<?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 Chaesang Jung on Medium]]></title>
        <description><![CDATA[Stories by Chaesang Jung on Medium]]></description>
        <link>https://medium.com/@chaesangjung?source=rss-08a1a95ef732------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*sD7K9r-TQ2pT4zKLE0LXPQ.jpeg</url>
            <title>Stories by Chaesang Jung on Medium</title>
            <link>https://medium.com/@chaesangjung?source=rss-08a1a95ef732------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 14:18:33 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@chaesangjung/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[Filling Forty Languages]]></title>
            <link>https://chaesangjung.medium.com/filling-forty-languages-cce8c31bb7ac?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/cce8c31bb7ac</guid>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[search]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 00:00:39 GMT</pubDate>
            <atom:updated>2026-11-19T22:19:59.218Z</atom:updated>
            <content:encoded><![CDATA[<p>One particle, one blank string. Shipping a product in forty languages meant holding out an empty vessel and hoping it filled in time.</p><p><em>The ninth piece in a series on how country and language were never one and the same.</em></p><p>This is the story of how a single line of text on a screen made its way into forty languages, and of filling the vessel that held it. When you built something at Google, the design generally assumed from the start that it could go out in more than forty languages and to more than a hundred countries. You left the vessel empty, so to speak, so that anything could ship anywhere once the strings were poured in. Which meant that whenever good translation failed to keep up, that became the bottleneck for the product or the feature. This was back when machine translation was nothing like what it is now, and far more of the work passed through human hands.</p><p>Every string that went on screen lived in a translation table. Where a name or a number belonged you left a placeholder, and the sentence got assembled around it. What came back was then reviewed by a linguist who knew the language. With more than forty languages to support, a single empty cell anywhere in that table could hold up an entire feature.</p><p>Most of those strings were handled as parameterized strings, the sprintf idea: leave a slot, drop a value in, and the sentence completes itself. Korean gives that approach a hard time. Whether a noun ends in a consonant decides whether the particle after it is 이 or 가, so you cannot mechanically staple one particle behind a placeholder. Word order flips against English constantly, too. The template system had no way to carry that final-consonant rule, so what we ended up writing was 이(가) and 을(를), both forms side by side, and when I run into 이(가) somewhere now I feel a small pang. There was a volunteer group inside the company that helped with Korean translation for products, and anyone who drew the short straw and got handed a disclaimer or some other piece of legal text had a rough time of it. Fonts and line breaking were bearable. The day an Arabic PM, working in a script that runs right to left, spotted something and filed a request, that was work of an entirely different kind.</p><p>Since people could not touch every string by hand, we also left machine translation switched on, and then forgot about it for long stretches, until some unidentifiable phrase surfaced somewhere on a screen. In screen templates the failure was usually a word rendered in the wrong sense from among synonyms and near-synonyms, and it tended to get caught only when a linguist, or a local employee in a country that language might ship to, reported it.</p><p>That covers the strings we made and put on screen, the ones Google took responsibility for translating. Users, of course, have no interest in who translated what. They want it to read naturally. Seen through that eye, even a translated result lands at different levels of satisfaction. At the bottom sits text that was never translated at all, and within that, Chinese characters showing up on a Korean screen raise discomfort immediately. English is comparatively fine, which made me wonder for a long time whether that was prejudice or something else, and we went at it with user interviews for quite a while. In my Play Store days our primary audience was Koreans living in Korea, and since Korea is not a country where mainland Chinese is in common use, Chinese characters rendering on screen was treated as a serious bug in its own right. A step above that are the strings that read like machine translation, obvious at a glance, always posing the question of whether that counts as effort. At the top sits translation someone cared about. All of that, though, is about text we prepared ourselves. What a content provider uploads with their own hands carries their intent, and while the Play Store was promoting automatic translation as a feature to be proud of at the time, you cannot simply translate and display what someone else posted without permission. That is where another story begins.</p><p>Promising forty languages is easy. Keeping the promise is not. We held out an empty vessel first and filled it as we went, waiting with a knot in the stomach over whether the pixels would break or the wrong contents would arrive, and quality sat pooled in that gap the whole time. A service can consider a translation done and a user will not necessarily receive it that way, so satisfaction varied from person to person, and gauging that difference stayed with me longer than moving words from one language into another did. I remember, now and then, the reviews from professional linguists and tech writers, which operated on another level.</p><p>That wall seems to have mostly come down since LLMs. Watching something that gave us so much trouble pass by now without comment leaves me feeling the distance of the years, and I send a quiet word of encouragement to whoever is out there filling blanks the way I once did.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/193"><em>brunch.co.kr/@chaesang/193</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/filling-forty-languages/"><em>https://ieumtech.net</em></a><em> on November 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cce8c31bb7ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Keeping the Country Out of the Code]]></title>
            <link>https://chaesangjung.medium.com/keeping-the-country-out-of-the-code-1a42811a79d0?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a42811a79d0</guid>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[localization]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 21:19:53 GMT</pubDate>
            <atom:updated>2026-11-11T21:22:55.514Z</atom:updated>
            <content:encoded><![CDATA[<p>A branch written for one country is a debt every other user helps pay down. So we kept the country in config instead of in the code.</p><p><em>The eighth piece in a series on how country and language were never one and the same.</em></p><p>Build a service that goes out to the whole world, search or a store, and sooner or later the sentence we treat each country differently comes down to how you write one line of code. Where you draw a country’s edge in that line, and how you handle it, turns out to be how the service sees the world.</p><p>The easy temptation is to write Korea into the code itself. If country is Korea, do this. But the moment you drop a branch for Korea alone into the middle of a path every user in the world walks through, the cost of that one simple comparison gets split among the 99% and more who have nothing to do with it. A handling for the few, paid for by the many, which is a Pareto of its own running the wrong way. It gets much worse once you reach locale-aware scoring, where each country carries its own weights. The maintenance side was no better. Google’s engineers worked inside a culture that treated code itself as a kind of debt, so everyone knew each of those branches was a balance to be paid down for years afterward. The strict and varied review process says as much.</p><p>It is not that such branches never existed. There was a good deal of if China in the code. The Chinese market had reasons, more than a few, and as each reason got its accommodation the branches kept multiplying. When Google withdrew from China, what surfaced in the reckoning was that all that accumulated if China was an awkward load, technically and in terms of the SDLC both, and development teams fought the latent code debt for a long while on either side of it. I watched that from fairly close up. With a case nowhere near China’s weight behind it, an argument for adding if Korea was short on standing before it started.</p><p>So the preferred shape was not if Korea but if this feature is on for this country. Rather than lifting Korea out for special treatment, you put every country on the same frame and kept the differences in configuration. Korea’s real-name verification, which the law there required, went in not as a country name in the code but as a feature flag: is this feature on for this locale. Which features went to which countries lived in a well-kept config off to the side. The code could then be refactored freely without the behavior drifting out of line. The hope behind it reached further, too. What if this same verification turns up in some other country, or the law changes and it has to be switched on or off differently.</p><p>That fastidiousness was not a matter of taste. It came out of a culture that treated merging something into shared code as adding to a commons, with strict review and approval standing in the way. Even a handling meant for one country was treated as a feature launch in its own right rather than an exception slipped in quietly, and even where the law demanded it, how far the effect reached was worked out through experiment and analysis before any of it was folded into the decision. For more on that culture, see <a href="https://abseil.io/resources/swe-book">Software Engineering at Google</a>, which Google has put online to read for free.</p><p>But before you can give someone something fitted to their country, you have to know what country they are in, and a user who has not even signed in leaves you few clues. You guess the language from what the browser sends in its settings, you place them by the IP they connected on, the geo IP, and for a time you read the door they came in through, a domain like google.co.kr. Some of it was conditions laid out as if statements. Some of it was probability models. That is how I remember it.</p><p>What is interesting is that the weight of that phrase, a signed-out user, turned over completely with time. Back when Chrome was uncommon and so were smartphones, nearly all the traffic pouring in from Korea came out of PC bangs, the internet cafes the country ran on, and nearly all of it was anonymous. Now, when everyone goes through the day signed in to their own account, what is left in that seat is the small number who deliberately opened an incognito window. The work of Google’s Chrome team and identity team had a doggedness worth respecting, and the few chances I had to work alongside them taught me a great deal.</p><p>Once someone signs in, things get easier. Single sign-on, living under one account wherever you go, was not a culture that existed before Chrome, and doing it at the browser level was an ambitious and enormous project even for Google. As it unfolded, Google began attaching probabilities to accounts. Look across what has passed through an account, mail and documents and the rest, and you can set a ratio for language. My personal account sat at roughly seven to three, Korean to English. My work account ran the other way, about two to eight. It was one of the ways a service spanning many products came to understand a user, and language, it turned out, could be had as a probability.</p><p>A country could not. The moment payment is involved a country has to be picked, exactly one, and I kept my personal account in Korea where the carrier and the bank were wired in, and my work account in the United States. Moving the personal account’s country to the U.S. after I left turned out to be more involved than I expected, and once I understood that the procedure connects to international cooperation on tracing things like terrorist financing, I decided the intent was worth following even where it was inconvenient. Taxes and payments, when you think about it, are the heaviest measure of which country a person belongs to. Language you can live across, seven to three. In front of money, and taxes above all, a country has to be one.</p><p>Domain and IP, account and taxes. Pull every clue together and the line called a country stays a little too blurred to write into code, while users cross it without giving it a thought. The effort to keep the country out of the code was, looking back, a way of admitting that blur. To everyone running a global service, a small measure of respect, and of sympathy.</p><p>Part of locale — Country and Language — how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization. See the full series: <a href="https://ieumtech.net/writing/series/locale">https://ieumtech.net/writing/series/locale</a></p><p>Adapted from my Korean essay on Brunch: https://brunch.co.kr/@chaesang/192</p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/keeping-the-country-out-of-the-code/"><em>https://ieumtech.net</em></a><em> on November 12, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a42811a79d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Crawler Has a Country, Too]]></title>
            <link>https://chaesangjung.medium.com/a-crawler-has-a-country-too-e12052e44a0e?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/e12052e44a0e</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[search]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 00:00:32 GMT</pubDate>
            <atom:updated>2026-11-03T13:26:15.255Z</atom:updated>
            <content:encoded><![CDATA[<p>Most of the web is no longer human traffic. The question of where a bot calls from is older than the AI crawlers now fighting over it.</p><p><em>The seventh piece in a series on how country and language were never one and the same.</em></p><p>Spread out the world’s web traffic and look at it, and the share that is not human comes as a surprise. One industry report put automated traffic past 53% of the web in 2025, up from 51% the year before. Not all of it is the same kind of bot. Some, like a search engine’s crawler, go around announcing who they are and what they want. Others, a DDoS flood among them, have to be turned away at the door.</p><p>For a long time a bot mostly meant crawl, a search engine going around reading the world’s documents in advance so it would have them ready. Now fetch is growing fast: someone asks a question, and the web gets searched on the spot to put an answer together. That is what LLM services are doing when they go looking while a user waits. Break AI crawling down by purpose and the shape is clear enough. In Cloudflare’s accounting for the third quarter of 2026, close to half of AI crawler requests, 47.6%, were collection for model training, 8.3% were for search indexing, a little under 40% were crawls with mixed purposes, and the fetches that actually arrive to answer a person’s question came to 3.7%. Scraping and answering will both keep growing for a while yet. What strikes me is the patience. In the old search, a result a beat late was a crisis. Now people wait, willingly, while something runs a live fetch on their behalf.</p><p>Asking where a bot comes from did not start with LLMs. When I was running Korean search, around the time google.co.kr and google.com were being folded into one, the question was already there. How far should the same query from the same person diverge depending on whether they are sitting in Korea or in the United States. Near the top of the priority list was a rule that sounds almost too plain. Click the result and you must not land on a not found or a not allowed. Content behind a wall, a post that was never set to public, a site that geo-blocks Korean visitors outright: we wanted those kept out of the results at roughly the same level as a dead link.</p><p>Geo-cloaking, where the same URL glances at where you are connecting from and hands you a different page, turned out to be more common than you would think. Hand a Korean user a page crawled in the United States and what they actually meet on the click can be a different thing entirely. So in principle the results served to Korean users had to come from a crawler that ran in Korea. For sites that treat every country alike it makes no difference. For sites with logic inside, the same Google crawler was treated one way arriving from the United States and another arriving from Korea, and the answer we settled on was to run crawlers per country. It was resources, which meant it was an investment, which meant rounds of persuasion, and there were plenty of moments where we came up short.</p><p>For a stretch my job was finding the good Korean sites and securing crawl quota one at a time, and most of what I remember is how hard it was. Back then the assumption that a foreign company was carting off Korean content sat even on sites that were already public, so knocking was delicate from the start. Search is not advertising. You do not pay for the slot; the value gets set only when a user chooses to click. That made the question of which sites to include heavier than it looks. And Korean sites often did not follow web standards closely enough, so a crawler reading what was inside them was not even straightforward.</p><p>Sites have their own reasons for opening the door selectively. A U.S. insurance site blocking connections from Korea, Netflix showing only what is licensed in each country: that is licensing and regulation doing its work. Turn the same gating around and apply it to public data, or to something you actually want known, and the story changes. We paid for this with our taxes, so foreigners should not see it. In an age when crawlers are how the world gets read, that comes close to barring your own gate. Not that you can crawl every site from every country. Which sites are worth the effort and the traffic was its own weighing, and it tangled naturally into the judgment of whether a site was worth it at all, which is to say site quality.</p><p>And once a name starts getting known, crawlers and bots from who knows where hammer at the door. The very crawl that is priceless to a search engine comes back to the site owner as maintenance cost and operational burden. Recent measurement bears it out. Fastly found that in January 2026 bots made up 49% of requests against humans at 51%, that only 1% of those bots were verified or wanted, and that 60% of all origin traffic came from bots. Cloudflare mitigated 27.8 million DDoS attacks in the first half of 2025 alone, 130% of the total for all of 2024, with hyper-volumetric attacks averaging 71 a day. The same traffic is a road to the world for one party and a bill to be paid by another.</p><p>SEO and web standards were a wall back then, hard for most people to clear, and in places a matter of tricks. With AI putting a standards-compliant page within anyone’s reach, that wall is neither a wall nor a trick anymore, and the names making the rounds in its place are AEO and GEO, answer engine and generative engine optimization. Looking back, the wariness Korean sites held toward Google’s crawler was a trailer for the fight the whole world is now having over crawlers that collect training data. What sites expect has probably not moved. What I regret is that in a season when sovereignty is the subject of the day, opening information up still gets counted as letting it leak abroad.</p><p>Now that bots travel the web more than people do, and the bots themselves split into crawl and fetch, that old question of where a crawler comes from and what it takes has only gotten heavier. When an AI builds its answer out of pages it fetched on the spot, and one of those pages cannot be reached from where the person asking is standing, whose world is that answer a view of. I have been away from traditional search a long while now. I still find myself wondering who is carrying the old rule, the one that said if you click it, it has to be there.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/191"><em>brunch.co.kr/@chaesang/191</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/a-crawler-has-a-country-too/"><em>https://ieumtech.net</em></a><em> on November 3, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e12052e44a0e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2026 Cloud Solutions Report, #10: Inference Infrastructure (2) — Token Supply]]></title>
            <link>https://chaesangjung.medium.com/2026-cloud-solutions-report-10-inference-infrastructure-2-token-supply-703db03fc617?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/703db03fc617</guid>
            <category><![CDATA[inference]]></category>
            <category><![CDATA[ai-infrastructure]]></category>
            <category><![CDATA[enterprise-technology]]></category>
            <category><![CDATA[2026-cloud-report]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 00:00:14 GMT</pubDate>
            <atom:updated>2026-11-01T09:02:58.631Z</atom:updated>
            <content:encoded><![CDATA[<p>Sold out before it is built, and 95% unused once rented. Reading the token factory’s P&amp;L has become the buyer’s job.</p><h3>Sold Out, and Sitting Idle</h3><p>This report calls a facility that puts GPUs in a data center, feeds them electricity, and sells nothing but tokens a token factory. The last report was about the side that buys tokens from that factory. This one is about the side that builds and runs it, the side that supplies tokens. The 2026 token factory can be summed up in two numbers that point in opposite directions. A neocloud is a newer cloud operator whose main business is renting GPUs and which is not a hyperscaler. One of them, Nebius, said on its second-quarter earnings call that newly energized GPU capacity was being contracted as fast as it came online, leaving nothing spare in the quarter, and that it was negotiating short-term contracts at $40 to $50 million per megawatt [1]. That same spring, a report from Cast AI, a Kubernetes cost optimization vendor, put average enterprise GPU utilization at 5% [2]. The same GPU sells out at a premium on the factory side while, on the side that rented it, 95% of the capacity goes unused.</p><p>Both numbers can be true in the same year because a token factory gets paid when a GPU is sold, not when it is used. A buyer on a committed contract who leaves the GPU idle does not reduce the factory’s revenue. The buyer-side average is one tenth of the 40 to 50% utilization that the last report treated as the line between a dedicated contract and serverless, so what happens inside the factory does not stay with the companies building factories.</p><p>How to measure a token factory, and how to describe it, is still being decided by the company that sells the chips. At his GTC keynote in March 2026, NVIDIA’s Jensen Huang said that “every data center, every single factory, by definition, is power constrained,” and that “a one gigawatt factory will never become two. It’s physically constrained” [3]. NVIDIA describes factory economics as revenue per megawatt, with the weighted product of throughput and price as the engine [4]. Put in the language of the people building these places, power is the term you have to lock in years ahead, and tokens per watt is the term you can compete on right now with chips and software. The companies in this report divide according to which of the two they chose to lean on.</p><p>This report asks three questions. What does a token factory’s income statement look like? Why is everyone trying to leave their own layer? And by what path do those conditions reach the contract and the unit price on the buyer’s side? The report takes them in order: first the income statement, then six players and where each placed its weight, and finally the points where all of this touches the buyer’s contract.</p><h3>The Token Factory’s P&amp;L</h3><h3>Revenue, Cost, Profit</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6YjJMStc-QNF4_-5.png" /><figcaption>Figure 1: The three layers of the inference industry and the scope of this report.</figcaption></figure><p>What this report looks at is the bottom two of the three floors the last report used to describe the inference industry: the inference infrastructure that squeezes more tokens out of the same GPU, and the AI data center that sells compute time. The two together are what this piece means by a token factory, which is a different thing from Token Factory, one of the product names used on the top floor where finished tokens are sold. The income statement reads in three lines. Revenue is contracted power multiplied by tokens per watt, by the price of a token, and by utilization, meaning the share of that power actually sold. Cost is depreciation, the practice of spreading a GPU’s purchase price over some number of years, plus power and interest. Profit is the difference. Of the four variables in revenue, power cannot be increased once the site is built, tokens per watt rises only by changing chips and software, and price is set by the market, which leaves utilization as the one variable a sales organization can move this quarter. The capital sunk into the building and the GPUs returns as revenue only to the extent the capacity is sold. In this piece, a factory’s utilization means the share of contracted power that is under contract and being billed, and a factory being sold out means that share has reached 100%.</p><p>What these lines look like in practice can be read off CoreWeave’s second quarter of 2026. Adjusted EBITDA margin, the margin before interest, taxes, and depreciation, was 59% [5]. Take depreciation out and operating margin falls to 5%, and once interest expense running at 25% of revenue is layered on, the quarter ends in a net loss. A little over half of revenue went to the equipment and what was left went to the debt. However good a token factory’s revenue looks, the costs decide the outcome.</p><h3>Prices Fall, Factories Sell Out</h3><p>Start with the price a buyer sees. Silicon Data tracks GPU rental rates daily the way a commodity desk tracks a raw material and publishes them as an index. In its October 8, 2026 values, the hyperscaler H100 rate is more than 2.5 times the neocloud-centered figure [6]. Over a longer horizon, the on-demand rate for the H100, the Hopper generation, has come down to less than half of its early-2024 level [7]. Yet the H100 one-year commitment index that SemiAnalysis publishes monthly, built only from signed one-year contract prices, rebounded about 40% from its October 2025 trough through March 2026 [8]. SemiAnalysis attributes this to delayed Blackwell deliveries leaving even August and September 2026 volume already booked. On-demand falling while contracts rise is the shape of the market in the first half of 2026. Table 1 puts the three indices side by side: the same H100 ranges from $1.70 to $7.15 an hour depending on where and on what contract you rent it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Gf8rxc4BD4O7sfPf.png" /><figcaption>Table 1: Hourly rental price for a single GPU, in dollars. Sources: Silicon Data (October 8, 2026), AIMultiple (September 2026), SemiAnalysis.</figcaption></figure><p>Rising contract prices are the premium a sold-out factory collects. Nebius’s $40 to $50 million per megawatt is one example. CoreWeave said in September 2026 that short-term contracts of three to six months signed in the third quarter worked out to roughly $40 million per megawatt on an annualized basis [9]. Run the arithmetic and that is about $7 an hour for a Blackwell-class GPU, above the neocloud on-demand rate in Table 1. Prices fall and factories still sell out because generations turn over quickly: older GPUs lose value every time a new generation ships, while the new one is scarce enough to carry a premium in the contract market. So the buyer’s question, whether to take an H100 cheaply now or wait for Blackwell, becomes a question of where on the factory’s depreciation curve you sign.</p><h3>Cost per Token, and Depreciation as a Floor</h3><p>Convert the hourly GPU rate into a cost per token using InferenceX, SemiAnalysis’s cost analysis tool, and the width of the margin shows. Serving one open model on GPUs a hyperscaler owns outright costs $0.05 to $0.15 per million tokens on a B200 and $0.17 to $0.40 on an H100 [10]. A GPU lowers its cost per token the more requests it batches together, at the price of a slower answer for any one person, so the width of those ranges comes from the speed a user feels, and cost climbs steeply as you buy speed. The serverless selling price for open models, as the last report found, runs $0.1 to $1 per million tokens. The gap between selling price and cost is the token factory’s margin, and at the bottom where the two ranges overlap, that margin nearly disappears.</p><p>The other half of cost is set by accounting, specifically by how many years a GPU is written down over. CoreWeave and Alphabet use six years. Amazon moved some servers back from six years to five in early 2025, citing the pace of AI technology [11]. On one side is the claim that a GPU’s useful life is two to three years; on the other, the argument that a GPU keeps earning as it moves down from training to inference and then to lighter internal workloads. This argument does not stay an accounting matter, because the depreciation schedule sets the floor below which a price cannot go without reporting a loss. The cash costs, power and interest, are the same for everyone, so a company on six years can compete profitably at a lower price than one on five. That profit rests on the premise that the GPU really earns for six years. If the useful life turns out to be shorter than the schedule, the companies on six years are also the ones who take the write-down first.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8Q07Fib6lhHDs4_B.png" /><figcaption>Figure 2: Annual revenue potential of a one-gigawatt factory by price tier and GPU generation. Source: NVIDIA Developer Blog (March 2026), NVIDIA’s own estimates.</figcaption></figure><p>Underneath all of these numbers sits power. NVIDIA claims that the Vera Rubin platform delivers up to 10 times higher inference throughput per megawatt than Blackwell, and about 10 times lower cost per token, for one workload it names [4]. If power is fixed, the way to raise revenue is to replace the GPUs on that same power with a newer generation, which means a token factory grows by generational turnover rather than by GPU count (Figure 2). But replacing a generation means the older GPUs have to vacate before they have served their six years, so this growth path runs straight into the depreciation argument above. That is part of why so many of the companies in the next section head for power first.</p><h3>Margin Upstairs, Power Downstairs</h3><p>Split the income statement by layer and an asymmetry appears: the potential margin is upstairs and the power is downstairs. That is why the token factories of 2026 are all trying to leave their own floor (Figure 3). Neoclouds buy inference API and software companies and move up. Companies selling tokens move down toward their own GPUs and power, as Together AI did in securing 500 megawatts. Chip companies open their own clouds and marketplaces on top of their own silicon, and hyperscalers build their own chips to pull the cost of the bottom floor off NVIDIA’s price list.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WMUKx0FGirrchJ05.png" /><figcaption>Figure 3: The four directions of vertical integration.</figcaption></figure><p>The four paths are not equally open. CoreWeave’s plan to acquire the data center operator Core Scientific and with it 1.3 gigawatts of power was voted down by the target’s shareholders in October 2025 [12]. The same company’s acquisitions in the upward direction, an experiment-tracking tool and a reinforcement learning startup, closed as planned [13]. Whoever holds power now knows what that asset is worth and does not sell it easily, so power has become something you build rather than something you buy. Which leaves the question of what token factories standing on the same GPUs under the same power constraint are supposed to compete on. Creating a reason the buyer cannot switch is also a way of holding onto that buyer.</p><h3>Vendor Deep Dives</h3><p>Six players, taken in order: three neoclouds, the chip company, the specialized chip camp, and the hyperscalers’ own silicon. For a reader in Korea the last section is the closest to daily practice, but the rest is worth following too, because part of hyperscaler capacity physically sits on neocloud sites, the floor under prices is set by the neocloud indices, and the chip company collects on both the silicon and the software. Each section starts with how that player sees the world and ends at the point where it touches the buyer’s contract.</p><h3>CoreWeave: Scale as Collateral</h3><p>In CoreWeave’s view of the world, this business is won on scale. Contract the power first, put the newest generation of GPUs on it before anyone else, sell that capacity long-term to large customers, and use those contracts to finance the next build, turning the cycle faster than the competition. At the end of the second quarter of 2026, backlog, meaning contracted revenue not yet recognized, stood at $104.2 billion [5]. The cycle rests on two anchor customers who commit to a large share of capacity on long terms. The OpenAI agreement runs up to $22.4 billion [14], and Meta has committed roughly $35 billion through the end of 2032 across two deals [15].</p><p>This factory has also stepped onto the floor above, launching Dedicated Inference, where the customer chooses the GPU and the serving software. Meanwhile the contracts keep getting longer: about 20% of backlog will not be recognized as revenue until after 48 months [16]. A token factory matching contract length to depreciation schedule is a factory proposing to share the risk of GPU lifespan with the buyer.</p><p>The weakness is that the cycle depends on debt and on exactly two customers. As of the end of June 2026, interest expense came to 25% of revenue, and 63% of first-half revenue came from the top two customers [16]. If either one moves to its own chips or its own data centers, that capacity has to find another buyer, and the interest keeps going out in the meantime. The buyer ends up in the position of having to worry about its factory’s customers.</p><h3>Nebius: Winning on Tokens per Watt</h3><p>In the view from Nebius, headquartered in the Netherlands, power and GPUs are commodities everyone buys at the same price, and the contest is over how many more tokens you can pull out of the same power. Its target for power under contract by the end of 2026 has been raised every quarter since November 2025 and now stands at 5 gigawatts [1], which is not power running today but the total the company intends to tie up by contract across many sites. Its anchors are Microsoft and Meta [17]. Where it parts from CoreWeave is in accounting: Nebius extended its depreciation schedule to five years in 2026 but still writes down faster than CoreWeave’s six [11], which means it books the profit on the same revenue more conservatively.</p><p>Where this factory places its weight is software. It took the phrase NVIDIA popularized and made it the name of its own inference service, Token Factory. In May 2026 it agreed to acquire the inference optimization startup Eigen AI, and in the same month brought over the core team from Clarifai and licensed its inference intellectual property [18][19]. The company’s explanation is that competitive advantage is shifting from the amount of compute to the inference efficiency that software determines, which in the language of the previous section is a declaration that it intends to win on tokens per watt rather than on gigawatts. On the buyer’s screen, GPU clusters and the token API sit in one console, so predictable workloads can be bought by the hour and exploratory ones by the token without changing vendors. The commitment terms still deserve a close reading.</p><p>Two risks run the other way. Having raised the contracted power target every quarter, the company now has to stand up more than a gigawatt a year from 2027, and a delay in either power or GPU delivery pushes the promise out. And most of its revenue comes from Microsoft and Meta, both of which build their own chips, so their silicon roadmaps are this factory’s demand risk.</p><h3>Crusoe: Building the Power First</h3><p>In the view from Crusoe, which started out mining bitcoin on stranded natural gas, the bottleneck is power. Microsoft’s Satya Nadella said much the same in a podcast reported in early November 2025: “the biggest issue we are now having is not a compute glut, but it’s power” [20]. Crusoe solves that problem from the other end, securing the site and the generation first and putting GPUs on top. In Abilene, Texas, after the campus leased by Oracle and provided to OpenAI’s Stargate [21], it is building a 900-megawatt campus for Microsoft with on-site generation of the same capacity [22]. Announcing a funding round in September 2026, Crusoe put total contracted capacity above 6 gigawatts, with 1 gigawatt live [23].</p><p>In November 2025 this factory added Managed Inference, selling per-token billing and reserved throughput, which gave it all three floors from the power plant to the token API [24]. What remains as risk is that a company still private has to lean on private capital to keep building at gigawatt scale, and that its anchor is concentrated in Stargate. What a buyer should check is whether power has actually been energized at the site where the contracted capacity will sit, and which assets under that vendor’s name are in fact the vendor’s own.</p><h3>NVIDIA: A Chip Company That Wants Every Layer</h3><p>In NVIDIA’s view of the world, token factories compete with one another, and whichever wins, that factory stands on NVIDIA silicon. But the NVIDIA of 2026 has built itself a position on every floor above the chip. In distribution, it placed itself between buyer and factory with DGX Cloud Lepton, a marketplace that gathers neocloud capacity alongside AWS and Azure GPU capacity on one screen for developers to choose from [25]. In software, it open-sourced the serving framework Dynamo to hold the standard, and sells NIM, the packaging that runs a model on top of it, under a paid license in production [26]. In silicon, the Groq technology it licensed non-exclusively for a reported $20 billion in December 2025 became the Groq 3 LPU, fabricated on a Samsung process, and the LPX systems that entered full production in August 2026 [27]. In model distribution, it filed in September 2026 for the $12.93 billion acquisition of Hugging Face, the repository where open models live [28].</p><p>NVIDIA is betting in particular on combinations of silicon. Paired with Groq 3 LPX, it says, Vera Rubin delivers up to 35 times higher throughput per megawatt for trillion-parameter, high-context workloads [4]. The stage that reads the whole input at once goes to the GPU, part of the stage that writes the answer one token at a time goes to the LPU, and the curve where cost climbs as you buy speed is meant to be flattened by the mix. While token factories compete on price, the metric, the marketplace, and the serving standard all belong to NVIDIA.</p><p>Widening across layers widens the risk too. Its largest customers, the hyperscalers, are simultaneously competitors building their own chips. The Hugging Face acquisition carries antitrust attention over the neutrality of the open model ecosystem. And the Lepton screen does not show which factory and which site the capacity a buyer picked actually sits on, or who answers when it fails. So the buyer’s questions are who the contracting counterparty is for capacity bought through a marketplace and who is accountable during an outage, and whether there is any layer of this industry NVIDIA is not in. If every layer runs through one company, switching vendors leaves the price floor and the software stack unchanged, and the switchable state the last report argued for is blocked at the chip company rather than at the token factory. The next two sections are about the attempts to find a way around that.</p><h3>The Specialized Chip Camp: Not Standing in NVIDIA’s Line</h3><p>In this camp’s view of the world, the GPU is an expensive way to make a token. Cerebras, which listed on Nasdaq in May 2026, took about 70% of its second-quarter revenue from cloud and services rather than hardware, which puts it closer to a token factory than to a chip company [29]. At the other end, Groq licensed its inference technology to NVIDIA and stayed independent, raising money in June and again in August 2026 and turning toward being a neocloud. The capacity it is adding runs on NVIDIA systems rather than its own LPUs, and its valuation halved over the year [30][31]. Among the newer entrants, Taalas, which etches a single model into a chip to make it very fast, was announced in August 2026 as an AMD acquisition [32]. d-Matrix, which uses general-purpose memory in place of the high-bandwidth memory GPUs need, entered production in June and acquired an inference deployment software company in August [33].</p><p>The companies that stay with their own silicon have a few things in common. They route around the two bottlenecks that constrain NVIDIA GPU supply, high-bandwidth memory and its packaging process. They keep a path open, as Cerebras has, to sell tokens through their own cloud in periods when chips are not selling. And they choose to sit beside the GPU rather than to displace it: AMD and Cerebras introduced a joint inference platform in July 2026 in which GPUs and wafer-scale chips divide the work [34]. The result is an income statement that increasingly resembles a neocloud’s. Revenue recognized from the 750-megawatt agreement Cerebras signed with OpenAI alone came to about 32% of quarterly revenue [29]. The risks resemble a neocloud’s too: Cerebras depends on two customers for 66% of revenue and on TSMC alone for chips [29], and a design like Taalas that fixes a model in silicon has to be refabricated when the model changes.</p><p>In front of this camp the buyer’s question becomes simple. How many of our workloads genuinely require generation speed measured per second, and what demand do they create? The answer decides whether this camp is a counterparty to contract with or a development to watch.</p><h3>Hyperscalers: The Biggest Customer, Building Its Own Chips</h3><p>In the hyperscalers’ view of the world, a large share of cost is NVIDIA’s price list, and all three have now moved their own chips up a generation. Google’s seventh-generation TPU, Ironwood, reached general availability in April 2026 [35]; AWS’s Trainium3 shipped in December 2025 [36]; and Microsoft unveiled Maia 200, an inference-only chip, in January 2026 [37]. What these three are aimed at is less external sales than pulling their own factory’s tokens per watt and depreciation curve off NVIDIA’s price list. Microsoft says Maia 200 delivers 30% better performance per dollar than the previous generation in its own fleet [37].</p><p>They differ in how open they are. Maia is internal only, Trainium can be rented by anyone inside AWS, and only Google has moved to sending its own silicon outside its own cloud. Anthropic secured gigawatt-scale TPU capacity through an expanded agreement in April 2026 [38], and Meta has been reported to be in talks to buy chips outright after leasing them [39]. At the same time the hyperscalers rent whole blocks of capacity from neoclouds to cover what they lack, which makes them both customers and competitors of the token factory.</p><p>Two things are worth a buyer’s attention. First, TPU and Trainium each demand their own software stack, particularly for training and custom optimization, so a workload optimized on them is harder to move than one on GPUs, and the buyer carries that portability cost. Second, nothing anywhere promises that the cost savings a hyperscaler gets from its own chips will show up in the buyer’s token price.</p><h3>Sovereign: Token Factories at National Scale</h3><p>Efforts to keep data and compute inside national borders go by the name sovereign AI, and the drive to build token factories is not only corporate. In July 2026 the European Union opened a call for AI gigafactories backed by €10 billion in public money and more than €20 billion in private investment [40][41]. In France, SoftBank has planned AI data center campuses [42]. HUMAIN, under Saudi Arabia’s sovereign wealth fund, agreed with AWS to build an AI zone in Riyadh [43]. Stargate UAE [44] and Microsoft’s Japan investment, announced with the explanation that it would help keep data processing inside Japan [45], sit on the same line.</p><p>In Korea, the National AI Computing Center broke ground on August 3, 2026 in Solaseado, Haenam, South Jeolla Province. At a project cost of roughly 2.5 trillion won, the plan is to install 15,000 advanced AI chips by 2028 and to grow that to 50,000 by 2030. Before completion, in 2027, some resources are to be opened early through private data centers so that startups, smaller companies, universities, and research institutes can use them through vouchers and discounted rates [46][47]. In the language of the previous section, this is public money securing power and GPUs first and opening the utilization above them to the private sector to fill. Whether the center will sell GPU hours or a token API has not been announced, and that decision will shape the demand that forms around it. The low utilization of enterprise GPUs cited earlier is the same warning for GPUs the public sector secures first.</p><p>Sovereign AI is an attempt to own three layers at national scale, but in none of these cases do the chips, the foundry, and the hyperscalers drop out of the picture. What a state can actually hold is the site, the power, where the data resides, and who operates what sits on top. The silicon is almost entirely imported. Serving software and models are the layer where a domestic share can be grown. A hyperscaler’s in-country region satisfies the residency condition but not the operator condition, so whether something counts as sovereign depends on which regulatory requirement it is meant to meet. And in places like Europe and Korea, where power costs more and sites are smaller, cost per gigawatt runs higher than at a global factory, which makes sovereign a product with a premium attached for the buyer.</p><h3>Closing: Recommendations for CIOs and CTOs</h3><p>If the last report argued for building a purchasing capability, this one argues for widening what that capability examines to include the vendor’s financial statements. A token factory’s circumstances enter the buyer’s contract along four paths: the order of capacity allocation, price and contract length, the vendor’s financial risk, and the premium on data residency. The recommendations follow that order. For organizations buying tokens through an API, the third and fourth apply first; for those contracting GPU hours or dedicated capacity, the first and second.</p><p>When a token factory is sold out, new customers go to the back of the allocation queue, and in a market where hyperscalers contract neocloud capacity in bulk, going through a hyperscaler can actually be faster. The trade is that it becomes harder to know which site your capacity sits on. An organization buying dedicated capacity or committed throughput is contracting for capacity that has not been built yet, so the first action is to get the site, the power energization schedule, and the GPU generation from the vendor in writing. A factory sells future capacity, so always allow for the possibility that the capacity in the contract has not been secured.</p><h3>2. If you accept a long lock-in, get a generation-change clause</h3><p>The last report’s advice to stay in a switchable state becomes, in a market of lengthening contracts, a matter of contract clauses rather than endpoint addresses. A long lock-in buys a lower price in exchange for carrying the GPU lifespan risk in place of the token factory, and a long contract signed into a sold-out market also locks in the scarcity premium. So if you accept a long lock-in, get it in a clause that you can move to a new generation within the same contract when Hopper gives way to Blackwell and Blackwell to Rubin, and at what price. The cost difference that follows from depreciation schedules, and NVIDIA’s claim of roughly a tenth of the cost per token on Rubin, are the grounds for that negotiation. Check whether your existing contracts already contain a clause adjusting price when a new generation is deployed, and prepare for the negotiation by matching contract length to the generation cycle or setting annual price step-downs in advance.</p><h3>3. Look at your vendor’s top customers, and at their chip roadmaps</h3><p>A token factory that takes more than 60% of revenue from its top two customers faces one of two outcomes if an anchor leaves: the freed capacity floods the market and prices collapse, or the vendor’s finances wobble and service continuity is at risk. Which one it is depends on that factory’s debt and interest expense and on who the anchor is, so read customer concentration, debt principal, and the identity of the anchors out of a listed vendor’s quarterly filings and put them in your contract review checklist. For a private vendor, read the public signals, the large contract announcements and the funding reports, and ask for clauses covering change of control, service discontinuation, and the right to move data and workloads out in a bankruptcy.</p><h3>4. Pay the sovereign premium only where data residency is a real constraint</h3><p>Sovereign is a product with a premium attached, and paying it on every workload does nothing for your costs beyond helping a token factory fill its utilization. Go workload by workload and write down whether the residency constraint comes from law, from a contract, or from habit. The more workloads that fall under habit, the larger the share that can sit outside the sovereign premium. For a public sector organization, that same distinction becomes the rule for dividing workloads between the national AI computing center and commercial clouds.</p><p>This report began with two numbers: token factories sold out before they are built, and 95% of GPUs sitting unused. That both appear in the same year means a GPU becomes a premium or leaves nothing but depreciation depending on which contract it is tied to. Closing out these two reports on inference infrastructure, the thing worth adding is that reading the token factory’s income statement, the document on the other side of that contract, has now become the buyer’s job.</p><h3>References</h3><p>[1] Nebius. “Nebius Group N.V. Q2 2026 Earnings Call Transcript.” August 12, 2026. <a href="https://www.tradingkey.com/news/transcripts/262106466-tradingkey">https://www.tradingkey.com/news/transcripts/262106466-tradingkey</a> · Results 6-K: <a href="https://www.sec.gov/Archives/edgar/data/0001513845/000110465926094844/nbis-20260812xex99d1.htm">https://www.sec.gov/Archives/edgar/data/0001513845/000110465926094844/nbis-20260812xex99d1.htm</a> · Contracted power target over time: Nebius Q3 2025 Letter to Shareholders, November 11, 2025. <a href="https://www.marketscreener.com/news/nebius-letter-to-shareholders-shareholder-letter-q3-ce7d5fd9df8df025">https://www.marketscreener.com/news/nebius-letter-to-shareholders-shareholder-letter-q3-ce7d5fd9df8df025</a></p><p>[2] Cast AI. “2026 State of Kubernetes Optimization Report.” April 21, 2026. <a href="https://cast.ai/press-release/2026-state-of-kubernetes-optimization-report/">https://cast.ai/press-release/2026-state-of-kubernetes-optimization-report/</a> · Cluster count: IT Brief. “Cast AI report finds 5% GPU use in Kubernetes clusters.” April 22, 2026. <a href="https://itbrief.news/story/cast-ai-report-finds-5-gpu-use-in-kubernetes-clusters">https://itbrief.news/story/cast-ai-report-finds-5-gpu-use-in-kubernetes-clusters</a> · VentureBeat, May 8, 2026. <a href="https://venturebeat.com/infrastructure/5-gpu-utilization-the-401-billion-ai-infrastructure-problem-enterprises-cant-keep-ignoring">https://venturebeat.com/infrastructure/5-gpu-utilization-the-401-billion-ai-infrastructure-problem-enterprises-cant-keep-ignoring</a></p><p>[3] NVIDIA. “GTC 2026 Keynote Transcript.” March 16, 2026. <a href="https://www.nvidia.com/gtc/keynote/transcript.md">https://www.nvidia.com/gtc/keynote/transcript.md</a></p><p>[4] NVIDIA Developer Blog. “Scaling Token Factory Revenue and AI Efficiency by Maximizing Performance per Watt.” March 25, 2026. <a href="https://developer.nvidia.com/blog/scaling-token-factory-revenue-and-ai-efficiency-by-maximizing-performance-per-watt/">https://developer.nvidia.com/blog/scaling-token-factory-revenue-and-ai-efficiency-by-maximizing-performance-per-watt/</a> · LPX product page: <a href="https://www.nvidia.com/en-eu/data-center/lpx/">https://www.nvidia.com/en-eu/data-center/lpx/</a></p><p>[5] CoreWeave. “CoreWeave Reports Strong Second Quarter 2026 Results.” August 11, 2026. <a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx">https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx</a></p><p>[6] Silicon Data. “Silicon Index: H100 / H200 / B200 rental price indices.” Retrieved October 8, 2026. <a href="https://www.silicondata.com/products/silicon-index?gpu=h100">https://www.silicondata.com/products/silicon-index?gpu=h100</a></p><p>[7] AIMultiple. “GPU Index: Cloud GPU Pricing.” Retrieved September 16, 2026. <a href="https://aimultiple.com/gpu-index">https://aimultiple.com/gpu-index</a> · Trend against early 2024: GPUSmith. “H100 Rental Price History and Trends.” July 13, 2026. <a href="https://gpusmith.com/articles/en/h100-rental-price-history-trends">https://gpusmith.com/articles/en/h100-rental-price-history-trends</a></p><p>[8] SemiAnalysis. “The Great GPU Shortage — Rental Capacity — Launching our H100 1 Year Rental Price Index.” April 2, 2026. <a href="https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity">https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity</a></p><p>[9] CoreWeave. “CoreWeave Continues to Contract New Compute Capacity at Higher Prices.” September 17, 2026. <a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx">https://investors.coreweave.com/news/news-details/2026/CoreWeave-Continues-to-Contract-New-Compute-Capacity-at-Higher-Prices/default.aspx</a></p><p>[10] SemiAnalysis InferenceX. “MiniMax M2.5/M2.7: B200 vs H100 cost per million tokens.” July 2026. <a href="https://inferencex.semianalysis.com/compare-per-dollar/minimax-m27-b200-vs-h100">https://inferencex.semianalysis.com/compare-per-dollar/minimax-m27-b200-vs-h100</a></p><p>[11] Deep Quarry. “Depreciation of GPUs: Between Useful Life and Wishful Thinking.” December 7, 2025. <a href="https://deepquarry.substack.com/p/depreciation-of-gpus-between-useful">https://deepquarry.substack.com/p/depreciation-of-gpus-between-useful</a> · Nebius four years to five: Nebius 6-K, August 12, 2026 (same document as [1]) · CoreWeave six years: Wccftech, July 7, 2025. <a href="https://wccftech.com/coreweave-crwv-depreciates-its-gpus-over-6-years-while-its-competitor-nebius-uses-a-4-year-depreciation-period/">https://wccftech.com/coreweave-crwv-depreciates-its-gpus-over-6-years-while-its-competitor-nebius-uses-a-4-year-depreciation-period/</a> · Alphabet six years: Alphabet Form 10-K FY2025. <a href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm">https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm</a></p><p>[12] Yahoo Finance. “Core Scientific shareholders vote down CoreWeave merger.” October 30, 2025. <a href="https://finance.yahoo.com/news/core-scientific-8217-merger-coreweave-144146009.html">https://finance.yahoo.com/news/core-scientific-8217-merger-coreweave-144146009.html</a> · Deal terms: Nasdaq. “CoreWeave to Acquire Core Scientific.” July 7, 2025. <a href="https://www.nasdaq.com/articles/coreweave-acquire-core-scientific">https://www.nasdaq.com/articles/coreweave-acquire-core-scientific</a></p><p>[13] CoreWeave. “CoreWeave to Acquire OpenPipe, Leader in Reinforcement Learning.” September 3, 2025. <a href="https://www.coreweave.com/news/coreweave-to-acquire-openpipe-leader-in-reinforcement-learning">https://www.coreweave.com/news/coreweave-to-acquire-openpipe-leader-in-reinforcement-learning</a></p><p>[14] CoreWeave. “CoreWeave Expands Agreement with OpenAI by up to $6.5B.” September 25, 2025. <a href="https://www.coreweave.com/news/coreweave-expands-agreement-with-openai-by-up-to-6-5b">https://www.coreweave.com/news/coreweave-expands-agreement-with-openai-by-up-to-6-5b</a></p><p>[15] CoreWeave. “CoreWeave and Meta Announce $21 Billion Expanded AI Infrastructure Agreement.” April 9, 2026. <a href="https://www.coreweave.com/news/coreweave-and-meta-announce-21-billion-expanded-ai-infrastructure-agreement">https://www.coreweave.com/news/coreweave-and-meta-announce-21-billion-expanded-ai-infrastructure-agreement</a> · September 2025 agreement (about $14.2 billion): AOL/Bloomberg, September 30, 2025. <a href="https://www.aol.com/articles/coreweave-extends-hot-streak-metas-155527134.html">https://www.aol.com/articles/coreweave-extends-hot-streak-metas-155527134.html</a></p><p>[16] CoreWeave. Form 10-Q, second quarter 2026. <a href="https://www.stocktitan.net/sec-filings/CRWV/10-q-core-weave-inc-quarterly-earnings-report-d3b596a1f167.html">https://www.stocktitan.net/sec-filings/CRWV/10-q-core-weave-inc-quarterly-earnings-report-d3b596a1f167.html</a></p><p>[17] Nebius. “Nebius Signs New AI Infrastructure Agreement with Meta.” Business Wire, March 16, 2026. <a href="https://markets.chroniclejournal.com/chroniclejournal/article/bizwire-2026-3-16-nebius-signs-new-ai-infrastructure-agreement-with-meta">https://markets.chroniclejournal.com/chroniclejournal/article/bizwire-2026-3-16-nebius-signs-new-ai-infrastructure-agreement-with-meta</a> · Microsoft agreement: Nebius 6-K, September 8, 2025. <a href="https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm">https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm</a></p><p>[18] Nebius. “Nebius agrees to acquire Eigen AI, strengthening Nebius Token Factory as a frontier inference platform.” May 1, 2026. <a href="https://nebius.com/newsroom/nebius-agrees-to-acquire-eigen-ai-strengthening-nebius-token-factory-as-a-frontier-inference-platform">https://nebius.com/newsroom/nebius-agrees-to-acquire-eigen-ai-strengthening-nebius-token-factory-as-a-frontier-inference-platform</a></p><p>[19] Nebius. “Nebius welcomes Clarifai’s core team and licenses inference IP to strengthen Nebius Token Factory.” May 12, 2026. <a href="https://nebius.com/newsroom/nebius-welcomes-clarifai-s-core-team-and-licenses-inference-ip-to-strengthen-nebius-token-factory">https://nebius.com/newsroom/nebius-welcomes-clarifai-s-core-team-and-licenses-inference-ip-to-strengthen-nebius-token-factory</a></p><p>[20] Yahoo Finance (republished from Tom’s Hardware). “Microsoft CEO says company doesn’t have a chip supply problem, it has a power problem.” November 2, 2025. <a href="https://finance.yahoo.com/news/microsoft-ceo-says-company-doesnt-122000891.html">https://finance.yahoo.com/news/microsoft-ceo-says-company-doesnt-122000891.html</a></p><p>[21] Crusoe. “Crusoe, Blue Owl Capital and Primary Digital Infrastructure Enter Joint Venture.” May 21, 2025. <a href="https://www.crusoe.ai/resources/newsroom/crusoe-blue-owl-capital-and-primary-digital-infrastructure-enter-joint-venture">https://www.crusoe.ai/resources/newsroom/crusoe-blue-owl-capital-and-primary-digital-infrastructure-enter-joint-venture</a> · Oracle 15-year lease reported (via WSJ): Yahoo Finance, May 30, 2025. <a href="https://finance.yahoo.com/news/openai-secures-11-6b-texas-211501983.html">https://finance.yahoo.com/news/openai-secures-11-6b-texas-211501983.html</a></p><p>[22] Crusoe. “Crusoe Announces New 900 MW AI Factory Campus in Abilene, Texas to Support Microsoft AI Infrastructure.” March 27, 2026. <a href="https://www.crusoe.ai/resources/newsroom/crusoe-announces-new-900-mw-ai-factory-campus-in-abilene-texas-to-support-microsoft-ai-infrastructure">https://www.crusoe.ai/resources/newsroom/crusoe-announces-new-900-mw-ai-factory-campus-in-abilene-texas-to-support-microsoft-ai-infrastructure</a></p><p>[23] Crusoe. “Crusoe Raises $3.9 Billion Series F for its Vertically-Integrated AI Infrastructure Platform.” September 17, 2026. <a href="https://www.crusoe.ai/resources/newsroom/crusoe-announces-series-f-funding">https://www.crusoe.ai/resources/newsroom/crusoe-announces-series-f-funding</a></p><p>[24] Crusoe. “Crusoe Launches Managed Inference, Delivering Breakthrough Speed for Production AI.” November 20, 2025. <a href="https://www.crusoe.ai/resources/newsroom/crusoe-launches-managed-inference-delivering-breakthrough-speed-for-production-ai">https://www.crusoe.ai/resources/newsroom/crusoe-launches-managed-inference-delivering-breakthrough-speed-for-production-ai</a></p><p>[25] NVIDIA Newsroom. “NVIDIA Announces DGX Cloud Lepton to Connect Developers to NVIDIA’s Global Compute Ecosystem.” May 18, 2025. <a href="https://nvidianews.nvidia.com/news/nvidia-announces-dgx-cloud-lepton-to-connect-developers-to-nvidias-global-compute-ecosystem">https://nvidianews.nvidia.com/news/nvidia-announces-dgx-cloud-lepton-to-connect-developers-to-nvidias-global-compute-ecosystem</a> · AWS and Azure participation: NVIDIA, June 11, 2025. <a href="https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-DGX-Cloud-Lepton-Connects-Europes-Developers-to-Global-NVIDIA-Compute-Ecosystem/default.aspx">https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-DGX-Cloud-Lepton-Connects-Europes-Developers-to-Global-NVIDIA-Compute-Ecosystem/default.aspx</a></p><p>[26] NVIDIA. “Dynamo: datacenter-scale distributed inference serving framework.” GitHub (Apache 2.0 license). <a href="https://github.com/ai-dynamo/dynamo">https://github.com/ai-dynamo/dynamo</a> · NVIDIA. “NIM Microservices.” Product page (free for development and testing; production requires an NVIDIA AI Enterprise license). <a href="https://www.nvidia.com/en-us/ai-data-science/products/nim-microservices/">https://www.nvidia.com/en-us/ai-data-science/products/nim-microservices/</a></p><p>[27] Dataconomy. “Samsung mass produces Nvidia’s Groq 3 inference chip on 4nm.” March 17, 2026. <a href="https://dataconomy.com/2026/03/17/samsung-mass-produces-nvidia-s-groq-3-inference-chip-on-4-na/">https://dataconomy.com/2026/03/17/samsung-mass-produces-nvidia-s-groq-3-inference-chip-on-4-na/</a> · LPX full production: Asia Economy (English edition), August 26, 2026. <a href="https://view.asiae.co.kr/en/article/2026082608011466535">https://view.asiae.co.kr/en/article/2026082608011466535</a> · Licensing amount reported: CNBC, December 24, 2025. <a href="https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html">https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html</a> · Licensing announcement: Groq Newsroom, December 24, 2025. <a href="https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale">https://groq.com/newsroom/groq-and-nvidia-enter-non-exclusive-inference-technology-licensing-agreement-to-accelerate-ai-inference-at-global-scale</a></p><p>[28] NVIDIA. Form 8-K. September 2, 2026. <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm">https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm</a> · NVIDIA Blog. “NVIDIA to Acquire Hugging Face.” September 3, 2026. <a href="https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/">https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/</a></p><p>[29] Cerebras. Form 10-Q, second quarter 2026. <a href="https://www.sec.gov/Archives/edgar/data/0002021728/000162828026056357/cbrs-20260630.htm">https://www.sec.gov/Archives/edgar/data/0002021728/000162828026056357/cbrs-20260630.htm</a></p><p>[30] TechCrunch. “AI chipmaker Groq confirms $650M raise, re-staffs after Nvidia’s $20B not-acqui-hire deal.” June 22, 2026. <a href="https://techcrunch.com/2026/06/22/ai-chipmaker-groq-confirms-650m-raise-re-staffs-after-nvidias-20b-not-acqui-hire-deal/">https://techcrunch.com/2026/06/22/ai-chipmaker-groq-confirms-650m-raise-re-staffs-after-nvidias-20b-not-acqui-hire-deal/</a></p><p>[31] TechCrunch. “Groq raises $350M to fuel its pivot from AI chips to neocloud.” August 17, 2026. <a href="https://techcrunch.com/2026/08/17/groq-raises-350m-to-fuel-its-pivot-from-ai-chips-to-neocloud/">https://techcrunch.com/2026/08/17/groq-raises-350m-to-fuel-its-pivot-from-ai-chips-to-neocloud/</a></p><p>[32] The Register. “AMD acquires AI chip startup Taalas to boost inference performance by etching models into silicon.” August 6, 2026. <a href="https://www.theregister.com/systems/2026/08/06/amd-acquires-ai-chip-startup-taalas-to-boost-inference-performance-by-etching-models-into-silicon/5284344">https://www.theregister.com/systems/2026/08/06/amd-acquires-ai-chip-startup-taalas-to-boost-inference-performance-by-etching-models-into-silicon/5284344</a></p><p>[33] d-Matrix. “d-Matrix Corsair AI Inference Platform Enters Full Production to Meet Customer Demand.” June 9, 2026. <a href="https://www.d-matrix.ai/announcements/d-matrix-corsair-ai-inference-platform-enters-full-production-to-meet-customer-demand/">https://www.d-matrix.ai/announcements/d-matrix-corsair-ai-inference-platform-enters-full-production-to-meet-customer-demand/</a> · Wallaroo acquisition: <a href="https://www.d-matrix.ai/announcements/d-matrix-acquires-wallaroo/">https://www.d-matrix.ai/announcements/d-matrix-acquires-wallaroo/</a></p><p>[34] The Register. “AMD and Cerebras join forces against Nvidia’s Groq LPUs.” July 23, 2026. <a href="https://www.theregister.com/systems/2026/07/23/amd-and-cerebras-join-forces-against-nvidias-groq-lpus/5277817">https://www.theregister.com/systems/2026/07/23/amd-and-cerebras-join-forces-against-nvidias-groq-lpus/5277817</a></p><p>[35] The Next Web. “Google’s Ironwood TPU goes GA for inference at Cloud Next.” April 22, 2026. <a href="https://thenextweb.com/news/google-ironwood-tpu-inference-cloud-next">https://thenextweb.com/news/google-ironwood-tpu-inference-cloud-next</a> · Eighth-generation timing reported: TrendForce, August 11, 2026. <a href="https://www.trendforce.com/news/2026/08/11/news-microsoft-eyes-maia-300-launch-in-sep-reportedly-seeks-300k-tsmc-capacity-as-google-aws-race-ahead/">https://www.trendforce.com/news/2026/08/11/news-microsoft-eyes-maia-300-launch-in-sep-reportedly-seeks-300k-tsmc-capacity-as-google-aws-race-ahead/</a></p><p>[36] AWS. “Amazon EC2 Trn3 UltraServers.” December 2, 2025. <a href="https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/">https://aws.amazon.com/about-aws/whats-new/2025/12/amazon-ec2-trn3-ultraservers/</a> · SDxCentral (republished from DCD). “AWS makes Trainium3 UltraServers generally available.” December 2, 2025. <a href="https://www.sdxcentral.com/news/aws-makes-trainium3-ultraservers-generally-available/">https://www.sdxcentral.com/news/aws-makes-trainium3-ultraservers-generally-available/</a></p><p>[37] Microsoft (Scott Guthrie). “Maia 200: The AI accelerator built for inference.” January 26, 2026. <a href="https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/">https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/</a></p><p>[38] Anthropic. “Expanding our partnership with Google and Broadcom.” April 6, 2026. <a href="https://www.anthropic.com/news/google-broadcom-partnership-compute">https://www.anthropic.com/news/google-broadcom-partnership-compute</a></p><p>[39] SiliconANGLE. “Google and Meta reportedly strike new multibillion-dollar AI chip deal.” February 26, 2026. <a href="https://siliconangle.com/2026/02/26/google-meta-reportedly-strike-new-multibillion-dollar-ai-chip-deal/">https://siliconangle.com/2026/02/26/google-meta-reportedly-strike-new-multibillion-dollar-ai-chip-deal/</a> · Direct purchase talks for 2027 reported: DIGITIMES, November 25, 2025. <a href="https://www.digitimes.com/news/a20251125PD236.html">https://www.digitimes.com/news/a20251125PD236.html</a></p><p>[40] European Commission. “EU launches AI Gigafactories call to boost Europe’s computing capacity and unlock more than EUR 30 billion.” July 30, 2026. <a href="https://digital-strategy.ec.europa.eu/en/news/eu-launches-ai-gigafactories-call-boost-europes-computing-capacity-and-unlock-more-eu30-billion">https://digital-strategy.ec.europa.eu/en/news/eu-launches-ai-gigafactories-call-boost-europes-computing-capacity-and-unlock-more-eu30-billion</a></p><p>[41] EuroHPC Joint Undertaking. “EuroHPC Joint Undertaking launches AI Gigafactories call.” July 30, 2026. <a href="https://eurohpc-ju.europa.eu/eurohpc-joint-undertaking-launches-ai-gigafactories-call-2026-07-30_en">https://eurohpc-ju.europa.eu/eurohpc-joint-undertaking-launches-ai-gigafactories-call-2026-07-30_en</a></p><p>[42] SoftBank Group. “SoftBank Group to invest in 5GW AI data center campuses in France.” May 31, 2026. <a href="https://group.softbank/en/news/press/20260531_0">https://group.softbank/en/news/press/20260531_0</a></p><p>[43] Business Wire (distributed via Newswire). “AWS and HUMAIN to deploy up to 150,000 AI accelerators in Saudi Arabia.” November 23, 2025. <a href="https://www.newswire.co.kr/newsRead.php?no=1023709">https://www.newswire.co.kr/newsRead.php?no=1023709</a></p><p>[44] The National. “Stargate UAE’s first phase to be completed in third quarter of 2026.” December 5, 2025. <a href="https://www.thenationalnews.com/business/2025/12/05/stargate-uaes-first-phase-to-be-completed-in-third-quarter-of-2026/">https://www.thenationalnews.com/business/2025/12/05/stargate-uaes-first-phase-to-be-completed-in-third-quarter-of-2026/</a></p><p>[45] Taipei Times (Bloomberg). “Microsoft to invest US$10 billion in Japan AI infrastructure.” April 4, 2026. <a href="https://www.taipeitimes.com/News/biz/archives/2026/04/04/2003854989">https://www.taipeitimes.com/News/biz/archives/2026/04/04/2003854989</a></p><p>[46] Byline Network. “National AI Computing Center breaks ground in Haenam.” August 3, 2026. <a href="https://byline.network/2026/08/3-439/">https://byline.network/2026/08/3-439/</a> · 50,000 chips by 2030 and early opening in 2027: KPI News, August 4, 2026. <a href="https://kpinews.kr/newsView/1065573132034612">https://kpinews.kr/newsView/1065573132034612</a> · Vouchers and discounted rates: Samsung SDS Newsroom, August 3, 2026. <a href="https://www.samsungsds.com/kr/news/sds-260803.html">https://www.samsungsds.com/kr/news/sds-260803.html</a></p><p>[47] MTN. “2.4 trillion won to build the National AI Computing Center at Solaseado, Haenam.” May 13, 2026. <a href="https://news.mtn.co.kr/news-detail/2026051322522099817">https://news.mtn.co.kr/news-detail/2026051322522099817</a> · Groundbreaking report (about 2.5 trillion won): Money Today, August 3, 2026. <a href="https://www.mt.co.kr/policy/2026/08/03/2026080316085726304">https://www.mt.co.kr/policy/2026/08/03/2026080316085726304</a></p><p><em>Part of the</em> <strong><em>2026 Cloud Solutions Report</em></strong> <em>— an ongoing series on the 2026 cloud tooling landscape.</em> <a href="https://ieumtech.net/writing/series/cloud-solutions-report"><em>See the full series →</em></a></p><p><em>Originally published in Korean by NIA Digital Service Issue Report, October 2026. Written as CEO of Ieum Tech.</em></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/2026-cloud-solutions-report-10-token-supply/"><em>https://ieumtech.net</em></a><em> on November 1, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=703db03fc617" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Country Is a Line Drawn by Law]]></title>
            <link>https://chaesangjung.medium.com/a-country-is-a-line-drawn-by-law-9bf68fb8909d?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/9bf68fb8909d</guid>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Sat, 31 Oct 2026 00:00:50 GMT</pubDate>
            <atom:updated>2026-10-31T06:59:31.577Z</atom:updated>
            <content:encoded><![CDATA[<p>Inside Search, a country was less a shape on the map than the reach of its laws. Where that line fell changed the answer.</p><p><em>The sixth piece in a series on how country and language were never one and the same.</em></p><p>When you take charge of one country on the Search team, the word country stops being a shape on the map or the language spoken on that ground and becomes something plainer. It is the reach of a set of laws. Running a service in a country means local law reaches into the product, and in our case that meant the results themselves and the features around them. The requirements were worked out line by line with the legal team, and the product and engineering teams took them as the first priority, ahead of whatever else was on the roadmap.</p><p>Of all of it, adult search took the most hands. How far you serve results when an adult types an adult query is a delicate question anywhere. In Korea it sat against heavy law protecting children and adolescents, so it was more delicate still. Past a certain threshold of keyword, an age verification screen had to come up first, and only after the age check did anything else happen.</p><p>One distinction mattered more than the rest. If a user typed the query with their own hands and something dark came back, there is at least an account you can give of it. That person asked for this. Suggestions are a different matter entirely. Related searches and query autocomplete are the service stepping forward and saying look at this, and if what it holds out is something foul, intent counts for very little. So the suggestion surfaces were built on a far more conservative standard than the result page, with several layers of blocking logic behind them.</p><p>It is old enough now that I can admit I broke it myself. Without understanding the machinery all the way down, I laid a shallow bit of cleverness on top of it and got it wrong, and for a stretch SafeSearch was forced on even for adult users in Korea. In my defense, I was not alone. A colleague who had reasoned their way to the same place reviewed the change with me, so there was an accomplice of sorts. An emergency went up at the end of the year, I wrote a letter of account for the company, and I was called in every so often for a couple of months after. The episode is still mounted on Namuwiki, a user-edited Korean wiki that forgets nothing. My apologies go to the SafeSearch team in Switzerland, who handled it in the middle of their holiday, and to the PR people who spent that season answering conspiracy theories and secondhand rumor.</p><p>Part of the difficulty is that what counts as a bad query has no single definition. Some things are taboo in every country, terrorism and child sexual abuse among them. Others hold only inside one border. What that leaves you with is not one global rule but a set of operational practices, each country carrying the law and regulation it has to honor.</p><p>I want to be careful here, because I have no interest in writing these rules off as bureaucratic convenience or as nothing but friction for the user. The intent behind them was usually right, and a company putting a service in front of people ought to meet it. But compliance is the floor, not the ceiling. Once the law is satisfied, an adult who has verified their age and is searching the same keyword in good faith deserves an answer that serves what they actually meant.</p><p>In those days the host decided a service’s nationality, so the same search carried different legal weight depending on whether it happened on google.co.kr or on google.com. The .com was read as sitting outside the reach of local law and was treated differently because of it. Which means the famous workaround, the search that was blocked in Korean and opened up in English, was never really a detour through language. It was a detour through a domain. Compared to the patched-together way we implemented that then, the domains have since merged into one and the gap has mostly closed. What remains is a person using a VPN to put themselves somewhere else, and that was never something the service side could reach anyway.</p><p>Inside Search, a country was drawn as a line of law rather than a line on a map, and where that line fell decided what the same query was allowed to return. Where to draw it, how, and who gets to hold the pen are questions that have not grown any lighter since. Lay sovereignty on top, asked country by country, and it gets heavier. You can tune the features, and the index and the result pages will do what they can. Whether what comes back actually satisfies what a sovereign asked for is another thing. Keeping something from happening is the easy half. Doing it well is very hard.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/190"><em>brunch.co.kr/@chaesang/190</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/a-country-is-a-line-drawn-by-law/"><em>https://ieumtech.net</em></a><em> on October 31, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9bf68fb8909d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Store Where Japan and Korea Came First]]></title>
            <link>https://chaesangjung.medium.com/the-store-where-japan-and-korea-came-first-f5610227ce4b?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/f5610227ce4b</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[search]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 00:00:47 GMT</pubDate>
            <atom:updated>2026-10-29T04:11:34.587Z</atom:updated>
            <content:encoded><![CDATA[<p>In Search, every count started with the United States. At the Play Store, the top of the revenue list read Japan and Korea.</p><p><em>The fifth piece in a series on how country and language were never one and the same.</em></p><p>After close to ten years on Google Search I moved over to the Play Store, the thing we used to call Android Market, and spent two years there with good people on work I still think about. In Search, when someone said “users,” the number that came up first was the 300 million in the United States, and Korea entered the conversation as a question mark. The Store ran on a different set of numbers. Japan and Korea were treated well there, because the two of them together were bringing in in-app purchase revenue that stood comparison with the United States. Most of what I worked on was one question asked many ways. How do you take a single global product and fit it to each country?</p><p>A search engine and a store are built on different foundations. Search gathers snapshots of pages crawled at some earlier point and hands them to the person asking, so you spend your days looking in one direction, toward users. In the Store, search is only one feature. You run a marketplace, and you stand beside the developer as a kind of publisher, introducing their product. The supply side, the studios shipping the apps, weighs as much as the demand side does. You end up close to the companies making the big games, and the standing homework was to stay close without tilting, because a marketplace has to hold its neutrality.</p><p>Around then in Japan, Monster Strike and Puzzle &amp; Dragons, two mobile games that printed money for their studios, were trading places at the top of the revenue charts. The team’s goal that stretch was to land Fate/Grand Order in Japan, and in Korea to land Lineage 2 Revolution and then Lineage M, the mobile descendants of a PC franchise Korean players had been living inside since the late nineties. Our part was to help the local teams in both countries so these titles would arrive with weight, and to hold off the other marketplaces reaching for the same releases. None of them were products you sold with advertising, but the fee revenue that followed an install was too heavy to set aside. At the same time we had to keep the rest of the world in view rather than letting Japan and Korea, where the money was, take all the attention. Neither half of that came easily.</p><p>The Play Store lived at one address, play.google.com, and almost everything inside it was still tied to a country. The device reported the country first, and the release rules were applied strictly on top of that. The same app could carry a different description and different screenshots in each country, and studios would swap out the icon and even the title. In games it went past translation into cultural code. The same game led with a young sorceress in the Korean and Japanese stores and with broad-shouldered male warriors in the United States and Europe. Campaigns built around a local celebrity went further still. None of these pictures decided a rank the way a search result does, but as the face advertising the app they mattered more than you would guess. These days the same slot carries anniversary badges and seasonal art as well.</p><p>Game studios had been doing this well for a long time. A large title had its own operations team in each country running its own campaigns and events, so what looked like one app from outside was effectively a different game running inside each market. The Store’s job was to let a developer ship to more than a hundred countries with a few clicks, then operate there, and make money doing it.</p><p>Earlier there had been apps split into a separate package per country, which meant one product existed as several. Over time those folded under a single package id that handled region internally. It is the same path google.co.kr and google.com walked before they finally became one address.</p><p>What went up per country was a mix. Some pages had been localized by hand, and on others the developer had simply switched auto-translation on and left it. How much weight to give each was a running argument. Judging translation quality directly by machine was hard then, so we read the evidence of care as a signal instead. Had someone prepared this page, or was the toggle just on? The two got different treatment. From the user’s side the difference shows immediately. My own ranking, for what it is worth, ran careful Korean first, then English, then machine translation, then Chinese.</p><p>Reviews and star ratings are their own story. An app ships bound to a country, but whether you use the per-country rating signal or the global total opens into something more complicated, and I will come back to it in a later piece.</p><p>In Search we divided countries by domain. In the Store it was package and country, applied strictly, and once something was bound to a country the language tangled itself back in. We were fortunate that Korea and Japan are both more than ninety percent single-language, which kept our corner simple to reason about. A global product had to carry far more than that. Even one app, in the end, never quite folded country and language into the same shape.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/187"><em>brunch.co.kr/@chaesang/187</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/the-store-where-japan-and-korea-came-first/"><em>https://ieumtech.net</em></a><em> on October 29, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f5610227ce4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Every Ranking Puts You Somewhere Else]]></title>
            <link>https://chaesangjung.medium.com/every-ranking-puts-you-somewhere-else-1e53840f0f1b?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e53840f0f1b</guid>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Sun, 25 Oct 2026 00:00:29 GMT</pubDate>
            <atom:updated>2026-10-25T14:52:13.036Z</atom:updated>
            <content:encoded><![CDATA[<p>Population, speakers, documents on the web. Korea’s place moved with every line it stood in, and those numbers decided the budget.</p><p><em>The fourth piece in a series on how country and language were never one and the same.</em></p><p>Working on the Google Search team, I covered Korea and Korean, but naturally the English-speaking user on the U.S. West Coast came first. When someone said “users,” the 300 million in the U.S. were what came to mind, and carrying something over from Korea meant meeting that reference point with 50 million in hand before anything else. Fifty million assumed 100% share, so even that was generous.</p><p>Lift the numbers a little, though, and there isn’t one way to line countries up. Rank by population and India and China stand at the front with 1.4 billion each while the U.S. sits well back at 300 million. Rank languages by how many people speak them and English stands first at 1.5 billion, though fewer than 400 million speak it as a first language and the rest learned it second. Rank the documents on the web by language and English takes close to half, followed by Spanish and German and Japanese, while Chinese, spoken by 1.4 billion, barely clears 1%. Population order, language order, and document order each stood in a different line. Share figures multiplied into all of it along with various strategies, and these are mixed memories, so treat the figures and rankings as approximate.</p><p>Here language and country get tangled once more. Because a locale is by nature one tag binding a language to a country, even inside the U.S. the people using English and the people using Spanish stood in separate lines as en_US and es_US, each a meaningful block of its own. There really are a great many people, in California especially, who speak only Spanish and no English, and es_US ranked as one of the largest Spanish locales after es_MX. Cross to Europe and countries with a firm language of their own, Germany or France, still have a substantial layer of English laid over the top. One country, one language rarely balanced out cleanly.</p><p>English dominated not only in people but more so in the volume of content piled onto the web, and beneath it the Latin languages sat heavily in several senses, with Spanish and Portuguese giving Brazil and Mexico numbers that were nothing to dismiss, and Southeast Asian figures just starting to climb around then. In that gap, Korean ranked somewhere around tenth in the world by number of speakers but dropped to a place that didn’t register at 1% when counted by content open on the web. It wasn’t that Korean content didn’t exist. The good material sat in private territory, invisible to anyone counting from outside. Machine translation wasn’t what it is now, and paired documents carrying the same content in several languages were rare, so I’ll come back to that when the series reaches translation and content.</p><p>What makes English formidable isn’t just size. Even en_US, one of the largest single locales at 250 million, is only about five times ko_KR’s 50 million. But because English is the hub every translation passes through and the reference language for search and for learning, the gap you felt whenever you compared search quality or document quality ran much wider than that. The distance to first place was always farther than the numbers said.</p><p>Take Korea alone and change the line it stands in, and the position swings every time. Around tenth in the world by GDP, outside the top thirty by population, and off the chart entirely by web content. Being roughly tenth by language was an awkward place to invest from and an awkward place to cut. Not high enough that touching it brought large numbers along immediately, not small enough to be worth folding away. The priority always landed a beat late. We tried using the low share against 50 million as an argument that there was room to grow, but apply that low share and Korea’s standing against other countries inside Google came back as a worse report card.</p><p>Covering Korea meant looking often at other countries and languages. For a while China, Japan, and Korea, with Hong Kong added, were grouped as CJK or CJHK and worked together out of shared circumstance. The ambition was fine, but how was that going to work. China fell out of search at some point, and Taiwan and Hong Kong were still before proper offices had been established, so seeing it through together was difficult. Japan went the other way: share around 40% until it merged with Yahoo Japan search and passed 90%, which is a different sort of departure. Korea was still four or five years away from Android phones and YouTube breaking through, so it was a stretch of fighting alone.</p><p>Numbers were certainly the starting point of every discussion, and 300 million in the U.S., 1.4 billion in India and China, 1.5 billion in English were laid out at the top of every meeting. But step one pace further in and a country’s position changed entirely depending on which line you made it stand in, and even those numbers were held separately by each service, which made gathering them in one place for comparison difficult. For a global company allocating investment by country this was a natural and reasonable discussion to have, strategically speaking, though I remember it wearing me down. Given another run at it I might do a little better, he says lightly.</p><p>Back when I was in the thick of it, Google had plans for India, Southeast Asia, Latin America, and Africa laid out ahead, and my own fighting ended here. With sovereignty the subject of the day, I recently saw Vietnamese and Indonesian sitting above Korean in CommonCrawl’s document rankings, and it left me with a good deal of regret. I’m not a player anymore, so my position is a little different. Still, before we reduce digital sovereignty to one measure of country and language, I worry about whether enough good Korean, enough good Korean documents, are visible in the places where AI and search engines are looking.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/185"><em>brunch.co.kr/@chaesang/185</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/every-ranking-puts-you-somewhere-else/"><em>https://ieumtech.net</em></a><em> on October 25, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e53840f0f1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Country Is This Service From?]]></title>
            <link>https://chaesangjung.medium.com/what-country-is-this-service-from-29e1613e12c5?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/29e1613e12c5</guid>
            <category><![CDATA[internet]]></category>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[internet-regulation]]></category>
            <category><![CDATA[korea]]></category>
            <category><![CDATA[google]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 00:00:08 GMT</pubDate>
            <atom:updated>2026-10-24T14:33:51.777Z</atom:updated>
            <content:encoded><![CDATA[<p>YouTube once blocked uploads if your country was set to Korea. Which country a service belongs to was never as simple as its domain.</p><p><em>The third piece in a series on how country and language were never one and the same.</em></p><p>There was a stretch when setting your country to Korea on YouTube meant you could neither upload a video nor leave a comment. Switch the setting to Global and it opened up again. The cause was a 2009 Korean rule people called the real-name internet law, formally the limited identity verification system.</p><p>What the rule required was that anyone posting to a bulletin board on a service above a certain size had to verify their real identity first. It was local law, and a company operating in Korea had to follow it, so Google tried to comply as I remember it. This was before the mobile era, and Google+ was being pushed with some ambition and would become its own embarrassment later, so the problem wasn’t really real names. It was how to get people to log in at all. Handing a difficult homework assignment to a guest you’d worked hard to bring in the door. In the end YouTube chose to block posting and uploading for users set to Korea.</p><p>Even inside one company the situation differed by product. Google Search, where users don’t create comments or content, slipped past the rule. But Gmail, which people used as part of daily life, raised a harder question. Gmail isn’t a bulletin board, and Google was running dozens of products at the time, so putting every feature through legal review one by one wasn’t workable. I was on Search then, so I had no direct or indirect responsibility for YouTube, and I remember watching these discussions go back and forth internally with a certain sadness. Around the same time Google Search was pulling out of China, so this was a period when the company spent a long while caught between local law and global policy. The Korean rule was struck down as unconstitutional in 2012 and disappeared, and since nearly every service now starts with a login of some kind, it may read as an old story.</p><p>The biggest locale question I carried then was this one. What country does this service belong to? Notice what the YouTube setting keyed on. It wasn’t Korean the language, it was Korea the country, and it wasn’t touching the user’s preference so much as tying the scope of the service to a country, which is to say to the reach of a law. Whether “the nationality of a service” is the right name for that, I’m not sure, but I haven’t found a better one.</p><p>An address like google.co.kr made everyone think of Korea and made local law easy to apply, but a domain never guaranteed a country. You can’t readily obtain the domain of every country in the world, and these days a country’s domain gets borrowed for an entirely different purpose, as with .tv and .io.</p><p>Even within one company the practice varied. YouTube used youtube.com/kr, Google used google.co.kr, Gmail used gmail.com. Looking back it was inefficient enough that a large company-wide project to consolidate everything onto google.com ran over the following years, and google.co.kr was left as a parking domain that forwards to google.com.</p><p>Set domains aside and the nationality of a service still doesn’t resolve into one legal answer. The question surfaces every time a commission dispute flares up around Google. It surfaces when people ask which country’s law applies to a service plainly hosted abroad. It surfaces when you ask how much meaning a physical server location carries in a cloud world, and again when a service doesn’t match the corporate entity behind it. Quite a few places need a definition of “country,” and once money moves through them it gets considerably more complicated.</p><p>In the days of “Beat Naver,” one of the team’s objectives was to grow query volume on google.co.kr. I understood the impulse to lift a single-digit share, but the moment the goal was pinned to the co.kr domain the arithmetic started to drift. People living in Korea who went to google.com and searched in English became invisible inside that target. We told ordinary users to use google.co.kr while, ironically, walking individual people through switching to google.com when they didn’t like what they saw.</p><p>Users now cross that boundary without a second thought when they need to. What used to be flipping a country setting to Global is now done with a paid or free VPN, far more commonly and within a fairly legal range. While I’m in Korea I can’t reach certain U.S. university sites, the DMV, or my insurance company, and from the U.S. I can’t see Coupang Play. World Cup season adds another round of hide-and-seek.</p><p>With sovereignty now a keyword, the nationality of a service doesn’t divide cleanly into domains and settings and numbers, and services keep adjusting against the people routing around them while trying to hold their own quality together. Settling on one country for a service is hard enough to grasp. Lay the country and language of the person using it on top, and then the country and language of the content it displays, and the difficulty climbs another level. Country settings happen in more than one place.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/184"><em>brunch.co.kr/@chaesang/184</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/what-country-is-this-service-from/"><em>https://ieumtech.net</em></a><em> on October 24, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=29e1613e12c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Illusion of Loving English]]></title>
            <link>https://chaesangjung.medium.com/the-illusion-of-loving-english-writing-chaesang-jung-6536173667b0?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/6536173667b0</guid>
            <category><![CDATA[localization]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[product-analytics]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[selection-bias]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 00:00:02 GMT</pubDate>
            <atom:updated>2026-11-01T08:52:14.114Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Half the queries from Korea came in English. The data wasn’t lying. We were misreading it.</strong></p><p><em>The second piece in a series on how country and language were never one and the same.</em></p><p>When I joined the Google Search team, my first assignment was google.co.kr: making good results for Korean users. This was the era of Naver and Daum, formidable rivals a clear step ahead, before the iPhone, before the Galaxy and Android, before Chrome, when Google’s results in Korea were bland and bare. That was where I started.</p><p>The first question I got while working with the Search team at headquarters was a strange one. Koreans are so good at English, so do we even need to localize separately? Back then, if you gathered the queries coming from Korea, filtered out the spam, and looked, more than half were in English. By the numbers alone it seemed Koreans loved English, a nation that searched in English, and from headquarters’ vantage point that was simply what the data said; there was no obvious reason to tend to Korean separately. So the question that came back to me was how to make the case that Koreans use their own language, and prefer it. Having to prove something so self-evident with data felt backwards, as if the cart had come before the horse.</p><p>But there was a trap hidden in that data. In those days Google’s share in Korea was under 5%. We chanted “Beat Naver,” but the road ahead was endless. The number-one query in Korea, by an overwhelming margin, was “Naver,” and number two was “Daum.” That meant even the Koreans who used Google Search were typing into its box on their way over to Naver. Strip those out and the share of English would climb even higher. The people using Google then weren’t Koreans at large but a minority: grad students, engineers, professionals who had just heard of it by word of mouth, people for whom English came naturally. English made up half the queries not because Koreans loved English, but simply because the small group using Google back then happened to use a lot of it. If you wanted to find something in Korean, you went to Naver, and few even bothered to try a Korean query on Google. Across all the searching happening in Korea, Korean was the overwhelming majority. It was only because we were peering through the one tiny window called Google that English looked like half, or more.</p><p>Follow the question of why even they typed in English, and you end up at content. Was there quality Korean material on the open web? No. It wasn’t that good Korean content didn’t exist. Naver had diligently, and genuinely, cultivated it through Knowledge-iN and blogs for Korean users, and I have to give them that. But those good things sat inside Naver’s walls, and search engines outside couldn’t see over the fence. Google’s ten blue links looked shabby in Korean, so people simply didn’t try, and those who needed English documents found it faster to search in English. The person typing English queries and the person living their life in Korean were the same person. From the very start, the language of the query and the language of the person were misaligned.</p><p>One phrase I kept reaching for in that work was lipstick on a pig. If what you have to show in Korean is thin and empty, can UI or features ever fix that? It’s the essential question. It’s like promoting a restaurant endlessly when the food isn’t good: customers leave fast. In Korea, Naver’s front page is rich and portal-like, but in the U.S. Google overtook Yahoo on search quality and speed, and arguing that Korean users were somehow different was genuinely hard to sustain. This wasn’t UI or usability. It was what we called quality, the domain of search quality itself, and so it was a far longer fight.</p><p>Later I learned this wasn’t only a Korean story. Even within en_US, the supposed heartland of English, a third were actually es_US, Spanish speakers, so the U.S. was hardly a country running on English alone. And yet the people holding over 90% of the share rarely looked even at that third beneath their own feet; a high share, it turns out, also means your own world looks like the whole world. By comparison, I was starting to work on the problems of a smaller place, Korea, a world they didn’t know.</p><p>It’s only that Korea and Japan are the rare cases where country and language line up almost perfectly. In most of the world, country and language were separate from the start. The obvious fact that Koreans use Korean was something I had to explain at length before that data, even running real user interviews to work it through. Sure, there were people good at English. But the ordinary Korean user never loved English over their own tongue. There were only people who, unable to find good things in their own language, fumbled at the search box in someone else’s.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/183"><em>brunch.co.kr/@chaesang/183</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/the-illusion-of-loving-english/"><em>https://ieumtech.net</em></a><em> on October 21, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6536173667b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Same Address, Different Face]]></title>
            <link>https://chaesangjung.medium.com/same-address-different-face-ad9a2220e45d?source=rss-08a1a95ef732------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad9a2220e45d</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[internationalization]]></category>
            <category><![CDATA[localization]]></category>
            <dc:creator><![CDATA[Chaesang Jung]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 00:00:48 GMT</pubDate>
            <atom:updated>2026-10-03T13:27:07.606Z</atom:updated>
            <content:encoded><![CDATA[<p>I assumed a URL showed everyone the same page. Working on a search engine taught me that who knocks, and from where, changes what opens.</p><p><em>The first piece in a series on how country and language were never one and the same.</em></p><p>I used to think a web page had one face. Same address, same page, no matter who opened it. I believed that as someone who used search engines, and I stopped believing it when I went to work at a company that builds one. Sites screen the traffic that comes in, for all kinds of reasons. Some show Googlebot nothing at all if it arrives from Korea. Others do the reverse and close the door on anyone arriving from outside Korea. One URL, and it puts on a different face depending on who is knocking and from where. I had seen the mechanism from the other side while implementing HTTP servers, and if you set it up wrong you learn about it the hard way, usually as a DDoS and a bill you didn’t expect.</p><p>That gives the people building a search engine a headache. The results you show someone searching from Korea ought to come from a crawler that ran in Korea. Hand a Korean user a page collected somewhere in the United States, and you are showing them something different from what they would see if they went to that address themselves. A search engine must not do that.</p><p>If a single address behaves like this, the results page is a harder question still. Two people type the same query. Should they see the same thing, or different things? And if different, what exactly is different about them that justifies it? The easy answer is that this is localization, the ordinary work of taking a global service to each market. But look closer and country and language are not sitting in one place. The side that ships the service has its own country and language. The user has theirs. And the content the service displays carries its own as well. Three domains, each holding a country and a language, meet on one screen, and they tangle and slip in every combination you can imagine.</p><p>Anyone who touches code has run into the word “locale.” ko_KR, en_US, ja_JP. An underscore, and language and country are fused into one token; the word treats the two as one from the start. We never found a proper Korean word for it, so we just said “locale” and got on with the work. For a long time I also assumed it was one thing. Every time we missed the moment the two came apart, the product gave a wrong answer.</p><p>On the service side, the work of solving this went by several names: localization, internationalization, and lately personalization. The words are long, so people count the letters in the middle and write l10n, i18n, p13n. Whatever you call it, the aim was the same. Don’t hand everyone the identical thing; hand each person, each country, each language something slightly different, with some thought behind it. Locale is the method for thinking about country and language. Personalization goes further and gets more complicated. Hyper-personalization, further and murkier still. To be honest, by that point it is less a method than a shrug: throw it out and see what sticks.</p><p>I wrestled with this for a long time on Google Search, and later met the same question again on the Play Store, this time not across the whole world but with Korea and Japan singled out for a close look. Slightly different products, examined seriously over quite a few years. What follows are a few technical stories and a few experiences from the user’s side, written down from memory.</p><p>Lately I see all of this bundled under one word, “sovereign,” as in sovereign cloud and digital sovereignty. Like locale, sovereign lumps several things together, and I find myself wondering whether, inside that word, we are really talking about the same thing. I have a modest hope that the record of someone who ran into it back then can serve as a reference for that question.</p><p>Korea and Japan are the unusual cases, places where one country and one language line up almost exactly. Elsewhere, country and language are not one. And the pair, the locale, never stays in one place. The side that ships the service has one, the user has one, the content has one, and spread across those three domains they came apart more often than not. It takes some nerve, but I am putting it together as a series.</p><p><em>Part of</em> <strong><em>locale — Country and Language</em></strong> <em>— how a service diverges by region, language, and country, seen from inside search, the Play Store, and localization.</em> <a href="https://ieumtech.net/writing/series/locale"><em>See the full series →</em></a></p><p><em>Adapted from my Korean essay on Brunch:</em> <a href="https://brunch.co.kr/@chaesang/182"><em>brunch.co.kr/@chaesang/182</em></a></p><p><em>Originally published at </em><a href="https://ieumtech.net/writing/same-address-different-face/"><em>https://ieumtech.net</em></a><em> on October 2, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad9a2220e45d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>