Back to all posts
EssaysJune 15, 202610 min read

A Golden Age of AI

For 72 hours the general public had a glimpse of the future with Fable 5. Highly anticipated and a step change in capability. Then it got pulled. What happens now? Will we get Fable back? When? Whatever the answer, it's clear that we are at a fork in the road. We reached the end of a distinct period of time in AI. From the time Opus 4.5 was released to the time Fable was pulled: I'll call this Act 2 of the Golden Age.

Act 1: November 2022 - November 2025

A clean 3 year stretch. ChatGPT entered public consciousness. Conversational. Text in, text out. RAG, vibe coding, "The hottest new programming language is English," What did Ilya see? etc. Fastest app to 100 million users. Shades of agentic AI. Useful in narrow applications.

Act 2: November 24, 2025 - June 12, 2026

Anthropic released Claude Opus 4.5. People went home for Thanksgiving and winter breaks and built things. Real things. Posted about it online. AI just started working, as advertised. Many claimed an inflection point. The frontier crossed into general applications and reliability. The hype was on! Then software 3.0, ralph loops, agents, OpenClaw, Codex, Hermes, sold out Mac Minis, Mythos rumors, etc. Remember Moltbook? That was fun. This was the last time many of you wrote a single line of code. Prompting became prompt engineering, which became context engineering, which became loop engineering.

This was an incredibly interesting time. Building things was fun, sleep was at a premium, and we embraced the AI psychosis. Busier than ever and having an incredibly fun time exploring the tech. Learning. Building. I know others feel the same. Why did it feel this way? This act was made possible by forces distinct to this time period:

  • The best state of the art models were broadly released to the public. Opus 4.5, 1M-token context, Gemini 3.1 Pro, GPT-5.5, Fable 5.
  • The world got hooked on cheap tokens, which were heavily subsidized through subscription plans. The $200/month plans' API equivalent values at full usage were reported to be $8,000 for Claude Max 20x and $14,000 for ChatGPT Pro 20x, based on SemiAnalysis work published June 10, 2026.
  • Just being early was a huge advantage. The information asymmetry between people who understood what modern AI was capable of vs those who did not was massive. The asymmetry between the AI pilled and everyone else was at its widest.
  • New tooling weekly, a real share of it from individual builders shipping open source. New startups, tools, and launches every week.

Then for 72 hours we popped our heads above the clouds, like Neo and Trinity heading to the machine city. We saw the jump in capability. What Anthropic was using internally for months.

Neo and Trinity rise above the clouds toward the Machine City
The Matrix Revolutions (2003) — the approach to the Machine City

This short window, in my opinion, and depending on what happens next, was the pinnacle of the Golden Age of AI.

Then it was gone. Pulled on June 12, 2026, as a result of the US government's export control directive stemming from Amazon reporting an alleged jailbreak of Fable 5 to the government. Anthropic disputes that it was a real jailbreak, calling it a narrow finding and a "misunderstanding" that is replicable with other models, including GPT-5.5. The irony is that Amazon is Anthropic's largest investor. Many lamented going back to the "old" models and weighing whether to stop work until Fable 5 returned. Those 72 hours made it clear how much better it was.

Act 3: June 13, 2026 - Where to from here?

No one knows. There are branching possibilities.

Acts 1 and 2 were about awareness and acceleration. How you use AI, which tools, what capabilities it unlocks for you, etc. Act 3 is about scale, sovereignty, and ACCESS.

Here are the forces that will shape Act 3 with a few predictions sprinkled in.

1. A Manhattan Project per month

The industrial AI buildout will continue at scale. One month of US data center construction now costs more than the entire Manhattan Project did, adjusted for inflation. ~$725B from five US hyperscalers in 2026 (more than Switzerland's GDP), over $1T in global data center capex, Goldman's ~$7.6T from 2026 to 2031, ~100GW of new capacity by 2030 (doubling the world's). Plans for data centers in space. The scale is hard to picture.

2. Geopolitics, regulation, and public markets

AI is too big to fail, too powerful to ignore, and too important to lose.

OpenAI's confidential IPO filing, with Anthropic close behind. $1.5T of new debt to finance the buildout. Nvidia books $81.6B per quarter, and AI-related investment has become a significant driver of US GDP growth. One sector holding up the economy.

Fable got banned. Frontier lab CEOs sitting at the negotiating table with G7 leaders. AI has hit the world stage. Every election from now on will have AI talking points. Every new capability now draws a regulator.

The frontier is a front in great-power competition. The spend is lopsided: $725B from five US firms against roughly $125B from China and €200B from the EU. Chips are the choke point, and rationing is the next phase. All of it runs through Taiwan.

3. The Cantillon effect for tokens

Access to frontier AI will be one of the most important things in the near future. On par with access to electricity, running water, and medical services. Gating access carries vast power.

The traditional Cantillon effect describes how when new money is created, it does not reach everyone at once. Whoever is closest to where it enters (the money printer) spends it first, before prices rise. Whoever is farthest gets it last, after everything has already gotten more expensive.

The Cantillon effect for tokens (Tokillon effect?) is when the best model is rationed, access forms rings, and nearness to the source becomes its own advantage.

The rings are already here, drawn by the product strategy. The newest model lands on paid tiers first. Free users wait. The best model costs the most. Then the lines get drawn by something other than price.

  • The center is the lab itself, using the frontier internally to run the business and train the next model. This is one lens through which to look at Steinberger to OpenAI and Karpathy to Anthropic. Unlimited token budgets and unreleased models. Closer to the source.
  • One ring out, Project Glasswing, 200 "trusted" organizations with early access.
  • Another ring out, US citizens and cleared domestic users (KYC), who get the frontier back first while everyone else waits.
  • Another ring out, allied nations under deals and treaties. At the edge, the rest of the world runs last year's model, or an open one a few months behind. The advantage fades the farther out you sit.

The lines are drawn by who you are and where you stand: your citizenship, your clearance, whose ally you are. Proximity becomes political, not financial. The Dark Mirror version of this idea adds rings for who you voted for, criticism of the government or AI lab, or donations to a political organization.

What erodes these rings is open models keeping pace with the frontier. Not catching up once, but staying close. If they do, the rings fade. If they fall behind, the rings harden.

4. Costs split in two

Frontier intelligence becomes expensive and gated. Everything else becomes cheap, fast, open source. Yet the market for frontier intelligence will outpace everything else by a wide margin. When the stakes are high enough, the best model is worth almost any price. You can't afford to be a few points of capability below your competitors.

5. Open source/Open weight models

If you don't own your AI stack, you're renting. Which means you can lose access at the whims of corporations, regulators, and governments.

The hardware to own it is getting harder to buy. The $599 Mac Mini is gone. The 512GB Mac Studio is gone. Both pulled this spring as AI data centers bought up the world's memory and roughly doubled the price of the chips those machines need.

Today, you can run your own AI. Privately, securely, with no subscriptions or API costs...

  • $700 - A base Mac Mini can run Qwen3 8B or Gemma 4, which is roughly equivalent to GPT-4o from 2024. Two to three steps behind the frontier.
  • $2,000 - A base Mac Studio can run Qwen3.6, a 35-billion-parameter model, roughly equivalent to the GPT-5 or Gemini 2.5 Pro tier from 2025. One to two steps behind the frontier.
  • $20k to $40k - Four Mac Studios wired together can run DeepSeek 671B and Kimi K2, trillion-parameter models, that rival GPT-5.5 and Opus on some benchmarks. The frontier from a few months ago.
  • $50k+ - A rack of GPUs runs the largest models fast. Unlike Macs, they can train models, not just run them. This is an actual data center in your house, and the point where you stop renting entirely.

Remember that frontier lab monthly subscriptions are highly subsidized. What happens when those subsidies get eroded away or disappear?

AI sovereignty is owning the hardware, compute, and model weights. This will be a developing theme in Act 3. Does open source proliferate, or get caught in the crosshairs of geopolitics and regulation?

Act 4: Possible futures

The outcomes of Act 3 will pave the way for the future of human civilization. We can reach for science fiction analogues. Which way?

  • A Star Trek future: Abundance and prosperity for human civilization. AI is used to its full potential to discover new science, technology, and physics. We can cure disease and travel the stars.
  • A Jetsons/Elysium future: A bifurcation of society. A small number with the capital, advanced AI, robots, medical and space technology break away while the 99% are left to fight for scraps. $10 per token for ASI. In the Jetsons series, they never show you the ground.
  • A Star Wars future: Monolithic forces determine the outcome of events. Advancements in science and technology are privatized. No open source. Umbrella Corp. Dystopia.

Human civilization and societal structures are only just beginning to grapple with AI, which is moving much faster than our institutions can keep up. We are the frog in the pot. How fast we normalized all of this.

Act 3 will be played out on the geopolitical theater, public markets, and industrial scale buildout of the infrastructure. Act 4 determines the direction of human civilization.

Which way?

Sources and further reading

If you're trying to figure out where your business sits in all this —
and what to actually do about it —
let's talk —

Related services

Turn the idea into shipped AI work.

AI strategy for fast-moving marketsDecide what to build now while AI capability, cost, and access keep shifting.AI training workshopsBuild team capability around the tools and workflows that are already here.

Build what's next.

Want help shipping the next one?

Let's talk