Who Got Rich Off Software Bloat (The Receipt Is Itemized)

Añadido: 2026-05-23

[00:00:00]A Macintosh 2 from 1990 had 2 megabytes of memory total. The empty white rectangle Slack draws on this screen before I type anything, before a single notification arrives before any work has been done, is sitting on 580 megabytes.

[00:00:15]The 1990 machine could have stored 290 copies of an empty Slack window in the RAM it had. The 1990 machine ran Photoshop.

[00:00:25]This is not the diagnosis. The diagnosis was the last video. This is the receipt.

[00:00:31]And the receipt is itemized. For the next 30 minutes, I'm going to do something the industry has spent 20 years avoiding. I am going to open the apps you open today, the ones you don't think about because they're just there.

[00:00:43]Okay? And I am going to bill them in megabytes, in dollars, in watt hours, in tons of e-waste. This is not a moral argument. It is an arithmetic one.

[00:00:53]The first video on this channel argued that software stopped trying. That over 40 years, every generation of code ate the performance gains of the hardware below it. That argument was the diagnosis. What you are about to watch is the bill that diagnosis generated.

[00:01:08]And I want to be careful about one thing before we start. I am not going to blame developers. Most developers I know are excellent. The point of this video is not to tell you who is bad at their job.

[00:01:18]The point of this video is to tell you who got rich because someone did.

[00:01:24]The methodology is simple and I will show it on camera so nobody can claim I gained the result. I open an app. I let it sit at idle for 60 seconds. I write down what it costs in RAM, in CPU, and network. I do this on the same machine for every app. A 6-month-old MacBook Pro with 16 GB of memory. The most ordinary developer laptop you can buy in 2026. No exotic hardware. No virtualization tricks. the machine you probably have on the desk in front of you. For each app, I will compare three things. What it cost in 1998 or 2000 when the same job was being done, what it costs today, and what the difference works out to in the only currency that matters at the end of the month, which is the one with a dollar sign in front of it. I will tell you when a number comes from the company that makes the app, and when it comes from someone independent. I will tell you when a number is a single measurement, and when it's a range. I will tell you when I cannot verify something, and I will mark it as such.

[00:02:14]The point of this video is to be useful as a citation, not to win an argument.

[00:02:19]The receipt is in eight items, then I will total it. And at the end, I will tell you exactly who collected this bill and how much they made. Item one, Slack.

[00:02:28]A 1995 IRC client took 1 megabyte to install and ran in under 10 megabytes of memory. It did exactly the same job Slack does. People type, other people read, there is a list of channels on the left. There are messages in the middle.

[00:02:41]This is the entire feature surface. It was solved in 1988. Slack by Slack's own engineering blog runs at a 10th percentile floor of 130 megabytes per workspace and a 99th percentile ceiling of 960 megabytes per workspace. That is Slack's own measurement in their own post. Titled and read this title twice, reducing Slack's memory footprint. The footprint is so notable that there is a blog post about reducing it. Two workspaces the way most people use it lands you somewhere between 400 megabytes and 2 GB. What did we get for the extra 1 billion bytes? We got reactions. We got an emoji picker. We got threads. We got a left sidebar that used to be a tree and is now a different tree. We did not get faster typing. We did not get a smaller install. We did not get a chat application that opens before you finish blinking, which IRC did on a 486. Slack's install footprint today is roughly 200 megabytes. WAMP 2, which played MP3s while running visualizations, could be trimmed to fit on a 1.44 megabyte floppy disc and still play music. The chat client now costs 140 times more disc space than a multimedia player did when disc space was actually expensive.

[00:03:53]This is item one on the receipt. Item two, Discord.

[00:03:57]In December of 2025, Discord shipped a feature. The feature was that the Discord desktop client would automatically restart itself when its memory usage exceeded 4 GB and the user had been idle for 30 minutes. They put this in a blog post. They wrote it as if it was a courtesy. Um, read that again. The team that ships the app shipped a feature that admits in production the app routinely uses 4 GB of memory to do nothing. 4 GB is the entire memory of a respectable laptop from 2014. 4 GB is 2,000 times the RAM that Photoshop 1.0 ran on in 1990. And Photoshop 1.0 was painting fullcolor images while Discord in this configuration is not doing anything at all. A voice call routinely takes Discord above 1 GB. Screen share 2 GB.

[00:04:45]The Mumble client from 2005, which solves the same problem with better audio latency, ships in under 30 megabytes and idles below 50. The math is the math.

[00:04:56]This is item two. Item three, Spotify.

[00:05:00]I'm going to give Spotify a little credit here because the comparison is rigged in their favor, and I'd rather you trust the rest of this video than score easy points.

[00:05:09]Win amp 2 played a local MP3 in 15 MGB of RAM. Spotify desktop IDLE sits at around 300 MGB. That is 20 times more memory. But Spotify is not just playing a song. It is also fetching a stream. It is also negotiating DRM. It is also rendering album art. It is also handling a playlist database the size of a small library. That said, here's the part that makes the receipt interesting.

[00:05:35]Spotify in a browser tab uses more memory than Spotify as a desktop app.

[00:05:40]Independently measured by Aleandro Cowuro. Desktop 212 megabytes. Web and Chrome with hardware acceleration off 382 megabytes. During playback, the gap widens. desktop holds at around 200 megabytes and 1% CPU. The web player jumps to 700 megabytes and almost 6% CPU. So, Spotify desktop is, and I want to be careful, actually fine as Electron apps go. The receipt I'm writing has a credit line. We'll come back to why that matters later. Item four, the text editor.

[00:06:14]UltraEdit on Windows XP in 2003 opened a 1 megabyte log file in under 300 milliseconds and used under 20 megabytes of memory. VS Code on a modern MacBook opening the same 1 megabyte log file with no extensions takes around 800 milliseconds and 300 to 500 megabytes of memory. With Copilot enabled, the floor climbs past a gigabyte with a single bracket colorizer extension that someone wrote in a weekend. GitHub has documented cases of 380 megabytes consumed on a 100line JavaScript file.

[00:06:47]VS code is Electron. Cursor is VS Code is Electron. Windsurf is VS Code is Electron. The entire generation of AI editors that the industry is currently using to write the next generation of software is built on the same Chromium runtime, paying the same fixed tax. For comparison, Sublime Text written in C++ in 2008 opens the same file in 100 milliseconds and uses 60 megabytes of memory. Helix and Zed written in Rust are faster than Sublime.

[00:07:17]The receipt for the editor isn't theoretical. It is the difference between 300 milliseconds and a second and a half multiplied by every file you open every day for the rest of your career. A working developer opens hundreds of files a day. The math on a 30-year career writes itself. We have collectively traded thousands of human lifetimes of waiting for a UI framework written in JavaScript. Item five, WhatsApp desktop. I am putting WhatsApp on the receipt because it shows you what happens when the receipt arrives and somebody in a meeting decides not to read it and in create by item five, WhatsApp desktop.

[00:07:51]WhatsApp had a native desktop app. It was a UWP app on Windows. It used about 240 megabytes of memory. It was by the standards of this video lean. In 2022 and 2023, Meta rewrote the WhatsApp desktop client. They did not rewrite it in Rust. They did not rewrite it in Zigg. They rewrote it as a web rapper.

[00:08:11]Today, by multiple user reports, the new desktop app uses around 1.2 GB of memory to do the same job the old one did in 240. Wow.

[00:08:20]The receipt on this one is short. They had it. They sold it for developer convenience. The new app is easier to ship updates for because it is a web page in a window. The old app was three to five times faster to open and used 1/5if the memory. We were the bill that paid for that decision. Item six, Microsoft Teams. Item six, Microsoft Teams.

[00:08:40]In March of 2023, Microsoft launched the new Teams and they led the keynote with their own performance numbers. Two times faster to launch, 50% less memory, 70% less disc space.

[00:08:52]I want you to read those numbers as what they are. Those are Microsoft's numbers about Microsoft's own product saying that the previous version was so heavy that they got to claim a two times improvement, a 50% improvement, and a 70% improvement by switching from one chromiumbased runtime Electron to a different Chromiumbased runtime web view 2 and shipping basically the same UI on top of it. Why?

[00:09:17]The improvement is real. Microsoft hired GigaM to benchmark it. The report is paywalled, which I want to note for the record. But Microsoft did publicly admit with their own marketing materials that the previous version of Teams was so bloated that switching to a slightly less bloated runtime was worth a whole launch event.

[00:09:34]Read that as a confession. The receipt has Microsoft's signature on it. Item seven, the browser. The browser is special because the browser is no longer an application. The browser is a second operating system and we keep pretending it isn't.

[00:09:48]I am citing Drew Devault for the next number and I want you to write it down because it is the single most clarifying statistic in this video. The W3C specifications that a modern browser is supposed to implement fully to be considered a complete browser. Those specifications total 114 million words.

[00:10:08]114 million. For comparison, that is more text than the C11 standard, plus the C++ 17 standard, plus the entire UEFI specification, plus the USB 3.2 specification, plus PZIX, plus all 8,754 IETF RFC's plus Wikipedia's list of longest novels combined. To write a browser today is to undertake a project on the scale of the Apollo program, Cowur.

[00:10:35]Building a new browser is functionally impossible, which means competition is impossible. Which means there are three engines controlled by two companies and every modern desktop app that uses one of them inherits the entire 114 millionword stack as attacks.

[00:10:50]In 2014, Bruce Dawson, who is a real Chromium engineer, documented that Chrome was setting the Windows system timer tick to 1 millisecond when it didn't need to. that raised CPU power consumption by roughly 25% on idle systems. That bug by Chrome's own bug tracker had been filed in 2010. Chrome shipped that battery draining bug for 4 years before anyone fixed it.

[00:11:15]In 2016, Microsoft published a head-to-head battery test on a Surface Book streaming Vimeo on a Loop. Edge 7 hours and 22 minutes. Chrome 4 hours and 19 minutes. Chrome in that specific test used 70% more battery than the same workload on Edge.

[00:11:33]The browser is the second largest line on the receipt. The largest line is what runs inside it. Item 8, the operating system itself.

[00:11:42]Windows 2000 ran cleanly on 64 megabytes of RAM. Windows XP when it was released ran on the same hardware. By Microsoft's own minimum specification, Windows 95 needed 4 megabytes of memory. 4 megabytes. You saw Windows 11 as measured by notebook check and hack a day on identical modern hardware idles at 2.2 to 3.3 GB of RAM.

[00:12:03]Same machine, just the operating system doing nothing. 3 GB, that is a 40,000fold increase in resting memory over 30 years. The OS now does more. A Wi-Fi stack, encryption, an updater, a notification system, but it does not do 400 times more, let alone 40,000. Most of the difference is not new capability.

[00:12:21]It is accumulated overhead and the hardware requirement is a downstream consequence. Windows 11 will not install on machines that lack TPM 2.0 and a recent enough processor. By canal's estimate repeated by the US public interest research group, this cutoff makes around 400 million functioning PCs ineligible for Windows 11. Machines that work, machines that boot, machines that run Windows 10 right now, huh?

[00:12:46]400 million PCs.

[00:12:49]Hold that number. We're about to total the bill. Eight items on the receipt.

[00:12:52]Slack, Discord, Spotify, the editor, WhatsApp, Teams, the browser, the operating system.

[00:12:58]Each one of them costs 10 to a thousand times more in RAM, in CPU, in disk, in time than the program that did the same job 25 years ago. None of them does 10 to times more. Most of them do roughly the same thing.

[00:13:12]Now, let's add it up. the bill first in dollars because the dollars are the easiest part to compute and the hardest part to dismiss.

[00:13:19]In 1995, 1 megabyte of computer memory costs $30.85.

[00:13:24]Today, 1 GBTE of memory costs around $3.

[00:13:27]That is roughly a 10 million times reduction in the cost of memory in three decades. This data is from John Macllum's memory price history, which is the canonical academic time series for memory pricing. It is the chart you are looking at right now where if you have 16 GB of RAM in your laptop, the chips you are sitting on would have cost in 1995 approximately $500,000.

[00:13:51]$500,000 of memory in a $1,000 laptop.

[00:13:55]That collapse, that 10 million times reduction is the most extraordinary technological deflation in human history. There is nothing else like it.

[00:14:03]Houses didn't get 10 million times cheaper. Food didn't. Books didn't. Cars didn't. Memory did.

[00:14:10]And we spent it not on capability, not on new ways for humans to work. We spent it on a 15 times heavier Slack and a 4 GB chat client and a desktop word processor that opens slower than its 1998 ancestor did on a Pentium 3.

[00:14:27]The deflation in hardware was the abundance. The bloat in software is the spending of that abundance. The receipt is itemized. The receipt has a payor and a payee and on this line the payor is you and the payee is the framework vendor whose hello world shipped 1,800 packages. The second line on the bill is energy.

[00:14:50]This is harder to measure on one laptop but easier to verify at the global scale. The International Energy Agency in their 2025 report on energy and AI estimates that data centers consumed approximately 415 terowatt hours of electricity in 2024. That is about 1 1.5% of all electricity generated on Earth by 2030. That figure is projected to roughly double to about 3% of global electricity.

[00:15:16]3% of global electricity to run software that for the most part is doing exactly what software was doing in 1998. sending messages, displaying documents, playing media, taking inputs, and producing outputs. The functional surface has not exploded. The substrate has.

[00:15:34]On your laptop, the receipt is more personal. The Edge versus Chrome benchmark we cited a moment ago, 70% more battery on the same Vimeo Stream, is not an abstraction. It is 1 and 1/2 hours of your laptop battery removed from your day every time you used Chrome instead of Edge in 2016. The 1 millisecond timer bug specifically wasted about a quarter of your CPU's idle power for four years across millions of users every day. Four years.

[00:16:02]That number works out to roughly the annual energy budget of a small country.

[00:16:08]This line on the bill is in W hours. The unit converts directly to grams of CO2 if you want to do that math. I am not going to do that math on camera because the conversion factor depends on which grid you are plugged into. But the unit converts. The third line on the bill is e-waste.

[00:16:24]The United Nations Global E-Waste Monitor in its 2024 edition reports that humans generated 62 million tons of electronic waste in 2022. That is up 82% from 2010. The projection for 2030 is 82 million tons per year. Of those 62 million tons in 2022, the UN estimates that 22.3%, fewer than 1 in 4 kg, was formally collected and recycled. The rest is in landfills, most of it in places like Abboglashi in Acra where teenagers strip copper from circuit boards over open fires.

[00:16:59]Now against that backdrop, the public interest research group estimates that the Windows 11 hardware cutoff that decision by Microsoft to require a recent processor in TPM 2.0 will at scale push around 400 million still functioning PCs into the obsolete column. Their projection puts the additional e-waste at approximately 725,000 tons. One software policy.

[00:17:25]The receipt has Microsoft's signature on this line, too. The PCs work. The PCs are functioning hardware. They are being thrown away because the operating system, the thing made of pure text written by humans, will not run on them.

[00:17:38]The hardware did not fail. The software refused.

[00:17:42]This is what I mean when I say you paid for hardware that the software spent for you. The fourth and final line on the bill is harder to measure, but in some sense the most important, the bill in hours of your life.

[00:17:55]Every time you wait for an app to open, you are paying a tax. The tax is in seconds. 3 seconds for Slack to draw a window. 8 seconds for Teams to be usable. 12 seconds for Chrome to launch with your 40 tabs restored. Half a second for VS Code to render a directory tree. 2 seconds for the spinner.

[00:18:15]Multiply by the number of times per day you launch each thing. Multiply by the number of working days in a year.

[00:18:21]multiply by the number of years you have left in your working life.

[00:18:25]If you are a developer who opens an editor 10 times a day and the editor takes two seconds longer than the editor we had in 2005, you will lose approximately 30 hours of your working life to that one decision to opening one app slower than it used to open to in the aggregate doing nothing.

[00:18:42]The receipt is itemized in seconds. The total is in years. That is the cost of accumulated overhead in the only currency that doesn't reset. So, we know what was lost. Time, dollars, watts, tons. And we know what was sold to us as the upside. Developer convenience, faster shipping, more features. Now, the part the industry never volunteers. Who collected the bill?

[00:19:03]Three groups, cloud, frameworks, hardware.

[00:19:07]The cloud first. Amazon Web Services in the first quarter of calendar 2025 recorded a 39.5% operating margin. That is for every dollar of revenue almost 40 cents fell to operating income.

[00:19:20]Costco, for context, operates at 3 and a half%. Walmart 4. AWS's rented memory and rented CPU marked up at roughly 10 times the margins of the companies that sell you food.

[00:19:32]For every gigabyte of RAM an Electron app burns at scale, somebody on the operations team at your company is paying Amazon to rent it. At 40% margin, the harder your stack works to render a button, the more money Amazon makes.

[00:19:45]This is not a moral judgment. It is the business model. The cloud industry structurally makes more money the worse your software performs. There is no incentive anywhere in the AWS revenue line for your application to use less memory. The pricing is per bite per hour. The bytes cost Amazon nothing. The hours cost you everything.

[00:20:04]Cloudflare's pingora, a rust rewrite of nginx, runs the same traffic on 70% less CPU and 67% less memory than the equivalent ngx deployment. Cloudflare's own engineering blog. The savings exist.

[00:20:19]The savings are real. The savings were just until somebody bothered to write Pingora being banked by Cloudflare's hosting bill, not their balance sheet.

[00:20:28]That is line one of the answer to who profited. Line two, the framework economy. Ma in September of 2025, Versel, the company that makes Nex.js, closed a series F at a $9.3 billion valuation from $3.25 billion to 9.3 billion in 16 months. The series F announcement disclosed that Nex.js downloads in the preceding 12 months exceeded all Nex.js downloads from 2016 through 2024 combined.

[00:20:56]Read that as a sentence. The framework's adoption tripled in a year. The framework's valuation tripled in a year.

[00:21:03]The framework's downloads tripled in a year. The performance of the median web page built on that framework by any measured benchmark did not.

[00:21:12]If you have ever wondered why every senior engineer on Twitter today seems vaguely furious about Nex.js, this curve is why. It is not a tooling question. It is a wealth transfer question. A $9.3 billion company is being built on top of the abstraction tax we have been documenting for 26 minutes.

[00:21:32]A 2016 incident is the cleanest illustration. A developer pulled an 11line package called left pad off npm.

[00:21:38]The entire JavaScript ecosystem briefly stopped working. Hundreds of major projects could not build the package padded strings with a character, 11 lines.

[00:21:47]A modern web application today depends on roughly 1,800 packages to display a hello world. Each one of them has a maintainer. Some of them are people.

[00:21:55]Some of them are weekend projects. None of them is being paid by the framework vendor whose $9.3 billion valuation depends on their software not breaking.

[00:22:04]The receipt has another payee that nobody is looking at. and the payee is the open-source contributor who is doing the framework's loadbearing work for free. Line three, the hardware industry, DAC, Nvidia's data center segment in the third quarter of fiscal 2026, that is calendar Q3 2025, recorded 51.2 billion in revenue in a single quarter, up 66% year-over-year.

[00:22:30]Most of that revenue is the inference workload. The cost of running, not training, the AI models that are now embedded inside every text editor, every spreadsheet, every browser tab. The same models that in many cases use three orders of magnitude more compute to produce the same paragraph that GPT3 produced 2 years ago.

[00:22:49]Apple's pitch for the M series silicon is performance per watt. The benchmarks back the pitch. The M4 is roughly 65% faster single core than the M1. But userfacing apps on those machines are not 65% faster. They are in many cases slower because the abstraction layers Electron, React, Webpack absorbed every performance gain Apple shipped exactly as they have done for two decades.

[00:23:14]The treadmill is the business model.

[00:23:16]Faster hardware is sold to you to compensate for slower software. The slower software is sold to you as developer convenience. The cloud is sold to you to run the slower software at scale. The framework is sold to the developer to write the slower software faster. Karox.

[00:23:31]Every layer is monetized. Every transition has a margin. You are paying for the staircase. So what changes when you read the bill?

[00:23:39]The first thing that changes is that you cannot unsee it. Once you know that Slack costs more memory than the operating system it runs on cost 25 years ago, you will see it every time you open Slack. Once you know that 400 million PCs are about to enter a landfill because of a software cutoff, you will see it the next time someone tells you Windows 11 is more secure.

[00:23:59]Once you know that AWS is making 40 cents on the dollar, renting you back the CPU cycles your Electron app burned, you will see it the next time your team's cloud bill goes up. So you can argue with vibes, you cannot argue with a stopwatch, you cannot argue with a receipt.

[00:24:15]The second thing that changes is that you start noticing the people who are not paying it. Cloudflare wrote Pingora because somebody read this bill. Discord rewrote their read state service in Rust because somebody read this bill.

[00:24:26]Microsoft is shipping Rust in the Windows kernel by their security team's own published numbers. Roughly 36,000 lines of new kernel code and 152,000 lines of direct because somebody read this bill. Bun exists because somebody read this bill. Zigg exists because somebody read this bill. Mojo exists because somebody read this bill.

[00:24:47]And that's the third thing. The receipt is also a permission slip. It is permission to refuse the layer count. It is permission to ask in the next architecture meeting what flows through the pipes. It is permission to delete the package you don't understand. It is permission when somebody tells you that you need a runtime to render a button to ask them why.

[00:25:07]And remember the credit line. Spotify desktop is fine. WhatsApp desktop used to be. The cost of these layers is not a law of nature. It is a choice. Other teams have made other choices. The receipt is not destiny. I told you at the start that the first video on this channel was the diagnosis. This one is the receipt. The next one is the treatment.

[00:25:27]The treatment is not nostalgia. The treatment is not learning to write assembly. The treatment is the rediscovery of a single principle. That every line of code you ship costs something in someone's RAM, on someone's electricity bill, in someone's hours at some point in the chain. And that the people who pretend otherwise are the people billing you.

[00:25:47]For two decades, the industry strategy was throw more hardware at it. That strategy is ending. Not because the hardware stopped improving. It didn't.

[00:25:55]It just stopped improving in the one direction the entire software industry had been quietly assuming would keep funding our laziness. Single thread performance is plateaued. Memory price reduction has slowed. The free lunch is over. And the bill is finally arriving.

[00:26:10]If you watch the first video on this channel, you watch the diagnosis. This was the receipt. The next video is what to do about it. What teams and individuals can actually start removing from their stack this week without rewriting their company. It will be on this channel in two weeks. Subscribe if you want to know when the bill stops coming due.