The AI Economy

What in the world is AI?

AI is everywhere. It’s plastered on billboards, listening to meetings, and even scattered in songs. And yet, what is it exactly? 96% of American adults are now aware of AI, but most still consider it to be things like chatbots, pictures and videos, or the work of science fiction.^¹ Have they heard of agentic systems, multimodal models, or tokenmaxxing? Probably not.

AI is accelerating. Whatever you know, and whether you like it or not, AI will stay in our lives. It’s been around for some time, we just didn’t have a common word for it.^[² Today, it feels like things are moving fast. Maybe too fast.³⁴

AI is expensive. To fund this growth, companies have committed over a trillion dollars in investment.⁵ And now it’s no longer just investors footing most of the bill. Last week, after negotiating with the US government on export controls, Anthropic re-released its most powerful and priciest models yet, Fable 5 and Mythos 5. Users on Reddit are already estimating it will cost them hundreds to thousands of dollars a week to run Fable.⁶

AI is exhausting. Keeping up with these advancements has been hard for many workers, business owners, and the media alike.⁷ Even for everyday use, it can be tiring. A typical model generates text at a rate 10 to 20 times faster than most people can actually read.⁸

AI is polarizing. Is it making us smarter? Dumber? Are we talking about it too much? Not enough? You might be thinking, Another piece about AI? Give me a break.

But still…AI is important. Technology, at its core, is simply a way of doing things. In that sense, AI is one way of doing things, one that is dramatically, slowly, quickly changing how we do lots of other things. How we choose where to eat, what to work towards, who to love. It’s been a few years since OpenAI’s GPT 3.5 made chatbots popular among consumers and seven months since Anthropic’s Opus 4.5 changed the way enterprises use agents. While it might feel like we’re in a new normal, we are still so early to understanding the capabilities and impacts of AI. A month ago, even Sergey Brin, the co-founder of Google, said he is still confused on how to prompt AI models well and that their team does not know the full potential of Gemini.⁹ Capital markets have been remarkably uncertain.^[¹⁰^] Regulators have both pushed on the gas and pulled the brakes.^[¹¹^] The most sophisticated investors, economists, and policymakers are all still debating: Are we building too much supply? How much demand is there, really? What is the impact on jobs? What does this mean for me? For society? For the future?

In this ongoing chaos, several simple questions remain:

1. What is the value of AI?

2. What is the cost of AI?

3. What are the limits?

4. Who decides?

It may seem like some people have all the answers, or at least are using models that have quite a lot of them. But I don’t think anyone, in any sector, actually knows how far ahead or behind they are. Instead of just bulldozing to superintelligence or cringing at the sound of more data centers, we owe it to ourselves to think about it. While the inner workings of AI might still be a blackbox^[¹²^], the ecosystem of resources around it doesn’t need to be a mystery. This project is my effort to spell this out in plain English.

Here, I will be making an atlas with simple maps of the AI economy. I’ll trace how we go from sand to silicon, to bits to tokens, to knowledge, decisions, and outcomes. Over time, I’ll write about how we discover the value of these outcomes, if any at all. And perhaps most importantly, I’ll explore how these shape the way we think, feel, and act. I don’t claim to be an expert, which is precisely why I want to learn more. And the real secret is: no one truly knows where the world is going. But we can try to see the present from different lenses and form our own views. Of course, every map is biased. A map can’t contain everything, otherwise it would be the whole world. So to choose what to cover, I’ll follow both my personal interests and the requests of readers. If you’re also wondering how to make sense of this world, ask me a question below. And know that you are talking to a human.

---

Note: AI was used to research facts, compile citations, create this website, and publish this post. AI was not used to outline, write, or edit the words above in this essay.

1. In a Pew Research Center survey of 5,119 U.S. adults conducted between February 17-23, 2026, 96% said they know at least a little about AI. When asked what technology first comes to mind when they think of AI, the most common answer was chatbots. About three-in-ten named chatbots ahead of robots, images and videos, generative AI and LLMs, the internet at large, and search engines. About half of adults now use AI chatbots, up from a third in 2024. Pew Research Center, “Americans and AI 2026: Chatbots, Smart Devices and Views on Impact,” June 17, 2026, https://www.pewresearch.org/internet/2026/06/17/what-do-americans-think-ai-is/ (accessed July 6, 2026). See also The National Desk, https://thenationaldesk.com/news/americas-news-now/americans-embrace-ai-but-with-skeptical-eye-pew-survey-shows.

2. The phrase ‘artificial intelligence’ first appears in the August 31, 1955 proposal for a Dartmouth summer workshop the following year, written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. AI was in daily life well before chatbots gave it a name (or a name other than Siri and Alexa). In a December 2022 Pew survey, only about three-in-ten U.S. adults could correctly identify all six common AI uses asked about, like product recommendations and email spam filters. Stanford University (John McCarthy’s pages), “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence,” August 31, 1955, https://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html (accessed July 6, 2026). See also Pew Research Center, https://www.pewresearch.org/science/2023/02/15/public-awareness-of-artificial-intelligence-in-everyday-activities/.

3. In the same February 2026 Pew survey from footnote 1, 63% of U.S. adults said AI is advancing too fast. Only 2% said too slowly, and about one-in-five said the pace is about right. Majorities of every age group said too fast. Pew Research Center, “Americans’ Views on AI Chatbots, Smart Devices and AI’s Impact,” June 17, 2026, https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/ (accessed July 6, 2026). See also The National Desk, https://thenationaldesk.com/news/americas-news-now/americans-embrace-ai-but-with-skeptical-eye-pew-survey-shows.

4. In ‘When AI builds itself,’ Anthropic’s Marina Favaro and Jack Clark wrote it would be good to have ‘the option to slow or temporarily pause frontier AI development’ and that Anthropic ‘would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner.’ The post ran three days after Anthropic’s confidential IPO filing. Anthropic, “When AI builds itself,” June 4, 2026, https://www.anthropic.com/institute/recursive-self-improvement (accessed July 5, 2026).

5. The Bank for International Settlements counts the five largest hyperscalers (Alphabet, Amazon, Meta, Microsoft, and Oracle) as set to spend over a trillion US dollars on AI-related capital expenditure across 2025 and 2026, outpacing their earnings and free cash flow. The model builders spend on top of that. OpenAI projects $50 billion on computing power in 2026 alone, per co-founder Greg Brockman’s court testimony, and Anthropic has committed $50 billion to American AI infrastructure along with multiple gigawatts of new compute with Google and Broadcom (Anthropic, April 6, 2026, https://www.anthropic.com/news/google-broadcom-partnership-compute). Bank for International Settlements, “Annual Economic Report 2026, Chapter I: Progress and Peril,” June 28, 2026, https://www.bis.org/publ/arpdf/ar2026e1.htm (accessed July 6, 2026). See also Reuters, https://www.reuters.com/technology/openai-projects-50-billion-spending-computing-power-this-year-brockman-says-2026-05-05/.

6. As of July 6, 2026, Anthropic’s pricing page lists Fable 5 and Mythos 5 at $10 per million input tokens and $50 per million output tokens, double Opus 4.8’s $5/$25 (the earlier Opus 4 and 4.1 are now listed at $15/$75). The models are included in paid Claude plans only through July 7, 2026, then are expected to be billed at those rates. Costs scale with agentic use. A Stanford-affiliated study (arXiv 2604.22750, April 2026) measured agentic coding at 4.17 million tokens per task on average, about a thousand times a typical chat query, across eight models. User estimates vary widely with usage. See the discussion at https://www.reddit.com/r/Anthropic/comments/1uom6kn/fable_5_api_costs/ and PCWorld’s coverage of early cost reports, July 1, 2026. Anthropic, “Pricing; Claude Platform Docs,” July 6, 2026, https://platform.claude.com/docs/en/about-claude/pricing (accessed July 6, 2026). See also Gizmodo, https://gizmodo.com/how-much-does-it-really-cost-to-use-claude-fable-gpt-5-5-and-gemini-3-5-flash-2000770837.

7. Microsoft’s 2026 Work Trend Index surveyed 20,000 knowledge workers in 10 countries from February to April 2026, finding that 65% of AI users fear falling behind if they do not adapt quickly. Newsrooms reported the same strain. In the Reuters Institute’s 2026 survey of 280 senior news executives and editors across 51 countries, fewer than four in ten said they are confident about journalism’s prospects this year, with the pace of AI-driven change described as difficult to predict. Microsoft WorkLab, “2026 Work Trend Index: Agents, human agency, and the opportunity for every organization,” May 19, 2026, https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization (accessed July 6, 2026). See also Reuters Institute for the Study of Journalism, https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026.

8. Frontier chat models serve output at roughly 45–175 tokens per second (Artificial Analysis and llm-stats benchmarks, July 2026). At the standard conversion of one token to about three-quarters of an English word, that is roughly 2,000 to 7,900 words per minute, against an average adult silent reading rate of 238 words per minute for non-fiction (Brysbaert’s 2019 meta-analysis of 190 studies, Journal of Memory and Language). 10x is the conservative end of this range. Typical flagship models run 14–20x faster. Artificial Analysis, “AI model output-speed benchmarks (GPT-5.5, Claude Opus, Gemini 3.1),” July 6, 2026, https://artificialanalysis.ai/models (accessed July 6, 2026). See also llm-stats.com (speeds); Brysbaert 2019, J. Memory & Language (reading rate), https://www.sciencedirect.com/science/article/abs/pii/S0749596X19300786.

9. At an unscripted AGI House fireside for Google DeepMind’s Build Day in June 2026, Google co-founder Sergey Brin said he is still figuring out how to prompt the models well and that the team does not know the full potential of Gemini, telling the room they had probably spent more time with the new model than he had. AGI House x Google DeepMind (fireside video), “Build Day fireside with Sergey Brin,” June 5, 2026, (accessed July 6, 2026). See also Search Engine Journal, https://www.searchenginejournal.com/googles-sergey-brin-sees-a-path-to-agi/578068/.

10. The count uses consecutive trading days where major press coverage attributed the dominant market narrative to AI, with a move of at least 2% in the Nasdaq or the SOX chip index or at least 5% in a mega-cap AI name. A rebound day ends an episode. Tariff-driven days are excluded. Dropping the mixed-driver June 9 episode makes it 6 selloffs. DeepSeek (January 2025) predates the window. Author’s analysis, July 5, 2026

11. The White House, “Fact Sheet: President Donald J. Trump Signs Historic Directive on AI in the National Security Enterprise,” June 2026, https://www.whitehouse.gov/fact-sheets/2026/06/fact-sheet-president-donald-j-trump-signs-historic-directive-on-ai-in-the-national-security-enterprise/ (accessed July 6, 2026).

12. Most AI models are black boxes whose inner workings even their builders cannot fully understand. A research field called interpretability aims to understand this phenomenon. The radiology paper cited here argues there is hope of understanding even the darkest black boxes, and works out how, why, and when to explain them in settings where explanations carry weight, like medicine. European Journal of Radiology (ScienceDirect), “Artificial intelligence and explanation: How, why, and when to explain black boxes,” 2024, https://www.sciencedirect.com/science/article/pii/S0720048X24001098 (accessed July 6, 2026). See also The New York Times, https://www.nytimes.com/2026/04/15/magazine/ai-black-box-interpretability-research.html.