs already squeezed by the price wars back home, and every other brand on the lot has to price against them.&lt;/p&gt;
&lt;p&gt;The same trail runs through almost everything else you buy. The phone in your pocket, the charger on your desk, the solar panel on a roof down the street, the parts inside a washing machine. Even when the box names another country, the path usually runs back through Chinese factories and Chinese refineries, and that is a real part of why those products cost what they do. Chinese carmakers are now opening plants in Hungary, Spain, Turkey, Thailand, Indonesia, and Brazil, which brings factory and engineering jobs that revolve around Chinese designs, Chinese processes, and Chinese supply chains. The lead is not an abstract statistic about a faraway country. It is already priced into the things on the shelf in front of you.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>china</category><category>business</category><category>economics</category><category>ai</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Microsoft Offered 60,000 People the Door. The Real Story Is in the Capex Line.</title><link>https://soursen.org/articles/microsoft-offered-60-000-people-the-door-the-real-story-is-in-the-capex/</link><guid isPermaLink="true">https://soursen.org/articles/microsoft-offered-60-000-people-the-door-the-real-story-is-in-the-capex/</guid><description>The headline reads as an AI layoff. Read the spending next to the revenue and a different reason shows up, one that will land at Google and Meta next.</description><pubDate>Sun, 01 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;On 26 October 2026, Microsoft told as many as 60,000 of its own employees they could leave voluntarily, with severance, if they wanted to. That is about 27% of a workforce of roughly 223,000. The number got written up as the biggest AI layoff yet, proof that the machines were finally eating the jobs. I think that reading is lazy, and I want to walk through why, because the actual driver is sitting in plain sight on Microsoft’s own earnings statement.&lt;/p&gt;
&lt;p&gt;A lot of what got posted about this was people &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.cnbc.com/2025/05/13/microsoft-is-cutting-3percent-of-workers-across-the-software-company.html&quot;&gt;making up numbers to fit a headline&lt;/a&gt;. So I am going to stick to figures Microsoft and the reporters covering it actually put on the record, and I will tell you where the number is a company projection versus a hard count.&lt;/p&gt;
&lt;h2 id=&quot;what-microsoft-actually-offered&quot;&gt;What Microsoft actually offered&lt;/h2&gt;
&lt;p&gt;This is a Voluntary Exit Offer, which Microsoft had never run before at this scale. Eligible staff sit at level 70 and below in the US, Canada, the UK, and India. Corporate vice presidents are out. There is no age, tenure, or performance gate. People decide by early December and leave on 31 December. The severance is Microsoft’s standard package, at least 8 weeks of pay in the US, and notably it drops the extended healthcare that a smaller spring program had included, per people familiar with the terms reported by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/30/microsoft-voluntary-exit-offer-buyout-60000-employees&quot;&gt;Fortune&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The framing from Microsoft is careful. A spokesperson said, “We don’t expect 60,000 people to walk out the door, and we would never want that.” CFO Amy Hood told staff the company expects very few to accept. Chief People Officer Amy Coleman called the offer “new for us” and said, “We’ll learn from it.”&lt;/p&gt;
&lt;p&gt;So the 60,000 is the size of the invitation, not a forecast of departures. That distinction is the whole thing, and most of the coverage blew past it.&lt;/p&gt;
&lt;h2 id=&quot;the-number-everyone-skipped&quot;&gt;The number everyone skipped&lt;/h2&gt;
&lt;p&gt;Microsoft has run a voluntary program like this once before, this spring. About 2,600 of roughly 8,750 eligible US employees took it, around 30% uptake, per &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/30/microsoft-voluntary-exit-offer-buyout-60000-employees&quot;&gt;Fortune’s reporting&lt;/a&gt;. The company booked an initial charge near $900 million on it.&lt;/p&gt;
&lt;p&gt;For this far larger offer, Microsoft is not giving a target. Its spokesperson instead pointed to a 3% to 5% participation range at peer companies, which would work out to somewhere between 1,800 and 3,000 people actually leaving out of the 60,000 invited. One analyst at TIKR, Gian Estrada, ran an illustrative model at 10% uptake, roughly 6,000 departures, and estimated about $1.9 billion in annual savings against a fiscal 2026 research and SG&amp;amp;A base of $70.23 billion.&lt;/p&gt;
&lt;p&gt;Sit with those two figures. The headline says 60,000. The realistic exit, on Microsoft’s own benchmark, is a few thousand. That gap is not an accident, and I will come back to why the company wants the big scary number out there anyway.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.DROcn_xk_ZQeovb.webp&quot; alt=&quot;How a 60,000 person offer becomes a few thousand exits. The invitation is wide, the expected uptake is narrow.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;How a 60,000 person offer becomes a few thousand exits. The invitation is wide, the expected uptake is narrow. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/30/microsoft-voluntary-exit-offer-buyout-60000-employees&quot;&gt;Fortune&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;now-look-at-the-spending&quot;&gt;Now look at the spending&lt;/h2&gt;
&lt;p&gt;Here is the part the layoff framing misses. Microsoft’s capital expenditure for the fiscal year that ended 30 June 2026 came to $115.9 billion, up from $64.6 billion the year before. That is roughly a 79% jump in one year. Over the same stretch the workforce shrank by about 5,000 people. The company spent almost double on data centers, chips, and networking gear while quietly getting smaller.&lt;/p&gt;
&lt;p&gt;The current pace is steeper still. For the quarter ending 30 September 2026, capex rose about 30% to $45.2 billion, and roughly three quarters of that went to short-lived assets such as chips and networking equipment, the kind that depreciate fast. Hood told investors she expects spending to top $50 billion in the following quarter. Put annually, Microsoft is pointing at something near $170 billion of capital spending across calendar 2026, per figures reported by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/28/microsofts-cloud-growth-accelerates-and-so-too-does-the-spending&quot;&gt;Fortune&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Against that, revenue is growing, but not at the same slope. Microsoft posted $92.3 billion in quarterly revenue, up 19% year over year, with Azure up 48% and net profit of $36.6 billion. Those are strong numbers. The problem is arithmetic. When your spending climbs around 30% and your top line climbs around 19%, the gap between the two lines widens every quarter, and that gap comes straight out of margin.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.BlkqtaP4_ZqE5kP.webp&quot; alt=&quot;Capex is climbing faster than revenue. The widening gap is the pressure behind the exit offer.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Capex is climbing faster than revenue. The widening gap is the pressure behind the exit offer. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3&quot;&gt;Microsoft investor relations&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;There is a second, slower problem buried in that capex number. Roughly three quarters of the September quarter’s $45.2 billion went to short-lived assets, the chips and networking gear that wear out or go obsolete in a few years, not the concrete and steel of the buildings that last decades. Short-lived assets depreciate fast, and depreciation is a real expense that lands on the income statement quarter after quarter whether or not the AI revenue shows up to cover it. The cash goes out now, and on top of that Microsoft signs up for years of rising depreciation charges that will press on margin long after the construction is done. Cutting payroll is one of the few levers that works on the same timeline.&lt;/p&gt;
&lt;h2 id=&quot;what-all-that-money-actually-buys&quot;&gt;What all that money actually buys&lt;/h2&gt;
&lt;p&gt;It helps to be concrete about where $170 billion a year goes. Microsoft said in July it expected to spend about $175 billion on AI infrastructure across 2026, later guided near $170 billion, and the physical form of that is data centers full of Nvidia and increasingly custom silicon, plus the power and cooling to run them. Hood has repeated a line all year, “We still have demand that exceeds supply,” which is the official reason the spending keeps climbing. The company is building capacity it believes it can rent out through Azure, where growth of 48% says the demand is real for now.&lt;/p&gt;
&lt;p&gt;The risk is timing. The buildings and chips have to be paid for today, and the return arrives only if enterprise AI demand keeps compounding at something like the current rate for years. If it slows, Microsoft is left carrying the depreciation on a very large pile of fast-aging hardware. That asymmetry, big certain cost now against uncertain return later, is exactly the setup that makes a finance team reach for every flexible cost it can find. Payroll is the most flexible one on the list.&lt;/p&gt;
&lt;h2 id=&quot;why-a-buyout-and-not-a-layoff&quot;&gt;Why a buyout and not a layoff&lt;/h2&gt;
&lt;p&gt;If the pressure is financial, why offer a voluntary exit instead of just cutting? Because a buyout sorts people for you, at lower legal and morale cost.&lt;/p&gt;
&lt;p&gt;Think about who says yes to an offer like this. The person already worried they are on a weak team, already sensing a reorg coming, already half out the door, takes 8 weeks of guaranteed pay over the gamble of being cut later with less. The person who believes they are valuable stays and bets on themselves. So a voluntary program tends to peel off exactly the staff who were least confident of their footing, without the company having to name names, run a formal performance process, or eat the headline of a mass layoff. It is cheaper, it is quieter, and it front-loads the easy departures before any harder decisions.&lt;/p&gt;
&lt;p&gt;Microsoft has been candid that layoffs remain on the table depending on business conditions. The spring program, Coleman noted, reduced the scale of the July 2025 layoffs that followed. So the pattern is a soft offer first, then sharper cuts for whatever gap remains. Read that way, the 60,000 figure functions as a signal to the whole org about where things are heading, which is why management is comfortable letting the alarming number circulate even while privately expecting almost no one to take it.&lt;/p&gt;
&lt;h2 id=&quot;what-this-says-about-the-ai-story&quot;&gt;What this says about the AI story&lt;/h2&gt;
&lt;p&gt;The clean version of this story is that AI made 60,000 workers redundant. Microsoft itself has pushed back on that, saying AI is not driving the buyout programs. I find the capex explanation far more convincing, and it is a subtler point. AI is not replacing these specific people. AI infrastructure is consuming the capital that would otherwise fund headroom in the budget, and payroll is one of the few large, flexible line items a company can pull to defend margin while it pours money into data centers.&lt;/p&gt;
&lt;p&gt;The data on attribution backs this up. When analysts tried to tag how many 2026 tech layoffs were actually caused by AI automation, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://invezz.com/news/2026/05/04/is-big-techs-725b-ai-splurge-being-funded-by-mass-layoffs/&quot;&gt;Nikkei Asia put it at 47.9% of first-quarter cuts&lt;/a&gt;, while a separate estimate from RationalFX landed at 20.4%. So even on the generous reading, more than half the cuts are something else, and on the stricter reading, four in five are. Sam Altman has acknowledged the “AI washing” at work here, companies pinning ordinary cost cuts on a more flattering story. A Cognizant executive put real AI productivity gains at six months to a year out, meaning a chunk of these cuts run ahead of any actual automation. The timing gives it away. You cut first, then hope the software catches up.&lt;/p&gt;
&lt;p&gt;Economist Jay Zagorsky, quoted by &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/30/microsoft-voluntary-exit-offer-buyout-60000-employees&quot;&gt;Fortune&lt;/a&gt;, read the offer as a company that wants fewer workers and more equipment, and expected other large tech firms to follow. He was right, and it is already happening across the industry.&lt;/p&gt;
&lt;h2 id=&quot;the-same-math-at-amazon-meta-and-the-rest&quot;&gt;The same math at Amazon, Meta, and the rest&lt;/h2&gt;
&lt;p&gt;This is not a Microsoft story. The four biggest spenders, Microsoft, Amazon, Meta, and Alphabet, are on track for about $725 billion in combined capital spending in 2026, up roughly 77% from $410 billion in 2025, per &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://invezz.com/news/2026/05/04/is-big-techs-725b-ai-splurge-being-funded-by-mass-layoffs/&quot;&gt;reporting compiled by Invezz&lt;/a&gt;. Microsoft is pointing at around $190 billion on a calendar basis, Amazon has committed about $200 billion, and Meta raised its full-year guidance to somewhere between $125 billion and $145 billion. Those are not numbers you fund out of petty cash.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.Cuf3KY3o_Z1HROWq.webp&quot; alt=&quot;Combined AI capex across the four biggest spenders jumped about 77% in a year, while the same firms cut staff.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;889&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Combined AI capex across the four biggest spenders jumped about 77% in a year, while the same firms cut staff. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://invezz.com/news/2026/05/04/is-big-techs-725b-ai-splurge-being-funded-by-mass-layoffs/&quot;&gt;Invezz&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Watch what each of them did to headcount alongside it. Amazon cut about &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.fox5ny.com/news/amazon-cutting-thousands-jobs-layoffs&quot;&gt;14,000 corporate jobs in October 2026&lt;/a&gt;, the first of a reported roughly 30,000 planned, which would make it the largest layoff in the company’s history, past the 27,000 of 2022. Meta cut about 8,000 roles earlier in the year. Microsoft offered 60,000 the door. In every case the company denied the obvious framing. Amazon CEO Andy Jassy said his cuts were not financially driven or tied to AI, blaming bureaucracy and “a lot more layers.” Microsoft said AI is not driving its buyouts. The denials rhyme because the real cause is the same, and it is not a story either company wants to tell out loud. Enormous, front-loaded AI spending against revenue that grows well but not that fast, with payroll pulled in to hold the margin line. The buyout is less about any one job being automated and more about rebalancing the whole cost structure toward silicon.&lt;/p&gt;
&lt;h2 id=&quot;where-the-cuts-are-not-landing&quot;&gt;Where the cuts are not landing&lt;/h2&gt;
&lt;p&gt;One detail cuts against the pure doom read, and it tells you what Microsoft actually values. The sales organization is not participating in this offer. Microsoft confirmed that directly. Reporting from &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://fortune.com/2026/10/30/microsoft-voluntary-exit-offer-buyout-60000-employees&quot;&gt;Fortune&lt;/a&gt; also indicates that some groups in data centers, marketing, cybersecurity, and quantum computing were left out, while Xbox, Windows, Copilot, finance, and AI experiences were fully in.&lt;/p&gt;
&lt;p&gt;Look at the pattern. The functions tied most directly to bringing in revenue or protecting it, sales and security, are the ones shielded. That is not sentiment. A sales team that misses its number gets cut on its own schedule, and a security team is what stands between the company and a compliance disaster. The groups asked to thin out are the ones further from the next dollar. If you are deciding where to place yourself inside a company like this, that is the signal worth reading, closer to revenue is safer than cooler to build.&lt;/p&gt;
&lt;h2 id=&quot;what-i-would-actually-do-with-this&quot;&gt;What I would actually do with this&lt;/h2&gt;
&lt;p&gt;The useful takeaway is not panic. It is that the companies spending the most on AI will keep leaning on their payroll as the shock absorber, and the roles that survive are the ones a finance team can draw a straight line from to money coming in or money being saved. If you are early in your career, that means weighting your time toward work that visibly moves revenue or cuts real cost, and being able to say so plainly, over work that is merely interesting. The demand for engineers is not gone. The demand is narrowing to the engineers a CFO can justify while the capex line keeps climbing.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>microsoft</category><category>layoffs</category><category>ai</category><category>artificial-intelligence</category><category>jobs</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>You Can Build a Serious Home Server for About $1,000, and It Solves a Problem Your Phone Cannot</title><link>https://soursen.org/articles/you-can-build-a-serious-home-server-for-about-1-000-and-it-solves-a/</link><guid isPermaLink="true">https://soursen.org/articles/you-can-build-a-serious-home-server-for-about-1-000-and-it-solves-a/</guid><description>A used enterprise server that cost a company thousands of dollars new sells for around $250 today, and with a few parts you have a private cloud that runs AI agents, stores your photos, records your cameras, and lets you walk through an airport with a phone that holds almost nothing. I went through the exact parts, the real prices in the current memory shortage, and what to actually run on it.</description><pubDate>Fri, 23 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Your phone is the weakest point in your privacy. At a border crossing, an officer in many countries can compel you to unlock it, and whatever is on it is theirs to read. One answer is to carry a phone that holds almost nothing, your data living on a server at home that you reach over an encrypted connection, and that you can disconnect from entirely while you travel. The piece that makes this work is a home lab server, and the surprising part is how cheap a genuinely powerful one has become. You can build it for around $1,000, even with memory and storage prices as high as they are right now.&lt;/p&gt;
&lt;p&gt;I run several of these. The machine at home is larger than most people need. What follows is the build I would recommend to someone starting out, with the actual parts, roughly current prices, and the software that turns a bare server into your own private cloud.&lt;/p&gt;
&lt;h2 id=&quot;why-keeping-data-off-the-phone-actually-works&quot;&gt;Why keeping data off the phone actually works&lt;/h2&gt;
&lt;p&gt;The strategy rests on a legal distinction that is easy to miss. A border officer’s search authority covers what is physically on the device, call history, texts, contacts, photos, videos, the files saved to the phone itself. It does not extend to data sitting only on a remote server. In the United States, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.nbcnews.com/news/us-news/border-patrol-says-its-barred-searching-cloud-data-phones-n782416&quot;&gt;CBP stated this directly&lt;/a&gt; in a 2017 letter from its acting commissioner to Senator Ron Wyden, confirming that a border search does not reach information located solely in the cloud, including email and social media content that lives on remote servers. Officers can search a phone without a warrant or suspicion in most cases, but only the content stored on it.&lt;/p&gt;
&lt;p&gt;That is the whole basis for moving your data home. If your photos, messages, and files live on a server you control and the phone holds almost nothing, then a device search finds almost nothing, because the data is not on the device to find. You also do not have to hand over passwords, though refusing can mean the phone gets detained. Rules differ by country and change over time, and the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.eff.org/issues/border-searches&quot;&gt;Electronic Frontier Foundation&lt;/a&gt; keeps current guidance worth reading before you travel. The point is that the architecture, data at home, thin phone in hand, lines up with where the legal line actually sits.&lt;/p&gt;
&lt;h2 id=&quot;why-a-used-server-not-a-new-pc&quot;&gt;Why a used server, not a new PC&lt;/h2&gt;
&lt;p&gt;A new computer with real power is painful to buy right now, because memory and SSD prices have gone vertical. DDR4 memory rose more than 50% in a single quarter in 2026, and SSDs are climbing on the same pressure, as manufacturers divert capacity to the chips feeding AI data centers. Building a fresh high spec desktop in this market is expensive.&lt;/p&gt;
&lt;p&gt;Used enterprise servers are the way around it. These machines were built to run in data centers, they are far more powerful than a consumer laptop of the same age, and large companies dump them on the secondhand market when their leases end, usually after 5 to 8 years. A server that cost thousands new sells for a few hundred dollars today. The catch is that a server is not built like a desktop. It has no monitor or keyboard, it runs headless in a closet or a garage, and it is loud, 8 fans at full tilt sound like a small aircraft. Put it somewhere the noise will not bother you and that stops mattering.&lt;/p&gt;
&lt;h2 id=&quot;the-base-machine&quot;&gt;The base machine&lt;/h2&gt;
&lt;p&gt;The server I would pick is the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.storagereview.com/review/dell-emc-poweredge-r640-review-intel-scalable-2nd-gen&quot;&gt;Dell PowerEdge R640&lt;/a&gt;, a 1U rack server from the 5 to 8 year old range. It takes 2 processor sockets, and the ones you want are a pair of Intel Xeon Gold CPUs, which together give you dozens of cores. It holds up to 24 memory modules for a maximum of 3 TB of RAM, and has up to 10 front bays for 2.5 inch drives. That is far more machine than a home setup needs, which is the point, there is enormous room to grow.&lt;/p&gt;
&lt;p&gt;On eBay you can find a bare R640 for around $238. Bare means it comes with the case, the 8 cooling fans, 2 power supplies rated 750 to 1,000 watts each, and the 2 CPUs, but no memory and no drives. Read each listing carefully, because what is included varies, and the network card it ships with is often a high speed 25 gigabit data center card that your home router cannot use. You want a plain 1 gigabit network card instead, which costs about $22 and works with any home router.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.Mgxmm27n_Z2jlV9L.webp&quot; alt=&quot;Figure 1. What goes into the roughly $1,000 build. The bare server with its 2 Xeon CPUs is a small share of the cost, and memory is the single most expensive line in the current shortage.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1165&quot; height=&quot;647&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. What goes into the roughly $1,000 build. The bare server with its 2 Xeon CPUs is a small share of the cost, and memory is the single most expensive line in the current shortage. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.ebay.com/&quot;&gt;Author chart based on example eBay listings, 2026&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;memory-the-expensive-part&quot;&gt;Memory, the expensive part&lt;/h2&gt;
&lt;p&gt;The R640 comes with no memory, so you buy modules that fit, and in this market this is the line that hurts most. The reasonable choice is 16 GB DDR4 RDIMM 2133 modules, which are plentiful on the used market and therefore still affordable at roughly $60 or more each, far below the jump in newer DDR5 prices. 6 of them give you 96 GB, a good starting point that runs plenty of virtual machines with room to expand.&lt;/p&gt;
&lt;p&gt;One rule matters here. The server has 2 CPUs, and each should get an equal amount of memory, so you install modules in matched sets, 3 sticks of 16 GB per CPU in this case. You can grow the same size module up to a maximum of 384 GB later, or move to 32 GB sticks if you want to go further. At current prices, budget around $420 for the starting 96 GB.&lt;/p&gt;
&lt;h2 id=&quot;storage-a-split-strategy&quot;&gt;Storage, a split strategy&lt;/h2&gt;
&lt;p&gt;The storage approach saves real money and avoids a common mistake. Use a small, brand new SSD as the boot drive, and cheap magnetic hard drives for the bulk storage. The reason is that SSDs wear out under heavy writes and have a limited lifespan, so a used one that has been hammered in a data center can fail on you. A magnetic drive is built to run 24/7 for years and does not wear the same way, so a used magnetic drive is a safe buy.&lt;/p&gt;
&lt;p&gt;For the boot drive you want a 2.5 inch SATA SSD, brand new. A 256 GB model runs about $50, and 500 GB if the budget allows. The boot drive only holds the operating system, which is mostly read and rarely written, so a modest SSD is all you need. For mass storage, use 2.5 inch SAS magnetic drives. A 2 TB drive is inexpensive, and an 8 TB is barely more, so you can start with one and add more into the empty bays over time. Each drive needs a caddy, a tray that slots it into the front of the server, at about $14 for the 2.5 inch kind, and you only need 2 to begin, one for the SSD and one for the hard drive. If you want the empty bays to look tidy, blank covers run about $5 for a set.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.CMrzRzAl_R38Ua.webp&quot; alt=&quot;Figure 2. The storage split that keeps cost down and reliability up. A brand new SSD holds the operating system, which is mostly read, while cheaper magnetic drives handle bulk storage and are built to run constantly for years.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1014&quot; height=&quot;550&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The storage split that keeps cost down and reliability up. A brand new SSD holds the operating system, which is mostly read, while cheaper magnetic drives handle bulk storage and are built to run constantly for years. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.storagereview.com/review/dell-emc-poweredge-r640-review-intel-scalable-2nd-gen&quot;&gt;Author diagram of the boot plus bulk storage approach&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-it-all-costs&quot;&gt;What it all costs&lt;/h2&gt;
&lt;p&gt;Add it up and the numbers land close to the target. The server is about $250, memory about $420, the 1 gigabit network card $22, the SSD $50, a SAS hard drive $150, a caddy $14, and an optional mounting kit $37. That totals $943. Add tax and you are still near $1,000. The power cord is effectively free, because you use any standard computer cord you have lying around, the data center cords the server ships with are not made for a home outlet.&lt;/p&gt;
&lt;p&gt;What you end up with is a rack server with 2 Xeon processors, dozens of cores, 96 GB of RAM, a 256 GB SSD, and 8 TB of hard drive storage. I would comfortably run 10 virtual machines on that at once without the machine straining, though it will still roar like a jet while it does.&lt;/p&gt;
&lt;h2 id=&quot;the-running-cost-nobody-mentions&quot;&gt;The running cost nobody mentions&lt;/h2&gt;
&lt;p&gt;The purchase price is only part of the bill. A 1U server like this runs 24/7, and these machines idle somewhere around 100 to 180 watts, climbing higher under load. At a US electricity rate near $0.14 per kWh, a used rack server left on all day costs roughly $20 to $90 a month to run, depending on the hardware and how hard you push it. Over a year that can quietly exceed what you paid for the server, so it belongs in the budget from the start. If your power is expensive, this changes the math more than the sticker price does.&lt;/p&gt;
&lt;p&gt;Noise is the other cost. The small, fast fans in a 1U chassis are loud, especially at boot, which is why a closet, garage, or basement is the right home for one. A 2U server with larger, slower fans is noticeably quieter for the same work, and a tower server quieter still, idling around 35 to 45 decibels, closer to a normal desktop. If the machine has to live in earshot, that tradeoff is worth paying for.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.D-2iqjb4_Z1ziVH3.webp&quot; alt=&quot;Figure 3. The 2 ongoing costs of a home server, power and noise. A 1U rack server is cheapest to buy and loudest to live with, while tower servers and mini PCs cost more per unit of power but are quiet enough for a room you use.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1150&quot; height=&quot;633&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The 2 ongoing costs of a home server, power and noise. A 1U rack server is cheapest to buy and loudest to live with, while tower servers and mini PCs cost more per unit of power but are quiet enough for a room you use. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://ovrclock.com/best-used-enterprise-servers-for-home-labs/&quot;&gt;Author chart based on used server power and noise ranges&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;when-a-quieter-smaller-machine-is-the-better-call&quot;&gt;When a quieter, smaller machine is the better call&lt;/h2&gt;
&lt;p&gt;A loud 1U server is not for everyone, and there are 2 honest alternatives. A tower server, something like a Dell T340, does similar work at living room noise levels and fits under a desk, for a somewhat higher price. A mini PC draws a fraction of the power, 15 to 30 watts against about 150 for the rack server, makes almost no noise, and handles a Nextcloud instance, backups, and a couple of light virtual machines without complaint. What a mini PC gives up is headroom, it will not run 10 virtual machines or record several 4K camera streams at once. If your goal is the travel privacy setup and a personal cloud, a mini PC may be all you need, and it will cost you a few dollars a month to run. The big rack server earns its place when you want compute to grow into, running local AI models, recording several cameras, and hosting many services at once.&lt;/p&gt;
&lt;h2 id=&quot;the-operating-system-that-makes-it-a-cloud&quot;&gt;The operating system that makes it a cloud&lt;/h2&gt;
&lt;p&gt;The software that turns this into a private cloud is &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.proxmox.com/en/proxmox-virtual-environment/overview&quot;&gt;Proxmox&lt;/a&gt;, a free Linux distribution built to run virtual machines. You install it once from a USB stick, using a display and keyboard only during that initial setup, and after that the server runs headless. You control everything from a browser on another computer. Proxmox lets you run many separate virtual machines on the one box, each behaving like its own computer, so a single server can act like 10.&lt;/p&gt;
&lt;p&gt;There is an easy path and a hobbyist path here. If you enjoy the tinkering, install each service by hand and learn how it works. If you would rather skip the deep configuration, set up one virtual machine running an AI agent, point it at a cloud model subscription for around $20 a month, and let the agent install and configure the rest for you. The only real effort then is standing up that first agent machine.&lt;/p&gt;
&lt;h2 id=&quot;what-to-actually-run-on-it&quot;&gt;What to actually run on it&lt;/h2&gt;
&lt;p&gt;A handful of free, self hosted applications cover most of what people want.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://nextcloud.com/&quot;&gt;Nextcloud&lt;/a&gt; is your own replacement for iCloud, Google Drive, and OneDrive, holding your photos and files on your own hardware, not a company’s. This is the heart of the travel strategy, because if your photos and messaging data live in Nextcloud, you can delete the Nextcloud app from your phone before a border and there is nothing on the device pointing to where your data is. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.home-assistant.io/&quot;&gt;Home Assistant&lt;/a&gt; runs your home automation and, more usefully, your security cameras, recording 24/7 footage from several 4K cameras straight to the server, which has plenty of space for it. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://syncthing.net/&quot;&gt;Syncthing&lt;/a&gt; handles continuous backups of your phone’s photos and files, so the device can be wiped at any time without losing anything. An AI agent running locally can tie these together, and paired with Home Assistant’s voice interface you can speak to it and have it control lights, cameras, and the rest.&lt;/p&gt;
&lt;h2 id=&quot;the-travel-security-workflow&quot;&gt;The travel security workflow&lt;/h2&gt;
&lt;p&gt;Here is how the pieces combine into the border crossing use case. Say your private conversations are on Signal. Before you fly, you back up Signal to your home server, verify the backup worked, then delete Signal from your phone, along with the media on the device, which your server has already been backing up automatically. You also remove the remote access app, so nothing on the phone reveals that a home server exists. You cross the border with a phone that holds almost nothing. On arrival, you reinstall Signal and restore it from the backup, and reconnect to your server.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.D1nRAE9o_Z17nYBC.webp&quot; alt=&quot;Figure 4. The travel workflow. Before crossing a border, your data moves to the home server and the apps come off the phone, so the device reveals nothing. After arrival, you reconnect and restore.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;936&quot; height=&quot;550&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The travel workflow. Before crossing a border, your data moves to the home server and the apps come off the phone, so the device reveals nothing. After arrival, you reconnect and restore. Source: &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tailscale.com/&quot;&gt;Author diagram of the travel privacy workflow&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;reaching-it-from-anywhere&quot;&gt;Reaching it from anywhere&lt;/h2&gt;
&lt;p&gt;The connection that makes all of this usable from the other side of the world is &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://tailscale.com/&quot;&gt;Tailscale&lt;/a&gt;. You register the server and each virtual machine on it with Tailscale, then install Tailscale on your phone, and every device behaves as though it is on the same local network no matter where you physically are. You get a private address for each machine, visible only to you, and you reach your private cloud over an encrypted link as if you were sitting at home. While traveling through a border, you remove the Tailscale app too, then reinstall it on the far side.&lt;/p&gt;
&lt;h2 id=&quot;if-you-are-not-ready-to-buy-a-server&quot;&gt;If you are not ready to buy a server&lt;/h2&gt;
&lt;p&gt;You do not strictly need to buy anything. If you have an old laptop sitting unused, you can install Proxmox on that and start smaller. The reason I still recommend the full server is value, given a choice between buying a new laptop and buying this far more powerful server for the same roughly $1,000, the server is the better machine by a wide margin, and it leaves room to grow into. The point is to lean less on your phone for your security, and a machine you control at home is how you do that.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>home-server</category><category>gpu</category><category>servers</category><category>ai</category><category>artificial-intelligence</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>A Gas Pipeline Is Breaking America’s $500 Billion AI Bet</title><link>https://soursen.org/articles/a-gas-pipeline-is-breaking-americas-500-billion-ai-bet/</link><guid isPermaLink="true">https://soursen.org/articles/a-gas-pipeline-is-breaking-americas-500-billion-ai-bet/</guid><description>Oracle just moved to defer rent on its flagship Stargate data center before it opens. The reason is not chips or compute, it is a permit fight over a pipe, and the borrowed money behind the project means one delay travels everywhere.</description><pubDate>Tue, 13 Oct 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;One of the biggest AI bets in America is wobbling, and the thing that cracked it is a gas pipeline stuck in a permitting fight in the New Mexico desert. The chips were never the problem. A permitting dispute over where a pipe can cross state land is.&lt;/p&gt;
&lt;p&gt;The site is Project Jupiter, a 1,400 acre campus near Santa Teresa in Dona Ana County, right on the US Mexico border. It is the flagship build of Stargate, the AI infrastructure push unveiled at the White House in January 2025 with OpenAI, Oracle, SoftBank, and MGX as partners. Oracle builds and runs the data centers, OpenAI fills them. On September 24, 2026, before the site is even finished, Oracle &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://techcrunch.com/2026/09/24/oracle-sends-force-majeure-notice-on-its-new-mexico-stargate-data-center/&quot;&gt;sent a force majeure notice&lt;/a&gt; to the developer, securing the right to hold back rent if the campus misses its 2028 target. I want to walk through why a single pipeline can threaten a project valued at up to $165 billion, and why the structure behind it means one delay does not stay in one place.&lt;/p&gt;
&lt;h2 id=&quot;a-very-expensive-shed-without-power&quot;&gt;A very expensive shed without power&lt;/h2&gt;
&lt;p&gt;A data center is only as useful as the power feeding it. Fill a building with the best chips on the planet and give it no electricity, and you have built a very expensive shed. Jupiter was meant to prove Stargate could actually work, gigawatts of compute running in the desert by 2028.&lt;/p&gt;
&lt;p&gt;The ambition kept climbing. The site was first discussed at about 1 GW, then pushed as high as 2.45 GW. No local grid can hand over a load that size, so the campus has to make its own power. In April the developers dropped their gas turbine plan and switched to an on site microgrid of Bloom Energy fuel cells running on natural gas. Fuel cells need fuel, the fuel needs a pipe, and that is where it fell apart.&lt;/p&gt;
&lt;p&gt;The pipeline is an Energy Transfer project meant to feed the site. In March 2026, New Mexico’s State Land Office denied it a right of way across state trust land. The developers asked for a reconsideration, and in July the office said no again. The federal Bureau of Land Management cleared one segment on emergency review in May, but the state refusal forced a reroute that pushed the in service date from this summer to February 1, 2027. The air permit for the fuel cells is still sitting with New Mexico’s environment department, with a decision due by November 23, 2026, and environmental groups have taken the project to court. Land rights, air permits, lawsuits, 3 separate fights, any one of which can stall the whole thing.&lt;/p&gt;
&lt;p&gt;The workaround Oracle is reportedly weighing shows how awkward the bind is. By early October, Bloomberg reported Oracle was considering trucking compressed natural gas to the site to bring early phases online before the pipeline arrives, the same trick it used for over a year at a data center near Salt Lake City. At Jupiter’s scale the math gets silly fast. One energy analyst cited by Bloomberg estimated that powering just 100 megawatts, roughly 4% of the site’s planned capacity, by trucked gas would burn through a large compressed gas trailer about every 40 minutes.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.Bl1Rmf5J_1FO519.webp&quot; alt=&quot;Figure 3. The timeline that turned a 2028 target into a rent deferral notice.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1185&quot; height=&quot;516&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 3. The timeline that turned a 2028 target into a rent deferral notice. Source: Bloomberg, DCD, Capacity, 2026&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;what-force-majeure-actually-says&quot;&gt;What force majeure actually says&lt;/h2&gt;
&lt;p&gt;Force majeure is a legal notice that events outside your control are getting in the way of a contract. Oracle sent it to the Blue Owl Capital unit developing the site, and it preserves the right to defer lease payments by up to 3 years if Jupiter misses its 2028 target. In plain terms, Oracle is saying there is a real chance it will not pay rent on schedule if the power situation does not change.&lt;/p&gt;
&lt;p&gt;Everyone involved is downplaying it. Oracle says Jupiter remains on schedule and that a force majeure notice does not, by itself, establish a delay. Blue Owl says the notice does not change the financial commitments to the project and that the 2 sides remain fully aligned. One person familiar with the financing &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://capacityglobal.com/news/165bn-project-jupiter-what-oracles-force-majeure-notice-means-for-ai-data-centre-risk/&quot;&gt;told Bloomberg&lt;/a&gt; it does not amount to an event of default. I think the notice still tells you something, because companies that are confident about a deadline do not usually paper the file with legal protection against missing it.&lt;/p&gt;
&lt;p&gt;The most hyped AI project in America is being held back by paperwork and a pipe, not by any shortage of compute. That is the part of the AI trade people keep forgetting. Software scales at the push of a button. Power and permits do not. They take time, sometimes years, and they answer to state land commissioners and judges, not to a product roadmap.&lt;/p&gt;
&lt;h2 id=&quot;who-is-actually-paying-for-this&quot;&gt;Who is actually paying for this&lt;/h2&gt;
&lt;p&gt;Follow the money and the fragility becomes obvious. Project Jupiter runs on borrowed cash. Blue Owl put in roughly $3 billion of equity, and a group of about 20 banks arranged an $18 billion construction loan.&lt;/p&gt;
&lt;p&gt;Then there is Oracle. To build capacity for OpenAI, Oracle’s spending has exploded. In its fiscal first quarter of 2027, the quarter that ended in August 2026, Oracle reported about $28 billion of gross capital expenditure, up from roughly $8.5 billion a year earlier, against just $19.3 billion of revenue. Free cash flow came in around negative $5 billion despite record operating cash flow, because the capex simply swamped it. Oracle’s total debt sits near $156 billion, and management has guided to $90 to $95 billion of gross capex for the full fiscal year. Oracle is funding the gap with debt, equity, and customer prepayments. The bet it is funding is enormous. In September 2025 Oracle and OpenAI signed a roughly $300 billion, 5 year compute contract set to ramp from 2027, and Oracle has reported a remaining performance obligation backlog of about $664 billion. That backlog is the number bulls point to. It is also a promise that only turns into cash if sites like Jupiter get built and switched on.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.B8FRdooB_28Qfgs.webp&quot; alt=&quot;Figure 1. Oracle is spending roughly $28B a quarter against $19B of revenue to build for OpenAI.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1185&quot; height=&quot;653&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 1. Oracle is spending roughly $28B a quarter against $19B of revenue to build for OpenAI. Source: Oracle fiscal Q1 2027 earnings, Sep 2026&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;Then there is the customer. OpenAI is the one meant to pay Oracle, and OpenAI runs largely on investor money rather than its own profits. The chain is simple to say and alarming to look at. Banks lend to the developer. Oracle borrows to build. OpenAI raises money to pay Oracle. Every link depends on the next one’s cash arriving on time, so a single delay anywhere pushes stress through the whole structure.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.1YnDjc1d_1pEwEY.webp&quot; alt=&quot;Figure 2. The money chain behind Project Jupiter, each link leaning on the next.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1185&quot; height=&quot;525&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 2. The money chain behind Project Jupiter, each link leaning on the next. Source: Bloomberg and Reuters reporting, 2026&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;A site that is not running earns nothing, while the interest on the loans keeps accruing every day. That is what the force majeure notice really does. If Oracle defers rent, the cost of the delay does not vanish, it moves down the chain to the lenders and equity holders. The lenders can see it coming. That $18 billion construction loan has been trading at roughly 89 to 91 cents on the dollar, below par, which is the market saying it is not sure the money comes back whole. Oracle’s own credit is feeling it too, with S&amp;amp;P at BBB, and the cost of insuring Oracle debt against default hit a record in late September.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-04.BodMtagi_1P7Xq4.webp&quot; alt=&quot;Figure 4. The strain is already visible in the loan price, the stock, and the founder’s paper wealth.&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1110&quot; height=&quot;587&quot;&gt;&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Figure 4. The strain is already visible in the loan price, the stock, and the founder’s paper wealth. Source: CNBC and Bloomberg, 2026&lt;/em&gt;&lt;/p&gt;
&lt;h2 id=&quot;the-founder-borrowed-against-the-thing-too&quot;&gt;The founder borrowed against the thing too&lt;/h2&gt;
&lt;p&gt;Oracle’s corporate debt is only half of it. Larry Ellison, who founded Oracle in 1977 and still controls more than 40% of it, has his own leverage tied to the same stock. As of September 21, 2026, per Oracle’s proxy filed that month, he had pledged 413 million Oracle shares as collateral for personal loans, up from 346 million a year earlier, roughly 36% of his stake and worth about $9.2 billion at the time. A pledge works like a mortgage where the house is Oracle stock. Oracle notes these are personal term loans, not margin accounts, and says its board reviews the arrangements. If the share price falls far enough, the lenders can demand more collateral. His fortune and Oracle’s stock price move together, and when the stock drops, both he and the company feel it at once.&lt;/p&gt;
&lt;p&gt;The timing of what happened next is the part I keep turning over. On June 22, 2026, Ellison adopted a Rule 10b5–1 plan, a prearranged schedule to sell shares automatically, this one for up to 50 million shares worth about $7.5 billion at the price at the time. When the plan surfaced in a filing on September 11, investors saw the founder lining up a multibillion dollar sale. One day later, on September 12, Oracle &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.cnbc.com/2026/09/12/larry-ellison-nixes-plan-to-offload-up-to-7point5-billion-worth-of-oracle-stock-.html&quot;&gt;announced the plan was cancelled&lt;/a&gt; with no shares sold and no reason given. Oracle does not explain these moves, and there are ordinary tax and estate reasons a person cancels a sale plan. Ellison’s net worth has also dropped by roughly $200 billion over the past year as Oracle stock fell from its peak, so every wobble in the company is a bigger wobble for him personally. Borrowed money stacked on borrowed money, again.&lt;/p&gt;
&lt;h2 id=&quot;why-these-are-not-3-separate-problems&quot;&gt;Why these are not 3 separate problems&lt;/h2&gt;
&lt;p&gt;Lay it out and it looks like 3 distinct issues. A site in the desert that cannot get its gas. A company borrowing tens of billions to build it. A founder borrowing against the shares that depend on it working. They are the same problem wearing 3 coats.&lt;/p&gt;
&lt;p&gt;The stalled site threatens the revenue. The revenue services the debt. The debt weighs on the stock. The stock backs the founder’s personal loans. Pull any one thread and the others move. A delay at a pipeline in New Mexico travels, through Oracle’s rent, through its bond prices, through its share price, all the way to the collateral behind one man’s personal borrowing. That is the AI boom’s weak point, and it is showing up most clearly in the project that was supposed to prove the boom works.&lt;/p&gt;
&lt;p&gt;Stargate was announced at the White House as a statement of intent worth up to $500 billion across its full build. Its flagship campus is now deferring rent before it opens, its construction loan trades below par, and its anchor tenant is spending $28 billion a quarter against $19 billion of revenue. The build may still get finished and fill up on schedule, and if it does, all this leverage stays invisible and everyone looks smart. If the gas shows up late, the cost lands on lenders and shareholders rather than on the companies that ordered the compute. The next data point to watch here is the air permit decision due on November 23, not any chip benchmark.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>ai</category><category>oracle</category><category>openai</category><category>data-centre</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Are Data Centers a Good Deal for the Communities That Host Them?</title><link>https://soursen.org/articles/are-data-centers-a-good-deal-for-the-communities-that-host-them/</link><guid isPermaLink="true">https://soursen.org/articles/are-data-centers-a-good-deal-for-the-communities-that-host-them/</guid><description>We have been talking about the cloud as though it were an abstraction for years. Clean, invisible, frictionless, somewhere up there. The actual picture is less romantic.</description><pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;We have been talking about the cloud as though it were an abstraction for years. Clean, invisible, frictionless, somewhere up there. The actual picture is less romantic. Data centers are physical buildings in real places, consuming electricity measured in terawatt-hours, drawing water from local sources, competing for grid capacity, and collecting tax breaks that cost local governments billions of dollars annually. The AI boom has made all of this more urgent, more expensive, and harder to ignore.&lt;/p&gt;
&lt;p&gt;The numbers are worth stating plainly. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai&quot;&gt;According to the International Energy Agency’s Energy and AI report&lt;/a&gt;, global data center electricity consumption is projected to roughly double from 415 TWh in 2024 to 945 TWh by 2030. That is approximately the entire current annual electricity consumption of Japan. AI-specific server electricity demand is growing even faster, at approximately 30% per year, well outpacing the 9% annual growth of conventional server demand. Electricity demand from data centers is now growing at about 15% per year, more than 4 times faster than the growth of total global electricity consumption from all other sectors combined.&lt;/p&gt;
&lt;p&gt;In the United States, the trajectory is sharper still. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/&quot;&gt;The Brookings Institution’s April 2026 analysis&lt;/a&gt; draws on Department of Energy projections showing US data centers consumed about 4.4% of total US electricity in 2023, with projections ranging from 6.7% to 12% by 2028 in absolute terms, a rise from 176 TWh to potentially 580 TWh. By 2030, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea&quot;&gt;S&amp;amp;P Global reports&lt;/a&gt; that data centers in the US are expected to consume more electricity processing data than all energy-intensive manufacturing combined, including aluminium, steel, cement, and chemicals production. This is not a niche infrastructure question anymore.&lt;/p&gt;
&lt;h2 id=&quot;the-promise-vs-the-reality-of-local-economic-benefits&quot;&gt;The Promise vs the Reality of Local Economic Benefits&lt;/h2&gt;
&lt;p&gt;Countries, regions, and cities are racing to attract data centers the way they once raced to attract factories. The scale of investment is genuinely staggering. Tech companies spent an estimated $375 billion on AI infrastructure in 2025, most of it building and expanding data centers, with approximately twice that amount projected for 2026. When a hyperscaler announces a new campus, local officials hold press conferences and describe transformative economic benefits. The numbers look impressive in the headline. They look different when you examine what they actually deliver.&lt;/p&gt;
&lt;p&gt;The core problem is one of labour intensity. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.brookings.edu/articles/new-evidence-on-data-center-employment-effects/&quot;&gt;Brookings published research in May 2026&lt;/a&gt; analysing approximately 770 US data center facilities linked to county-level employment data from 2003 to 2024. The findings are sobering. In a typical county with 98,000 workers, a data center generates approximately 2,000 to 4,000 additional jobs after 6 years, and that figure depends on having multiple facilities in the same area. Single facilities produce modest employment gains. Meanwhile, the most automated hyperscale campuses can operate with as few as 20 to 30 permanent staff per 100 megawatts of capacity.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://goodjobsfirst.org/cloudy-with-a-loss-of-spending-control-how-data-centers-are-endangering-state-budgets/&quot;&gt;Good Jobs First, a nonprofit subsidy watchdog&lt;/a&gt;, has documented the subsidy cost per job across data center projects. The average cost of data center megadeals is $1.95 million per permanent job. In one extreme case, a New York data center promised 125 jobs in exchange for $1.4 billion in public subsidies, or $11 million per job. Compare that to $137,000 per job for economic development incentives in other sectors. Virginia, the largest data center market in the world, generates just 1 permanent job for every $13 million invested in data center infrastructure, according to a January 2026 analysis by Food and Water Watch.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CZf_kvKV_Z1eCj5H.webp&quot; alt=&quot;Illustration for the section “The Promise vs the Reality of Local Economic Benefits”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;808&quot;&gt;&lt;/p&gt;
&lt;p&gt;Public subsidy cost per permanent job created. Data center megadeals average $1.95M per job vs $137K for other economic development programmes. Source: Good Jobs First, June 2026 / Brookings Institution, May 2026.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;The average data center megadeal costs $1.95 million per permanent job 14 times more than normal economic development incentives, often with no job creation requirement at all. By 2030, US data centers are expected to consume more electricity processing data than all energy-intensive manufacturing combined including aluminium, steel, cement, and chemicals. This is not a footnote. It is a policy emergency.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;The construction jobs that accompany data center builds are real and meaningful. They are also temporary. The ribbon-cutting photo opportunity and the years of permanent employment that follow it are very different things, and the incentive packages being offered to hyperscalers are mostly calibrated around the former rather than the latter.&lt;/p&gt;
&lt;h2 id=&quot;the-subsidy-numbers-nobody-wanted-you-to-see&quot;&gt;The Subsidy Numbers Nobody Wanted You to See&lt;/h2&gt;
&lt;p&gt;The opacity surrounding data center tax arrangements is not accidental. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://goodjobsfirst.org/cloudy-data-costly-deals-how-poorly-states-disclose-data-center-subsidies/&quot;&gt;Good Jobs First’s November 2025 report&lt;/a&gt; found that at least 36 states have crafted specific subsidies for data center projects. Only 11 of those states disclose which companies receive those benefits. Virginia, the largest data center market on earth, forgoes close to $1 billion in state and local sales and use tax revenue every year without telling the public which companies benefit or how much each one receives. 14 states with relevant exemptions do not properly disclose their revenue losses at all.&lt;/p&gt;
&lt;p&gt;The dollar amounts being forgone are not minor budget lines. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://goodjobsfirst.org/even-cloudier-with-a-greater-loss-of-spending-control-how-data-center-tax-abatements-undermine-public-budgets/&quot;&gt;Good Jobs First’s June 2026 report&lt;/a&gt; found that Virginia’s data center sales and use tax exemption cost $1.94 billion in fiscal year 2025. Texas projects annual losses exceeding $1.3 billion in FY 2026, rising from $1 billion in FY 2025, with cumulative losses potentially reaching $9 billion between 2025 and 2030. Ohio discovered in May 2026 that its data center tax exemption cost $1.6 billion in 2025, more than 10 times higher than its own earlier projection. Georgia raised its FY 2026 cost projection by 664%, to $2.5 billion. Indiana lost $700 million to data center tax exemptions in 2025, with more than 83% of that subsidy going to a single company.&lt;/p&gt;
&lt;p&gt;The companies stacking these subsidies are not struggling startups. Amazon, Google, Microsoft, and Meta collectively spent an estimated $360 billion on capital expenditures in 2025, mostly building data centers. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.brookings.edu/articles/new-evidence-on-data-center-employment-effects/&quot;&gt;As the Brookings research confirmed&lt;/a&gt;, location decisions for hyperscale facilities are driven by power availability, land, and fibre infrastructure, not by tax breaks. The subsidies are not attracting investment that would not otherwise have come. They are subsidising investment that was already going to happen anyway.&lt;/p&gt;
&lt;p&gt;The backlash to this dynamic is now visible. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.brookings.edu/articles/new-evidence-on-data-center-employment-effects/&quot;&gt;Brookings reports&lt;/a&gt; that more than 100 local communities have enacted moratoriums on data center construction. More than 300 state data center bills were filed in the first 6 weeks of 2026 alone. Virginia, Georgia, and Oklahoma, states that once competed to offer the largest tax incentives, are now reconsidering those programmes entirely. In the PJM grid region, which serves 65 million people across 13 states, power supply costs jumped from $2.2 billion to $14.7 billion in a single year, with data centers accounting for nearly two-thirds of that increase. Residential electricity rates nationally rose approximately 32% between July 2020 and July 2025. The bill is landing with people who did not agree to pay it.&lt;/p&gt;
&lt;h2 id=&quot;the-content-question-nobody-is-asking-loudly-enough&quot;&gt;The Content Question Nobody Is Asking Loudly Enough&lt;/h2&gt;
&lt;p&gt;There is a dimension to the data center expansion story that the energy and jobs debate tends to skip over. The AI models being trained and served inside these facilities did not arrive empty. They were built on the accumulated output of millions of writers, researchers, artists, programmers, journalists, and institutions who created that content without knowing it would be used as training material, and without receiving any compensation for it.&lt;/p&gt;
&lt;p&gt;The same companies now requesting land, electricity, water, and fast-track regulatory approval from local governments are the ones that absorbed this content. The argument that AI represents a genuine public good becomes considerably harder to make when the raw material for that good was taken from the public without its consent, the physical infrastructure to exploit it is being subsidised by public budgets, and the value generated flows to a small number of shareholders in a handful of companies.&lt;/p&gt;
&lt;p&gt;Cloudflare has begun addressing the traffic side of this with its new tools for controlling AI crawlers, described in its &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://blog.cloudflare.com/content-independence-day-ai-options/&quot;&gt;Your Site, Your Rules framework,&lt;/a&gt; and its pay-per-crawl proposal to allow content owners to charge AI crawlers for access. The principle is straightforward: the raw material of AI should not be treated as though it has no owner, no cost, and no rights. Whether that principle gets embedded into regulation or remains a technical opt-in mechanism will determine how equitable the AI economy actually becomes.&lt;/p&gt;
&lt;h2 id=&quot;what-responsible-policy-actually-looks-like&quot;&gt;What Responsible Policy Actually Looks Like&lt;/h2&gt;
&lt;p&gt;The European Commission has a framework in place for the energy performance of data centers, linked to the &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficiency-targets-directive-and-rules/energy-efficiency-directive_en&quot;&gt;Energy Efficiency Directive&lt;/a&gt;. Measurement is a necessary start. It is not sufficient. Setting conditions on what data centers must deliver in exchange for public support is a different, and harder, policy task.&lt;/p&gt;
&lt;p&gt;The questions any government should be asking before approving a major data center project are not complicated. How much electricity capacity does this project consume, and what does that mean for local grid availability and residential rates? How much water does it require, and is that consistent with regional supply constraints? How many permanent jobs does it create, with what wages and under what employment conditions? What tax revenue does the local government forgo, and does that figure appear anywhere in a public document? If the renewable energy being directed to this facility could have been used to decarbonise local homes or industry, what is the opportunity cost of routing it to a proprietary AI system instead?&lt;/p&gt;
&lt;p&gt;The Sanders and Ocasio-Cortez AI Data Center Moratorium Act, introduced in March 2026, proposed pausing all new large-scale AI data center construction until a federal review framework is in place. Whether or not that specific bill advances, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.brookings.edu/articles/new-evidence-on-data-center-employment-effects/&quot;&gt;the Brookings research&lt;/a&gt; makes the case for the underlying principle clearly. Policy is currently moving faster than evidence. Decisions about billions of dollars in public subsidies, grid capacity, and land use are being made on the basis of industry projections that overstate employment benefits by a factor of 3, according to the May 2026 academic research.&lt;/p&gt;
&lt;p&gt;Good Jobs First’s recommendation is direct. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://goodjobsfirst.org/even-cloudier-with-a-greater-loss-of-spending-control-how-data-center-tax-abatements-undermine-public-budgets/&quot;&gt;States should sunset or substantially reform their data center sales tax exemptions&lt;/a&gt;, add caps on total subsidy exposure per company, require full disclosure of costs and benefits, and stop treating capital-intensive automation as equivalent to the manufacturing employment it superficially resembles. The 36-state subsidy framework for data centers was written for a different era. It was not designed to hand billion-dollar annual tax forgiveness to some of the most profitable companies in history.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;AI can be a genuinely useful technology. Precisely for that reason, it should not be deployed under an extractive, opaque model subsidised by the public. If a company wants to train models using publicly created content, it should pay for that content. If it wants to build data centers in a community, it should cover the full costs. If it wants to consume water, electricity, and grid capacity, it should demonstrate that the social return justifies that consumption. That is not anti-innovation. It is the basic accounting that any other major industry is expected to provide.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;The cloud has a geography. It consumes real resources in real places and creates real winners and losers. The communities asked to host it deserve a clear-eyed answer to a simple question: what do we actually get in return?&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>data-center</category><category>ai</category><category>artificial-intelligence</category><category>gpu</category><category>web3</category><author>hello@soursen.org (Shivashish Yadav)</author></item><item><title>Moore’s Law Is Dying.</title><link>https://soursen.org/articles/moores-law-is-dying/</link><guid isPermaLink="true">https://soursen.org/articles/moores-law-is-dying/</guid><description>2 nanometers. That is approximately 10 to 20 atoms wide, depending on the element. TSMC began volume production of 2nm chips in Q4 2025, ramping hard through 2026, with initial yield rates of approximately 70%.</description><pubDate>Fri, 10 Jul 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;2 nanometers. That is approximately 10 to 20 atoms wide, depending on the element. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://heqingele.com/blog/tsmc-2nm-yield-rates-mass-production-status-2026/&quot;&gt;TSMC began volume production of 2nm chips in Q4 2025&lt;/a&gt;, ramping hard through 2026, with initial yield rates of approximately 70%. The entire production capacity of TSMC’s first 2 dedicated 2nm fabs is booked through the end of the year. Apple holds more than half of that initial allocation, with its A20 processor for the next iPhone generation, its M6 chip for Mac, and the Vision Pro R2 wearable chip all queued for the N2 node. AMD, Qualcomm, Nvidia, and MediaTek are directly behind.&lt;/p&gt;
&lt;p&gt;&lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://drrobertcastellano.substack.com/p/tsmcs-2-nm-era-technology-leadership&quot;&gt;TSMC’s 2nm capacity is expanding fast&lt;/a&gt;, from 40,000 wafers per month at the end of 2025 to a target of 100,000 per month in 2026 and 200,000 per month by 2027. The node uses gate-all-around nanosheet transistors combined with backside power delivery, a combination that resets the power-performance curve in ways that FinFET scaling could not. This is genuinely impressive engineering.&lt;/p&gt;
&lt;p&gt;Then, in the space of about a week in late June 2026, 3 separate organisations announced things that collectively suggest TSMC’s achievement, extraordinary as it is, may be the last act of one era rather than the opening of another. Huawei, IBM, and OpenAI each announced something different, from 3 different directions, and each one is saying the same underlying thing: shrinking transistors sideways is no longer the primary lever of progress.&lt;/p&gt;
&lt;p&gt;What is striking is that none of them sound alarmed. They sound like engineers who have already figured out what comes next.&lt;/p&gt;
&lt;p&gt;Before getting into the 3 announcements, you need to understand why Moore’s Law is actually running out, because the explanation in most coverage is wrong. The problem is not that transistors are getting too small to shrink further. The problem is where the energy goes.&lt;/p&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-01.CG7gYehp_l4fye.webp&quot; alt=&quot;Illustration for the section “Moore’s Law Is Dying.”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;984&quot; height=&quot;892&quot;&gt;&lt;/p&gt;
&lt;p&gt;More than 80% of the energy consumed by a modern high-performance chip does not go toward computation. It goes toward moving information between components inside the chip. This is the number that the product launch presentations never mention, because it is not a marketable feature. It is a fundamental constraint.&lt;/p&gt;
&lt;p&gt;Think of a chip as a city. Transistors are the buildings where the actual work happens. The microscopic metal wires connecting them are roads. A modern chip has more than 20 layers of those roads stacked on top of each other. Every signal traveling through those wires runs into 2 forces working against it: the resistance of the metal, and what engineers call parasitic capacitance, an invisible charge effect along the wire that has to cycle with every piece of information that passes through. The longer the path, the worse both effects compound. Engineers call this the RC delay, and it is the actual bottleneck in modern chips, not transistor size.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Processor clock speeds have stagnated for 15 years not because transistors stopped shrinking but because the paths between them are still too long.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;This is why processor clock speeds have stagnated for roughly 15 years despite transistor counts continuing to grow. You can pack more transistors onto the chip, but if the signals still have to travel long distances to reach each other, you gain density without gaining speed. The solution that Huawei, IBM, and OpenAI are each converging on is the same one: stop building wider and start building upward, compressing the distances signals need to travel.&lt;/p&gt;
&lt;h2 id=&quot;huaweis-7-year-detour-that-might-redefine-the-industry&quot;&gt;Huawei’s 7-Year Detour That Might Redefine the Industry&lt;/h2&gt;
&lt;p&gt;In June 2026, at the International Circuit and Systems Symposium in Shanghai, He Tingbo, president of HiSilicon, Huawei’s chip design arm, presented a new architectural approach called Logic Folding. This was He Tingbo’s first major public appearance in years, during a period when Huawei has been largely cut off from the equipment that TSMC uses to make chips at advanced nodes. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://en.wikipedia.org/wiki/Extreme_ultraviolet_lithography&quot;&gt;ASML’s EUV lithography machines&lt;/a&gt;, which cost hundreds of millions of dollars each and require export licences that US sanctions have blocked, are essential for patterning features below roughly 7nm with a single exposure pass. SMIC, Huawei’s foundry partner, has been compensating by making multiple passes over the same pattern: 3, 4, sometimes 5 successive exposures where TSMC does it in 1. This works up to a point. Every additional pass adds defects and manufacturing complexity.&lt;/p&gt;
&lt;p&gt;That ceiling forced a genuinely important question. Is horizontal transistor shrinkage actually the right direction? Huawei’s answer is no, and Logic Folding is their alternative. The concept takes a circuit that normally spreads flat across silicon and folds it vertically, like folding a piece of paper in half. Stack the logic layers on top of each other and components that were separated by millimetres of horizontal distance are now separated by microns of vertical distance. The signal path collapses. The RC delay that was limiting performance collapses with it.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;Huawei claims 55% higher transistor density and 41% better energy efficiency without a single EUV machine. If the thermal challenge is solved, sanctions may have accidentally accelerated the next era of chip design.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Alongside Logic Folding, Huawei introduced what they call the Tau law, a new framework for measuring chip progress not by transistor width but by signal propagation time across the system. Progress measured in time rather than in nanometres. The claimed performance numbers are striking: up to 55% higher transistor density and 41% better energy efficiency without a single EUV machine in the manufacturing chain. Their upcoming Kirin 2026 processor, the first commercial chip using this architecture and expected in late 2026, is reportedly targeting 238 million transistors per square millimetre. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://semiwiki.com/forum/threads/tsmcs-2nm-chips-the-results-are-out.24329/&quot;&gt;That figure would be comparable to TSMC’s first-generation 3nm node&lt;/a&gt;, achieved through vertical integration rather than lateral shrinkage.&lt;/p&gt;
&lt;p&gt;The legitimate engineering challenge embedded in this claim is worth naming. Communication between 2 stacked active logic layers at the required speeds depends on millions of tiny copper connections called bonding interconnects. Huawei says they are targeting a bonding pitch of 1.5 microns, compared to roughly 9 microns at the current industry best. That is 6 times tighter than anything that exists today in production. Managing the heat generated by 2 stacked active logic layers inside a smartphone form factor, without the liquid cooling that server racks can use, is the other constraint that needs to be solved.&lt;/p&gt;
&lt;p&gt;The historical precedent for constraints producing innovation is strong enough that this does not feel like a reason to dismiss the approach. When the US cut GPS signals to Europe during the Kosovo conflict in 1999, Europe built the Galileo satellite navigation system. When the 1973 oil embargo hit Japan, Japan became the world leader in fuel efficiency. Huawei’s CEO has said publicly that the sanctions forced a direction Huawei would not have taken voluntarily.&lt;/p&gt;
&lt;h2 id=&quot;ibms-07-nanometre-chip-and-what-the-number-actually-means&quot;&gt;IBM’s 0.7 Nanometre Chip and What the Number Actually Means&lt;/h2&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-02.zYpbD4Rk_Z1xCeuJ.webp&quot; alt=&quot;Illustration for the section “IBM’s 0.7 Nanometre Chip and What the Number Actually Means”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1488&quot; height=&quot;811&quot;&gt;&lt;/p&gt;
&lt;p&gt;On June 25, 2026, &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://newsroom.ibm.com/2026-06-25-ibm-debuts-worlds-first-sub-1-nanometer-chip-technology&quot;&gt;IBM announced the world’s first sub-1 nanometre semiconductor technology&lt;/a&gt;, presented at the VLSI 2026 Symposium in Kyoto. The node is called 0.7nm, or 7 angstroms. 7 angstroms is approximately 7 atoms wide.&lt;/p&gt;
&lt;p&gt;Before going further, the naming caveat is worth stating clearly. A node name in semiconductors is no longer a physical measurement. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://thenextweb.com/news/ibm-sub-1-nanometer-chip-nanostack&quot;&gt;The Next Web’s analysis of the announcement&lt;/a&gt; notes this explicitly: a 0.7nm node does not mean any feature on the chip is 0.7 nanometres across. It is a generation label. What the number signals is relative progress against prior generations, and on that basis the IBM announcement is significant.&lt;/p&gt;
&lt;p&gt;The architecture IBM unveiled is called Nanostack. Instead of continuing to shrink transistors laterally, Nanostack builds them in 2 stacked vertical layers on the same wafer, one directly on top of the other, fabricated sequentially. IBM demonstrated a working CMOS inverter, confirming the structure can be physically built and can switch. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://newsroom.ibm.com/2026-06-25-ibm-debuts-worlds-first-sub-1-nanometer-chip-technology&quot;&gt;IBM’s press release confirmed&lt;/a&gt; approximately 100 billion transistors on a fingernail-sized die, roughly twice the density of IBM’s own 2nm chip from 2021, with up to 50% higher performance or 70% better energy efficiency compared to that 2nm predecessor.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;IBM’s Nanostack delivered 40% SRAM scaling in a single step solving a problem the industry had quietly given up on since TSMC’s 3nm node.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;The SRAM result may be the most practically important finding for AI specifically. SRAM is the ultra-fast on-chip memory that lives inside a processor and serves as the buffer for the most computationally intense operations. It had essentially stopped shrinking at TSMC’s 3nm node. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://letsdatascience.com/news/ibm-debuts-07nm-nanostack-with-nearly-100b-transistors-f05b42f2&quot;&gt;IBM’s Nanostack delivers 40% SRAM scaling&lt;/a&gt; in a single step. IBM projects Nanostack-based AI accelerators could reach 7,000 TOPS, compared to approximately 1,500 TOPS in current leading-edge chips, and could cut large language model training time from roughly 3 months to 2 weeks.&lt;/p&gt;
&lt;p&gt;The honest context is that this is a laboratory result, not a product. IBM does not manufacture logic chips at commercial volume. It licenses its architectures to partners. The 2nm chip IBM announced in 2021 has only recently approached broad commercial deployment, 5 years after its announcement. Nanostack production is probably 5 years from commercial silicon. But laboratory demonstrations in this industry are the map, and this map extends below 1 nanometre for the first time with validated hardware.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;IBM Research director Jay Gambetta framed it directly:&lt;/em&gt; “With our new nanostack architecture, we’re not just making smaller transistors, we’re reinventing how chips are built to deliver dramatically more power and energy efficiency.”&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;openais-jalapeño-the-chip-that-ai-designed-in-9-months&quot;&gt;OpenAI’s Jalapeño: The Chip That AI Designed in 9 Months&lt;/h2&gt;
&lt;p&gt;On June 24, 1 day before IBM’s announcement, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom-built inference processor. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.bloomberg.com/news/articles/2026-06-24/openai-broadcom-custom-ai-chip-jalape%C3%B1o&quot;&gt;Bloomberg reported that Broadcom CEO Hock Tan confirmed a 50% reduction in cost per inference&lt;/a&gt; compared to current Nvidia GPU-based inference, with first deployments planned before the end of 2026.&lt;/p&gt;
&lt;p&gt;&lt;em&gt;Jalapeño went from initial concept to manufacturing tape-out in 9 months. A design of this complexity typically takes 2 to 3 years. AI compressed the timeline by a factor of 3 by helping design the chip it will run on.&lt;/em&gt;&lt;/p&gt;
&lt;p&gt;What is genuinely novel about Jalapeño is not just the performance specification. It is the design timeline. A chip of this complexity would typically require 2 to 3 years from initial design to manufacturing tape-out. Jalapeño went from initial concept to tape-out in 9 months. The acceleration came from using OpenAI’s own models to assist with significant portions of the design and optimisation process. AI accelerated the design of the chip that AI will run on.&lt;/p&gt;
&lt;p&gt;The loop this creates is worth pausing on. AI designs a chip. The chip runs AI more efficiently. That AI helps design the next chip more quickly. Each generation compresses the timeline of the next. If the 50% inference cost reduction holds in production, and if the design cycle continues to compress, the economics of running large language models at scale change faster than the conventional semiconductor roadmap would suggest.&lt;/p&gt;
&lt;p&gt;OpenAI is also not alone in this direction. Google has been designing its own TPU chips for over a decade and is now on its 7th generation. Amazon has its Trainium and Inferentia lines. Microsoft is developing its Maia series. Meta has released details of its MTIA inference chip. The hyperscalers collectively represent an enormous shift of AI compute demand away from commodity GPU hardware toward purpose-built silicon.&lt;/p&gt;
&lt;h2 id=&quot;what-all-3-are-actually-saying&quot;&gt;What All 3 Are Actually Saying&lt;/h2&gt;
&lt;p&gt;&lt;img src=&quot;https://soursen.org/_astro/img-03.CQ9fq98T_Z1Fs9TH.webp&quot; alt=&quot;Illustration for the section “What All 3 Are Actually Saying”&quot; loading=&quot;lazy&quot; decoding=&quot;async&quot; width=&quot;1600&quot; height=&quot;818&quot;&gt;&lt;/p&gt;
&lt;p&gt;Set the 3 announcements side by side and the common thread is unmistakable. Huawei is folding logic vertically to collapse signal paths without EUV access. IBM is stacking transistors at the transistor level to break below 1 nanometre. OpenAI is designing custom chips optimised for the specific matrix operations that large language models perform, compressing the design cycle with the same intelligence the chip will eventually run.&lt;/p&gt;
&lt;p&gt;None of them is competing on transistor width. All of them are competing on signal travel time, on vertical density, on design intelligence, and on system-level integration. The era where progress was reliably measured in nanometres of width is closing.&lt;/p&gt;
&lt;p&gt;The broader industry is converging on the same direction. &lt;a rel=&quot;noopener noreferrer&quot; target=&quot;_blank&quot; href=&quot;https://www.amd.com/en/products/processors/server/epyc.html&quot;&gt;AMD has been stacking memory on processors for 3 years&lt;/a&gt; with its 3D V-Cache technology. Intel is routing power delivery through the back of the wafer on its latest process designs. TSMC is preparing its own version of vertical integration for the nodes after N2.&lt;/p&gt;
&lt;p&gt;For AI specifically, this matters in a direct economic way. The cost of running a large language model is determined almost entirely by GPU compute efficiency. If Jalapeño delivers on its 50% inference cost reduction when deployed before year-end, the economics of AI inference shift immediately. If IBM’s Nanostack architecture makes it to production in 5 years, the projected AI accelerator performance of 7,000 TOPS would represent roughly a 5x improvement over current leading hardware. If Huawei’s Logic Folding proves its thermal management thesis, it demonstrates that the path to competitive chip performance no longer requires access to the world’s most expensive lithography equipment.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;The companies that understand what is happening are designing the full stack: the silicon, the memory layout, the packaging, and the power architecture together rather than separately. The advantage in the next decade of AI will not belong to whoever has the largest model. It will belong to whoever can run intelligence most efficiently at scale, and that is now a hardware problem as much as a software one.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;These 3 announcements arrived in the same week by coincidence of timing. That coincidence makes them more interesting, not less. When Huawei, IBM, and OpenAI each independently pivot toward the same architectural direction at roughly the same moment, it is a reasonable signal that the direction is correct.&lt;/p&gt;
&lt;p&gt;The physics that limited horizontal transistor scaling has been known for years. The engineering responses to those limits have been arriving throughout this decade. Last week, 3 of them arrived together and pointed in the same direction clearly enough that paying attention seemed warranted.&lt;/p&gt;</content:encoded><dc:creator>Shivashish Yadav</dc:creator><category>Infrastructure</category><category>huawei</category><category>ibm</category><category>openai</category><category>semiconductors</category><category>nvidia</category><author>hello@soursen.org (Shivashish Yadav)</author></item></channel></rss>