Latest Posts (5 found)
Wreflection 1 weeks ago

Command Lines

In the early 1950s, Grace Hopper coined the term “compiler” and built one of the first versions with her A-0 system 1 . The compilers that followed abstracted away machine code, letting programmers focus on higher-level logic instead of lower-level hardware details. Today, AI coding assistants 2 are enabling a similar change, letting software engineers focus on higher-order work by generating code from natural language prompts 3 . Everyone from big tech to well-funded startups is competing to capture this shift. Yesterday Google announced Antigravity, their new AI coding assistant, and the day before, AWS announced the general availability of their AI coding tool, Kiro. Last week, Cursor, the standout startup in this space, raised $2.3B in their series-D round at a valuation of $29.3B. Two lines in Cursor’s press release stood out to me. The first: We’ve also crossed $1B in annualized revenue, counting millions of developers. This disclosure means Anysphere Inc. (Cursor’s parent company) is the fastest company in history to reach $1B in annual recurring revenue (ARR). Yes, faster than OpenAI, and faster than Anthropic 4 . Source: Yuchen Jin, Twitter/X, 2025 Engineers are trying every new AI coding tool. As a result, the AI-coding tool market is growing exponentially (+5x in just over a year) 5 . But it’s still early. As I wrote in Why Some AI Wrappers Build Billion-dollar Businesses , companies spend several hundred billion dollars a year on software engineering, and AI has the potential to unlock productivity gains across that entire spend. Software developers represent roughly 30% of the workforce at the world’s five largest market cap companies, all of which are technology firms as of October 2025. Development tools that boost productivity by even modest percentages unlock billions in value. In my view, this nascent market is splitting based on three types of users. Source: Command Lines, wreflection.com, 2025 On one end is Handcrafted Coding . These are engineers who actively decline to use LLMs, either because of skepticism about quality or insistence on full control of every code. They argue that accepting AI suggestions creates technical debt you cannot see until it breaks in production. This segment continues to decline as the quality of AI coding models improves. The opposite end is Vibe Coding . These are typically non-engineers, who use AI to build concepts and prototypes. They prompt the model hoping for an end-to-end solution, accept the output with minimal review, and trust that it works. The user describes what they want and lets the model figure out the implementation details of how to build it. In the middle sits Architect + AI Coding . The engineer uses the AI/LLM as a pair programmer exploring system designs, analyzing data models, and reviewing API details. When the work is something entirely new or something that needs careful handling, the human programmer still codes those pieces by hand. But for boilerplate code, package installations, generic User Interface (UI) components, and any kind of code that is typically found on the internet, they assign it to the model 6 . The engineer stays in command of what is important to them and delegates what is not. Based on the user types, I think, the AI coding market splits into two. Source: wreflection.com based on SemiAnalysis estimate, 2025 Hands-off: Non-engineers (product managers, designers, marketers, other internal employees) use these tools to vibe code early product concepts. They look to AI as the lead engineer to spin-up concepts/prototypes of apps, websites, and tools by simply prompting the AI to make something for them. Lovable, Vercel, Bolt, and Figma Make fit here 7 . Code from these users, as of now, are not typically pushed to prod. Hands-on: Professional software engineers use these tools in their existing workflow to ship production code. They use AI as an assistant to write boilerplate code, refactor existing services, wire new features or UI screens, and triage bugs in codebases. Cursor, Claude Code, OpenAI Codex, Github Copilot, Cline, AWS Kiro play here. These products live where the work is done , and integrate into the engineer’s workflow. This is, at least as of now, the bigger market segment. To see an evaluation of all the major AI coding tools currently in the market, checkout this breakdown by Peter Yang, who runs the newsletter Behind The Craft . That brings me to the second thing in Cursor’s press release that stood out to me: Our in-house models now generate more code than almost any other LLMs in the world. While I am not convinced about that claim 8 , what I am convinced about is that Cursor is still growing despite its previous reliance on foundation models. From Why Some AI Wrappers Build Billion-dollar Businesses again: But Cursor and other such tools depend almost entirely on accessing Anthropic, OpenAI and Gemini models, until open-source open-weight and in-house models match or exceed frontier models in quality. Developer forums are filled with complaints about rate limits from paying subscribers. In my own projects, I exhausted my Claude credits in Cursor mid-project and despite preferring Cursor’s user interface and design, I migrated to Claude Code (and pay ten times more to avoid rate limits). The interface may be better, but model access proved decisive. Cursor’s new in-house model Composer-2, which just launched last month, is a good example of how this model versus application competition is evolving. Cursor claims (without any external benchmarks, I must say) that Composer-2 is almost as good as frontier models but 4x faster. It’s still early to say how true that is. Open-source models have not yet come close to the top spots in SWE-bench verified or in private evals 9 . Source : Introducing Claude Sonnet 4.5, Anthropic, 2025. To me, model quality is the most decisive factor in these AI coding wars. And in my view, that’s why Claude Code has already overtaken Cursor, and OpenAI’s Codex is close behind, despite both having launched a year or so later. Source : SemiAnalysis, 2025 Even though the newcomers Cursor, Claude Code, and OpenAI Codex are the talk of the (developer) town, incumbents such as Microsoft with Github Copilot, AWS with Kiro, and Google with Antigravity, can utilize their existing customer relationships, bundle their offerings with their existing suites, and/or provide their option as the default in their tech stack to compete. As an example, Cursor charges $20–$40 monthly per user for productive usage, while Google Antigravity launched free with generous limits for individual users. Github Copilot still leads this market, proving once again that enterprise bundling and distribution has structural advantages. This is the classic Microsoft Teams vs. Slack Dynamic 10 . One way for startups to compete is by winning individual users who may use a coding tool with or without formal approval, and then be the tool’s advocate inside the organization. That organic interest and adoption eventually forces IT and security teams to officially review the tool and then eventually sanction its usage. Yet, even as these newer tools capture developer mindshare, the underlying developer tools market is changing. Both the IDEs developers choose and the resources they we consult have changed dramatically. StackOverflow, once the default for programmers stuck on a programming issue, has seen its traffic and number of questions decline dramatically since ChatGPT’s launch, suggesting that AI is already replacing some traditional developer resources. Source : Developer Tools 2.0, Sequoia, 2023 Just as compilers freed programmers from writing assembly code, AI tools are freeing software engineers from the grunt work of writing boilerplate and routine code, and letting them focus on higher-order thinking. Eventually, one day, AI may get so good that it will generate applications on demand and create entire software ecosystems autonomously. Both hands-off and hands-on AI coding tools, as well as incumbents and newcomers, see themselves as the path to that fully autonomous software generation, even if they are taking different approaches. The ones who get there will be those who deliver the best model quality that ships code reliably, go deep enough to ship features that foundation models can’t care enough to replicate, and become sticky enough that users will not leave even when they can 11 . If you enjoyed this post, please consider sharing it on Twitter/X or LinkedIn , and tag me when you do. Thanks for reading Wreflection! Subscribe for free to receive new posts and support my work. Hopper’s A-0 system and her definition of the term compiler is different from what we consider a compiler today, but it established the foundational concept. In the context of coding assistants, most products labeled as AI tools are powered by LLMs, and so I use AI and LLM interchangeably in this article despite the actual difference. https://x.com/karpathy/status/1617979122625712128 A better comparison might be at the product level rather than company level. In that case, ChatGPT and Claude both reached $1B faster than Cursor did. https://newsletter.semianalysis.com/p/microsofts-ai-strategy-deconstructed I would argue that the vast majority of productive code is hidden behind company firewalls. Current foundation models are trained on publicly available data on the internet, and do not have access to proprietary codebases. We are yet to see breakthrough solutions where a company augments their confidential private data to generate production-ready code using current LLMs. While Retrieval-Augmented Generation has shown some promise, it has not yet delivered transformative results. Companies such as Glean are actively working on this problem. Replit and Cognition probably appeal to both segments. To me, Replit leans hands-off with its rapid prototyping focus. Cognition’s agent-based approach, though hands-off, lets engineers still control the code directly, making it lean hands-on. I was curious how Cursor knows how much code is generated by other LLMs outside Cursor? When I asked this on hackernews, swyx suggested that they “ can pretty much triangulate across openrouter x feedback from the top 3 model labs to compare with internal usage and figure that out ”. To me, triangulation makes sense for internal estimates. but for external publication, I’m surprised Cursor didn’t include “we estimate” or similar qualifying language. My understanding is that FTC policy requires substantiation before making definitive comparative claims (like more than, better than etc). All that to say, I’m not fully convinced about their claims. SWE-bench is a benchmark for evaluating large language models (LLMs) on real world software engineering tasks and issues collected from GitHub. Performance against public benchmarks can be gamed by the model builders. Currently after any new model launch, we see people using the model in the wild and forming a consensus around how the model performs which is a better indicator than these benchmarks. Microsoft bundled Teams into Office 365 subscriptions at no extra cost, using its dominant enterprise distribution to surpass Slack’s paid standalone product within three years despite Slack’s earlier launch and product innovation. See https://venturebeat.com/ai/microsoft-teams-has-13-million-daily-active-users-beating-slack Natasha Malpani , Twitter/X, 2025 Source: Yuchen Jin, Twitter/X, 2025 Engineers are trying every new AI coding tool. As a result, the AI-coding tool market is growing exponentially (+5x in just over a year) 5 . But it’s still early. As I wrote in Why Some AI Wrappers Build Billion-dollar Businesses , companies spend several hundred billion dollars a year on software engineering, and AI has the potential to unlock productivity gains across that entire spend. Software developers represent roughly 30% of the workforce at the world’s five largest market cap companies, all of which are technology firms as of October 2025. Development tools that boost productivity by even modest percentages unlock billions in value. In my view, this nascent market is splitting based on three types of users. Source: Command Lines, wreflection.com, 2025 On one end is Handcrafted Coding . These are engineers who actively decline to use LLMs, either because of skepticism about quality or insistence on full control of every code. They argue that accepting AI suggestions creates technical debt you cannot see until it breaks in production. This segment continues to decline as the quality of AI coding models improves. The opposite end is Vibe Coding . These are typically non-engineers, who use AI to build concepts and prototypes. They prompt the model hoping for an end-to-end solution, accept the output with minimal review, and trust that it works. The user describes what they want and lets the model figure out the implementation details of how to build it. In the middle sits Architect + AI Coding . The engineer uses the AI/LLM as a pair programmer exploring system designs, analyzing data models, and reviewing API details. When the work is something entirely new or something that needs careful handling, the human programmer still codes those pieces by hand. But for boilerplate code, package installations, generic User Interface (UI) components, and any kind of code that is typically found on the internet, they assign it to the model 6 . The engineer stays in command of what is important to them and delegates what is not. The Market Split Based on the user types, I think, the AI coding market splits into two. Source: wreflection.com based on SemiAnalysis estimate, 2025 Hands-off: Non-engineers (product managers, designers, marketers, other internal employees) use these tools to vibe code early product concepts. They look to AI as the lead engineer to spin-up concepts/prototypes of apps, websites, and tools by simply prompting the AI to make something for them. Lovable, Vercel, Bolt, and Figma Make fit here 7 . Code from these users, as of now, are not typically pushed to prod. Hands-on: Professional software engineers use these tools in their existing workflow to ship production code. They use AI as an assistant to write boilerplate code, refactor existing services, wire new features or UI screens, and triage bugs in codebases. Cursor, Claude Code, OpenAI Codex, Github Copilot, Cline, AWS Kiro play here. These products live where the work is done , and integrate into the engineer’s workflow. This is, at least as of now, the bigger market segment.

business Business ai AI programming Programming
0 views
Wreflection 1 months ago

AI's Dial-Up Era

It is 1995. Your computer modem screeches as it tries to connect to something called the internet. Maybe it works. Maybe you try again. For the first time in history, you can exchange letters with someone across the world in seconds. Only 2000-something websites exist 1 , so you could theoretically visit them all over a weekend. Most websites are just text on gray backgrounds with the occasional pixelated image 2 . Loading times are brutal. A single image takes a minute, a 1-minute video could take hours. Most people do not trust putting their credit cards online. The advice everyone gives: don’t trust strangers on the internet. People split into two camps very soon. Optimists predict grand transformations. Some believe digital commerce will overtake physical retail within years. Others insist we’ll wander around in virtual reality worlds. “I expect that within the next five years more than one in ten people will wear head-mounted computer displays while traveling in buses, trains, and planes.” - Nicholas Negroponte, MIT Professor, 1993 Pessimists call the internet a fad and a bubble. Source : Did Paul Krugman Say the Internet’s Effect on the World Economy Would Be ‘No Greater Than the Fax Machine’s’? Snopes, 2018. Original quote in Red Herring magazine, 1998 If you told the average person in 1995 that within 25 years, we’d consume news from strangers on social media over newspapers, watch shows on-demand in place of cable TV, find romantic partners through apps more than through friends, and flip “don’t trust strangers on the internet” so completely that we’d let internet strangers pick us up in their personal vehicles and sleep in their spare bedrooms, most people would find that hard to believe. We’re in 1995 again. This time with Artificial Intelligence. And both sides of today’s debate are making similar mistakes. One side warns that AI will eliminate entire professions and cause mass unemployment within a couple of years. The other claims that AI will create more jobs than it destroys. One camp dismisses AI as overhyped vaporware destined for a bubble burst, while the other predicts it will automate every knowledge task and reshape civilization within the decade. Both are part right and part wrong. Subscribe now Geoffrey Hinton, who some call the Father of AI, warned in 2016 that AI would trigger mass unemployment. “People should stop training radiologists now,” he declared, certain that AI would replace them within years. Source : Twitter/X, Andy Walters/Geoffrey Hinton, 2023 Yet as Deena Mousa, a researcher, shows in “ The Algorithm Will See You Now ,” AI hasn’t replaced radiologists, despite predictions. It is thriving. In 2025, American diagnostic radiology residency programs offered a record 1,208 positions across all radiology specialties, a four percent increase from 2024, and the field’s vacancy rates are at all-time highs . In 2025, radiology was the second-highest-paid medical specialty in the country, with an average income of $520,000, over 48 percent higher than the average salary in 2015 . Source : The algorithm will see you now, Deena Mousa, 2025 Mousa identifies a few factors for why the prediction failed - real-world complexity, the job involves more than image recognition, and regulatory/insurance hurdles. Most critical she points is Jevons Paradox , which is the economic principle that a technological improvement in resource efficiency leads to an increase in the total consumption of that resource, rather than a decrease. Her argument is that as AI makes radiologists more productive, better diagnostics and faster turnaround at lower costs mean more people get scans. So employment doesn’t decrease. It increases. This is also the Tech world’s consensus. Microsoft CEO Satya Nadella agrees , as does Box CEO Aaron Levie, who suggests : “The least understood yet most important concept in the world is Jevons Paradox. When we make a technology more efficient, demand goes well beyond the original level. AI is the perfect example of this—almost anything that AI is applied to will see more demand, not less.” They’re only half right. First, as Andrej Karpathy, the computer scientist who coined the term vibe coding , points out , radiology is not the right job to look for initial job displacements. “Radiology is too multi-faceted, too high risk, too regulated. When looking for jobs that will change a lot due to AI on shorter time scales, I’d look in other places - jobs that look like repetition of one rote task, each task being relatively independent, closed (not requiring too much context), short (in time), forgiving (the cost of mistake is low), and of course automatable giving current (and digital) capability. Even then, I’d expect to see AI adopted as a tool at first, where jobs change and refactor (e.g. more monitoring or supervising than manual doing, etc).” Second, the tech consensus that we will see increased employment actually depends on the industry. Specifically, how much unfulfilled demand can be unlocked in that industry, and whether this unfulfilled demand growth outpaces continued automation and productivity improvement . To understand this better, look at what actually happened in three industries over a 200-year period from 1800 to 2000. In the paper Automation and jobs: when technology boosts employment , James Bessen, an economist, shows the employment, productivity, and demand data for textile, iron & steel, and motor vehicle industries. Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 After automation, both textile and iron/steel workers saw employment increase for nearly a century before experiencing a steep decline. Vehicle manufacturing, by contrast, holds steady and hasn’t seen the same steep decline yet. To answer why those two industries saw sharp declines but motor vehicle manufacturing did not, first look at the productivity of workers in all three industries: Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 Then look at the demand across those three industries: Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 What the graphs show is a consistent pattern ( note: the productivity and demand graphs are logarithmic, meaning productivity and demand grew exponentially ). Early on, a service or product is expensive because many workers are needed to produce it. Most people can’t afford it or use them sparingly. For example, in the early 1800s, most people could only afford a pair of pants or shirt. Then automation makes workers dramatically more productive. A textile worker in 1900 could produce fifty times more than one in 1800. This productivity explosion crashes prices, which creates massive new demand. Suddenly everyone can afford multiple outfits instead of just one or two. Employment and productivity both surge ( note: employment growth masks internal segment displacement and wage changes. See footnote 3 ) Once demand saturates, employment doesn’t further increase but holds steady at peak demand. But as automation continues and workers keep getting more productive, employment starts to decline. In textiles, mechanization enabled massive output growth but ultimately displaced workers once consumption plateaued while automation and productivity continued climbing. We probably don’t need infinite clothing. Similarly, patients will likely never need a million radiology reports, no matter how cheap they become and so radiologists will eventually hit a ceiling. We don’t need infinite food, clothing, tax returns, and so on. Motor vehicles, in Bessen’s graphs, tell a different story because demand remains far from saturated. Most people globally still don’t own cars. Automation hasn’t completely conquered manufacturing either (Tesla’s retreat from full manufacturing automation proves the current technical limits). When both demand and automation potential remain high, employment can sustain or even grow despite productivity gains. Software presents an even more interesting question. How many apps do you need? What about software that generates applications on demand, that creates entire software ecosystems autonomously? Until now, handcrafted software was the constraint. Expensive software engineers and their labor costs limited what companies could afford to build. Automation changes this equation by making those engineers far more productive. Both consumer and enterprise software markets suggest significant unmet demand because businesses have consistently left projects unbuilt 4 . They couldn’t justify the development costs or had to allocate limited resources to their top priority projects. I saw this firsthand at Amazon. Thousands of ideas went unfunded not because they lacked business value, but because of the lack of engineering resources to build them. If AI can produce software at a fraction of the cost, that unleashes enormous latent demand. The key question then is if and when that demand will saturate. So to generalize, for each industry, employment hinges on a race between two forces: The magnitude and growth of unmet market demand, and Whether that demand growth outpaces productivity improvements from automation. Illustrative matrix for different industries Different industries will experience different outcomes depending on who’s winning that demand and productivity race. The second debate centers on whether this AI boom is a bubble waiting to burst . The dotcom boom of the 1990s saw a wave of companies adding “.com” to their name to ride the mania and watch their valuations soar. Infra companies poured billions into fiber optics and undersea cables - expensive projects only possible because people believed the hype 5 . All of this eventually burst in spectacular fashion in the dotcom crash in 2000-2001 . Infrastructure companies like Cisco briefly became the most valuable in the world only to come tumbling down 6 . Pets.com served as the poster child of this exuberance raising $82.5 million in its IPO, spending millions on a Super Bowl ad only to collapse nine months later 7 . But the dotcom bubble also got several things right . More importantly, it eventually bought us the physical infrastructure that made YouTube, Netflix, and Facebook possible. Sure, companies like Worldcom, NorthPoint, and Global Crossing making these investments went bankrupt, but they also laid the foundation for the future. Although the crash proved the skeptics right in the short term, it proved the optimists were directionally correct in the long term. Today’s AI boom shows similar exuberance. Consider the AI startup founded by former OpenAI executive Mira Murati, which raised $2 billion at a $10 billion valuation, the largest seed round in history 8 . This despite having no product and declining to reveal what it’s building or how it will generate returns. Several AI wrappers have raised millions in seed funding with little to no moat. Yet some investments will outlast the hype and will likely help future AI companies even if this is a bubble. For example, the annual capital expenditures of Hyperscalers 9 that have more than doubled since ChatGPT’s release - Microsoft, Google, Meta, and Amazon are collectively spending almost half a trillion dollars on data centers, chips, and compute infrastructure. Regardless of which specific companies survive, this infrastructure being built now will create the foundation for our AI future - from inference capacity to the power generation needed to support it. Source : Is AI a bubble, Exponential View, 2025 The infrastructure investments may have long-term value, but are we already in bubble territory? Azeem Azhar, a tech analyst and investor, provides an excellent practical framework to answer the AI bubble question. He benchmarks today’s AI boom using five gauges: economic strain (investment as a share of GDP), industry strain (capex to revenue ratios), revenue growth trajectories (doubling time), valuation heat (price-to-earnings multiples), and funding quality (the resilience of capital sources). His analysis shows that AI remains in a demand-led boom rather than a bubble, but if two of the five gauges head into red, we will be in bubble territory. The demand is real. After all OpenAI is one of the fastest-growing companies in history 10 . But that alone doesn’t prevent bubbles. OpenAI will likely be fine given its product-market fit, but many other AI companies face the same unit economics questions that plagued dotcom companies in the 1990s. Pets.com had millions of users too (a then large portion of internet users), but as the tech axiom goes, you can acquire infinite customers and generate infinite revenue if you sell dollars for 85 cents 11 . So despite the demand, the pattern may rhyme with the 1990s. Expect overbuilding. Expect some spectacular failures. But also expect the infrastructure to outlast the hype cycle and enable things we can’t yet imagine. So where does this leave us? We’re early in the AI revolution. We’re at that metaphorical screeching modem phase of the internet era. Just as infrastructure companies poured billions into fiber optics, hyperscalers now pour billions into compute. Startups add “.ai” to their names like companies once added “.com” as they seek higher valuations. The hype will cycle through both euphoria and despair. Some predictions will look laughably wrong. Some that seem crazy will prove conservative. Different industries will experience different outcomes. Unlike what the Jevons optimists suggest, demand for many things plateaus once human needs are met. Employment outcomes in any industry depend on the magnitude and growth of unmet market demand and whether that demand growth outpaces productivity improvements from automation. Cost reduction will unlock market segments. Aswath Damodaran, a finance professor, (in)famously undervalued Uber assuming it would only capture a portion of the existing taxi market 12 . He missed that making rides dramatically cheaper would expand the market itself as people took Ubers to destinations they’d never have paid taxi prices to reach. AI will similarly enable products and services currently too expensive to build with human intelligence. A restaurant owner might use AI to create custom supply chain software that say at $100,000 with human developers would never have been built. A non-profit might deploy AI to contest a legal battle that was previously unaffordable. We can predict change, but we can’t predict the details. No one in 1995 predicted we’d date strangers from the internet, ride in their ubers , or sleep in their airbnbs . Or that a job called influencers would become the most sought-after career among young people. Human creativity generates outcomes we can’t forecast with our current mental models. Expect new domains and industries to emerge. AI has already helped us decode more animal communication in the last five years than in the last fifty. Can we predict what jobs a technology that allows us to have full-blown conversations with them will unlock? A job that doesn’t exist today will likely be the most sought-after job in 2050. We can’t name it because it hasn’t been invented yet. Job categories will transform. Even as the internet made some jobs obsolete, it also transformed others and created new categories. Expect the same with AI. Karpathy ends with a question: About 6 months ago, I was also asked to vote if we will have less or more software engineers in 5 years. Exercise left for the reader. To answer this question, go back to 1995 and ask the same question but with journalists. You might have predicted more journalists because the internet would create more demand by enabling you to reach the whole world. You’d be right for 10 or so years as employment in journalism grew until the early 2000s. But 30 years later, the number of newspapers and the number of journalists both have declined, even though more “journalism” happens than ever. Just not by people we call journalists. Bloggers, influencers, YouTubers, and newsletter writers do the work that traditional journalists used to do 13 . The same pattern will play out with software engineers. We’ll see more people doing software engineering work and in a decade or so, what “software engineer” means will have transformed. Consider the restaurant owner from earlier who uses AI to create custom inventory software that is useful only for them. They won’t call themselves a software engineer. So just like in 1995, if the AI optimists today say that within 25 years, we’d prefer news from AI over social media influencers, watch AI-generated characters in place of human actors, find romantic partners through AI matchmakers more than through dating apps (or perhaps use AI romantic partners itself), and flip “don’t trust AI” so completely that we’d rely on AI for life-or-death decisions and trust it to raise our children, most people would find that hard to believe. Even with all the intelligence, both natural and artificial, no one can predict with certainty what our AI future will look like. Not the tech CEOs, not the AI researchers, and certainly not some random guy pontificating on the internet. But whether we get the details right or not, our AI future is loading. Approximately 2,879 websites were established before 1995, expanding to 23,500 by June 1995. See https://en.wikipedia.org/wiki/List_of_websites_founded_before_1995 https://www.fastcompany.com/91140068/how-the-internet-went-mainstream-in-1994 Historical data from the Industrial Revolution shows that even as aggregate textile employment grew, workers shifted between job types within the industry as some roles became redundant. And correspondingly, some jobs saw wages collapse while others saw increases. For example, domestic hand-loom weavers were displaced by power looms and saw their wages collapse, while self-acting mule spinners (a newly created role) and factory workers saw stable employment and steady compensation growth. Additionally, Britain’s deflationary period (1815-1850) saw food prices fall by half, meaning real purchasing power often rose even when nominal wages declined. Despite all this, the psychological reality was harsh. Even with falling prices, watching your paycheck shrink while neighbors lost jobs and debts grew harder to pay (lower wages but fixed obligations) created real instability regardless of aggregate statistics improvements. Also see Acemoglu & Johnson, 2024 . 4 Ask any tech product leader about their roadmap planning, and they’ll all universally report far more worthwhile projects than resources to build them, forcing ruthless prioritization to decide what gets built. https://en.wikipedia.org/wiki/Dot-com_bubble#Bubble_in_telecom https://nypost.com/2000/03/28/ciscos-market-cap-tops-microsofts/ https://www.begintoinvest.com/lessons-from-pets-tech-bubble/ https://techcrunch.com/2025/07/15/mira-muratis-thinking-machines-lab-is-worth-12b-in-seed-round/ Hyperscalers are large cloud computing companies like Microsoft, Google, Meta, and Amazon that operate massive data centers and provide the computing infrastructure necessary to train and run AI models at scale. https://epoch.ai/gradient-updates/openai-is-projecting-unprecedented-revenue-growth https://25iq.com/2016/10/14/a-half-dozen-more-things-ive-learned-from-bill-gurley-about-investing/ Damodaran valued Uber at $6B, assuming a 10% market share of the then $100B taxi market. Uber’s market cap as of October 2025 is $190B. Job transformation doesn’t guarantee comparable compensation. Much of this new “journalism” or content creation happens for free or at rates far below what traditional news organizations paid, separating the work from stable employment. Source : Did Paul Krugman Say the Internet’s Effect on the World Economy Would Be ‘No Greater Than the Fax Machine’s’? Snopes, 2018. Original quote in Red Herring magazine, 1998 If you told the average person in 1995 that within 25 years, we’d consume news from strangers on social media over newspapers, watch shows on-demand in place of cable TV, find romantic partners through apps more than through friends, and flip “don’t trust strangers on the internet” so completely that we’d let internet strangers pick us up in their personal vehicles and sleep in their spare bedrooms, most people would find that hard to believe. We’re in 1995 again. This time with Artificial Intelligence. And both sides of today’s debate are making similar mistakes. One side warns that AI will eliminate entire professions and cause mass unemployment within a couple of years. The other claims that AI will create more jobs than it destroys. One camp dismisses AI as overhyped vaporware destined for a bubble burst, while the other predicts it will automate every knowledge task and reshape civilization within the decade. Both are part right and part wrong. Subscribe now The Employment Paradox: Why Automation’s Impact Depends On The Industry Geoffrey Hinton, who some call the Father of AI, warned in 2016 that AI would trigger mass unemployment. “People should stop training radiologists now,” he declared, certain that AI would replace them within years. Source : Twitter/X, Andy Walters/Geoffrey Hinton, 2023 Yet as Deena Mousa, a researcher, shows in “ The Algorithm Will See You Now ,” AI hasn’t replaced radiologists, despite predictions. It is thriving. In 2025, American diagnostic radiology residency programs offered a record 1,208 positions across all radiology specialties, a four percent increase from 2024, and the field’s vacancy rates are at all-time highs . In 2025, radiology was the second-highest-paid medical specialty in the country, with an average income of $520,000, over 48 percent higher than the average salary in 2015 . Source : The algorithm will see you now, Deena Mousa, 2025 Mousa identifies a few factors for why the prediction failed - real-world complexity, the job involves more than image recognition, and regulatory/insurance hurdles. Most critical she points is Jevons Paradox , which is the economic principle that a technological improvement in resource efficiency leads to an increase in the total consumption of that resource, rather than a decrease. Her argument is that as AI makes radiologists more productive, better diagnostics and faster turnaround at lower costs mean more people get scans. So employment doesn’t decrease. It increases. This is also the Tech world’s consensus. Microsoft CEO Satya Nadella agrees , as does Box CEO Aaron Levie, who suggests : “The least understood yet most important concept in the world is Jevons Paradox. When we make a technology more efficient, demand goes well beyond the original level. AI is the perfect example of this—almost anything that AI is applied to will see more demand, not less.” They’re only half right. First, as Andrej Karpathy, the computer scientist who coined the term vibe coding , points out , radiology is not the right job to look for initial job displacements. “Radiology is too multi-faceted, too high risk, too regulated. When looking for jobs that will change a lot due to AI on shorter time scales, I’d look in other places - jobs that look like repetition of one rote task, each task being relatively independent, closed (not requiring too much context), short (in time), forgiving (the cost of mistake is low), and of course automatable giving current (and digital) capability. Even then, I’d expect to see AI adopted as a tool at first, where jobs change and refactor (e.g. more monitoring or supervising than manual doing, etc).” Second, the tech consensus that we will see increased employment actually depends on the industry. Specifically, how much unfulfilled demand can be unlocked in that industry, and whether this unfulfilled demand growth outpaces continued automation and productivity improvement . To understand this better, look at what actually happened in three industries over a 200-year period from 1800 to 2000. In the paper Automation and jobs: when technology boosts employment , James Bessen, an economist, shows the employment, productivity, and demand data for textile, iron & steel, and motor vehicle industries. Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 After automation, both textile and iron/steel workers saw employment increase for nearly a century before experiencing a steep decline. Vehicle manufacturing, by contrast, holds steady and hasn’t seen the same steep decline yet. To answer why those two industries saw sharp declines but motor vehicle manufacturing did not, first look at the productivity of workers in all three industries: Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 Then look at the demand across those three industries: Source : Automation and jobs: when technology boosts employment, James Bessen, 2019 What the graphs show is a consistent pattern ( note: the productivity and demand graphs are logarithmic, meaning productivity and demand grew exponentially ). Early on, a service or product is expensive because many workers are needed to produce it. Most people can’t afford it or use them sparingly. For example, in the early 1800s, most people could only afford a pair of pants or shirt. Then automation makes workers dramatically more productive. A textile worker in 1900 could produce fifty times more than one in 1800. This productivity explosion crashes prices, which creates massive new demand. Suddenly everyone can afford multiple outfits instead of just one or two. Employment and productivity both surge ( note: employment growth masks internal segment displacement and wage changes. See footnote 3 ) Once demand saturates, employment doesn’t further increase but holds steady at peak demand. But as automation continues and workers keep getting more productive, employment starts to decline. In textiles, mechanization enabled massive output growth but ultimately displaced workers once consumption plateaued while automation and productivity continued climbing. We probably don’t need infinite clothing. Similarly, patients will likely never need a million radiology reports, no matter how cheap they become and so radiologists will eventually hit a ceiling. We don’t need infinite food, clothing, tax returns, and so on. Motor vehicles, in Bessen’s graphs, tell a different story because demand remains far from saturated. Most people globally still don’t own cars. Automation hasn’t completely conquered manufacturing either (Tesla’s retreat from full manufacturing automation proves the current technical limits). When both demand and automation potential remain high, employment can sustain or even grow despite productivity gains. Software presents an even more interesting question. How many apps do you need? What about software that generates applications on demand, that creates entire software ecosystems autonomously? Until now, handcrafted software was the constraint. Expensive software engineers and their labor costs limited what companies could afford to build. Automation changes this equation by making those engineers far more productive. Both consumer and enterprise software markets suggest significant unmet demand because businesses have consistently left projects unbuilt 4 . They couldn’t justify the development costs or had to allocate limited resources to their top priority projects. I saw this firsthand at Amazon. Thousands of ideas went unfunded not because they lacked business value, but because of the lack of engineering resources to build them. If AI can produce software at a fraction of the cost, that unleashes enormous latent demand. The key question then is if and when that demand will saturate. So to generalize, for each industry, employment hinges on a race between two forces: The magnitude and growth of unmet market demand, and Whether that demand growth outpaces productivity improvements from automation.

ai AI business Business
0 views
Wreflection 1 months ago

Wrapping Your Head Around AI Wrappers

“That’s just an AI wrapper.” The put‑down is now familiar for anyone developing something new using Artificial Intelligence. The push-back is just as familiar. “Everything is a wrapper. OpenAI is a wrapper around Nvidia and Azure. Netflix is a wrapper around AWS. Salesforce is an Oracle database wrapper valued at $320 billion,” says Perplexity CEO Aravind Srinivas 1 . For those not familiar with the term “AI Wrapper,” here’s a good definition 2 . It is a dismissive term that refers to a lightweight application or service that uses existing AI models or APIs to provide specific functionality, typically with minimal effort or complexity involved in its creation. A popular example of an AI wrapper are apps that enable users to “chat” with a PDF. This type of AI application allows users to upload a PDF document, such as a research paper, and interact with an AI model to quickly analyze and obtain answers about the specific content. In the early days of ChatGPT, uploading documents as part of the prompt or creating a custom GPT was not possible, so these apps became very popular, very fast. AI Wrapper Meme: An API call to OpenAI under the hood. In my view, this AI wrapper debate misses a larger point. Wrappers are not all the same. Thin tricks enjoy a brief run and last only until big platforms bundle them into their suites. But products that live where users already work, write back to a proprietary system of record , and/or can make use of proprietary data can endure. The wrapper label is a distraction from what I think actually matters: (1) Is it a feature or a product, and (2) How big is the market segment. Thanks for reading Wreflection! Subscribe for free to receive new posts and support my work. Begin with the earlier example of a wrapper that lets you chat with a PDF. Such a tool solves one narrow problem - answering questions about a document. It does not create new documents or edit existing ones. It typically does not capture any unique data, or learn from user behavior. It is a means to an end; a capability rather than an end-to-end solution. As a result, this kind of feature belongs inside a document viewer or editor, or in the flagship applications of model providers. So when the foundation models themselves (OpenAI/ChatGPT, Anthropic/Claude, Google/Gemini) bundle this feature natively, the standalone tool becomes redundant. This is classic feature behavior - easy to copy, no end-to-end job, no moat or long-term defensibility. One caveat though; even those that are features can be an interesting indie businesses that make money until the platforms build it into their apps 3 . PDF.ai $500K MRR, PhotoAI $77K MRR, Chatbase $70K MRR, InteriorAI $53K MRR 4 . Jenni AI went from $2,000 to over $333,000 MRR in just 18 months 5 . Some wrappers are genuine products but live in market segments so large that model builders and big tech platforms cannot ignore them. Two vectors of competition come into play: (1) model access, and (2) distribution. Coding assistants illustrate both. Tools like Cursor turned a wrapper into a development environment that reads the repo, edits files, writes code, reverts changes, runs agents, and reimagines the developer experience for the AI-era. The market justifies the attention. Software developers represent roughly 30% of the workforce at the world’s five largest market cap companies, all of which are technology firms as of 2025 6 . Development tools that boost productivity by even modest percentages unlock billions in value. That makes this segment a prime target for both model builders and incumbents that already own distribution channels. But Cursor and other such tools depend almost entirely on accessing Anthropic, OpenAI and Gemini models. Developer forums are filled with complaints about rate limits from paying subscribers. In my own experiences, I exhausted my Claude credits in Cursor mid-project and despite preferring Cursor’s user interface and design, I migrated to Claude Code (and pay ten times more to avoid rate limits). The interface is good, but model access proved decisive. This foundation model competition extends to every category that OpenAI Applications CEO flagged as strategic (Knowledge/Tutoring, Health, Creative Expression, and Shopping) as well as other large market segments such as Writing Assistants, Legal Assistants, etc. Distribution poses the second threat. Even where model builders stay out, startups face a different race - can they build a user base faster than incumbents with existing products and distribution can add AI features? This is the classic Microsoft Teams vs. Slack Dynamic 7 . The challenge is in establishing a loyal customer base before Microsoft embeds Copilot in Excel/PowerPoint, or Google weaves Gemini into Workspace, or Adobe integrates AI across its creative suite. A standalone AI wrapper for spreadsheets or presentations must overcome not just feature parity but bundling/distribution advantages and switching costs. This distribution competition from incumbents also holds in other large markets such as healthcare and law. In these markets, regulatory friction and control of systems of record favor established players such as Epic Systems in healthcare. For e.g. A clinical note generator that cannot write to the Electronic Health Record (EHR) will likely come up against Epic’s distribution advantages sooner or later. Three caveats here: (1) First, speed to market can create exit options even without long-term defensibility; tools like Cursor may lack control over its core dependency (model access), but rapid growth make them attractive targets for model builders seeking instant market presence. (2) Second, superior execution occasionally beats structural advantage; Midjourney’s product quality convinced Meta to use it despite Meta’s substantially larger budget and distribution power. (3) Third, foundation models may avoid certain markets despite their size; regulatory burden in healthcare and legal, or reputational damage from AI companions or pornographic adult content may provide opportunities for operators willing to face extreme regulatory scrutiny or controversy. The opportunity remains large 8 , but competition (and/or acquisition) can come knocking. Cursor went from zero to $100 million in recurring revenue in 18 months, and became the subject of recurring OpenAI acquisition rumors. Windsurf , another coding assistant, received a $2.4B acquisition licensing deal from Google. Gamma reached $50 million in revenue in about a year. Lovable hit $50 million in revenue in just six months. Galileo AI acquired by Google for an undisclosed amount. Not every market gap attracts model builders or big tech. A long tail of jobs exists that are too small for venture scale but large enough to support multimillion-dollar businesses. These niches suit frugal founders with disciplined scope and lean operations. Consider those Manifestation or Horoscopes or Dream Interpreter AI apps. A dream interpreter that lets users record dreams each morning, generates AI videos based on them, maintains some kind of dream journal, and surfaces patterns over time solves a complete job. Yes, users could describe dreams to ChatGPT and it even stores history/memory, but a dedicated app can structure the dream capture with specific fields (recurring people, places, things, themes etc.) and integrate with sleep tracking data in ways a general chatbot likely cannot. Such a niche is small enough to avoid model attention but large enough to sustain a profitable indie business. While the previous categories frame opportunities for new ventures, incumbents face their own strategic choices in the wrapper debate when model builders arrive. Those that navigate model builder competition share two characteristics. First, they own the outcome even when they don’t own the model. Applications already embedded in user workflows (Gmail/Calendar, Sheets, EHR/EMR, Figma) require no new habit formation, and building these platforms from scratch is much harder than adding AI capability to existing ones. When these applications ship actions directly into a proprietary system of record (managing the calendar, filing the claim, creating the purchase order, and so on), “done” happens inside the incumbent’s environment. AI becomes another input to an existing workflow rather than a replacement for it. Second, successful incumbents build proprietary data from customer usage. Corrections, edge cases, and approvals become training data that refines the product over time, that a frontier model will not have access to. Cursor, though not an incumbent and despite its dependence on external models, plans to compete by capturing developer behavior patterns as CEO Michael Truell notes in his Stratechery interview : Ben: Is that a real sustainable advantage for you going forward, where you can really dominate the space because you have the usage data, it’s not just calling out to an LLM, that got you started, but now you’re training your own models based on people using Cursor. You started out by having the whole context of the code, which is the first thing you need to do to even accomplish this, but now you have your own data to train on. Michael: Yeah, I think it’s a big advantage, and I think these dynamics of high ceiling, you can kind of pick between products and then this kind of third dynamic of distribution then gets your data, which then helps you make the product better. I think all three of those things were shared by search at the end of the 90s and early 2000s, and so in many ways I think that actually, the competitive dynamics of our market mirror search more than normal enterprise software markets. Both critics and defenders of AI wrappers have a point, and both miss something crucial. The critics are right that some wrappers lack defensibility and will disappear when platforms absorb their features. The defenders are right that every successful software company wraps something. But the real insight lies between these positions. Even if a new application starts as a wrapper, it can endure if it embeds itself in existing workflows, writes to proprietary systems of record, or builds proprietary data and learns from usage. These are the same traits that separate lasting products from fleeting features. Perplexity AI CEO, Aravind Srinivas pushing back on criticism about the business potential of Perplexity: https://medium.com/@alvaro_72265/the-misunderstood-ai-wrapper-opportunity-afabb3c74f31 https://ai.plainenglish.io/wrappers-win-why-your-ai-startup-doesnt-need-to-reinvent-the-wheel-6a6d59d23a9a https://aijourn.com/how-ai-wrappers-are-creating-multi-million-dollar-businesses/ https://growthpartners.online/stories/how-jenni-ai-went-from-0-to-333k-mrr Microsoft bundled Teams into Office 365 subscriptions at no extra cost, using its dominant enterprise distribution to surpass Slack’s paid standalone product within three years despite Slack’s earlier launch and product innovation. See https://venturebeat.com/ai/microsoft-teams-has-13-million-daily-active-users-beating-slack https://a16z.com/revenue-benchmarks-ai-apps/ AI Wrapper Meme: An API call to OpenAI under the hood. In my view, this AI wrapper debate misses a larger point. Wrappers are not all the same. Thin tricks enjoy a brief run and last only until big platforms bundle them into their suites. But products that live where users already work, write back to a proprietary system of record , and/or can make use of proprietary data can endure. The wrapper label is a distraction from what I think actually matters: (1) Is it a feature or a product, and (2) How big is the market segment. Thanks for reading Wreflection! Subscribe for free to receive new posts and support my work. Feature Or Product Begin with the earlier example of a wrapper that lets you chat with a PDF. Such a tool solves one narrow problem - answering questions about a document. It does not create new documents or edit existing ones. It typically does not capture any unique data, or learn from user behavior. It is a means to an end; a capability rather than an end-to-end solution. As a result, this kind of feature belongs inside a document viewer or editor, or in the flagship applications of model providers. So when the foundation models themselves (OpenAI/ChatGPT, Anthropic/Claude, Google/Gemini) bundle this feature natively, the standalone tool becomes redundant. This is classic feature behavior - easy to copy, no end-to-end job, no moat or long-term defensibility. One caveat though; even those that are features can be an interesting indie businesses that make money until the platforms build it into their apps 3 . PDF.ai $500K MRR, PhotoAI $77K MRR, Chatbase $70K MRR, InteriorAI $53K MRR 4 . Jenni AI went from $2,000 to over $333,000 MRR in just 18 months 5 . Cursor went from zero to $100 million in recurring revenue in 18 months, and became the subject of recurring OpenAI acquisition rumors. Windsurf , another coding assistant, received a $2.4B acquisition licensing deal from Google. Gamma reached $50 million in revenue in about a year. Lovable hit $50 million in revenue in just six months. Galileo AI acquired by Google for an undisclosed amount. First, they own the outcome even when they don’t own the model. Applications already embedded in user workflows (Gmail/Calendar, Sheets, EHR/EMR, Figma) require no new habit formation, and building these platforms from scratch is much harder than adding AI capability to existing ones. When these applications ship actions directly into a proprietary system of record (managing the calendar, filing the claim, creating the purchase order, and so on), “done” happens inside the incumbent’s environment. AI becomes another input to an existing workflow rather than a replacement for it. Second, successful incumbents build proprietary data from customer usage. Corrections, edge cases, and approvals become training data that refines the product over time, that a frontier model will not have access to. Cursor, though not an incumbent and despite its dependence on external models, plans to compete by capturing developer behavior patterns as CEO Michael Truell notes in his Stratechery interview : Ben: Is that a real sustainable advantage for you going forward, where you can really dominate the space because you have the usage data, it’s not just calling out to an LLM, that got you started, but now you’re training your own models based on people using Cursor. You started out by having the whole context of the code, which is the first thing you need to do to even accomplish this, but now you have your own data to train on. Michael: Yeah, I think it’s a big advantage, and I think these dynamics of high ceiling, you can kind of pick between products and then this kind of third dynamic of distribution then gets your data, which then helps you make the product better. I think all three of those things were shared by search at the end of the 90s and early 2000s, and so in many ways I think that actually, the competitive dynamics of our market mirror search more than normal enterprise software markets.

api API business Business ai AI
0 views
Wreflection 2 months ago

Platform Pricing Comes At A Cost

“In construction, a platform is something that lifts you up and on which others can stand. The same is true in business. By building a digital platform, other businesses can easily connect their business with yours, build products and services on top of it, and co-create value.” - Harvard Business Review, Three Elements of a Successful Platform Platforms and Marketplaces 1 are one of the defining business models of the internet age. Amazon matches sellers with buyers, Uber connects drivers and riders, Airbnb brings together hosts and guests, and the list goes on. Their strength lies in network effects - more users attract more suppliers, and more suppliers attract ever more users. But network effects alone do not guarantee success. Pricing is just as critical. Venture Capitalist Bill Gurley argued in his article on marketplace dynamics that the take rate or rake a platform takes shapes who joins, who stays, and whether the business endures. Building on that, I classify marketplaces and platforms by their business model into four kinds. Their pricing power—what they can charge, and what they must give up— depends on where they land in this classification. Paid Cross-side Platforms – These platforms create value primarily by enabling direct exchanges between its consumers and producers, and benefit from cross-side network effects, i.e. the volume and nature of merchants attracts users, and more users attract more merchants. Interaction is cross-directional, from sellers to customers and customers to sellers. Customers do not typically interact with other customers (and merchants do not necessarily interact with other merchants). Examples are Airbnb, eBay, Amazon Marketplace, Alibaba’s Tmall, Uber, Lyft etc. I call them paid platforms because they charge merchants a rake (also known as take rate, revenue share, commission, or transaction fees) when they sell to customers. Paid platforms are typically rake-elastic 2 - if fees rise too high, sellers will look for cheaper alternatives. With no hardware or infrastructure lock-in, competitors can lure sellers away by offering lower fees. Managed platforms are an exception. They can sustain higher rakes if customers or merchants see value in the oversight and service they provide. Infra/Hardware Cross-side Platforms – Cross-side platforms or marketplaces built on top of a infrastructure or hardware layer are unique because these platforms do not typically attract users by the volume and nature of the merchants on their platforms; rather the hardware attracts users, and the users then attract merchants. In some cases, unique infrastructure (warehouses, delivery mechanisms, curated storefronts) pulls in merchants on its own. In either case, because demand is captive to a locked-in installed base and/or the platform offers infrastructure that sellers can’t get elsewhere, these platforms are more rake-inelastic , and can charge higher rakes without losing their best suppliers. Examples include Apple Appstore, Xbox Gamestore, Google Playstore. Theoretically, even a physical marketplace in a unique location that allows merchants to sell their wares, and charges them on a revenue share model falls in this category. Fulfillment by Amazon fits here too. Its vast warehouses and logistics network are costly to copy, giving Amazon structural bargaining power to set storage/fulfillment fees. Free Cross-side Platforms – These platforms charge neither consumers nor producers at the point of exchange. Exchanges between their consumers and producers are almost always free. Instead, they turn attention into currency - more users bring more engagement, which advertisers then pay to access. They benefit from cross-side network effects. Google and Yelp are examples of this kind of platform. Pricing power comes not from rake, but from audience quality, auction dynamics, and measurable ROI. Platforms with affluent, high-intent users sell scarcer attention, and if ads convert and targeting is superior, bids rise. Same-side Platforms – These platforms create value by enabling direct exchange among its users, and are monetized either by charging the users directly (rare), or by charging advertisers to sell services to users (most often). Messaging apps, social networks, and communities are examples, and these platforms benefit from same-side effects (more users -> more users) and once there is a valuable installed base, one-way cross-side effects (more users -> more advertisers). But note that more advertisers do not directly mean more customers. In fact, more ads -> bad customer experience -> fewer customers. WhatsApp, Facebook, WeChat, Skype are examples. Some companies sit across multiple models or blend different characteristics but products typically pertain to one model. For example, Google Search is a Free Cross-side Platform for search, but Gmail is a Same-side Platform . YouTube is a Paid Cross-side Platform despite being free to users, as it takes a portion of advertising revenue before paying creators. Based on the above classification and given a platform business model, how much rake can the marketplace bear? Paid Cross-side Platforms are more rake-elastic because a competing platform with lower rake can first attract sellers on fees, and then eventually win customers as sellers pass the savings on as lower prices. That’s why Airbnb must keep its rake around 15%. Because attempting to increase it would invite competitors to undercut on fees, attract its hosts, and ultimately capture its customers with lower prices. Infra/Hardware Platforms get away with higher rake because of their hardware differentiation. That’s why Apple’s Appstore can get away (and they have for 15+ years) with charging developers a high rake (the famous 30% Apple Tax ) because even if a competing platform like Android offered a lower rake (say 5%), users are unlikely to switch as long as Apple’s hardware remains attractive. Here’s a thought exercise - if Apple increased its take rate to 50%, would the quality and quantity of apps materially suffer, or would customers switch platforms to Android? Change is more likely to come from regulators than from competitors or customers. In Free Platforms , prices charged to advertisers don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price (charged to advertisers) of the dominant platform and capture more users. Similar to Free Platforms , prices charged to advertisers in Same-side Platforms don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price of the dominant platform and capture more users. The way to capture marketshare is to offer differentiated service (Facebook vs. Myspace, TikTok vs. Facebook). From a cost point of view, there are three costs at play in a platform business. An upfront capital expense incurred in developing the platform (primarily in software development, in some cases physical assets and hardware costs), fixed operating costs that do not vary with each unit of output (i.e. expenses that are largely constant over a range and only step changes as the platform grows such as office rent, most employee salaries), and variable costs that vary with each unit of output (i.e. costs that vary with the number of sellers and customers serviced such as platform maintenance costs, customer service, payment processing) 3 . A balanced marketplace pricing strategy recognizes four critical factors - the platform’s position on the pricing-power spectrum above, the underlying cost structure, end-customer price elasticity, and access to patient long-term investors. In early stages, especially if access to such long-term investment capital is scarce, a platform may need to charge a rake high enough to account for capital expenses, and fixed and variable costs as a percentage of Gross Merchandise Value (GMV). This ensures sufficient cash flow to keep the business viable. But it will also mean slower seller (and thus customer) acquisition for the platform. Eventually, as upfront investments are depreciated and fixed costs are spread over more customers and orders, capital expense falls as a share of GMV. So, at scale, the optimal rake should sit only a touch above variable costs with a buffer to cover fixed costs. Such a low but sustainable rake attracts sellers and is prohibitively hard for new entrants to match/undercut. Importantly, it keeps prices low for end-customers. Sellers typically pass through fee increases to customers as higher prices, which, if the product is price elastic, suppresses transaction volume. If the profit loss from reduced sales exceeds the profit gain from the higher rake, the rate increase destroys value. Fig 1: Rake, CapEx, VC & FC expressed as a % of GMV but # of Sellers is absolute. Table 1: Rakes must be set based on Variable Costs + Fixed Op Costs buffer. High rakes can be undercut. Target rake ≈ Variable Costs + Residual CapEx. Complementary Services, Concentration, and Contribution Beyond the base take rate, platforms can raise their effective take rates by offering value-added services like payments processing, quality control, insurance, and advertising. These services let suppliers who want more distribution or trust signals pay extra while keeping base prices accessible for others. Consider Amazon’s advertising business, which generated $56B in 2024. Amazon Marketplace’s base rake sits at around 15-20% , but suppliers who pay for prominent search placement raise Amazon’s effective take rate to an estimated 40-50% if you include advertising and other value added services. Tanay Jaipuria, in his newsletter, explores a few more factors that influence take rates, two of which I find particularly compelling: Supplier (and buyer) concentration. Fragmented suppliers have little bargaining power and accept higher rakes; concentrated suppliers can negotiate lower rates or threaten to withhold supply. The same logic applies to buyer concentration, though most consumer marketplaces naturally have fragmented buyers. The best example to highlight this is by digging into Expedia, which sells both flights and hotels. On the flight side, there are typically only ~4 airlines that matter for a given geography, so the supply is very concentrated. Meanwhile, on the hotel side, the supply is much more fragmented. While they don’t breakdown the take rate by segment, most estimates suggest that on the hotel side, their take rate is in the 15-20% range, while on the airlines side it is in the 3-5% range. The concentration is a key driver in this. Think about if Expedia didn’t have United when you performed a search. If you wanted the best price, you could search on Expedia, and then separately search on United before making your purchase. Incremental sales justify higher take rates. Platforms that generate sales that suppliers wouldn’t have captured on their own can charge more for that incremental value. And the more a platform contributes to a transaction, the higher the rake it can justify. One way to see this is to consider the difference between Shopify and Amazon. Shopify provides merchants the tools to set up a store and process transactions but isn’t necessarily bringing them sales and traffic. Meanwhile, Amazon has aggregated 100s of millions of buyers, and sellers lose the ability to reach them on a given search (and potentially make an incremental sale) if they aren’t on Amazon’s third-party marketplace. Shopify’s effective take rate is ~3% while Amazon’s is 10-15% depending on the category of product. In fact, some marketplaces go as far as charging different take rates on transactions depending on their level of contribution to driving the transaction. Take Udemy’s pricing structure, which highlights the importance of distribution: If the course is sold through the instructor’s link → instructor keeps 97% If the course is sold organically on Udemy → instructor keeps 50% If the course is sold through Udemy’s partners’ → instructor keeps 25% Source : Marketplace take rates factors, Tanay Jaipuria, 2021 The chart above raises an obvious question, what’s going on with Shutterstock? Unlike most marketplaces that recognize only their take rate as revenue, Shutterstock reports the full customer payment because it controls the transaction before transferring content. So, per its 2024 financial statements , GMV was approximately reported revenue of $935M. Sales & Marketing (S&M) was $223M (24% of GMV). What this means is that just to cover S&M, Shutterstock needs 24 percentage points of their take rate. Add in other operating costs and their operating profit margin sits at just 7%. In comparison, Airbnb generated $81.8B in GMV, $11.1B in revenue, and only spends $2.1B (2.5% of GMV) in S&M for an operating profit of $2.5B, or 22.5%. Unlike Airbnb with its direct traffic and brand strength, Shutterstock competes in fiercely competitive paid search markets for search terms such as “stock photo for”. This discussion raises a related question about YouTube. How does the platform sustain a 45% take rate ? Alphabet/Google does not break out YouTube’s financials separately ( technically compliant with legal requirements but debatable given YouTube’s scale and distinct creator business model 4 ). This means we cannot know their true cost structure. Assuming YouTube operates more like Airbnb with its 2.6% S&M spend rather than Shutterstock with its 24% paid search dependency, the take rate likely sits well above the optimal rate. The answer is Game theory dynamics . Potential competitors must calculate whether triggering a price war justifies the investment. In YouTube’s case, if a new video platform (say Netflix) launched by offering better rates for creators, YouTube would immediately match it. At equal rates, creators would not move to the new platform since they have already uploaded their entire video portfolios, and built reputations on YouTube. The new entrant would have invested time and capital building the platform for minimal market share gain. From a potential entrant’s perspective, this market is not worth entering. The result is oligopolistic behavior where incumbents charge above-equilibrium rates 5 . The game theory deterrent breaks when competitors exploit technological shifts (AI-generated videos), or pursue different models (shorter video formats like TikTok). More likely, regulators will need to step in when they see platforms earning outsized profits. Even then, meaningful competition takes years to materialize. New entrants must raise capital, build platform technology, recruit sellers, acquire customers, and critically, help sellers build the reputation and reviews that make them credible. For platforms with geographic constraints like Uber or Airbnb, competitors must light up city by city, or region by region. Eventually, once these competitors reach scale, market forces push take rates toward that equilibrium of variable costs plus a sustainable buffer. But until they do, extracting maximum profit through high rakes is rational. In summary, four forces determine take rates in practice. Platform type (including complementary services, concentration, and contribution) define pricing power, cost structure sets the minimum viable take rate, end-customer price elasticity determines how rake changes affect transaction volume and total profit, and game theory and regulatory action governs how long take rates above equilibrium last. Pricing just a bit above variables costs + fixed costs buffer is the sweet spot where the platform balances seller acquisition, low prices for end-consumers, and long-term defensibility 6 . In traditional tech-speak, platforms are those where other software developers can build applications and software on top of the platform application/software. I take a broader view of platforms, one that includes marketplaces for sellers and buyers to connect, or advertisers and customers to interact. And so I use the terms platform and marketplace interchangeably despite the actual difference. I define “Rake elasticity” as the sensitivity of seller participation to changes in a platform’s take rate or commission, with pass-through to end prices; it parallels price elasticity of supply and price elasticity of demand but applies to platform fees rather than to product price, and to seller participation rather than product demand or supply. In other words, the percentage change in seller participation (and eventually GMV) for a 1% percent change in the platform’s take rate. I use variable costs to mean all costs that vary with each transaction. These include performance marketing that scales with sales (search ads tied to conversions, affiliate commissions, sales incentives), even though GAAP classifies these in Sales & Marketing rather than Cost of Revenue. When analyzing platform unit economics from financial statements, you cannot rely solely on the Cost of Revenue line. Shutterstock’s financials illustrate this distinction. Alphabet began disclosing YouTube advertising revenue separately starting in 2020 (it was $36.1B in 2024 of Google’s total $348B) but does not report YouTube as a standalone operating segment with separate costs and profitability. FASB’s ASC 280 requires separate segment reporting when revenue, profit/loss, or assets exceed 10% of consolidated totals. YouTube clearly exceeds the revenue threshold and its creator revenue sharing model operates fundamentally differently from Search or Maps, making a strong case for separate segment treatment. This dynamic does not hold in most traditional businesses because matching prices leads to market share splitting rather than complete incumbent retention. For example, when an incumbent retailer matches a new entrant’s prices, customers simply split between both stores based on convenience and preference. The game theory deterrent only works in businesses with network effects (platforms), high switching costs (telecom), contractual lock-in (medical equipment, enterprise software), or prohibitive infrastructure requirements (airlines, semiconductors). This is an updated version of an essay I first published in 2017 on my now-defunct blog. Paid Cross-side Platforms – These platforms create value primarily by enabling direct exchanges between its consumers and producers, and benefit from cross-side network effects, i.e. the volume and nature of merchants attracts users, and more users attract more merchants. Interaction is cross-directional, from sellers to customers and customers to sellers. Customers do not typically interact with other customers (and merchants do not necessarily interact with other merchants). Examples are Airbnb, eBay, Amazon Marketplace, Alibaba’s Tmall, Uber, Lyft etc. I call them paid platforms because they charge merchants a rake (also known as take rate, revenue share, commission, or transaction fees) when they sell to customers. Paid platforms are typically rake-elastic 2 - if fees rise too high, sellers will look for cheaper alternatives. With no hardware or infrastructure lock-in, competitors can lure sellers away by offering lower fees. Managed platforms are an exception. They can sustain higher rakes if customers or merchants see value in the oversight and service they provide. Infra/Hardware Cross-side Platforms – Cross-side platforms or marketplaces built on top of a infrastructure or hardware layer are unique because these platforms do not typically attract users by the volume and nature of the merchants on their platforms; rather the hardware attracts users, and the users then attract merchants. In some cases, unique infrastructure (warehouses, delivery mechanisms, curated storefronts) pulls in merchants on its own. In either case, because demand is captive to a locked-in installed base and/or the platform offers infrastructure that sellers can’t get elsewhere, these platforms are more rake-inelastic , and can charge higher rakes without losing their best suppliers. Examples include Apple Appstore, Xbox Gamestore, Google Playstore. Theoretically, even a physical marketplace in a unique location that allows merchants to sell their wares, and charges them on a revenue share model falls in this category. Fulfillment by Amazon fits here too. Its vast warehouses and logistics network are costly to copy, giving Amazon structural bargaining power to set storage/fulfillment fees. Free Cross-side Platforms – These platforms charge neither consumers nor producers at the point of exchange. Exchanges between their consumers and producers are almost always free. Instead, they turn attention into currency - more users bring more engagement, which advertisers then pay to access. They benefit from cross-side network effects. Google and Yelp are examples of this kind of platform. Pricing power comes not from rake, but from audience quality, auction dynamics, and measurable ROI. Platforms with affluent, high-intent users sell scarcer attention, and if ads convert and targeting is superior, bids rise. Same-side Platforms – These platforms create value by enabling direct exchange among its users, and are monetized either by charging the users directly (rare), or by charging advertisers to sell services to users (most often). Messaging apps, social networks, and communities are examples, and these platforms benefit from same-side effects (more users -> more users) and once there is a valuable installed base, one-way cross-side effects (more users -> more advertisers). But note that more advertisers do not directly mean more customers. In fact, more ads -> bad customer experience -> fewer customers. WhatsApp, Facebook, WeChat, Skype are examples. Some companies sit across multiple models or blend different characteristics but products typically pertain to one model. For example, Google Search is a Free Cross-side Platform for search, but Gmail is a Same-side Platform . YouTube is a Paid Cross-side Platform despite being free to users, as it takes a portion of advertising revenue before paying creators. Platform Type Determines Rake Elasticity Based on the above classification and given a platform business model, how much rake can the marketplace bear? Paid Cross-side Platforms are more rake-elastic because a competing platform with lower rake can first attract sellers on fees, and then eventually win customers as sellers pass the savings on as lower prices. That’s why Airbnb must keep its rake around 15%. Because attempting to increase it would invite competitors to undercut on fees, attract its hosts, and ultimately capture its customers with lower prices. Infra/Hardware Platforms get away with higher rake because of their hardware differentiation. That’s why Apple’s Appstore can get away (and they have for 15+ years) with charging developers a high rake (the famous 30% Apple Tax ) because even if a competing platform like Android offered a lower rake (say 5%), users are unlikely to switch as long as Apple’s hardware remains attractive. Here’s a thought exercise - if Apple increased its take rate to 50%, would the quality and quantity of apps materially suffer, or would customers switch platforms to Android? Change is more likely to come from regulators than from competitors or customers. In Free Platforms , prices charged to advertisers don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price (charged to advertisers) of the dominant platform and capture more users. Similar to Free Platforms , prices charged to advertisers in Same-side Platforms don’t (directly) affect the prices on customers, so theoretically new platforms cannot undercut the price of the dominant platform and capture more users. The way to capture marketshare is to offer differentiated service (Facebook vs. Myspace, TikTok vs. Facebook). From a cost point of view, there are three costs at play in a platform business. An upfront capital expense incurred in developing the platform (primarily in software development, in some cases physical assets and hardware costs), fixed operating costs that do not vary with each unit of output (i.e. expenses that are largely constant over a range and only step changes as the platform grows such as office rent, most employee salaries), and variable costs that vary with each unit of output (i.e. costs that vary with the number of sellers and customers serviced such as platform maintenance costs, customer service, payment processing) 3 . A balanced marketplace pricing strategy recognizes four critical factors - the platform’s position on the pricing-power spectrum above, the underlying cost structure, end-customer price elasticity, and access to patient long-term investors. In early stages, especially if access to such long-term investment capital is scarce, a platform may need to charge a rake high enough to account for capital expenses, and fixed and variable costs as a percentage of Gross Merchandise Value (GMV). This ensures sufficient cash flow to keep the business viable. But it will also mean slower seller (and thus customer) acquisition for the platform. Eventually, as upfront investments are depreciated and fixed costs are spread over more customers and orders, capital expense falls as a share of GMV. So, at scale, the optimal rake should sit only a touch above variable costs with a buffer to cover fixed costs. Such a low but sustainable rake attracts sellers and is prohibitively hard for new entrants to match/undercut. Importantly, it keeps prices low for end-customers. Sellers typically pass through fee increases to customers as higher prices, which, if the product is price elastic, suppresses transaction volume. If the profit loss from reduced sales exceeds the profit gain from the higher rake, the rate increase destroys value. Fig 1: Rake, CapEx, VC & FC expressed as a % of GMV but # of Sellers is absolute. Table 1: Rakes must be set based on Variable Costs + Fixed Op Costs buffer. High rakes can be undercut. Target rake ≈ Variable Costs + Residual CapEx. Complementary Services, Concentration, and Contribution Beyond the base take rate, platforms can raise their effective take rates by offering value-added services like payments processing, quality control, insurance, and advertising. These services let suppliers who want more distribution or trust signals pay extra while keeping base prices accessible for others. Consider Amazon’s advertising business, which generated $56B in 2024. Amazon Marketplace’s base rake sits at around 15-20% , but suppliers who pay for prominent search placement raise Amazon’s effective take rate to an estimated 40-50% if you include advertising and other value added services. Tanay Jaipuria, in his newsletter, explores a few more factors that influence take rates, two of which I find particularly compelling: Supplier (and buyer) concentration. Fragmented suppliers have little bargaining power and accept higher rakes; concentrated suppliers can negotiate lower rates or threaten to withhold supply. The same logic applies to buyer concentration, though most consumer marketplaces naturally have fragmented buyers. The best example to highlight this is by digging into Expedia, which sells both flights and hotels. On the flight side, there are typically only ~4 airlines that matter for a given geography, so the supply is very concentrated. Meanwhile, on the hotel side, the supply is much more fragmented. While they don’t breakdown the take rate by segment, most estimates suggest that on the hotel side, their take rate is in the 15-20% range, while on the airlines side it is in the 3-5% range. The concentration is a key driver in this. Think about if Expedia didn’t have United when you performed a search. If you wanted the best price, you could search on Expedia, and then separately search on United before making your purchase. Incremental sales justify higher take rates. Platforms that generate sales that suppliers wouldn’t have captured on their own can charge more for that incremental value. And the more a platform contributes to a transaction, the higher the rake it can justify. One way to see this is to consider the difference between Shopify and Amazon. Shopify provides merchants the tools to set up a store and process transactions but isn’t necessarily bringing them sales and traffic. Meanwhile, Amazon has aggregated 100s of millions of buyers, and sellers lose the ability to reach them on a given search (and potentially make an incremental sale) if they aren’t on Amazon’s third-party marketplace. Shopify’s effective take rate is ~3% while Amazon’s is 10-15% depending on the category of product. In fact, some marketplaces go as far as charging different take rates on transactions depending on their level of contribution to driving the transaction. Take Udemy’s pricing structure, which highlights the importance of distribution: If the course is sold through the instructor’s link → instructor keeps 97% If the course is sold organically on Udemy → instructor keeps 50% If the course is sold through Udemy’s partners’ → instructor keeps 25%

business Business
0 views
Wreflection 2 months ago

Hello World!

Welcome to Wreflection! “The act of writing is the act of discovering what you believe.” This quote captures the essence of why I am starting this. Through my writing, I seek to discover and validate (or invalidate) my beliefs, and share what I discover. An experiment in thinking out loud. You can expect to find theories, frameworks, analysis, and opinion on business, tech, and the business of tech, primarily around four areas: Business Strategy - Why do some win while others fail. What trade-offs shape product, pricing, and market dynamics? Product Lessons - Stories from past experiences building and scaling products. Lessons from leading tech teams. Recent experiments, successes, and failures. What works, what doesn't? Tech Analysis - Making sense of Artificial Intelligence, Platforms, Marketplaces, and whatever's changing our world, and specifically what it means for business. Personal reflection - Occasional essays about my life and other interests. I am an electrical engineer by training, with a masters in business. I most recently spent 10 years at Amazon building warehouse robots, and launching new categories and products. Because I spent 10 years of my career there, I expect to write a lot about my views on Amazon; but all are my personal views and does not represent the views of Amazon or its management. Data and information will never represent confidential Amazon information. I look forward to making this worth your time. Subscribe now Business Strategy - Why do some win while others fail. What trade-offs shape product, pricing, and market dynamics? Product Lessons - Stories from past experiences building and scaling products. Lessons from leading tech teams. Recent experiments, successes, and failures. What works, what doesn't? Tech Analysis - Making sense of Artificial Intelligence, Platforms, Marketplaces, and whatever's changing our world, and specifically what it means for business. Personal reflection - Occasional essays about my life and other interests.

business Business Tech Analysis
0 views