<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Derek Yuan on Medium]]></title>
        <description><![CDATA[Stories by Derek Yuan on Medium]]></description>
        <link>https://medium.com/@derekyuan10000?source=rss-c9338594785d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*bXNgISVyguJnmUBjrNMGcw.png</url>
            <title>Stories by Derek Yuan on Medium</title>
            <link>https://medium.com/@derekyuan10000?source=rss-c9338594785d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 07:29:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@derekyuan10000/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Smarter, Not Faster: How Chess Engines Really Beat Humans]]></title>
            <link>https://medium.com/@derekyuan10000/smarter-not-faster-how-chess-engines-really-beat-humans-a4bf6da10d16?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4bf6da10d16</guid>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[chess]]></category>
            <category><![CDATA[stockfish]]></category>
            <category><![CDATA[chessengine]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 18:09:06 GMT</pubDate>
            <atom:updated>2026-10-12T18:09:06.769Z</atom:updated>
            <content:encoded><![CDATA[<h4>Deep Blue searched 200 million positions a second. A 2006 PC searched 8 million and still beat the world champion. The gap between those two numbers is most of the story of computer chess.</h4><p>Deep Blue could evaluate 200 million chess positions a second when it beat Garry Kasparov in 1997 (<a href="https://www.ibm.com/history/deep-blue">IBM</a>). Nine years later, Deep Fritz beat world champion Vladimir Kramnik 4–2 on a PC that managed only 8 million positions a second (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>). That’s 25 times slower, and it still won.</p><p>That’s not a hardware story. That’s an algorithm story, and it’s the best argument I know that clever beats fast.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ywnIxSRlwdYH-cFD" /><figcaption><em>Garry Kasparov and Deep Blue’s Joe Hoane at the 1997 rematch.</em> Source: <a href="https://www.ibm.com/history/deep-blue">IBM</a></figcaption></figure><h3>What Is an Engine Actually Doing?</h3><p>Strip away the mystique and a chess engine is two things: a search and a scoring function. From any position it generates the legal moves, plays each one, and repeats to a chosen depth; at the bottom of the tree it asks the scoring function “who’s better here?”; then it picks the move that leaves it best off assuming the opponent replies as well as possible. That’s minimax, and the problem is that the tree explodes. A typical middlegame position offers around 35 legal moves (a rule-of-thumb figure from general knowledge, not a sourced number), which means ten plies deep is roughly 35 to the power of ten, about 2.8 quadrillion positions, by my arithmetic.</p><p>The fix is alpha-beta pruning: stop exploring a branch the moment you’ve proved the opponent has a better option elsewhere, because they’d never let you reach it. Deep Blue used custom chips to run alpha-beta search in parallel (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>), and Stockfish’s own website still describes its search as highly optimised alpha-beta with aggressive pruning (<a href="https://stockfishchess.org/">Stockfish</a>). The core idea fits in a few lines:</p><pre>FUNCTION alphabeta(position, depth, alpha, beta)<br>    IF depth = 0 THEN<br>        RETURN evaluate(position)                 // the scoring function<br>    FOR EACH move IN ordered_moves(position)      // best-looking moves first<br>        score ← -alphabeta(play(position, move), depth - 1, -beta, -alpha)<br>        IF score ≥ beta THEN<br>            RETURN beta                           // opponent avoids this line; stop looking<br>        alpha ← MAX(alpha, score)<br>    RETURN alpha<br>// mate and stalemate handling omitted</pre><p>The detail that matters is the comment on the loop line. Alpha-beta prunes more when it examines good moves first; with perfect ordering, the tree you actually visit shrinks to roughly the square root of the full one, which is a classic theoretical result rather than something I’ve measured. In practice that means you can search about twice as deep for the same effort. Levy and Newborn estimated that each extra ply of depth adds 50 to 70 Elo points (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>), so pruning is not a micro-optimisation; it’s playing strength.</p><p>Even Deep Blue’s depth is oddly hard to pin down. Wikipedia says it typically searched six to eight moves ahead, and twenty or more in some situations; <a href="https://chess.fandom.com/wiki/Deep_Blue?oldid=4806">Chess Wiki</a> quotes six to twelve plies, up to forty; the <a href="https://www.computerhistory.org/chess/art-43305f13ef377/">Computer History Museum</a> says at times up to thirty moves. Those aren’t the same claim, so I’d hold the exact figure loosely. What’s not in dispute is the other side of the comparison: Deep Fritz averaged 17 to 18 plies in the middlegame “thanks to heuristics” (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>), which on two of those three Deep Blue figures means it was looking further ahead, on hardware a 25th as quick.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hJMIVIhCkEu8TxUU" /><figcaption><em>Two of Deep Blue’s processor cabinets.</em> Source: <a href="https://www.ibm.com/history/deep-blue">IBM</a></figcaption></figure><h3>Where Does the Opening Book Fit?</h3><p>Here’s a myth worth clearing up, because it floats around: engines don’t lose to opening books. A book is the opposite of search, a table of moves worked out in advance. Deep Blue’s opening book encapsulated more than 4,000 positions and 700,000 grandmaster games, with an “extended book” summarising whole games, and its opening library came from grandmasters Miguel Illescas, John Fedorowicz and Nick de Firmian (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>).</p><p>My explanation for why that works is simple. Early positions are the furthest from any concrete result, so search has the least to say about them and accumulated human knowledge has the most. The book isn’t beating the engine; it’s standing in for it at the one stage of the game where it’s weakest. These days, as far as I can tell, a lot of top-level opening theory is checked against engines anyway, so the book is where human work ends and machine work begins. I’d expect that to matter in the world championship match in Geneva this winter, where both camps will arrive with machine-checked preparation.</p><h3>The Referee Got Replaced</h3><p>Search tells you which positions to look at; the scoring function decides what they’re worth. Deep Blue’s was a masterpiece of human labour. It started as a generalised function with parameters set by analysing thousands of master games, was split into 8,000 parts, and was fine-tuned by grandmaster Joel Benjamin before the rematch (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>). Wikipedia calls it an expert system, relying on rules defined by chess masters and computer scientists.</p><p>The modern replacement is NNUE, an “efficiently updatable neural network” that originated in computer shogi and was ported to Stockfish, which released it in Stockfish 12 on 2 September 2020. That version won at least ten times more game pairs than it lost against Stockfish 11 (<a href="https://stockfishchess.org/blog/2020/stockfish-12/">Stockfish</a>). By July 2023 the hand-written evaluation had been removed entirely (<a href="https://en.wikipedia.org/wiki/Stockfish_(chess)">Wikipedia</a>); Leela Chess Zero, meanwhile, trains a network to develop its own internal logic rather than rely on human-defined rules, and its team supplied billions of positions to help train Stockfish’s net (<a href="https://chessprogramming.org/Stockfish">Chess Programming Wiki</a>).</p><p>There’s a programming lesson sitting underneath all of this. The “efficiently updatable” part is incremental computation: Stockfish’s own notes say the network is cheap to run on a CPU because only parts of it need updating after a typical chess move (<a href="https://sources.debian.org/src/stockfish/12-2/README.md">Stockfish README, via Debian</a>). One move changes very little of the board, so you recompute only what changed. It’s the same instinct behind incremental builds and caching; the fastest code is the code you didn’t run.</p><h3>The Test Suite That Never Sleeps</h3><p>The least glamorous reason Stockfish keeps winning is its engineering process. For Stockfish 15, the developers tested nearly 13,000 different changes and kept the best 200 (<a href="https://newreleases.io/project/github/official-stockfish/Stockfish/release/sf_15">Stockfish release notes</a>). That’s a rejection rate of about 98.5% by my arithmetic; every patch has to prove, statistically, that it beats the incumbent in play. As of October 2026 the testing framework has used more than 20,700 years of CPU time to play over 10.3 billion games (<a href="https://en.wikipedia.org/wiki/Stockfish_(chess)">Wikipedia</a>).</p><p>The latest result of that grind is Stockfish 19, released on 5 September 2026. A release summary reports gains of up to 44 Elo over Stockfish 18 and more than three times as many game pairs won as lost (<a href="https://zeli.app/de/story/49599992">Zeli</a>); that’s a secondary summary, so check it against the official release notes before quoting it.</p><h3>What the Comparison Can’t Prove</h3><p>I don’t want to oversell the Deep Blue versus Deep Fritz contrast. It isn’t a controlled experiment; it spans nine years of hardware and software progress, and Deep Blue’s own makers describe it as deriving its strength from brute force computing power (<a href="https://www.ibm.com/history/deep-blue">IBM</a>). The 1997 match was also narrow, 3.5–2.5, and still disputed: Kasparov suggested humans had intervened on the machine’s behalf, IBM denied it, and IBM dismantled the machine rather than grant a rematch (<a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Wikipedia</a>).</p><p>So the honest version of my claim is narrower than the headline. Brute force got computers into the room with the world champion; pruning, better scoring and relentless testing are what turned narrow wins into routine ones. Speed buys you depth. Judgement decides what the depth is worth.</p><p>For more on chess, algorithms and AI, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4bf6da10d16" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chess’s New Superpower Has a Budget Line: How Uzbekistan Won the Olympiad]]></title>
            <link>https://medium.com/@derekyuan10000/chesss-new-superpower-has-a-budget-line-how-uzbekistan-won-the-olympiad-4acd0e4f139d?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/4acd0e4f139d</guid>
            <category><![CDATA[worldchesschampion]]></category>
            <category><![CDATA[grandmaster]]></category>
            <category><![CDATA[uzbekistan]]></category>
            <category><![CDATA[chess]]></category>
            <category><![CDATA[chess-news]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 18:01:56 GMT</pubDate>
            <atom:updated>2026-10-12T18:01:56.602Z</atom:updated>
            <content:encoded><![CDATA[<h4>Twenty out of twenty-two match points, $200,000 a head and a 500-seat chess palace on the way. Uzbekistan’s Olympiad gold was planned years ago; the open question is whether that plan is a model or a monoculture.</h4><p>Uzbekistan won ten of its eleven matches at the Chess Olympiad in Samarkand, beating the top-seeded United States, defending champions India and China along the way, and finished on 20 of a possible 22 match points (<a href="https://www.euronews.com/2026/10/02/uzbekistan-wins-home-gold-at-record-breaking-chess-olympiad-in-samarkand">Euronews</a>). Its only defeat came against Germany.</p><p>That’s not a lucky home run. That’s a budget line paying out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dlV1DY4xgvkVTMAs.jpg" /><figcaption><em>Uzbekistan’s team celebrates Open gold in Samarkand.</em> Source: <a href="https://www.euronews.com/2026/10/02/uzbekistan-wins-home-gold-at-record-breaking-chess-olympiad-in-samarkand">Euronews</a></figcaption></figure><h3>How Did They Actually Win It?</h3><p>The final round was a controlled finish. A draw against Ukraine would probably have been enough, but reserve-board debutant Mukhiddin Madaminov won on board four to make it 2.5–1.5, ending the event unbeaten on 5.5/7 (<a href="https://www.fide.com/uzbekistans-dream-finish-olympiad-gold-in-samarkand/">FIDE</a>). On board one, Nodirbek Abdusattorov went the whole Olympiad without losing a game and posted a 2869 performance rating, which earned him the individual gold; four of the five players from the 2022 Chennai winners were back in the same team (<a href="https://www.fide.com/uzbekistans-dream-finish-olympiad-gold-in-samarkand/">FIDE</a>).</p><p>The hinge, at least in hindsight, was round six against India. In Spectator Australia’s account, Nihal Sarin had played a strong game before a blunder on move 37 handed Uzbekistan a spectacular finish (<a href="https://www.spectator.com.au/?p=10943686">Spectator Australia</a>); the president’s office adds that Nodirbek Yakubboev, who had been in a difficult position, fought on and punished the mistake (<a href="https://president.uz/en/lists/view/9675">president.uz</a>). Uzbekistan then beat China, lost its only match to Germany, and closed with a 3.5–0.5 win over Armenia in round ten (<a href="https://en.chessbase.com/post/chess-olympiad-2026-10">ChessBase</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*a72QCh-lJBNhheB0.jpg" /><figcaption><em>Nodirbek Abdusattorov, board one gold medallist.</em> Source: <a href="https://www.fide.com/uzbekistans-dream-finish-olympiad-gold-in-samarkand/">FIDE</a></figcaption></figure><p>Abdusattorov was not shy afterwards, saying Uzbekistan was now “unrivalled in the world of chess” (<a href="https://www.euronews.com/2026/10/02/uzbekistan-wins-home-gold-at-record-breaking-chess-olympiad-in-samarkand">Euronews</a>). Javokhir Sindarov, in the same report, said winning the title at home felt different from winning it in Chennai. The scoreboard says he’s entitled to both sentiments; the more interesting question is how a country of about 34 million got here.</p><h3>Who Paid for It?</h3><p>The short answer is the state, and it has never pretended otherwise. On 14 January 2021, President Shavkat Mirziyoyev signed decree 4954, approving a State Program for Chess Development for 2021–2025 and creating a chess promotion fund under the national federation (<a href="https://echecs.org/blog/uzbekistan-set-become-world-chess-leader-thanks-presidents-decree">Chess’n Math Association, via echecs.org</a>). The targets were not modest: at least 3% of the population playing regularly by 2025, a top ten place in FIDE’s rankings by the same year, and chess taught in grades two to four across 1,000 elementary schools.</p><p>The same Canadian scholastic body put the first-year price tag at roughly 10 billion soms (about $1 million) for the Chess in Schools project, 7.5 billion (about $750,000) for equipment and 20 billion (about $2 million) to teach chess inside PE lessons (<a href="https://echecs.org/blog/uzbekistan-set-become-world-chess-leader-thanks-presidents-decree">Chess’n Math Association, via echecs.org</a>). That’s around $3.75 million by my addition. These are 2021 conversions from an advocacy group, so treat the dollar figures as ballpark.</p><p>Now set that against the victory bonuses. Each Olympic champion receives $200,000, their coaches $50,000, plus housing and cars (<a href="https://president.uz/en/lists/view/9675">president.uz</a>). On a five-man squad, which is how <a href="https://www.fide.com/uzbekistans-dream-finish-olympiad-gold-in-samarkand/">FIDE</a> describes the team, that’s $1 million for the players alone, by my arithmetic; roughly a quarter of the first-year schools budget, announced at a single meeting.</p><p>Here’s my read: this is not a talent story with a funding footnote. It’s a funding story that happens to produce talent. That’s not to say the talent is manufactured; Sindarov became the youngest winner of the Chess World Cup in 2025 and then won the 2026 Candidates unbeaten, with six wins and eight draws (<a href="https://www.uzbekembassy.in/?p=37303">Embassy of Uzbekistan in India</a>). But a system that reliably puts talent in front of coaches, boards and tournaments is how that kind of player gets found early.</p><p>One caution on the evidence: much of the headline data on Uzbek chess comes from the Uzbek state itself. The president’s office counts 17 new grandmasters in the past decade, five of them women, and quotes his own verdict: “This is a school, this is a system” (<a href="https://president.uz/en/lists/view/9675">president.uz</a>). Plausible, and impressive if true; also not an independent audit. As far as I can tell, whether the 3% participation and top-ten targets were actually hit by 2025 is still unanswered in the sources I could find.</p><h3>What Happens After the Fireworks?</h3><p>The closing ceremony was a statement of intent in its own right: a drone show and fireworks over Samarkand, staged at the Amphitheatre of the Eternal City for more than 2,000 guests (<a href="https://www.fide.com/fireworks-drones-and-champions-samarkands-grand-olympiad-finale/">FIDE</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*9qpWEsb91PqkfjWO.jpg" /><figcaption><em>Fireworks over Samarkand at the Olympiad closing ceremony.</em> Source: <a href="https://www.fide.com/fireworks-drones-and-champions-samarkands-grand-olympiad-finale/">FIDE</a></figcaption></figure><p>The sequel is already scheduled. Branches of the Tashkent International Chess Academy are due in Samarkand, Khorezm and Fergana; the state will cover half of private club fees for players rated above 1,500; and a “Grand Chess” tournament brand will run regular international events (<a href="https://zamin.uz/en/sport/223018-a-500-seat-chess-palace-to-be-built-in-new-tashkent.html">Zamin.uz</a>). A chess palace is proposed for New Tashkent, with a 500-seat hall, a residential block and sports grounds. The scouting brief is to find talented children early in every region, district and mahalla, then pair each with a strong mentor (<a href="https://president.uz/en/lists/view/9675">president.uz</a>).</p><p>Even the headline numbers need holding loosely. Euronews counts almost 400 teams from more than 190 countries, made up of 205 Open and 189 Women’s sides (<a href="https://www.euronews.com/2026/10/02/uzbekistan-wins-home-gold-at-record-breaking-chess-olympiad-in-samarkand">Euronews</a>); the president’s office speaks of around 2,000 players from more than 200 countries (<a href="https://president.uz/en/lists/view/9675">president.uz</a>). The team count isn’t in dispute; the country count, I’d say, is.</p><h3>Is This a Monolith or Microservices?</h3><p>There’s a programming lesson sitting underneath all of this. Uzbekistan runs chess like a monolith: one funder, one federation, one decree, one deployment pipeline. That’s fast to ship and easy to reason about, and it has the classic weakness, a single point of failure. Everything depends on that budget line staying funded and politically fashionable.</p><p>India’s chess boom, as far as I can tell, has far more independent moving parts, from families and private academies to online platforms and sponsors; closer to microservices, messier to coordinate but able to degrade gracefully when one piece fails. That’s my read, not a measured finding. And the messy model is hardly losing: India took Open silver and also the Gaprindashvili Cup, awarded for the best combined Open and Women’s result (<a href="https://www.fide.com/fireworks-drones-and-champions-samarkands-grand-olympiad-finale/">FIDE</a>).</p><h3>Geneva Is the Real Exam</h3><p>The next test is individual rather than institutional. Sindarov challenges Gukesh Dommaraju for the world title in Geneva, a 14-game classical match in which the first to 7.5 points wins (<a href="https://www.fide.com/geneva-to-host-fide-world-championship-match-2026/">FIDE</a>). Both men are 20, making it the youngest world championship match in history.</p><p>Even the dates are not settled. FIDE’s original Geneva announcement says 25 November to 15 December, World Chess now reports an official window of 22 November to 13 December with Game 1 on the 24th (<a href="https://worldchess.com/news/fide-quietly-changes-dates-for-gukeshsindarov-world-championship">World Chess</a>), and Euronews gives 22 November to 12 December. Check your hotel booking against all three.</p><p>Olympiad form is a weak guide to match form. Gukesh lost on board four to Hungary’s Gleb Dudin in India’s final round, finishing with silver on his board (<a href="https://www.fide.com/uzbekistans-dream-finish-olympiad-gold-in-samarkand/">FIDE</a>); one game at the end of a team event tells you very little about a three-week duel. Gukesh also has the one thing Sindarov doesn’t, which is experience of winning a world title match.</p><p>Still, the shape of the story is clear. Uzbekistan didn’t discover talent in Samarkand; it built the machinery that finds it, and the machinery has now produced a team champion and a world title challenger within six months of each other.</p><p>Talent is common. Pipelines are expensive.</p><p>For more on chess, finance and politics, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4acd0e4f139d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Are AI Stocks Actually in a Bubble, or Is This 1999 vs 2005?]]></title>
            <link>https://medium.com/@derekyuan10000/are-ai-stocks-actually-in-a-bubble-or-is-this-1999-vs-2005-0ac89ae81a92?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ac89ae81a92</guid>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 26 Sep 2026 16:01:02 GMT</pubDate>
            <atom:updated>2026-09-26T16:01:02.036Z</atom:updated>
            <content:encoded><![CDATA[<h4>Oracle just invoked a legal clause usually reserved for hurricanes and wars to get out of paying for a data centre. Michael Burry is short half the sector. Warren Buffett’s favourite valuation gauge just did something it’s only done twice in 155 years. None of that tells you whether the story ends in March 2000 or keeps compounding for another decade.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*0UBJAUSSbXJbNwCe" /></figure><p>On Thursday, Oracle sent a force majeure notice to the developer of its $165 billion Project Jupiter data centre in New Mexico, a legal move usually reserved for acts of god, aimed at delaying payment if the site isn’t operational by 2028. The stock fell 4%, and the debt financing the project was already trading below 90 cents on the dollar before the notice even went out (<a href="https://www.cnbc.com/2026/09/24/oracle-data-center-force-majeure.html">CNBC</a>; <a href="https://247wallst.com/investing/2026/09/24/oracle-declares-force-majeure-on-165-billion-project-jupiter-data-center/">24/7 Wall St.</a>).</p><p>That’s one data point, from one company, on one Thursday. It’s also exactly the kind of story that makes the “is this 1999 again” question impossible to ignore, because Oracle isn’t some speculative pre-revenue startup. It’s booked $664 billion in remaining performance obligations and still burned $23.7 billion in free cash flow last fiscal year funding the buildout (<a href="https://247wallst.com/investing/2026/09/24/oracle-declares-force-majeure-on-165-billion-project-jupiter-data-center/">24/7 Wall St.</a>). If the AI leaders themselves are hitting cash flow problems this severe, what does that say about everyone building on top of them?</p><h3>The Bear Case Has Real Numbers Behind It</h3><p>Start with concentration. The ten largest stocks in the S&amp;P 500 now account for roughly 41% of the entire index, according to Bank of America (<a href="https://mygreenknight.substack.com/p/2026-vs-1999-an-honest-scorecard">Substack, citing Bank of America data</a>). Hyperscaler capital spending on data centres and chips is heading toward three quarters of a trillion dollars this year, a figure with no real precedent outside wartime industrial mobilisation.</p><p>Then there’s the valuation gauge Warren Buffett himself watches most closely. The cyclically adjusted price-to-earnings ratio, which divides the S&amp;P 500’s price by average inflation-adjusted earnings over the past decade, has averaged around 17 across the past 155 years. It hit 35 just before the Great Depression, spiked to an all-time high of 44 at the dot-com peak, and has sat above 40 since May of this year (<a href="https://www.fool.com/investing/2026/09/14/warren-buffett-is-sending-a-stark-warning-to-inves/">The Motley Fool</a>). The separate Buffett Indicator, which measures total market value against GDP, has climbed to roughly 238% (<a href="https://www.foreignpolicyjournal.com/2026/09/04/warren-buffetts-stock-market-warnings-a-timeline-of-the-casino-years/">Foreign Policy Journal</a>). Buffett has spent this year calling the market a casino with a church attached, and saying the casino side has gotten more crowded.</p><p>Michael Burry, who made his name shorting mortgage bonds before 2008, is now short Nvidia, Palantir, Oracle, Tesla, Micron and Nebius, along with bearish positions against the Nasdaq 100 and semiconductor ETFs. His Palantir short is up 35% since he entered it, and Oracle has fallen 51% from its Q3 2025 peak (<a href="https://finance.yahoo.com/news/michael-burry-shorts-doing-far-174839022.html">Yahoo Finance</a>). His argument isn’t that AI doesn’t work. It’s that hyperscalers aren’t earning an adequate return on what they’re spending, and that a stock this loved by Wall Street, where roughly 95% of analysts rate Nvidia a buy, is cheap to bet against precisely because almost nobody else is doing it (<a href="https://www.bnnbloomberg.ca/ticker-take/2026/09/04/what-michael-burrys-been-buying-and-shorting-jon-erlichman/">BNN Bloomberg</a>).</p><h3>Except the Bull Case Also Has Real Numbers</h3><p>Here’s where the comparison to 1999 actually breaks down, and it breaks down on the number that matters most: what you’re paying per dollar of profit. The seven largest US stocks entered 2026 trading around 26 times forward earnings, according to Fidelity’s calculations. The seven largest stocks in 1999 traded at roughly 66 times (<a href="https://mygreenknight.substack.com/p/2026-vs-1999-an-honest-scorecard">Substack, citing Fidelity data</a>). The Nasdaq-100’s forward multiple hit around 60 in March 2000, with Cisco alone trading near 100 times forward earnings. Nothing in the current Magnificent Seven is close to that.</p><p>The composition of the returns matters just as much as the level. In the 1995-to-1999 run-up, roughly 314 percentage points of the S&amp;P 500 technology sector’s 488% total return came from multiple expansion, meaning investors paying more for the same dollar of earnings, with the rest from actual earnings growth. In the four years from the end of 2021 to the end of 2025, the technology sector’s 92% return has been about 78% earnings growth and only 9% multiple expansion (<a href="https://blog.roundhillinvestments.com/case-against-the-ai-bubble">Roundhill Investments</a>). Put simply: 1999’s rally was mostly a story about hope. This one is mostly a story about profit that’s already arrived, from Nvidia’s $216 billion in fiscal 2026 revenue on down.</p><h3>The Part That Actually Does Rhyme With 1999</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iiwySnaPaRNnmxWk" /></figure><p>If there’s one thread connecting this cycle to the last one, it’s not the multiples. It’s the circularity of who’s paying whom.</p><p>In September, Nvidia agreed to invest up to $100 billion in OpenAI, structured so Nvidia’s cash flows back to Nvidia as OpenAI uses much of that same money to buy Nvidia chips. Analysts flagged the circular elements immediately; Nvidia’s own response was that it treats every customer with the same priority whether or not it holds a stake in them, which is exactly the kind of reassurance a company gives when the structure genuinely does raise the question (<a href="https://heise.de/-10666776">heise online</a>). It’s not the only deal like it. Microsoft, Amazon and SoftBank all show up on both sides of similar arrangements across the industry, as investors, chip customers and infrastructure landlords to the same small set of AI labs simultaneously.</p><p>That’s not identical to 1999, when telecoms lent each other money to buy each other’s fibre capacity and called it revenue. But it rhymes closely enough that it’s the single metric worth tracking harder than any P/E ratio: how much of the “demand” underpinning these valuations is genuinely external customers paying for AI products, and how much is a small number of companies financing each other’s purchases of each other’s hardware.</p><h3>So, 1999 or 2005?</h3><p>Neither, exactly, which is a less satisfying answer than either camp wants but the more honest one. The froth is real and it’s concentrated in specific places: private valuations that have doubled in a year, leveraged retail money chasing single stocks, and financing structures where the customer and the investor are the same company wearing two hats. Oracle’s debt trading below 90 cents on the dollar on a $165 billion project is what that froth looks like when it finally shows up in a price.</p><p>But the companies actually running the show, the Nvidias and Microsofts and Googles, are nowhere near 1999’s multiples and are funding an enormous share of this buildout out of real, current profit rather than pure story. That’s closer to the picture a few years after a bubble popped than the picture right before one did, minus the part where the popping already happened. If you want a single date to anchor the comparison to, most of the actual data lines up closer to 1998 than 2000: past the point where sceptics could credibly say the technology wasn’t real, deep into a capital spending mania, but with revenue still accelerating into the capex rather than rolling over ahead of it (<a href="https://liqiqiii.github.io/VLCC-Analysis-2026/ai_bubble/report_en.html">VLCC Super Cycle Analysis</a>).</p><p>What ends that comparison, and turns 1998 into 2000, is the circular financing catching up with itself before the revenue does. Watch that gap, not the headline multiple.</p><p>A boom being right about the technology and wrong about the price aren’t mutually exclusive. The fibre glut of 2000 wiped out a generation of telecom investors and then quietly powered the next two decades of the internet built on top of the cable they’d overpaid for. You can be entirely correct that AI changes everything and still lose money owning the stock that proves it, if you bought it at the wrong multiple from the wrong lender.</p><p>Worth watching closely either way.</p><p>For more on where AI capital is actually going, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ac89ae81a92" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Disaster for the CDU. Rise of radicalisation in Germany once again?]]></title>
            <link>https://medium.com/the-geopolitical-economist/disaster-for-the-cdu-rise-of-radicalisation-in-germany-once-again-e15cddbd9706?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/e15cddbd9706</guid>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[usdc]]></category>
            <category><![CDATA[germany]]></category>
            <category><![CDATA[afd]]></category>
            <category><![CDATA[german-politics]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 19:24:46 GMT</pubDate>
            <atom:updated>2026-09-25T19:24:46.172Z</atom:updated>
            <content:encoded><![CDATA[<h4>Two state elections, two wins for the extremes, and a party that just got voted out of a parliament it has sat in since reunification. Germany’s centre didn’t lose ground on Sunday night. It got pincered.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rFkH9oHjxoH82YNE.jpg" /><figcaption>Merz addressing CDU on Tuesday <a href="https://d3i6fh83elv35t.cloudfront.net/static/2025/02/2025-02-24T131230Z_315569602_RC211DAGI1LM_RTRMADP_3_GERMANY-ELECTION-CDU-MERZ-1200x800.jpg">Source: PBS News</a></figcaption></figure><p>In Mecklenburg-Vorpommern, the CDU polled 4.9%. That’s not a bad night; that’s below the 5% threshold, which means zero seats in a Landtag it has held continuously since German reunification. In Berlin, the party fared better and still lost nearly a quarter of its vote share, dropping from 28.2% to 18.8% and shedding 24 of its 52 seats (<a href="https://en.wikipedia.org/wiki/2026_Mecklenburg-Vorpommern_state_election">Wikipedia, citing state returning officer data</a>; <a href="https://en.wikipedia.org/wiki/2026_Berlin_state_election">Wikipedia, citing state returning officer data</a>).</p><p>Chancellor Friedrich Merz didn’t hedge about it. “We cannot sugar coat the result in Mecklenburg-Vorpommern,” he said. “It is a disaster. The CDU fought hard, but in the end it was the CDU that got crushed in the polarization between the SPD and the AfD” (<a href="https://edition.cnn.com/2026/09/20/europe/berlin-mecklenburg-vorpommern-election-afd-intl">CNN</a>).</p><p>That’s not a coincidence I’m willing to write off as two unrelated bad nights. It’s the same mechanism twice: a CDU squeezed out from both directions at once, by a far right in the east and a far left in the west of the country’s own capital.</p><h3>Mecklenburg-Vorpommern: The AfD’s Biggest Win Yet</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_KHyl867GlUCjJ-X" /><figcaption>Mecklenburg-Vorpommern 2026 State Election result <a href="https://www.tagesschau.de/inland/innenpolitik/wahlkreiskarte-ltw-mecklenburg-vorpommern-2026-100.html">Source: Tagesschau</a></figcaption></figure><p>The headline number is the AfD’s. Leif-Erik Holm’s party took 38.2% of the vote, up 21.5 points on 2016, and 32 of the Landtag’s 71 seats, up 18. The SPD held on for first place in seats with 29, but its vote share slid to 35.5%, four points down and its narrowest margin over the AfD in the state’s post-reunification history (<a href="https://en.wikipedia.org/wiki/2026_Mecklenburg-Vorpommern_state_election">Wikipedia, citing state returning officer data</a>).</p><p>Turnout jumped 7.3 points to 78.1%. That’s the tell that this wasn’t apathy finding a new outlet; it was a genuinely larger electorate choosing sides, and choosing the AfD in numbers that dwarf anything the party has managed in the state before.</p><p>The CDU came fifth, behind even the Greens, on 4.9% of the vote. A party that held 12 seats going in now holds none. There’s a programming lesson sitting almost too neatly underneath this one: a 5% electoral threshold is a hard boolean, not a sliding scale, and being on the wrong side of it by a tenth of a point doesn’t cost you a few seats, it costs you every seat. The CDU didn’t have a bad night in Mecklenburg-Vorpommern by any normal political measure. It had a bad night that happened to land on the wrong side of an if-statement, and the consequences were total.</p><p><em>Friedrich Merz, Federal Chancellor of Germany and CDU leader.</em> Source: <a href="https://www.bundeskanzler.de/bk-en/friedrich-merz">Federal Chancellor’s Office</a></p><h3>Berlin: A Different Extreme, A Different Story</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/330/0*5GogdhUn9jTMHQ-I" /><figcaption>Berlin State Election Map</figcaption></figure><p>Two hundred kilometres east, the same night produced a mirror image. Die Linke’s Elif Eralp took 25.7% of the vote, first place for the party in the capital for the first time since reunification, and 38 of 130 seats, up 16. The CDU came second on 18.8%, and the AfD, under Kristin Brinker, jumped 7.2 points to 16.3% and third place (<a href="https://www.aljazeera.com/news/2026/9/20/two-german-states-go-to-the-polls-after-far-right-gains">Al Jazeera</a>; <a href="https://en.wikipedia.org/wiki/2026_Berlin_state_election">Wikipedia, citing state returning officer data</a>).</p><p>The SPD’s result is the one that gets lost in the AfD-versus-Linke framing, and it shouldn’t. Steffen Krach’s party fell to 12.1%, down 6.3 points and into fourth place behind the Greens, its worst Berlin result of the modern era. Krach spent the final fortnight of the campaign fending off a Hanover prosecutor’s investigation into a €9,900 donation tied to a hospital sale he’d overseen; he returned the money and called the implication baseless, but the timeline kept shifting under him as reporters found new documents, and the story never really left the front pages before polling day (<a href="https://en.wikipedia.org/wiki/2026_Berlin_state_election">Wikipedia, citing Der Tagesspiegel and RBB reporting</a>).</p><p>The CDU had its own scandal to survive and didn’t. Outgoing mayor Kai Wegner was hit with sustained criticism over his handling of the 3 January arson attack on Berlin’s power grid, after it emerged he’d spent part of that day playing tennis rather than coordinating the response; he stepped down as the party’s lead candidate in July, replaced by finance senator Stefan Evers with barely two months to campaign.</p><p>Two parties walked into election night carrying scandal, and it was the CDU, not the SPD, that lost nearly a quarter of its vote. Voters had two different corruption stories on offer and picked the one that actually punished the incumbent.</p><h3>Why the Left and the Right Are Winning the Same Argument</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g8YNtJlutjvHVnnf.jpg" /><figcaption>Die Linke celebrating Berlin win <a href="https://www.turkishminute.com/2026/09/20/die-linke-wins-berlin-election-putting-elif-eralp-in-contention-for-mayor/">Source: Turkish Minute</a></figcaption></figure><p>Both winners are telling voters the same underlying story, just with opposite villains. Eralp’s central pledge is expropriating housing from large corporate landlords, a policy Berliners already backed in a non-binding 2021 referendum that the outgoing CDU-SPD coalition quietly shelved. Holm’s pitch in Mecklenburg-Vorpommern leaned on immigration and a “blue wende”, accusing the federal government of prioritising asylum seekers and family reunification over a struggling regional economy.</p><p>Different targets, same message: the people currently in charge aren’t fixing the thing you’re angry about, so vote for someone who says they will.</p><p>That message lands harder in the east. A 2023 Infratest dimap survey found 43% of people in Germany’s eastern states feel like second-class citizens, and the Elitenmonitor study, run jointly by universities in Leipzig, Jena and Zittau-Görlitz, found easterners held just 12.1% of the country’s top institutional positions in 2024, barely up from 10.9% in 2018 despite making up a much larger share of the population (<a href="https://www.riotimesonline.com/germany-state-elections-mecklenburg-vorpommern-berlin-afd-linke-2026/">Rio Times</a>). Mecklenburg-Vorpommern’s result wasn’t produced in a vacuum. It came two weeks after the AfD’s landslide in Saxony-Anhalt, a result that reporting attributed at the time to mobilising turnout for progressive parties elsewhere in the country, in the SPD’s case in Mecklenburg-Vorpommern and in Die Linke’s case in Berlin.</p><p>It worked, in the narrow sense that both of those parties did better than they otherwise might have. It also, on the numbers, wasn’t nearly enough to stop the pattern the AfD’s Saxony-Anhalt result set in motion.</p><h3>The Firewall Is Squeezing the CDU From Both Sides Now</h3><p>Here’s where it gets structurally awkward for Merz specifically. The CDU maintains a firewall against governing with the AfD, and the federal party separately passed a “resolution of incompatibility” ruling out cooperation with Die Linke too. That’s a defensible position on its own terms; it’s a much harder one to hold when both parties on the other side of that wall are the ones actually winning elections.</p><p>In Mecklenburg-Vorpommern, the firewall doesn’t matter this time; the CDU isn’t in the Landtag to enforce it. In Berlin, it means the CDU is shut out of any coalition with the party that just came first, while the SPD and Greens are now weighing whether to govern with Eralp on her own terms, expropriation pledge included.</p><p>The AfD’s legal status complicates the eastern half of that picture further. A Cologne court ruled in February that Germany’s domestic intelligence agency may not, for now, formally classify the national AfD as confirmed right-wing extremist, even as several state branches, Saxony-Anhalt’s among them, remain classified that way. No senior CDU figure has publicly proposed breaking the firewall in response to Sunday’s results. But the arithmetic behind that restraint is getting less comfortable with every state election that goes this way.</p><h3>Where This Actually Leaves Merz</h3><p>I don’t think this is a story about the CDU losing an argument on the merits so much as a story about the CDU losing the argument on both flanks simultaneously, in different states, for different reasons, in the same week. That’s a harder problem than either result looks like in isolation. A party that’s bleeding voters to the right in the east and to the left in the capital doesn’t have an obvious single message that wins both back, and “stay the course” clearly isn’t it after a night this bad.</p><p>Merz’s own quote gives away how the CDU is choosing to read this. He framed Mecklenburg-Vorpommern as the CDU getting crushed between the SPD and the AfD, not as a verdict on his own government. That’s not unreasonable as far as the seat math goes; the SPD held the state. But it also sidesteps the part of the story that should worry the chancellor most, which is that the party of government just got mathematically excluded from a state parliament entirely, and came second in the capital to a party promising to expropriate landlords.</p><p>Both of those are the kind of results a governing party spends years explaining away, not weeks.</p><h3>So What Actually Happens in Mecklenburg-Vorpommern?</h3><p>Simply, its highly unlikely that the AFD get to govern, and the maths is the least interesting part of why not. They won the most votes and the most seats, 32 of 71, but the SPD, the Left and the Greens have all separately ruled out sitting in government with the AfD, and that firewall has held in every eastern state so far regardless of how large the AfD’s plurality gets.</p><p>Once you take an AfD government off the table, three outcomes are actually on it.</p><p>The first, and by far the likeliest, is that Schwesig simply widens the coalition she already runs. Her outgoing SPD-Left government holds 34 seats between them, two short of the 36 needed for a majority. Add the Greens’ five and a SPD-Left-Green government reaches 39, a working majority built without the CDU or the AfD anywhere near it. Nothing about that requires new ground to be broken; the SPD and Left have governed together in Schwerin for years, and the Greens have spent this campaign to the left of where they were in 2021. This is the coalition talks will most likely land on, and probably without much drama.</p><p>The second is a minority SPD-Left government, the same two parties as before but without the Greens, relying on abstentions from the Greens or, less plausibly, from a CDU with zero seats to abstain with. That’s messier and weaker than the three-party option, and there’s no obvious reason for Schwesig to choose instability over a five-seat cushion that’s sitting right there.</p><p>The third is the one nobody wants and everybody’s aware of: talks collapse, no combination can agree a programme, and Mecklenburg-Vorpommern ends up back at the polls within the year. Given how comfortably the majority option works on paper, that scenario is a tail risk rather than a real prediction. But it’s also exactly the kind of tail risk that gets more likely every time an eastern state election produces an AfD plurality this large, because it raises the stakes on the smaller parties actually holding together once the cameras are gone.</p><p>The state the AfD nearly quadrupled its seat count in is, for now, going to be governed by everyone except the AfD. That’s the firewall working as designed. Whether it’s still working the next time this happens, with an even bigger AfD plurality and an even smaller field of parties left to build a majority from, is the question Merz’s party should actually be losing sleep over.</p><h3>And in Berlin?</h3><p>Alot more could happen in Berlin: A Left-Green-SPD coalition, Eralp’s party plus Werner Graf’s Greens plus Krach’s SPD, commands 78 of 130 seats. A CDU-Green-SPD “Kenya” coalition, the same three-party shape that currently runs the federal government, would clear the bar too, with 68. Both are mathematically fine. Politically, only one of them is currently on the table.</p><p>Krach effectively took the SPD out of the running as kingmaker on election night itself, telling reporters “we’ve been voted out” and that the mandate to govern lay with other parties, not the CDU-SPD coalition he’d just been part of. That leaves the Greens holding the actual decision, and Graf had already signalled before polling day that he wouldn’t work with a CDU that maintains its resolution ruling out cooperation with Die Linke. The morning after the count, he added a condition of his own, calling on Die Linke to give a “consistent reckoning with antisemitism in its own ranks”, a pointed reference to a campaign event in Neukölln where militant pro-Palestinian supporters aligned with a local Linke branch called for violence against Israeli forces from the stage.</p><p>That’s the negotiation to watch, not the coalition math. The seats support a red-red-green government under a Linke lead candidate for the first time in the party’s history; whether it actually forms depends on Eralp giving Graf something concrete enough to answer that condition, on top of an expropriation pledge the SPD has already said it won’t accept as a governing red line either. If those talks stall, the CDU-Green-SPD alternative is sitting right there with a real majority of its own, and Berlin ends up governed by the second-place party after all.</p><p>For more on where German and European politics is actually heading, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e15cddbd9706" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-geopolitical-economist/disaster-for-the-cdu-rise-of-radicalisation-in-germany-once-again-e15cddbd9706">Disaster for the CDU. Rise of radicalisation in Germany once again?</a> was originally published in <a href="https://medium.com/the-geopolitical-economist">The Geopolitical Economist</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Party That Shouldn’t Exist: What Germany’s BSW Meltdown Says About the Political Compass]]></title>
            <link>https://medium.com/@derekyuan10000/a-party-that-shouldnt-exist-what-germany-s-bsw-meltdown-says-about-the-political-compass-13ac760eb9aa?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/13ac760eb9aa</guid>
            <category><![CDATA[bsw]]></category>
            <category><![CDATA[germany]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[global-politics]]></category>
            <category><![CDATA[german-politics]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 12 Sep 2026 17:31:31 GMT</pubDate>
            <atom:updated>2026-09-12T17:31:31.426Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1DXGdUnXTqTD2NI4.png" /></figure><p>Picture a party promising a €15 minimum wage, rent caps and a bigger welfare state. Now give it tighter immigration controls, a rejection of gender-inclusive language, and a demand to restart the Nord Stream pipelines to Russia. On the left-right scale most people carry around in their heads, a party like that shouldn’t exist; the wiring doesn’t allow it. In Germany, it has existed since January 2024. This month, in the one state where it actually won power, it started tearing itself apart.</p><p>That’s not really a story about one small party losing an argument with itself. It’s what happens when someone actually moves into a gap political scientists had been pointing at for a decade.</p><h3>What the BSW Actually Is</h3><p>Sahra Wagenknecht split from Die Linke in October 2023 and formally launched the Bündnis Sahra Wagenknecht two months later, taking a slice of the party’s most prominent figures with her (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>). The results came fast: 6.2% in the 2024 European elections, then double digits in three eastern state elections that September: 11.8% in Saxony, 13.5% in Brandenburg, and, most strikingly, 15.8% in Thuringia, where it finished third (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>).</p><p>Then came the wobble. In February 2025’s federal election, the BSW managed 4.981% of the vote, missing the 5% threshold for Bundestag seats by 9,529 votes out of tens of millions cast; the Bundestag’s own election-review committee later confirmed the figure and rejected the party’s request for a nationwide recount (<a href="https://www.bundestag.de/dokumente/textarchiv/2025/kw51-de-wahlpruefung-1129218">Deutscher Bundestag</a>). It’s among the closest a party has ever come to the hurdle without clearing it. Wagenknecht stepped down as party leader in November 2025, handing the role to MEP Fabio De Masi while making clear she wasn’t disappearing from the scene (<a href="https://www.politico.eu/article/top-german-left-wing-politician-steps-down-leader-bsw-sahra-wagenknecht/">Politico</a>). The party is even set to rename itself this October, from Bündnis <em>Sahra Wagenknecht</em> to Bündnis <em>Soziale Gerechtigkeit und Wirtschaftliche Vernunft</em>, keeping the initials but quietly dropping her name from the front door (<a href="https://www.bpb.de/themen/parteien/wer-steht-zur-wahl/mecklenburg-vorpommern-2026/580580/bundnis-sahra-wagenknecht-vernunft-und-gerechtigkeit/">Bundeszentrale für politische Bildung</a>).</p><h3>Left on Everything Except the Things That Aren’t</h3><p>The policy mix is the whole story. Economically, the BSW sits well to the left of Germany’s mainstream: a higher minimum wage, rent controls, an expanded welfare state, stronger collective bargaining, and nationalisation of what it calls key industries (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>). Culturally and on foreign policy, it reads almost like a mirror image: tighter immigration controls, open scepticism of gender-inclusive language, opposition to what it considers wasteful green-energy targets, a demand to reopen the Nord Stream pipelines, and a foreign policy built around NATO scepticism and opposition to further arms deliveries to Ukraine (<a href="https://www.bpb.de/themen/parteien/wer-steht-zur-wahl/mecklenburg-vorpommern-2026/580580/bundnis-sahra-wagenknecht-vernunft-und-gerechtigkeit/">bpb.de</a>).</p><p>Political scientists have a name for this combination, and it predates the BSW by a decade: left-authoritarian, the pairing of left-wing economics with right-wing cultural politics (<a href="https://www.sciencedirect.com/science/article/pii/S0261379425000927">ScienceDirect</a>). Research on the German electorate has repeatedly found that around a fifth of voters hold roughly this exact bundle of positions; what changed in January 2024 wasn’t the voters, it was that a party finally showed up to represent them (<a href="https://www.sciencedirect.com/science/article/pii/S0261379425000927">ScienceDirect</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i8ahgo41Les-0iRkef_KIQ.png" /><figcaption>Compass of German political parties</figcaption></figure><h3>Why the Chart Predicted This, and Missed It Too</h3><p>Most people’s mental model of politics is a single axis, left to right, where your position on the economy is assumed to predict your position on everything else. The BSW breaks that assumption immediately, which is why it reads as a contradiction to anyone still using that model.</p><p>Political scientists dropped the single-axis model decades ago for exactly this reason. The standard replacement, developed by Liesbet Hooghe, Gary Marks and Carole Wilson, crosses an economic left-right axis with a separate cultural axis running from GAL (green, alternative, libertarian) to TAN (traditional, authoritarian, nationalist) (<a href="https://hooghe.web.unc.edu/wp-content/uploads/sites/11492/2023/04/2023-Dassonneville-Hooghe-Marks-Transformation-of-the-political-space-A-citizens-perspective.pdf">Hooghe &amp; Marks</a>). On paper, the two axes are independent; a party could in theory land anywhere on the grid.</p><p>In practice, decades of European election data show parties clustering on a diagonal. Left-wing parties gravitate toward the GAL end of the cultural axis and right-wing parties toward TAN, while the other two corners stay almost empty (<a href="https://arxiv.org/pdf/2607.01464">comparative political-scaling research</a>). Researchers had a name for the empty corner before the BSW existed: the “left-authoritarian supply gap” (<a href="https://www.tandfonline.com/doi/full/10.1080/09644008.2025.2502100">West European Politics</a>). The BSW isn’t an anomaly the model failed to predict. It’s the model’s own prediction, arriving twenty years late, in the one corner of the chart that had been sitting empty the whole time.</p><p><em>The corner German party researchers had been calling the “left-authoritarian supply gap” for a decade. BSW is the first party to actually sit in it.</em></p><p>It’s not unlike training a classifier on a dataset where one class has almost no examples in it. The category is theoretically well-defined and the label exists in the schema, but the model has nothing to learn from until an actual instance turns up. Then it has to work out, on the fly, whether that instance behaves anything like its neighbours in the other corners, or breaks in ways nobody had data for.</p><h3>Even Two Axes Wasn’t Enough</h3><p>The two-dimensional model doesn’t fully capture the BSW either. Its domestic mix, economically left and culturally conservative, maps cleanly onto GAL-TAN. Its foreign policy doesn’t. Research on BSW voters found that its NATO scepticism, hostility to Ukraine arms shipments and preference for détente with Moscow function as “an issue dimension on its own,” only loosely tied to its left-authoritarian domestic profile (<a href="https://www.sciencedirect.com/science/article/pii/S0261379425000927">ScienceDirect</a>).</p><p>Even the researchers who build these charts for a living are still arguing about how many axes European politics actually needs. A study published this year split the usual GAL-TAN cultural axis into two separate dimensions, one tracking “modernisation” values and the other tracking attitudes to globalisation and national sovereignty, specifically because a growing number of parties won’t sit still on a single cultural axis anymore (<a href="https://www.tandfonline.com/doi/full/10.1080/13501763.2025.2501091">Journal of European Public Policy</a>).</p><p>The chart keeps needing more dimensions. The BSW is one of the reasons why.</p><h3>Thuringia, September 2026: The Stress Test</h3><p>Thuringia was the BSW’s best-ever result: 15.8% in September 2024, third place behind the AfD and CDU. By December that year it was in government, junior partner in Germany’s first-ever “Brombeer-Koalition,” black CDU, red SPD and purple BSW, under Minister-President Mario Voigt (CDU), with the BSW’s Thuringia leader Katja Wolf as deputy premier and finance minister (<a href="https://taz.de/BSW-in-Thueringen/!6210389/">taz</a>). It’s still the only coalition of its kind in the country.</p><p>The arrangement never had its own majority. It held exactly 44 of the Landtag’s 88 seats, a bare tie that worked only because neither the AfD nor Die Linke could outvote it either (<a href="https://www.zdfheute.de/politik/deutschland/thueringen-bsw-brombeer-koalition-sahra-wagenknecht-katja-wolf-100.html">ZDF</a>).</p><p>The trigger arrived in August 2026, when Voigt’s doctoral title was formally revoked. The BSW’s national leadership in Berlin demanded he resign and the coalition end. Wolf and the Thuringia branch refused and kept governing, defying their own federal party (<a href="https://www.zdfheute.de/politik/deutschland/bsw-thueringen-fraktion-austritte-100.html">ZDF</a>). What followed was a cascade. One MP quit in protest at the national leadership’s tolerance for an AfD-adjacent government forming in neighbouring Saxony-Anhalt, a direct break with the convention against cooperating with the far right that most German parties still observe. Three more MPs left days later to form a breakaway parliamentary group. Then Katja Wolf herself, along with the rest of Thuringia’s state executive committee, walked out of the party entirely, taking most of the remaining delegation with her: thirteen of the original fifteen BSW members of the Landtag are now gone from the party, though most say they’ll keep propping up Voigt’s government regardless (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>; <a href="https://taz.de/BSW-in-Thueringen/!6210389/">taz</a>).</p><p>The party that occupied the gap the chart said was empty is now the party that can’t hold itself together inside it.</p><p>The fault line is almost too on-the-nose. Wolf had long been seen as the BSW’s pragmatic wing, the one actually running a ministry and negotiating a budget; Wagenknecht represented its purer, more Berlin-based ideological core (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>). Thuringia’s AfD, sensing opportunity, called the rump BSW “unwählbar” and demanded snap elections (<a href="https://jungefreiheit.de/politik/deutschland/2026/schon-wieder-austritte-bsw-fraktion-in-thueringen-bricht-auseinander/">Junge Freiheit</a>).</p><p>This isn’t the first time. Brandenburg’s SPD-BSW coalition, itself a first for German politics, collapsed in January 2026 in almost identical fashion: the state’s BSW deputy premier and finance minister announced at a press conference that he was quitting the party, arguing it couldn’t handle the responsibility of actually governing; the rest of the cabinet’s BSW ministers followed him out within days (<a href="https://en.wikipedia.org/wiki/Sahra_Wagenknecht_Alliance">Wikipedia</a>).</p><p>Two coalitions, two collapses, eight months apart. That’s not bad luck. That’s a pattern.</p><h3>What the Chart Still Can’t Tell You</h3><p>The chart wasn’t wrong. The gap was real, the polling on left-authoritarian voters backs that up, and the BSW’s own early numbers, double digits in three states and a near-miss nationally on its first attempt, prove there were genuine voters waiting for exactly this bundle of positions.</p><p>What the chart couldn’t tell you is whether a party built to occupy a demographically real but historically empty corner of political space can hold together once it has to do the boring part of politics: negotiate a budget, share power, survive a scandal that has nothing to do with ideology. A left-authoritarian party has no natural coalition partner by design. The mainstream left won’t touch its cultural politics; the mainstream right won’t touch its economics. That leaves it dependent on ad hoc, first-of-their-kind arrangements like the Brombeer-Koalition, held together mostly by seat arithmetic rather than shared values, and about as durable as that implies.</p><p>Unlike a Green or a Christian Democrat, a BSW politician also doesn’t have decades of settled party culture to fall back on when the founder’s own instincts, tolerating an AfD-led government two states over, say, over an actual coalition’s survival, start pulling in a direction the people who have to govern don’t want to follow.</p><p>Sparse classes in a dataset are real categories, not noise. They’re also the ones a model is worst equipped to generalise about, precisely because there’s no dense cluster of prior examples to check new behaviour against. The BSW is Germany’s first real data point in a corner of the chart that used to be purely theoretical. Two years in, the data point is still being generated, and on the current evidence, it isn’t looking especially stable.</p><p>The chart said the seat was empty. Germany’s finding out that empty and stable were never the same thing.</p><p>For more on where European politics is actually heading, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=13ac760eb9aa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Britain Turned Its AI Policy Into Microservices. Nobody Wrote the API.]]></title>
            <link>https://medium.com/@derekyuan10000/britain-turned-its-ai-policy-into-microservices-nobody-wrote-the-api-1546f6f607f9?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1546f6f607f9</guid>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sun, 06 Sep 2026 15:48:04 GMT</pubDate>
            <atom:updated>2026-09-06T15:48:04.216Z</atom:updated>
            <content:encoded><![CDATA[<h4>Andy Burnham scrapped Britain’s tech department and gave AI a Cabinet seat on the same day. Six weeks on, five different offices share the job, and Westminster still can’t agree who’s actually in charge.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/465/1*CvU2JxDo41MZXi9bPpL3rA.png" /><figcaption>Andy Burnham pictured talking to the DSTI <a href="https://i.guim.co.uk/img/media/64081e82285e9a5c36a14d08f38f212d354a2024/362_0_4659_3728/master/4659.jpg?width=1300&amp;dpr=2&amp;s=none&amp;crop=none">Source: TheGuardian</a></figcaption></figure><p>On 20 July, Andy Burnham became Prime Minister and did two things in the same breath: he gave Britain its first Cabinet-level AI minister, and he abolished the department that minister used to work for. That’s not really the contradiction it sounds like. What it actually produced is stranger: at least five separate offices across Whitehall now carry a piece of AI policy, and six weeks in, nobody in Westminster can fully agree who calls the final shot.</p><h3>What DSIT Actually Did</h3><p>The Department for Science, Innovation and Technology was only three and a half years old. Rishi Sunak created it in February 2023 to put science, digital policy and AI under one Secretary of State, and in that short life it stood up the AI Opportunities Action Plan, the AI Security Institute, a Â£500 million Sovereign AI Fund, and the Advanced Research and Invention Agency, while acting as the tech sector’s single point of contact in government (<a href="https://www.timeshighereducation.com/news/dsits-likely-abolition-criticised-university-and-tech-leaders">Times Higher Education</a>). It also carried a Â£1.1 billion AI Hardware Plan through to delivery (<a href="https://techround.co.uk/news/the-dbist-has-business-innovation-science-and-trade-but-wheres-the-technology/">TechRound</a>).</p><p>None of that saved it. Burnham’s team had reportedly asked officials to prepare options for breaking the department apart before he’d even moved into Downing Street, and the announcement landed on his first day in office (<a href="https://www.computing.co.uk/news/2026/government/tech-leaders-warn-burnham-against-plans-to-scrap-dsit">Computing</a>).</p><h3>The Reshuffle, In Order</h3><p>Here’s what actually happened to DSIT’s old brief, in the order it happened. On 20 July, most of it moved into a newly enlarged business department, rebranded the Department for Business, Innovation, Science and Trade and led by Jonathan Reynolds, while digital transformation and online harms went to a renamed Department for Digital, Culture, Media and Sport under Lisa Nandy (<a href="https://www.thinkdigitalpartners.com/news/2026/07/21/government-abolishes-dsit-as-ai-gains-a-seat-at-the-cabinet-table/">THINK Digital Partners</a>). Kanishka Narayan, previously a junior DSIT minister, was promoted to Minister of State for Artificial Intelligence, made a full Privy Counsellor, and given a seat at Cabinet, jointly working across the Cabinet Office and DBIST.</p><p>Four days later, on 24 July, the government stood up a new AI Taskforce inside a freshly created Office for the Prime Minister and the Cabinet. Lord Patrick Vallance, who had just resigned as science minister, was named chair, reporting directly to Burnham; the Taskforce itself reports to Cabinet Secretary Dame Antonia Romeo and to Narayan, who leads the unit day to day (<a href="https://www.gov.uk/government/news/ai-to-power-change-at-the-heart-of-government-as-lord-vallance-appointed-chair-of-new-pm-ai-taskforce">GOV.UK</a>). The AI Security Institute, DSIT’s flagship safety body, moved with it into the OPMC.</p><h3>Five People, One Job</h3><p>Add it up and AI policy in Britain now runs through Reynolds at DBIST, Nandy at DDCMS, Narayan straddling both the Cabinet Office and DBIST, Romeo as the civil servant the Taskforce reports to, and Vallance chairing it with a direct line to the Prime Minister. Commentators tallying this at the time landed on the same number: five people now hold some form of AI responsibility across government (<a href="https://grahamlovelace.substack.com/p/creatives-call-for-clarity-on-ai">Charting Gen AI</a>).</p><p>Lord Vaizey, a Conservative peer on the Communications and Digital Committee, wrote that Narayan had been “set up to fail” without a single coherent department behind him, adding a line that lands a bit differently coming from a human commentator: “I defy Claude or ChatGPT to detect anything approaching a sensible pattern in this mess” (<a href="https://grahamlovelace.substack.com/p/creatives-call-for-clarity-on-ai">Charting Gen AI</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/260/0*f4nQgaQNufyKwOSm" /><figcaption><em>Kanishka Narayan, Britain’s first Cabinet-attending Minister for Artificial Intelligence.</em> Source: <a href="https://members.parliament.uk/member/5039/portrait">UK Parliament</a></figcaption></figure><h3>The Distributed Monolith Problem</h3><p>As a developer, this org chart is a familiar shape. DSIT was a monolith: one team, one codebase, one deploy pipeline for AI policy, science funding and digital regulation together. Burnham’s reorg didn’t cleanly split it into microservices with defined boundaries and a single owner apiece. It scattered the same responsibilities across five stakeholders who all still have to coordinate on every release, without ever publishing the interface between them. That’s not microservices. That’s a distributed monolith, the specific anti-pattern where you’ve paid the coordination cost of splitting a system apart without earning any of the independence you split it apart to get.</p><p>Peers in the House of Lords made almost exactly this point without the software vocabulary. Crossbench life peer Lionel Tarassenko warned that a reorganisation on this scale typically takes twelve months to settle, which he noted is “equivalent to three generations of frontier AI models” (<a href="https://www.computerweekly.com/news/366645889/House-of-Lords-questions-Andy-Burnhams-AI-plans">Computer Weekly</a>). That’s the whole argument in one line: the org chart moves on a Whitehall clock, and the technology it’s meant to govern doesn’t wait for it.</p><h3>The Case Burnham’s Team Would Make</h3><p>To be fair to Downing Street, this isn’t as reckless as the backlash made it sound. For most of the period from 1995 to 2023, British science policy lived inside a business department anyway; DSIT was the three-year exception, not the historical norm (<a href="https://theconversation.com/what-the-sudden-demise-of-the-uks-department-of-science-means-for-the-future-of-research-288121">The Conversation</a>). Reynolds has framed the merger as bringing technology closer to industrial strategy rather than sidelining it, and Olivier Usher of Challenge Works has made the same case from outside government, arguing that linking innovation policy directly with business could produce more joined-up decisions on funding, export support and regulation than DSIT’s standalone model did (<a href="https://techround.co.uk/news/the-dbist-has-business-innovation-science-and-trade-but-wheres-the-technology/">TechRound</a>).</p><p>Former government AI adviser Matt Clifford wasn’t buying it, calling the whole move “a big mistake” at a moment he views as critical for tech as both an economic and national security issue (<a href="https://www.cityam.com/burnham-risks-big-mistake-on-ai-with-tech-department-shake-up/">City AM</a>). AI Security Institute chair Ian Hogarth went further, arguing that reorganisations like this distract departments for a year and drive out the best civil servants, who join government to have impact rather than to sit through machinery-of-government paperwork (<a href="https://sifted.eu/articles/andy-burnham-dsit-department-science-technology-overhaul">Sifted</a>).</p><h3>So Who’s Actually Steering This?</h3><p>Strip away the politics of whether tech belongs inside a business department, and the governance question stands on its own. The AI Hardware Plan, the Sovereign AI Fund and the AI Security Institute all have to keep running through a transition where sign-off now passes through Reynolds, Nandy, Narayan, Romeo and Vallance rather than a single Secretary of State. Government’s own comparison for this Taskforce is the Vaccines Taskforce, which worked precisely because it had one accountable lead reporting straight to the top (<a href="https://www.gov.uk/government/news/ai-to-power-change-at-the-heart-of-government-as-lord-vallance-appointed-chair-of-new-pm-ai-taskforce">GOV.UK</a>). Vallance chairing this one and reporting directly to Burnham is a genuine attempt to recreate that; it’s just competing with four other people who also have a legitimate claim on the same brief.</p><p>I’d want to see what happens the first time one of those five offices needs a fast decision, a funding call, a security incident response, that the others don’t immediately agree on. That’s the actual test of whether this is a distributed monolith or something that eventually settles into working microservices. Six weeks isn’t long enough to know. Twelve months, per Lord Tarassenko’s own arithmetic, is three AI model generations away.</p><p>Ownership diagrams matter more than org chart titles. Right now, Britain’s AI policy has five names on the diagram and no arrow pointing to which one wins the argument.</p><p>For more on where AI, politics and finance actually meet, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1546f6f607f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SpaceX Just Bought the Second-Best Coding Tool for a First-Place Price]]></title>
            <link>https://medium.com/@derekyuan10000/spacex-just-bought-the-second-best-coding-tool-for-a-first-place-price-1256a6078c7a?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1256a6078c7a</guid>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[personal-finance]]></category>
            <category><![CDATA[elon-musk]]></category>
            <category><![CDATA[spacex]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 17:16:26 GMT</pubDate>
            <atom:updated>2026-08-29T17:16:26.369Z</atom:updated>
            <content:encoded><![CDATA[<h4>Sixty billion dollars, fifteen times revenue, and a market share that’s been sliding for a year. The Cursor acquisition is either the smartest infrastructure play in AI or the same cash-burn story wearing a different balance sheet.</h4><p>$60 billion. That’s what SpaceX paid to close its acquisition of Cursor on 14 August, an all-stock deal issuing roughly 391 million SpaceX Class A shares, confirmed via a regulatory filing (<a href="https://www.bloomberg.com/news/articles/2026-08-14/spacex-completes-its-60-billion-cursor-acquisition">Bloomberg</a>). That’s also about fifteen times Cursor’s reported annualised revenue of $4 billion (<a href="https://www.fool.com/investing/2026/06/29/spacexs-60-billion-cursor-deal-changes-everything/">The Motley Fool</a>). And according to enterprise spending data tracked by Ramp, Cursor’s own market share has slid from 41% in June last year to around 26% by May this year, as rivals ate into its lead (<a href="https://www.cnbc.com/2026/06/16/spacex-spcx-cursor-acquisition-ipo.html">CNBC</a>).</p><p>That’s the puzzle underneath the biggest AI acquisition of the year so far. Elon Musk just spent enough to buy a small country’s GDP on a company that was, by the numbers, already losing the market it was bought to win.</p><h3>How the Deal Came Together</h3><p>The structure goes back to April, when SpaceX and Cursor announced a technology partnership that quietly included an option letting SpaceX buy Cursor outright for $60 billion. SpaceX went public two months later in a historic IPO; the deal moved forward soon after, and formally closed on 14 August (<a href="https://techcrunch.com/2026/06/16/spacex-to-acquire-cursor-for-60b-in-stock-days-after-blockbuster-ipo/">TechCrunch</a>). Cursor’s own blog confirmed the news the same day, framing it as access to “the largest fleet of GPUs in the world” (<a href="https://cursor.com/blog/joining-spacex">Cursor</a>).</p><p>What happens if a deal like this falls through is oddly unclear. TechCrunch reported the original April terms included a $10 billion break-up fee if SpaceX walked away; a separate report citing SpaceX’s own IPO filings put the number at a $1.5 billion termination fee plus $8.5 billion in committed computing resources (<a href="https://www.cnbc.com/2026/06/16/spacex-spcx-cursor-acquisition-ipo.html">CNBC</a>). Both can’t be exactly right, and neither outlet has corrected the other, so I’d hold the precise figure loosely; what’s not in dispute is that SpaceX built in a lot of downside protection for a deal it clearly wanted to happen either way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5OddRU2dheSf931p.png" /><figcaption><em>Cursor’s announcement confirming the close of the SpaceX acquisition.</em> Source: <a href="https://cursor.com/blog/joining-spacex">Cursor</a></figcaption></figure><h3>The Math That Doesn’t Obviously Work</h3><p>Before SpaceX came knocking, Cursor was reportedly in the middle of raising $2 billion at a $50 billion valuation from investors including Andreessen Horowitz and Nvidia; one source told TechCrunch at the time that even that capital wasn’t going to be enough for the company to break even (<a href="https://techcrunch.com/2026/06/16/spacex-to-acquire-cursor-for-60b-in-stock-days-after-blockbuster-ipo/">TechCrunch</a>). SpaceX paid a 20% premium on top of that valuation, for a company whose own spend-share numbers were already heading the wrong way.</p><p>Cursor didn’t win the market it was bought to dominate. It’s still losing it.</p><p>The market’s initial verdict on SpaceX itself was mixed at best. Shares ticked up around 1.5% on the day the deal closed (<a href="https://finance.yahoo.com/technology/ai/articles/spacex-completes-record-60-billion-131311785.html">Investing.com, via Yahoo Finance</a>). By 26 August, though, SpaceX stock had pulled back 39% from its peak, according to the Motley Fool’s later coverage, a reminder that a friendly opening-day pop doesn’t settle the argument on its own.</p><h3>Why Buy It Anyway?</h3><p>The steelman case is about compute, not code. SpaceX has spent the past couple of years building out data centre capacity that it already rents to outside AI labs, Anthropic and Google among them (<a href="https://techcrunch.com/2026/08/15/spacex-officially-closes-its-cursor-acquisition/">TechCrunch</a>). Owning Cursor turns that rental business into a vertically integrated one: instead of just selling GPU time to whoever builds the next coding agent, SpaceX now owns a coding agent with roughly $4 billion of existing revenue and access to its own Colossus supercomputer to run it.</p><p>There’s a genuinely odd detail sitting inside that logic. Anthropic is one of the rivals this deal is explicitly meant to help SpaceX catch in the coding-agent race; Anthropic is also, right now, a paying customer of SpaceX’s compute (<a href="https://thenextweb.com/news/spacex-cursor-acquisition-completed-gpu-fleet">The Next Web</a>). Musk is trying to beat a company that pays him rent.</p><p>The bull case has real numbers behind it too. Morgan Stanley kept a $300 base price target on SpaceX stock and projected the Cursor deal could add up to $13 billion in SpaceX revenue by 2027 (<a href="https://finance.yahoo.com/technology/ai/articles/spacex-completes-record-60-billion-131311785.html">Investing.com, via Yahoo Finance</a>). Nvidia is reportedly among Cursor’s enterprise customers, which suggests the erosion in Ramp’s spend-share data hasn’t yet reached the accounts that matter most (<a href="https://www.fool.com/investing/2026/06/29/spacexs-60-billion-cursor-deal-changes-everything/">The Motley Fool</a>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/678/0*nwf2Qgn38FORDwJe" /><figcaption><em>Elon Musk, founder and CEO of SpaceX.</em> Source: <a href="https://fortune.com/img-assets/wp-content/uploads/2026/02/GettyImages-2217127913-e1770664209262.jpg?format=webp&amp;w=576&amp;quality=75">Fortune</a></figcaption></figure><h3>The Field That’s Actually Squeezing Cursor</h3><p>The competitive picture explains the share slide better than anything specific to Cursor’s product. Claude Code and Codex CLI have spent 2026 building the terminal-agent category alongside Cursor’s editor-first approach; GitHub Copilot is rolling out a new Polaris model aimed squarely at agentic refactors; and Meta shipped Muse Code on 5 August, described at the time as the first credible fourth entrant to the terminal-agent field (<a href="https://codersera.com/blog/ai-coding-agents-complete-guide-2026/">Codersera</a>). As a developer myself, the more telling detail is that most engineering teams now run two or three of these tools side by side and route between them by task rather than standardising on one, which makes any single company’s “market share” a slippery number to justify a $60 billion price tag with.</p><h3>What happens next</h3><p>I don’t think this is a value investment, and I don’t think Musk is pretending it is. It’s a strategic land grab that only pays off if owning the GPUs underneath Cursor lets it out-execute the market on cost per task in a way this year’s spend-share numbers haven’t caught up to yet. That’s the same growth-over-current-profitability logic behind pretty much every AI capital expenditure figure I’ve covered here, Alphabet’s earnings included; the difference is that SpaceX is financing the bet with freshly minted stock rather than burning cash directly.</p><p>SpaceX’s own AI division isn’t obviously in a position of strength to make that bet from. It reportedly posted a $2.5 billion operating loss on $818 million of revenue in the same quarter this deal was completing, and the company is separately facing legal scrutiny over emissions from the gas turbines powering its data centre (<a href="https://www.fool.com/investing/2026/06/29/spacexs-60-billion-cursor-deal-changes-everything/">The Motley Fool</a>).</p><p>I’d want a year of enterprise renewal numbers, not launch-week press quotes, before calling this smart money.</p><p>Sixty billion dollars buys you compute, customers and a headline. It doesn’t buy back the market share you’d already lost.</p><p>For more on where AI capital is actually going, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1256a6078c7a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Burnham Bounce: Real, Fragile, or Both?]]></title>
            <link>https://medium.com/@derekyuan10000/the-burnham-bounce-real-fragile-or-both-12736efd55be?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/12736efd55be</guid>
            <category><![CDATA[nigel-farage]]></category>
            <category><![CDATA[labour-party]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[uk-politics]]></category>
            <category><![CDATA[andy-burnham]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 22 Aug 2026 18:01:01 GMT</pubDate>
            <atom:updated>2026-08-22T18:01:01.069Z</atom:updated>
            <content:encoded><![CDATA[<h4>Five weeks into his premiership, one pollster gives Andy Burnham the best rating for any incoming PM since Boris Johnson’s early pandemic peak. Another has Reform UK still narrowly ahead. Both are correct.</h4><p>Five weeks ago, Andy Burnham walked into Downing Street with the best opening approval rating any Labour prime minister has recorded in modern polling. This week, YouGov put his net favourability at +13, the highest for any political leader since Boris Johnson’s early Covid-era surge (<a href="https://yougov.com/en-gb/articles/55382-political-favourability-ratings-august-2026">YouGov</a>). Ask a different pollster the same week and Reform UK is still narrowly ahead in the number that actually decides elections.</p><p>That’s not really a contradiction. It’s two different questions getting collapsed into one headline: is Burnham popular, and is Labour winning? Five weeks in, the honest answer to both is that it depends which poll you’re reading and how long you’re prepared to wait before the picture settles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*W5DE2w_y69KQ6hsl.jpg" /><figcaption><em>Prime Minister Andy Burnham arriving at 10 Downing Street.</em> Source: <a href="https://ichef.bbci.co.uk/ace/standard/960/cpsprodpb/563b/live/8171af70-8754-11f1-b2ab-0dd01740f9f6.jpg.webp">BBC</a></figcaption></figure><p>How We Got Here</p><p>You already know the mechanics from the last piece: the Makerfield by-election, the leadership contest, the DSIT reshuffle that put AI at the Cabinet table and scrapped the standalone tech department in the same breath. Burnham was sworn in on 20 July. What’s worth revisiting now is what’s happened to the numbers since, because five weeks is enough time for a honeymoon to either harden into something real or start showing cracks.</p><h3>Just How Good Are His Numbers?</h3><p>Genuinely good, and not by a small margin. Burnham’s first favourability rating as prime minister, taken in his opening days, was +3; six points higher than Keir Starmer’s opening score after the 2024 election, though still short of Theresa May’s +12 in 2016 (<a href="https://yougov.com/en-gb/articles/55228-political-favourability-ratings-july-2026">YouGov</a>). By August that had climbed to +13, with 46% of Britons holding a favourable view of him.</p><p>Opinium’s separate job-approval tracker, which asks a different question, tells a similar story with one complication. On 5–8 August, Burnham scored net +16, with 37% approving and 21% disapproving of his handling of the job (<a href="https://www.opinium.com/resource-center/voting-intention-5-august-2026/">Opinium</a>). A week later, on 12–14 August, that had fallen to +11 (<a href="https://www.opinium.com/resource-center/voting-intention-12-august-2026/">Opinium</a>).</p><p>That’s a five-point slide in a single week.</p><p>He’s still, by some distance, the only party leader with a positive rating on either measure. Farage sits at -25, Badenoch at -5, Davey at -4. But a leader whose approval is already sliding five weeks into the job isn’t in the same position as one whose approval is still climbing, even if both numbers are technically “strong.”</p><h3>But Is Labour Actually Winning?</h3><p>Here’s where it gets genuinely messy. On 2–3 August, YouGov had Reform on 23% and Labour on 22%, extending a run of three weeks in which Reform hadn’t led outside the margin of error (<a href="https://yougov.com/en-gb/articles/55293-voting-intention-2-3-august-2026-ref-23-lab-22-con-19-grn-13-ld-12">YouGov</a>). A week later, on 9–10 August, Labour had overtaken them, 24% to 22%, the party’s best YouGov figure since July of the previous year (<a href="https://yougov.com/en-gb/articles/55331-voting-intention-9-10-august-2026-lab-24-ref-22-con-21-grn-12-ld-11">YouGov</a>). A week after that, on 16–17 August, Reform were back in front, 24% to 22% (<a href="https://yougov.com/en-gb/articles/55381-voting-intention-16-17-august-2026-ref-24-lab-22-con-19-grn-13-ld-12">YouGov</a>).</p><p>Same pollster. Same methodology. One week apart. The two front-runners swapped places twice.</p><p>Other pollsters have been steadier, and mostly kinder to Labour. Opinium’s 12–14 August poll had Labour on 27% to Reform’s 25%, unchanged from the week before (<a href="https://www.opinium.com/resource-center/voting-intention-12-august-2026/">Opinium</a>). BMG, polling for The i in early August, called it plainly: “the first Labour lead in our series since March 2025”, sixteen months of Reform dominance broken (<a href="https://bmgresearch.com/news/bmg-and-the-i-poll-voting-intention-the-burnham-bounce-and-farages-approval-ratings/">BMG Research</a>). LBC’s Ipsos poll had Labour on 28% to Reform’s 25%, with 36% naming Burnham the most capable person to run the country against 15% for Farage (<a href="https://www.lbc.co.uk/article/burnham-bounce-labour-ahead-reform-exclusive-lbc-polling-ipsos-5HjdfnT_2/">LBC</a>).</p><p>If you average across all of it rather than chasing whichever headline broke that week, the picture is clearer, if less dramatic. PollCheck’s seven-poll moving average, as of 21 August, has Labour on 26.0% to Reform’s 25.0% (<a href="https://www.pollcheck.co.uk/gb-polls">PollCheck</a>). A real lead. Not a landslide, and not remotely settled.</p><p>If that reads like trying to make sense of noisy telemetry rather than a clean signal, that’s because it is. Any single poll is one sample with its own house effect and its own margin of error sitting on top of the true number; reading it in isolation is like trusting a single spike on a monitoring dashboard instead of the smoothed trend line behind it. You wait for the average across sources before you act on it, not the loudest data point.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nqDUWCJJgGEIRO5s.jpg" /><figcaption><em>The Prime Minister visiting fire-damaged homes in Stourbridge, 14 August.</em> Source: <a href="https://www.flickr.com/photos/number10gov/55464125873/">Number 10 Downing Street</a></figcaption></figure><h3>What’s Actually Driving It</h3><p>Some of this is retail politics done well. A £2 cap on bus fares and cuts to energy costs both landed in Burnham’s first week, alongside the scrapped digital ID scheme he’d promised to kill. He’s spent much of August on a “listening mode” tour of the country on the cost of living, styling himself explicitly as a circuit breaker rather than a continuation of the last government (<a href="https://www.lbc.co.uk/article/burnham-bounce-labour-ahead-reform-exclusive-lbc-polling-ipsos-5HjdfnT_2/">LBC</a>).</p><p>Some of it is coalition repair. Net favourability among Green voters has risen 72 points in a single month, from -45 to +27, the single sharpest shift in the entire August YouGov dataset (<a href="https://yougov.com/en-gb/articles/55382-political-favourability-ratings-august-2026">YouGov</a>). That’s not a coincidence; it’s Burnham’s positioning on devolution and public ownership doing exactly what it was designed to do, pulling back voters Labour bled to the Greens under Starmer.</p><h3>Meanwhile in Clacton</h3><p>Farage’s own by-election, forced by his resignation, was held on 13 August. He won it by a majority of 12,784 against 34 other candidates, on a 44% turnout (<a href="https://members.parliament.uk/member/5091/electionresult">UK Parliament</a>). The seat that made him remains entirely safe, even as his national ratings fall to their lowest point since the election.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/260/0*pL6xrZprl5Rw9Ndi" /><figcaption><em>Nigel Farage, Reform UK leader and MP for Clacton.</em> Source: <a href="https://members.parliament.uk/member/5091/portrait">UK Parliament</a></figcaption></figure><p>Worth sitting with, that split. A leader can lose a national argument while keeping an iron grip on the one seat where he’s actually on the ballot. It’s the mirror image of Burnham’s problem: strong on the constituency that knows him best, weaker on the country that’s still deciding.</p><h3>So, Is It Real?</h3><p>Here’s my actual read. The personal bounce is real, and it’s broader than a single friendly pollster; it shows up in YouGov, Opinium, BMG, Freshwater Strategy and Ipsos alike, and it reaches into demographics, Green and Lib Dem voters especially, that a Labour PM doesn’t usually touch. That’s not noise.</p><p>The party lead over Reform is also real, but it’s thin and it’s fragile; a one-point average that has flipped direction in the same pollster’s tracker within a single week tells you it hasn’t hardened into anything yet. I’d hold both of those readings loosely rather than pick a side.</p><p>What I’d actually watch is the gap underneath both numbers. Opinium’s early-August poll found just 29% of Britons think Burnham has the right priorities for the country, against 21% who think he doesn’t and 30% who think it’s a mix of both (<a href="https://www.opinium.com/resource-center/voting-intention-5-august-2026/">Opinium</a>). People like him well ahead of buying his programme. That gap closes one of two ways: the programme catches up to the personal warmth, or the warmth erodes down to meet it. His first Budget, expected between mid-October and late November, is where that gets tested for real (<a href="https://www.cityam.com/burnham-bounce-pm-gets-popularity-boost-at-farages-expense/">City AM</a>).</p><p>Being liked got him through August. Being believed is a September problem.</p><p>For more on UK politics and where the numbers actually point, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12736efd55be" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fifteen Beats Twenty: How Young Chess Just Got]]></title>
            <link>https://medium.com/@derekyuan10000/fifteen-beats-twenty-how-young-chess-just-got-0c6c628c0cb0?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/0c6c628c0cb0</guid>
            <category><![CDATA[gukesh-d]]></category>
            <category><![CDATA[gukesh]]></category>
            <category><![CDATA[young-people-in-chess]]></category>
            <category><![CDATA[chess-india]]></category>
            <category><![CDATA[chess]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 15 Aug 2026 16:01:02 GMT</pubDate>
            <atom:updated>2026-08-15T16:01:02.119Z</atom:updated>
            <content:encoded><![CDATA[<h4>The world champion just got overtaken by someone who can’t legally drive.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yi3Y4EtRJm3hP1vG.png" /><figcaption><em>The August 2026 FIDE rating list.</em> Source: <a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a></figcaption></figure><p>Gukesh Dommaraju is twenty years old, the reigning world chess champion, and, as of last week, no longer even the highest-rated teenager in the game. The August FIDE list dropped him 14 points to 2703 after three straight losses at the Chennai Grand Masters, his first tournament in India since winning the title. Fifteen-year-old Yagiz Kaan Erdogmus, sitting at 2716, went the other way and passed him (<a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a>).</p><p>Twenty used to be the prodigy age. Right now it’s the age you get overtaken at.</p><h3>A Rough Second Year on the Throne</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/0*Xt6B0Y0xKQG4rPRB.png" /><figcaption><em>World chess champion Gukesh Dommaraju.</em> Source: <a href="https://www.chess.com/players/gukesh-dommaraju">Chess.com</a></figcaption></figure><p>Chennai wasn’t a one-off. Gukesh finished joint-bottom at the Prague Masters in March, losing three games and dropping 19.5 points in a single tournament. He then finished last at Norway Chess. Then last again at Chennai, without winning a single game across seven rounds, narrowly avoiding dropping below 2700 entirely thanks to a flawed but unrefuted swindle against Hans Niemann in the final round (<a href="https://www.espn.com/chess/story/_/id/49417280/gukesh-dommaraju-nightmare-2026-threatens-derail-world-championship-defence">ESPN</a>; <a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a>).</p><p>The whiplash is the point. This is the same player who became the youngest undisputed world champion in history at 18, beating Garry Kasparov’s 1985 record by four years, and who missed the outright youngest-grandmaster record by just 17 days as a 12-year-old (<a href="https://www.chess.com/players/gukesh-dommaraju">Chess.com</a>). Former world champion Viswanathan Anand, watching from Chennai, still backs him to recover before the Olympiad. His coach points to age itself as the explanation, telling ESPN that finding new motivation after achieving a lifelong goal at 18 “can be difficult for someone so young” (<a href="https://www.espn.com/chess/story/_/id/49417280/gukesh-dommaraju-nightmare-2026-threatens-derail-world-championship-defence">ESPN</a>).</p><p>He’ll defend the title later this year against Javokhir Sindarov, the 2005-born winner of this year’s Candidates. Sindarov replaced Gukesh on the full Grand Chess Tour roster back in March, after Gukesh scaled back his schedule to focus on training (<a href="https://worldchess.com/news/world-chess-champion-gukesh-scales-back-schedule-after-poor-run-of-form">World Chess</a>).</p><h3>The Boy Who Broke Wei Yi’s Record</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/0*NCbWrEYkbQjtvAFC.png" /><figcaption><em>Turkish grandmaster Yagiz Kaan Erdogmus.</em> Source: <a href="https://www.chess.com/players/yagiz-kaan-erdogmus">Chess.com</a></figcaption></figure><p>Erdogmus isn’t just riding someone else’s bad month. In April, he broke 2700 for the first time, becoming the youngest player in history to do so, beating the previous record holder Wei Yi’s mark by close to a year (<a href="https://en.wikipedia.org/wiki/Ya%C4%9F%C4%B1z_Kaan_Erdo%C4%9Fmu%C5%9F">Wikipedia</a>). He got there partly through a series of exhibition matches billed as “Clash of Generations,” in which he’s beaten former world championship contenders Peter Svidler, Maxime Vachier-Lagrave, and, most emphatically, former world champion Veselin Topalov 5–1 with a 2990 performance rating (<a href="https://www.chess.com/news/view/erdogmus-makes-history-again-to-become-youngest-ever-2700-player">Chess.com</a>). Magnus Carlsen has called him the best 14-year-old the world has ever seen; Erdogmus was still 14 when he earned that line.</p><p>On the August list he’s world number 27, junior number one, and, per Chess.com’s Prodigy Watch metric, the single most statistically impressive teenager in the game right now (<a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a>). He’s coached by former world number two Shakhriyar Mamedyarov, who has stated the goal out loud: become the first player ever to reach 2900. Backed financially by Turkish entrepreneur Evren Ucok, Erdogmus is now chasing a spot in the world top 10 before American legend Gata Kamsky’s record for youngest-ever top-10 finish expires in mid-2027, which by ChessBase’s math requires roughly another 40 rating points (<a href="https://en.chessbase.com/post/youngest-player-in-the-top-ten">ChessBase</a>).</p><h3>India’s Assembly Line</h3><p>None of this is happening in a vacuum, and it’s worth remembering that Gukesh himself is a product of the same machine now producing his rivals. India had exactly one grandmaster in 1988: Viswanathan Anand. It has 93 today, with a 94th pending FIDE ratification, and Tamil Nadu alone, Gukesh’s home state, accounts for 35 of them (<a href="https://kingdomofchess.com/all-grandmasters-from-india/">Kingdom of Chess</a>).</p><p>The depth is the real story, not any single name. Arjun Erigaisi sits at world number 9 (2759), Praggnanandhaa Rameshbabu at number 12 (2750), Nihal Sarin at number 25 (2718). On the junior list alone, six Indians place inside the world’s top 25: Pranesh M at 3, Pranav V at 4, Leon Luke Mendonca at 8, Bharath Subramaniyam at 16, Aditya Mittal at 20, and Pranav Anand at 21 (<a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a>). On the women’s side, Divya Deshmukh became India’s fourth female grandmaster in 2025, joining Koneru Humpy, Harika Dronavalli, and Vaishali Rameshbabu; India swept both open and women’s gold at the 2024 Olympiad in Budapest (<a href="https://www.mappr.co/chess-grandmasters-by-country/">Mappr</a>).</p><p>Gukesh, Praggnanandhaa, and fellow grandmasters Murali Karthikeyan and Aravindh Chithambaram all came up through the same Chennai institution, the Velammal school system. That’s not a coincidence. It’s a pipeline that keeps producing overlapping generations of contenders, which means the “changing of the guard” happening to Gukesh right now is, in part, India’s own conveyor belt catching up to its earliest product.</p><h3>It’s Not Just India</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/0*6QxaE90Va0CDLkcj.jpg" /><figcaption><em>Argentine chess prodigy Faustino Oro.</em> Source: <a href="https://www.chess.com/players/faustino-oro">Chess.com</a></figcaption></figure><p>-Erdogmus is Turkish. The single biggest youngest-ever milestone of the year belongs to an Argentine. In May, 12-year-old Faustino Oro became the second-youngest grandmaster in history at 12 years, 6 months, and 26 days, trailing only American Abhimanyu Mishra’s 2021 record. Oro beat Carlsen himself in an online bullet game at age 10 and is now nicknamed “the Messi of chess” back home (<a href="https://www.fide.com/faustino-oro-the-messi-of-chess-becomes-second-youngest-gm-in-history/">FIDE</a>).</p><p>Meanwhile 11-year-old English junior Bodhana Sivanandan became the sixth-youngest player ever to beat a 2600-rated grandmaster this July, and Russian 11-year-old Roman Shogdzhiev currently sits second on Chess.com’s Prodigy Watch, closing in on Nodirbek Abdusattorov’s record for highest rating achieved before age 11.5 (<a href="https://www.chess.com/news/view/august-2026-fide-rating-list">Chess.com</a>). The floor for what counts as a remarkable young talent is dropping in Turkey, Argentina, England, and Russia at the same time it’s dropping in India. That’s a pattern, not a coincidence, and it’s hard to separate from the fact that any of these kids can now train daily against engine-strength opposition without leaving their bedroom, a kind of always-on sparring partner no generation before them had access to.</p><p>Gukesh’s bad year is real, but it’s also just one data point in a much bigger shift; teenage and early-20s slumps have happened to plenty of champions who went on to recover fine. What isn’t a blip is the structural trend underneath it: the age at which a player is expected to be a serious contender keeps falling, across federations that have nothing else in common, and it’s very unlikely to reverse.</p><p>Twenty isn’t old. In this sport, right now, it just isn’t young enough anymore.</p><p>For more on finance, politics, chess and programming, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0c6c628c0cb0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Follow the Money: The Only AI Company Actually Turning a Profit]]></title>
            <link>https://medium.com/@derekyuan10000/follow-the-money-the-only-ai-company-actually-turning-a-profit-6c4e2c3d8e17?source=rss-c9338594785d------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c4e2c3d8e17</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[stocks]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[personal-finance]]></category>
            <dc:creator><![CDATA[Derek Yuan]]></dc:creator>
            <pubDate>Sat, 08 Aug 2026 16:01:01 GMT</pubDate>
            <atom:updated>2026-08-08T16:01:01.913Z</atom:updated>
            <content:encoded><![CDATA[<h4>Everyone in AI is spending like they’ve already won. Only one of them is making money like it.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*H_JMFXAb_AGKr689.jpg" /><figcaption>Source: <a href="https://cdn.arstechnica.net/wp-content/uploads/2023/07/ai-logos.jpg">Arstechnica</a></figcaption></figure><p>Alphabet spent $44.9 billion in a single quarter building AI infrastructure and still posted a $40.8 billion operating profit. OpenAI spent less than that in the same three months and still lost $3.7 billion (<a href="https://news.creeta.com/en/openai-q1-2026-financials-burn-rate/">The Information, via Creeta</a>).</p><p>That gap is the whole story.</p><p>Every AI company is telling investors the same thing: spend now, profit later. The infrastructure race has pushed Microsoft, Amazon, Alphabet and Meta to a combined $725 billion in 2026 capital expenditure, up 77% from roughly $410 billion the year before (<a href="https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion">Tom’s Hardware</a>). Global AI investment this year is projected to clear $2.5 trillion (<a href="https://intellectia.ai/blog/ai-investment-bubble-2026">Intellectia</a>). But “later” is doing an enormous amount of work in that sentence, and the further you dig into who’s actually converting that spending into profit, the more the picture narrows to basically one company.</p><h3>The Trillion-Dollar Question Nobody’s Answering</h3><p>Start with the company burning cash fastest. OpenAI’s internal projections show roughly $14 billion in losses for 2026 alone, with some outside estimates running as high as $27 billion (<a href="https://www.rdworldonline.com/facing-14b-losses-in-2026-openai-is-now-seeking-100b-in-funding-but-can-it-ever-turn-a-profit/">RD World</a>; <a href="https://valueaddvc.com/blog/openai-revenue-2026-20b-arr-4b-month-path-to-profitability">Value Add VC</a>). Leaked audited financials put 2025’s actual booked revenue at $13.07 billion against a $20.9 billion operating loss (<a href="https://valueaddvc.com/blog/openai-revenue-2026-25b-arr-a-20-9b-leaked-loss-and-why-anthropic-just-passed-it">Value Add VC</a>). In the first quarter of 2026, the company brought in $5.7 billion and burned $3.7 billion doing it, an operating margin of roughly negative 122% (<a href="https://news.creeta.com/en/openai-q1-2026-financials-burn-rate/">Creeta, citing The Information</a>).</p><p>Put plainly: OpenAI currently loses about $1.22 for every dollar it makes.</p><p>The company’s own forecasts don’t project sustained profitability until 2029 or 2030, by which point cumulative losses are projected to hit $115 billion (<a href="https://www.rdworldonline.com/facing-14b-losses-in-2026-openai-is-now-seeking-100b-in-funding-but-can-it-ever-turn-a-profit/">RD World</a>). To be fair, revenue is genuinely growing, and a company choosing to burn cash while scaling isn’t automatically a red flag; that’s the entire venture-backed playbook. But the scale here is unusual even by that standard, and OpenAI has now filed confidentially for an IPO, which means these numbers stop being private soon.</p><h3>Google’s Unfair Advantage</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ct7ssx6QlbWz-ZZv" /><figcaption><em>Inside a Google data center</em> Source: <a href="https://datacenters.google/discover-more/photo-gallery/">Google Data Centers</a></figcaption></figure><p>Now compare that to Alphabet. Google Cloud brought in $24.8 billion in the second quarter of 2026, up 82% year-on-year, and cloud operating income more than tripled to $8.81 billion. Operating margin expanded from 20.75% to 35.6% in a single year (<a href="https://finance.yahoo.com/technology/ai/articles/forget-ai-capex-concerns-buy-131800018.html">Yahoo Finance/Zacks</a>). The cloud backlog sits at $514 billion, up more than $50 billion in one quarter alone (<a href="https://finance.yahoo.com/technology/ai/articles/alphabet-inc-goog-q2-2026-050111614.html">Yahoo Finance</a>). Total company revenue hit $119.8 billion, up 24%, beating analyst estimates comfortably.</p><p>Yes, Alphabet also raised its 2026 capex guidance to $195–205 billion and posted negative free cash flow for the quarter (<a href="https://www.cnbc.com/2026/07/22/google-earnings-q2-goog-live-updates.html">CNBC</a>; <a href="https://www.investing.com/news/company-news/alphabet-q2-2026-slides-24-revenue-growth-cloud-surges-despite-capex-93CH-4807148">Investing.com</a>). The market wasn’t thrilled; shares dropped nearly 4% on the guidance hike. But there’s a fundamental difference between Alphabet’s capex and OpenAI’s: Alphabet is spending record sums <em>on top of</em> a search-and-advertising business that still throws off enormous cash, and building the chips it spends that money on itself. Its Tensor Processing Units mean Alphabet isn’t paying Nvidia’s margin on every GPU the way its rivals are.</p><p>That’s the unfair advantage. Google doesn’t need outside financing to build its AI stack. Almost nobody else in this race can say that.</p><h3>The Circular Money Problem</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*VUKYh3KfbbY2urKd" /><figcaption><em>NVIDIA’s GB200 NVL72 hardware, the rack-scale system at the centre of its data-centre deals.</em> Source: <a href="https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing">NVIDIA Newsroom</a></figcaption></figure><p>Here’s where it gets uncomfortable. A huge amount of the capital funding the AI buildout isn’t coming from customers paying for AI products; it’s coming from AI companies funding each other.</p><p>Nvidia is reportedly in talks to guarantee up to $250 billion in lease payments for an OpenAI data centre in Ohio, on top of a separate $350 billion financing arrangement under discussion for OpenAI’s chip purchases (<a href="https://www.axios.com/2026/07/27/nvidia-openai-financing-ai-jensen-huang-ssi">Axios</a>). That’s not a one-off. Nvidia has announced more than $540 billion in similar deals across the AI ecosystem this year alone, taking stakes in the very companies that buy its chips: OpenAI, CoreWeave, Nebius, and others (<a href="https://finance.yahoo.com/technology/ai/articles/nvidia-750-billion-deals-revive-102003935.html">Bloomberg, via Yahoo Finance</a>). Google has done a version of the same thing, backstopping roughly $35 billion in lease payments across five data centres to help Anthropic secure financing on better terms than it could alone (<a href="https://en.bloomingbit.io/feed/news/117116">Bloomberg</a>).</p><p>Nvidia’s Jensen Huang has repeatedly pushed back on the idea that this is circular, arguing it’s a small share of what these companies ultimately have to raise elsewhere. Maybe. But when the chipmaker is guaranteeing the debt of the companies buying its chips, and those companies are burning billions a quarter, “circular” is a fair word for what’s happening. CoreWeave is the cleanest cautionary example: $2.08 billion in Q1 2026 revenue, a $740 million net loss (<a href="https://startupfortune.com/nvidias-750-billion-ai-financing-web-is-drawing-scrutiny-over-circular-money-flows/">Startup Fortune</a>), and a stock that fell more than 50% from its peak on exactly these concerns (<a href="https://finance.yahoo.com/news/coreweave-stock-soars-ceo-rebuts-142135107.html">Yahoo Finance</a>).</p><p>None of this means the whole thing collapses tomorrow. But it does mean the industry’s downside case isn’t “one company misses earnings.” It’s a web of interlinked debt where a demand slowdown at any single node stresses several others at once.</p><h3>What Anthropic’s Numbers Actually Tell You</h3><p><em>Anthropic co-founder and CEO Dario Amodei.</em> Source: <a href="https://www.weforum.org/people/dario-amodei/">World Economic Forum</a></p><p>I’d be doing you a disservice if I let Anthropic off easy here just because it happens to make the model I’m writing this in. The company’s revenue growth is genuinely extraordinary by any standard: from roughly $1 billion in annualised revenue at the end of 2024 to a disclosed $47 billion run-rate at its May 2026 Series H, which valued the company at $965 billion (<a href="https://sacra.com/c/anthropic/">Sacra</a>). Independent trackers now estimate ARR closer to $69–70 billion as of July (<a href="https://www.nextbigfuture.com/2026/07/reports-of-69-billion-annualized-revenue-rate-for-anthropic.html">NextBigFuture</a>).</p><p>But the reported figures don’t fully line up with each other. Anthropic told the press it had hit $14 billion in ARR in February 2026 (<a href="https://www.saastr.com/anthropic-just-hit-14-billion-in-arr-up-from-1-billion-just-14-months-ago/">SaaStr</a>). Two weeks later, in a March 2026 court filing, CFO Krishna Rao stated under oath that the company’s revenues were “exceeding $5 billion to date” (<a href="https://www.wheresyoured.at/anthropics-profitability-swindle/">Where’s Your Ed At</a>). Those two figures are difficult to reconcile, and it’s worth being honest about that rather than picking whichever number makes the better headline. Anthropic also books revenue from cloud resellers like AWS and Google Cloud on a gross basis, which inflates the top line relative to companies that report net (<a href="https://aibusinessweekly.net/p/anthropic-statistics">AI Business Weekly</a>). That’s not necessarily misleading, but it does mean the $47 billion figure isn’t directly comparable to a typical SaaS company’s ARR.</p><p>None of this means Anthropic is lying about its growth. It almost certainly is growing faster than any enterprise software company in history. But “fastest-growing” and “reliably measured” aren’t the same claim, and a company heading toward an IPO with a confidential S-1 already filed is going to face public-market scrutiny that private funding rounds simply don’t apply.</p><h3>Where This Goes Next</h3><p>Here’s my actual read, not just the numbers.</p><p><strong>Google is underrated by a market that’s still pricing in search disruption.</strong> Alphabet’s stock dropped on its own earnings beat because investors are fixated on capex, but capex funded by a genuinely profitable core business, spent on infrastructure you own outright, is a completely different animal to capex funded by venture debt and Nvidia guarantees. The market is treating Alphabet like it has OpenAI’s balance sheet risk. It doesn’t.</p><p><strong>OpenAI and Anthropic are heading toward the same reckoning at different speeds.</strong> Private markets tolerated ARR mythology because the only people checking the maths were the investors writing the cheques. Public markets don’t work that way. When OpenAI and Anthropic both eventually list, the gap between headline run-rate and audited, reconciled revenue is going to matter a great deal more than it does today, and I’d expect at least one of the two to face a genuinely rough first year as a public company on exactly this point.</p><p><strong>The circular financing web is the single biggest tail risk in this market right now, and it’s underpriced.</strong> Not because any one deal is inherently fraudulent, but because Nvidia, OpenAI, Oracle, CoreWeave and now Google-via-Anthropic are all leaning on each other’s balance sheets at the same time that several of them are burning cash at historic rates. That’s not a 2000-style bubble where the underlying revenue doesn’t exist; the revenue is real. It’s closer to 2008: real assets, real cash flows, but leverage and interconnection dense enough that a shock in one place doesn’t stay contained to one place.</p><p><strong>Nvidia’s margin is the multi-year story nobody’s pricing correctly.</strong> A 53% net margin and 44x earnings looks nothing like the dot-com peak, and that comparison gets made constantly to wave away bubble concerns. Fair enough for now. But Google’s TPUs, Amazon’s Trainium and Microsoft’s Maia chips all exist specifically to reduce hyperscaler dependence on Nvidia’s margin, and every dollar of that $725 billion in 2026 capex that goes to custom silicon instead of Nvidia GPUs is a dollar Nvidia doesn’t see. That shift is still early. It won’t be by 2028.</p><p>Capability isn’t the bottleneck anymore. Cash flow is, and right now there’s exactly one company in this race that has both.</p><p>For more on finance, politics, chess and programming, you can find my other writing at: blog.derekyuan.co.uk</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c4e2c3d8e17" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>