The question worth asking in June 2026 is not whether open source AI is sustainable. That question is too coarse to answer. The sharper one is: is open source dying at the top frontier of AI — and if it is, under what conditions does that hold, and what would break the bet?

DeepSeek V4 shipped in April 2026 under MIT — 1.6 trillion total parameters, one-million-token context, at one-thirtieth the cost of Claude Opus 4.7 and GPT-5.5. It is the closest open weights have ever come to the frontier. It is also, on the hardest benchmarks, three to six months behind. The gap has been stable for two years. By some measures it is widening.

Two readings of that fact are possible. The first: open source is converging on the frontier, just slowly. The second: open source has reached an asymptote, and the gap from here is structural, not temporal. Both readings are defensible from the data alone. Which one is right depends on conditions the data cannot answer — conditions about capital, governance, the architecture of the safety stack, the mechanics of how open source has been catching up, and the strategic posture of Chinese labs once they reach the frontier.

The argument of this piece is conditional, not declarative. If four gates hold — and if China’s open-weights moment turns out to be strategic rather than structural — open source is dying at the top frontier. If any one of these breaks, the bet breaks. Naming the gates, naming what would break them, and naming the Chinese question is the work.

Stay Ahed of The AI Supercycle →

Read more

What makes this moment unique is that it is often misunderstood as a story about automation. In reality, it is increasingly a story about augmentation.

The emerging model is not AI replacing humans. It is humans equipped with AI, and increasingly with fleets of AI agents, becoming capable of orchestrating work that …

Read more

In that regard, Jensen Huang’s ability to zoom out, connect the dots, and articulate the broader transformation underway is truly exceptional.

I highly recommend watching the talk below, along with the accompanying article it inspired.

The piece builds on Huang’s perspective while incorporating several frameworks and mental models I’ve been developing to better understand and navigate the AI Supercycle.

Every industrial age is defined by a machine that consumes one substrate and emits another. The substrate that gets emitted — not the machine that emits it — is what reshapes the economy. Three hundred years of industrial history compress into three of these conversions, and the third one is being installed in real time, in concrete and copper, on every continent simultaneously.

The machine is the dynamo. The substrate is intelligence. The cocoon is forming now.

Stay Ahed of The AI Supercycle →

The Substrate Doctrine — Atoms, Electrons, Tokens

Stay Ahed of The AI Supercycle →

A substrate becomes infrastructure when three conditions hold. It must be generable on demand (a machine can produce it). It must be transportable at near-zero marginal cost (a wire, a fiber, a rack). It must be universally consumable (everything plugs in). Three substrates have ever passed all three tests at planetary scale.

Electrons. A coil rotated through a magnetic field emits an invisible force. Within a century, that force cocoons every continent in a grid. Total addressable market: every joule of useful work humanity wants done.

Bits. Electrons routed through switches at near-luminal speed emit packets. Within three decades, those packets cocoon the planet in a network. Total addressable market: every act of human coordination.

Tokens. Electrons fed into massively parallel arithmetic emit numbers that encode meaning — text, code, images, video, protein structures, control signals, decisions. The substrate is intelligence. The economy is reorganizing for the third time. Total addressable market: every act of cognition humanity wants performed, plus the cognition of the machines themselves.

The implication of the lineage is not that AI is “like” electricity. The implication is that AI is on the same curve as electricity — same generational capital requirements, same regulatory eventualities, same commodity endpoint, same hundred-year reorganization of everything downstream. We are not in the chatbot phase. We are in the dynamo phase. The chatbot is one light bulb.

A second, deeper claim sits underneath the lineage: anything with structure is convertible to tokens. Proteins fold predictably. Cells behave predictably. Genes express predictably. Physics is predictable. Materials are predictable. The token is therefore not just a unit of language — it is the universal carrier of any learnable structure in the universe. The substrate is wider than language by orders of magnitude.

The Primitive Inversion — From Retrieval to Generation

Stay Ahed of The AI Supercycle →

For sixty years, computing was a retrieval architecture. You wrote, you stored, you fetched. Every screen ever looked at served something pre-recorded by someone, somewhere, earlier. The most valuable real estate of the prior internet was the index — the catalog of what had already been made.

That primitive has inverted. Output is now produced originally, in real time, contextually, for one consumer, once. Every pixel, every word, every frame, every recommendation — generated, not retrieved. This is not a refinement of the old computing. It is a different machine emitting a different substrate. Hence the rename: data centers do not produce intelligence; AI factories do.

The unit economics of an AI factory are unlike anything in software history. One rack carries seventy-two accelerators, weighs two tons, costs around four million dollars, and contains roughly one and a half million parts. One gigawatt of factory capacity costs around fifty billion dollars to build and emits three hundred to four hundred billion dollars of token output per year — a revenue-to-capex multiple of six to eight times at the asset level. The total capital flowing into the stack in 2026 is on the order of one trillion dollars, against an implied steady-state ecosystem of roughly twenty trillion dollars per year. This is not a software business — it is a fab. This is the unit economic that breaks every capex-normalization forecast in the market. Forecasters are pricing buildings; operators are pricing production lines.

The second-order consequence is retrieval decay. Every business priced on retrieval economics — search advertising, content aggregation, social feeds, recommendation networks — is in slow runoff. Retrieval businesses paid pennies per query because they re-served stored content with marginal cost approaching zero. Generation businesses meter compute per query because each output is originally produced. The cost curve is different. The pricing model is different. The defensibility is different.

The third-order consequence is the agentic demand shock. Today, intelligence is consumed by roughly one billion humans through chat interfaces. In the agentic phase — already underway — intelligence will be consumed by roughly 100 billion agents running continuously, talking to each other, executing tasks, calling tools, and coordinating workflows. That is a hundredfold demand step change layered on top of the existing consumer demand. The factories under construction are not overbuilt. They are underbuilt relative to a demand curve that has not yet materialized.

The Nine-Layer Cake — A Deeper Reading

Stay Ahed of The AI Supercycle →

The standard investor framing of the AI stack has five layers — energy, compute, infrastructure, models, applications. It is a useful public-facing compression, but it hides at least four separations that have become structurally consequential. A more accurate reading of the stack has nine layers.

The five-layer framing collapses what are now separate profit pools. Networking has separated from compute — rack-scale interconnects consume more than twenty percent of system cost and are growing roughly two hundred percent year-on-year, with one company holding monopoly economics on the interconnect.

Foundries have separated from silicon — allocation power at the leading-edge fab and the high-bandwidth memory oligopoly is now more consequential than chip design itself.

Governance has separated from generic policy and trust — it is now a priced variable in enterprise procurement and a binding constraint on cross-border deployment. Silicon itself has bifurcated into a generalist substrate, hyperscaler custom silicon, and merchant custom challengers. Each of these separations is its own market with its own incumbents, moat, and bottleneck.

Bottom-to-top, the nine layers are:

1 · Energy & Physical. Generation, transmission, water, cooling, land, permits. The first serious grid investment cycle in roughly a century. Power is the binding constraint for everything above it. Rare-earth supply, transformer steel, and grid-scale storage chemistry are the physical chokepoints that govern the pace of the entire stack. Capital intensity is extreme; margin is regulated; the moat is physical and political.

2 · Foundries & Packaging. The fab and the package — the physical conversion of designs into silicon. Allocation power at the leading-edge foundry and the HBM oligopoly is now a more consequential bottleneck than chip design. Advanced packaging (CoWoS and equivalents) is the binding constraint on supply for the next two generations of accelerators. Capital intensity is the highest in the entire economy. Margin is high at the frontier and structurally protected by twenty-plus-year capability gaps to the next-best competitor.

3 · Silicon. Chip design itself — generalist accelerators, hyperscaler custom silicon (custom TPUs, training accelerators, inference accelerators), and merchant custom challengers. The layer has bifurcated into three structurally different positions: the dominant generalist who sets the software stack, the captive customs who optimize for one workload, and the merchant customs who specialize on inference or specific architectures. Margin concentrates at the generalist; volume disperses to the customs.

4 · Networking & Protocols. Rack-scale interconnect, optical fabric, switching silicon, and the protocols stitching factories together. This layer carries monopoly economics on the leading edge — the interconnect is the most under-priced asset in the public market. Networking is now growing faster than compute itself because every additional accelerator demands a non-linear increase in interconnect bandwidth.

5 · Compute Capacity. The actual deployable rack-and-power-contract bundle — what hyperscalers and neoclouds sell. The difference between owning silicon and being able to deploy it at scale is now its own market. Capital intensity is extreme. Margin is high while scarcity holds and will normalize toward utility-like over a decade as the physical layer catches up.

6 · Foundation Models. The visible language models cover one slice. The industrial layer is everything else — protein, cell, genome, materials, physics, climate, robotics control, autonomous driving, code, design, audio. Anything with structure is learnable; anything learnable becomes a model; every model is a vertical market. The layer is shaped by three structural philosophies: maximalist capability frontiers (capability is the moat), trust and safety frontiers (refusal and reliability are the moat), and integrated platform frontiers (distribution into existing ecosystems is the moat). These are not interchangeable strategies — they produce structurally different products and command different multiples. The safety-positioned frontier in particular commands a refusal premium — enterprise buyers pay materially more for a model whose refusal behavior they can underwrite.

7 · Agentic Harness. The orchestration substrate that turns model output into operational output — tool calling, memory, planning, retries, evaluation, guardrails, sandboxing, multi-agent coordination. This is the layer that most application-layer companies confuse with their product. It is in fact a horizontal infrastructure layer in its own right, comparable to the cloud-native infrastructure layer of the prior decade. The companies that own the harness — and the persistent state that agents accumulate inside it — will become the next generation of infrastructure incumbents.

8 · Distribution. Applications, surfaces, and channels — the layer that touches the user. $100B of venture capital deployed in 2025 alone, the largest year in the history of the asset class. It is also the most fragmented layer and the one with the highest failure rate. Most application-layer companies built on a single model capability are absorbed by the next model release. The breakouts here own a workflow, not a feature, and reprice from per-seat to per-token or per-outcome.

9 · Governance. Policy, sanctions, export controls, deemed-export rules, AI safety regulation, model accountability, antitrust. This was treated for years as a soft layer — a horizontal constraint, not a market. It has now become a control plane perpendicular to the entire stack. A single jurisdictional directive can switch off a frontier model across every customer on the planet in seventy-two hours. Governance does not sit on top of the stack — it cuts through it. This is the most consequential reframe in the deeper reading: Layer 9 was never a layer. It is the dimension along which every other layer is conditional.

Two reading rules govern the stack. Each layer is constrained by the one below it — you cannot ship a model without compute, you cannot ship compute without silicon, you cannot ship silicon without foundries, you cannot ship foundries without energy. Each layer is commercialized by the one above it — silicon needs models to monetize, models need a harness, the harness needs distribution. Read top-down to see what gets priced. Read bottom-up to see what gets bottlenecked.

Where profit pools concentrate. Read the stack as a column to see one company’s full vertical posture — only a handful of operators in the world are deeply present across more than three layers, and those that are command the largest market capitalizations on Earth. Read the stack as a row to see who owns one layer in isolation. The diagonal between vertical integration and horizontal monopoly is where the supercycle’s profit pools live. Capex concentrates at L1–L5. Margin concentrates at L3 (silicon), L4 (networking), L6 (models), L7 (harness), and L8 (distribution). Demand concentrates at L8. Bottleneck concentrates at L1. Control concentrates at L9.

Any portfolio that touches only one layer carries dependency risk on all the others. The complete investment thesis specifies which layer is being financed, which adjacent layer the position depends on, and which jurisdictional governance regime the position assumes.

Task vs Purpose — The Cognitive Jevons Paradox

Stay Ahed of The AI Supercycle →

There is a single mechanism that controls how the labor market reorganizes in the intelligence age. It is the same mechanism that controlled the steam age, the electrification age, and the computerization age, and it has been misread every time. The mechanism is the distinction between task and purpose.

A task is the thing a person does in a given hour. A purpose is the reason the task exists. Automation collapses tasks. It does not collapse purposes. And when the underlying purpose is demand-unconstrained — when there is more of it wanted than is currently supplied — automating the task expands employment in the role, not contracts it.

Twelve years ago, a leading voice forecast the extinction of the radiology profession on the strength of superhuman computer vision. Computer vision did saturate radiology. Every reading is now augmented by it. And the number of radiologists is up, scan volume is up, departmental profitability is up. The mechanism: AI made each radiologist more productive, which lowered the cost per scan, which expanded the volume of patients who could be admitted, which increased departmental revenue, which justified hiring more radiologists. The same forecast made about software engineers in 2024 — “ninety percent of coding will be automated” — is being falsified in real time by record hiring at the foundation model labs.

This is a cognitive version of Jevons’ paradox: in 1865, William Stanley Jevons observed that more efficient steam engines did not reduce British coal consumption — they expanded it, because lower cost per unit of work expanded the universe of work being done. Cognitive labor sits in the same regime. Cheaper cognition expands the demand for cognition. The plumber becomes a bathroom designer. The carpenter becomes a home architect. The salesperson becomes an interior consultant. The accountant becomes a strategic advisor. The radiologist becomes a clinician with infinite scan capacity.

Three predictions follow directly. First, aggregate cognitive employment grows, not shrinks — the roles change shape, the headcount expands. Second, roles defined by task automate while roles defined by purpose elevate — a clear definition of purpose is now a labor market moat. Third, the geographic and educational concentration of cognitive work flattens — the prior dominant programming language was known by roughly two percent of humanity, the new programming language is human language known by roughly one hundred percent, and the population of programmers therefore becomes the population of humans. Returns to scale move from technical populations to user populations.

Practical Implications

Stay Ahed of The AI Supercycle →

Read more

Not as a passing trend or productivity hack, but as a signal of what’s coming next for many of us, whether we’ve realized it yet or not.

What I left out, however, was the most important question: did I choose this transformation? Was it a deliberate decision? A reality I consciously created?

I wish I could say yes.

The truth is that I didn’t arrive here entirely by choice. This shift is largely the result of forces much bigger than me. It is the natural consequence of how the most powerful industries in the world are evolving. And because I’ve spent my career operating at the edge of the AI industry, being “in the harness” wasn’t a tactical choice. It became a strategic constraint.

Am I disappointed by that?

Not at all.

In fact, this is the most exciting professional transformation I’ve experienced in my life.

Today, I can do things I only dreamed about a decade ago. I can fully embrace my hyper-generalist nature. I can reconnect disciplines that were previously separated by time, expertise, and organizational boundaries. I can become the kind of business polymath that modern specialization had almost made impossible.

AI has given me leverage to think broadly again.

And that’s why this moment is both terrifying and exhilarating.

Being in the harness is infinitely more interesting than being locked inside a gaming console. The real world becomes your environment. Every industry becomes your playground. Every inefficiency becomes an opportunity.

The ability to create business arbitrage, to identify opportunities between what is possible and what is currently being done, is becoming one of the most valuable skills of our era.

For anyone working in business, the defining skill of the coming decades will be the ability to transform intelligence into outcomes.

To turn tokens into decisions.

To turn decisions into execution.

To turn execution into business results.

That is the new game.

Welcome to the new world.

The only question is:

Are you ready to embrace it?

Access my harness with the exec plan!

I want to tell you how I ended up living inside a harness of agents instead of in front of a computer. To make that story land — and to make clear it wasn’t a personal preference but a forced move — I have to walk you through what happened to AI between late 2022 and now.

Not as a list of models. As a sequence of inflection points, each one breaking the rule the previous era ran on, each one forcing the next. There are four such breaks. They land roughly one a year. Together, they are the reason it is no longer possible to keep working the way I, or anyone, used to.

Access my harness with the exec plan!

Read them in order, because the logic only resolves at the end. By the time you reach the fourth break, the personal arc will look like what it actually was: a delayed response to an exponential the rest of the system had already absorbed.

2022 → 2024: pre-training scaling, and the chat era it produced

In November 2022, ChatGPT became public. Within months, the rule that had been quietly running AI since 2020 became visible to everyone: make the model bigger, feed it more data, and it gets smarter.

Access my harness with the exec plan!

This was pre-training scaling — the original scaling law. Dario Amodei put it plainly in 2024: from 2020 to 2023, the thing being scaled was pre-trained models trained on increasing amounts of internet text, with a tiny bit of fine-tuning on top. The recipe was simple, the gains were stunning, and the product was equally simple — chat:

• One model, one turn, one human. The user prompted; the model answered. The user prompted again. Every step was driven from the outside.

• The interface was a window. Operating happened by hand, in real time, at the chat surface.

• What it rewarded was operating skill. Knowing how to structure a prompt, work the window, read the output, push it again. Being good at chat was a real edge. It carried a lot of careful, capable people for two years.

Underneath, though, the engine driving the gains was already running into a wall. By 2024, the conversation in papers and at NeurIPS had shifted to “diminishing returns”:

• Doubling pre-training compute no longer doubled the improvement. The curve was bending, not breaking — but it was unmistakably bending.

• High-quality training data had largely been exhausted. The internet had been scraped. The books had been ingested. Code had been processed. The cheap, dense tokens were gone.

• The labs knew it before the public did, and started saying so by the end of 2024.

This is the structural fact everything else is downstream of. Pre-training scaling hit a wall. But the frontier didn’t stop — it never does. It had to find a new place to put compute. The era of “just make it bigger” was ending, and scaling itself had to relocate.

Hold that move in mind: when one axis flattens, the frontier moves outward to the next layer. It happens four times in this story.

September 2024: the frontier moved to test-time, and reasoning began

On September 12, 2024, OpenAI released o1 — a model designed to spend more compute thinking at the moment you used it, not at the moment it was trained. The full version shipped in December as part of the “12 Days of OpenAI” event. Within months, every major lab had reasoning models in production: OpenAI o3-mini (January 2025), o3 and o4-mini (April 2025), Anthropic’s reasoning Claude, Google’s Gemini Thinking, DeepSeek R1.

Access my harness with the exec plan!

The technical detail mattered less than the structural fact. Scale had migrated. The new way to make models smarter was not to grow them, but to let them deliberate. Compute that used to be spent in training was now being spent in inference — the budget moved from before deployment to during use.

That single migration kicked off the second era, and it had three immediate consequences:

• The unit of work expanded from a turn to a task. A reasoning model could chew on a problem for minutes, not seconds. The human’s role shifted from “operate every step” to “pose the problem, judge the result.”

• Smaller models became serious contenders. A small model with a sufficient thinking budget could outperform a much larger model with none. The “biggest model wins” rule of pre-training quietly stopped applying.

• Latency stopped being a vice. The market accepted models that took longer to answer, because the answers were structurally different — they showed work, reasoned across steps, caught their own errors mid-stream.

But the real consequence wasn’t visible until you looked one step further. A model that can think step by step can also act step by step. It can call a tool, check the result, decide what to do next, and call another tool. It can run for minutes, not seconds. It can chase a goal across many turns without a human between them.

Reasoning, in other words, was the precondition for agency. The chat era’s defining shape — human in the loop, every turn — was about to break, and o1 was the first crack.

2025: the agent era, and the protocol that made it real

The agent era didn’t arrive because models suddenly got smart enough. It arrived because two things landed in the same window:

Access my harness with the exec plan!

• Reasoning unlocked autonomous step-taking from the inside. A model could now run a chain of thought, evaluate intermediate results, and adapt its plan mid-task.

• The Model Context Protocol gave it hands. In November 2024, Anthropic open-sourced MCP — a single standard for how an AI model connects to external tools, data, and applications. The industry called it “the USB-C of AI.”

MCP went from announcement to infrastructure faster than any AI standard before it:

• March 2025 — OpenAI adopted MCP across the Agents SDK, Responses API, and ChatGPT desktop.

• By late 2025 — 97 million monthly SDK downloads, 10,000+ active servers, first-class support in ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code.

• December 2025 — Anthropic, OpenAI, and Block jointly formed the Agentic AI Foundation under the Linux Foundation and donated MCP into it, with AWS, Google, Microsoft, Cloudflare, and Bloomberg as supporting members.

The plumbing had become industry infrastructure. What that combination produced was a fundamentally different thing in your computer. Not a chat companion. An agent — a system that takes a goal and runs with it, calls tools on its own, reads its own results, corrects its own errors, and comes back when the work is done.

Through 2025, every major lab shipped autonomous agent products:

• OpenAI — ChatGPT Agent mode (operator-style autonomous browsing, scheduling, multi-step tasks).

• Anthropic — Computer Use (the model controlling a real desktop, clicking, typing, navigating).

• Microsoft, Google, Meta, Amazon — all shipping agentic equivalents through 2025.

The shape of work changed underneath my hands while I was still operating. Where the chat era was operate every turn, the new way was:

• Hand a goal to the system.

• Let it run.

• Judge the result.

That is not the same job. It is a different job entirely. The most productive people quietly stopped operating. I had not.

This is the period — roughly mid-2025 onward — when I started getting out-produced by people working the same hours with the same tools. The gap widened week by week. I just hadn’t yet identified the layer the gap lived on.

2026: swarms, AGaaS, and the form factor leaves the desk

The shift from chat to a single agent was the visible break. The less visible one — the one happening as I write this — is from a single agent to a swarm.

Access my harness with the exec plan!

Once orchestration is standardized over MCP, an agent stops being a thing you launch and becomes a unit you compose:

• One agent becomes many specialists. Narrow agents — one for content, one for analytics, one for outreach, one for review — each with a single job.

• They run in parallel, on independent schedules. Not a serial pipeline; a swarm.

• They share a single source of intelligence. A common memory layer, a cached source of truth, gated approvals where stakes are high.

The unit of work climbed again — from a turn (chat), to a task (agent), to an operation (swarm). And the scaling axis migrated one more rung outward: not the model, not the loop, but the orchestration around them.

Capability is now grown by adding agents, loops, gates, and memory in coordination. The harness is the new scaling layer.

The business model mutates to match, because the unit of value travels with the frontier:

• SaaS sold a tool and charged for the seat. A human did the work inside it. The pricing assumed human operation as the rate-limiting step.

• AGaaS — agents-as-a-service — sells the work itself. You don’t buy the software and operate it. You buy the outcome an agent or a swarm delivers.

• The thing being priced shifts from an instrument a person uses to a result a system produces. AWS’s leadership has stated openly that agentic AI applied to e-commerce could be “the next multibillion-dollar business for AWS.” OpenAI has projected that roughly one-third of its 2025 revenue would come from agentic AI tooling. Outcome-based pricing is moving from edge case to default.

And the form factor follows the same vector outward:

• Chat needed a screen. You sat at it; the work happened where you were.

• Agents tolerate a phone. The work runs in a runtime elsewhere; you steer through a thin handle.

• Swarms can be steered from anywhere a goal can be uttered. Voice. Glasses. An ambient surface. The screen shrinks because the work it used to hold has moved off it.

The next surface won’t be a laptop. It will be an AI-native object — closer to an always-on personal-agent appliance than a PC — that you speak goals into and tap approvals on.

That is where we are in June 2026. Not at the end of any cycle, but visibly running the fourth scaling era after pre-training, post-training, and test-time: agency and orchestration as the place where capability is now grown.

The forcing function: one rule, four iterations

Step back from the four breaks and a single move repeats:

Access my harness with the exec plan!

When the current scaling axis hits diminishing returns, the frontier doesn’t stop — it relocates outward, one layer further from the model. Each relocation creates a new unit of work, a new business model, and a new form factor, and renders the prior layer’s defining skill a commodity.

The cycle, named: new axis → new unit of work → new business model → new form factor → old layer commoditizes → repeat.

Run that against the actual record:

• Pre-training (2020–2024) scaled the model.

  • Unit: a turn. Business: SaaS. Form: a desktop chat window. Skill rewarded: operating it well.

  • What broke it: data exhaustion + diminishing returns from naive scaling.

• Test-time (Sept 2024 → 2025) scaled the thinking.

  • Unit: a task. Business: reasoning-tier subscriptions ($200/mo Pro, dedicated reasoning APIs). Form: still a screen, but the human stepped further back. Skill rewarded: framing and judgment.

  • What broke it: reasoning made agency feasible; the loop wanted to close.

• Agency (2025) scaled the loop.

  • Unit: a goal. Business: outcome delivery. Form: agent + phone. Skill rewarded: setting frames the loop runs inside.

  • What broke it: once agents work, you want many of them, coordinated.

• Orchestration / swarms (2026 →) scales the harness.

  • Unit: an operation. Business: AGaaS. Form: ambient surface. Skill rewarded: authorship across a swarm.

Each iteration is faster than the last:

• Pre-training ran ~4 years (late 2020 → late 2024).

• Test-time ran ~1 year (Sept 2024 → late 2025).

• Agency ran ~1 year (2025).

• Orchestration is the era we are inside right now.

The compression isn’t rhetoric or vibe; it is the visible spacing of inflection points since 2020, and it is the part that hits humans hardest. You no longer get years to learn the new layer before it commoditizes.

This is what I mean when I say the transition was forced, not chosen. I didn’t pick a workflow. I got carried by a scaling law that kept relocating capability outward — and the only way to stay near where value was being made was to move with it.

What this did to me

For about eighteen months — while the second and third eras were happening — I worked the way I always had. In the loop, turn by turn, operating well. I was good at it. I had been good at it for two years. I had built habits around it.

And I was being out-produced. At a widening rate. By people working the same hours, on the same models, with the same access.

For a long time, I misdiagnosed it:

• First I told myself it was a discipline problem. I tried longer hours. Sharper focus. Better prompts. The gap kept widening.

• Then I told myself it was a tooling problem. I tried better models, better wrappers, better workflows. The gap kept widening.

• Eventually I realized it was a layer problem. The rung I was standing on — operating — had quietly become the part the machine does for itself. Every hour I spent operating by hand was an hour spent doing something the system now does at scale, for almost nothing.

Access my harness with the exec plan!

That is the punchline of the four-year arc seen from inside one person’s career. Each scaling relocation deprecates the previous layer’s defining skill. The people who tripled their output in 2025 had not found a better tool. They had stepped one rung outward — from operating to framing, from doing the work to authoring it — while I was still working harder at the part that no longer mattered.

When the world’s software output roughly tripled while the number of people doing the work stayed flat (GitHub’s own April 2026 figures), the people producing three times as much were not typing three times faster. They had restructured how the work happens — one person directing a system, instead of one person performing the steps.

Access my harness with the exec plan!

And the moment enough people did that, the market quietly reset its expectation of what a single person produces. Producing careful, excellent work at the old volume stopped being a stable position; it became a slow exit, because the baseline had moved.

So I moved. Not because the new way was exciting, but because the alternative was to keep doing careful work at a level the market had already priced as not enough.

• I stopped operating and started framing.

• I stopped working inside a chat window and started directing a harness from wherever I happened to be.

• The frame I set in the morning steered a swarm through the day.

• The work stopped running at my desk and started running in a remote runtime I held a thin handle to.

I didn’t leave the desk on purpose. The work left, and I followed it.

The bedrock underneath the four eras

A reasonable objection follows immediately, and it deserves an honest answer.

Access my harness with the exec plan!

If every rung commoditizes, and the cycle keeps accelerating, what stops the rung I just moved to from sinking next? Won’t framing get automated the way operating did?

Yes — as a craft, it will. Models will draft problems, propose constraints, generate evaluation criteria. The skill of framing will get machine help, and within a year or two it will get a lot of it. So if framing-as-a-skill were what I was relying on, I would just be one rung higher on a sinking ladder, with the floor coming up faster.

But the four-era arc points at something deeper than any single skill. Every rung the frontier has climbed has been a capability — and every capability eventually commoditizes when scale relocates. Operating fell. Framing-as-craft will fall. Orchestration-as-craft will fall after that.

The thing that can never be automated, because it isn’t a capability at all, is authorship.

• Wanting a particular outcome. Not the ability to produce a particular outcome — the underlying wanting of it.

• Choosing which tradeoffs are acceptable. Not the analytical work of comparing them — the moment of deciding, with the weight of consequence.

• Answering for the result when it ships. Not the work of producing it — the liability for what it does in the world.

The machine can take on any competence given enough capability. It cannot take on wanting, or liability. It can draft a thing; it cannot be the one who owns it.

That is the bedrock the churn can’t reach, because it isn’t a skill that can be learned faster by something faster than you. It is a question of whose will, and whose responsibility.

So the deep “why” of my transition isn’t “use agents.” It is a migration of where I stand:

• From the person who performs the work

• To the person who authors it

That is the only ground the next scaling relocation cannot reach, because it isn’t on the ladder. Everything else I do now — the frames, the gates, the swarm — sits on top of that one position. I stopped trying to be faster than the machine and started being the thing it answers to.

That is the why.

What it actually looks like to live there — what life in the harness feels like day to day — is the next piece.

Key Takeaways & Mental Models

Access my harness with the exec plan!

The Scaling Migration — AI’s scaling laws didn’t stop; they keep relocating outward from the model. Pre-training (2020–2024) → test-time reasoning (Sept 2024 → 2025) → agency (2025) → orchestration / swarms (2026). Each relocation grows capability at a new layer.

The Forcing Cycle — Each new scaling axis creates a new unit of work, a new business model, and a new form factor, and turns the prior layer’s defining skill into a commodity. Run it four times: chat → swarms, SaaS → AGaaS, screen → ambient surface.

The Compression — Pre-training ran four years. Test-time ran about one. Agency ran about a year. Orchestration is now. The compression is the visible spacing of inflection points since 2020 — not rhetoric, geometry.

The Operating Trap — Continuing to compete on the rung the frontier just left feels like diligence. It is the fastest way to fall behind, because the system now does that rung for free, at scale, and without rest.

The Output Asymmetry — When output triples on flat headcount, the gap isn’t effort — it’s restructuring. Once the baseline resets, holding the old layer isn’t a preference; it’s a slow exit.

The Accountability Floor — Every rung on the ladder is a capability, and every capability eventually commoditizes. The one position that doesn’t is authorship — wanting, choosing, answering for outcomes. It’s not on the ladder, which is exactly why the churn can’t reach it.

With massive ♥️ Gennaro Cuofano, The Business Engineer

As an analyst, a student of history, and someone deeply interested in structural realities, I see it as my responsibility to make a deliberate effort to contextualize, abstract, compress, and build mental models that help explain the world around us.

That is what The Business Engineer is about. For me, it is a lifelong quest.

So when ChatGPT was released, I felt it was essential to provide a framework, a mental model capable of making sense of the emerging AI era.

At the time, despite the relentless pace of progress in AI, it is easy to forget how many people believed everything would change within a single year. The expectations were immense. In that environment, I felt a better frame was needed.

That is where the AI Supercycle Thesis emerged.

And a word of caution: the term supercycle should not be interpreted as a purely economic concept. It goes far beyond economics. It describes the intersection of three powerful forces: geopolitics, cultural transformation, and economic change.

The economy is only one layer of the story. In many ways, it is the instrument through which deeper geopolitical and cultural shifts express themselves.

The AI Supercycle is ultimately an attempt to understand the thread connecting these forces and how, together, they shape the next era of human development.

For that reason, the AI Supercycle is not a forecast. It is a structure. A thirty-to-fifty-year economic transformation comparable in scope to the Industrial Revolution, running on three nested clocks, exhibiting the paradoxical dual character of bubble and revolution at the same time, and operating through a nine-layer industrial stack that begins in mineral extraction and ends in geopolitical control. This piece is the map.

Stay Ahed of The AI Supercycle →

Subscribe To The Podcast

The Argument in One Paragraph

We are inside the first decade of a transformation that will run for half a century.

It has three nested cycles operating at different time horizons, a dual character that is simultaneously a bubble and a supercycle, and a layered industrial stack of nine layers, running from energy at the bedrock to governance at the ceiling.

The ceiling does not behave like a layer. It behaves like a control plane perpendicular to the whole stack — one that answers to geopolitics rather than engineering.

The floor does not behave like a market. It behaves like a geopolitical chokepoint — a small handful of countries deciding what the rest of the stack is allowed to build. Everything between the floor and the ceiling is where the next decade of competition lives.

What the Supercycle Actually Is

Stay Ahed of The AI Supercycle →

The launch of ChatGPT on November 30, 2022 was not the arrival of AI. AI had been arriving for years before that night.

What changed is that AI reached the point where it could scale to billions of users on the same day. That capability threshold — distribution at consumer-internet scale — is what triggered the supercycle. It forced every layer of the existing technology economy to absorb a new substrate simultaneously, and substrate-level shifts of that magnitude are what produce multi-decade economic transformations.

Three cycles run inside the supercycle, each on its own clock:

• Short cycle (5–10 years): rapid capability improvements and market disruptions. Where the bubble lives. Where most pundits spend their attention.

• Medium cycle (10–20 years): industry restructuring, business-model evolution, the end of incumbent dominance in software and search.

• Long cycle (30–50 years): civilizational transformation. Comparable to the half-century electricity took to reorganize industrial production around itself between 1890 and 1940.

Conflating these clocks produces the most common analytical errors of the moment. The “AI is overhyped” camp and the “AI changes everything” camp are usually arguing about different cycles without realizing it.

The dual character is the most useful piece of intellectual scaffolding for navigating the present. AI exhibits genuine bubble characteristics — excessive hype, overvaluation, capex commitments that exceed the unit economics of any single application — at the same time as it exhibits genuine supercycle characteristics — real productivity gains, fundamental restructuring of profit pools, civilizational impact. Both are true.

The mistake is treating them as competing claims. They operate on different cycles. The bubble can pop on the short cycle while the supercycle continues uninterrupted on the long cycle, exactly the way the dot-com bust did not stop the internet from becoming the substrate of the next thirty years.

The three phases of integration unfold over the supercycle’s full arc:

• Phase 1 — Linear Expansion. AI as a layer on top of every internet-era industry. Web Squared. E-commerce, social, fintech, digital marketing, software-as-a-service all get amplified rather than replaced. This is the decade we are in.

• Phase 2 — Non-Linear Restructuring. Industries get rebuilt around AI-native architectures rather than retrofitted onto pre-AI ones. SaaS gives way to its successor architecture. Search gives way to assistants. Customer service gives way to autonomous agents.

• Phase 3 — Civilizational Reorganization. What the supercycle is actually for. Machine cognition becomes a general-purpose industrial input, comparable to what electricity became after 1920.

The canonical transition signal is the “BlackBerry moment.” Google’s search revenue is hitting record highs at the exact moment AI assistants pose an existential threat to its core business — the same pattern as BlackBerry peaking as the iPhone arrived. Supercycle transitions look like strength right before they look like collapse. This is the pattern to watch for across every industry the supercycle touches.

The Nine Layers

Stay Ahed of The AI Supercycle →

The AI economy is not a list of companies racing each other. It is a layered industrial stack — and at any given moment the binding constraint sits at a different layer than it did six months ago. Reading the map well means knowing where the constraint currently lives and where it is rotating to next.

The nine layers, top to bottom:

Governance · Distribution · Agentic Harness · Foundation Models · Compute Capacity · Networking & Protocols · Silicon · Foundries & Packaging · Energy & Physical.

Two rules govern the stack:

• Each layer is constrained by the one below it. You cannot ship a model without compute. You cannot ship compute without silicon. You cannot ship silicon without foundries. You cannot ship foundries without energy. The bedrock dictates what the upper floors can do.

• Each layer is commercialized by the one above it. Silicon needs a model to monetize. Models need a harness. Harnesses need distribution. The roof dictates what the lower floors can earn.

Read the map two ways. Read a column to see one company’s full vertical posture — Google’s nine-layer column shows the closed-loop integration that nobody else has matched. Read a row to see who owns a single layer — the networking row, for instance, reveals NVIDIA’s monopoly and the fact that everyone else in the AI economy is paying rent on it. The diagonals where dominance concentrates are where the supercycle’s profit pools live.

Layer by layer, what each one is and why it matters:

• Layer 9 — Governance. The rules that determine whether a model can be released, deployed, sold, or used. Capability gating inside labs, procurement frameworks inside enterprises, export controls between nation-states, litigation over contested deployments. No longer a passive layer; an active priced variable.

• Layer 8 — Distribution. The surfaces through which AI reaches end users. ChatGPT, Gemini, Copilot, the Apple-Gemini handshake, Meta’s social graph. The decade’s most expensive land grab.

• Layer 7 — Agentic Harness. The wrapper that turns a foundation model into a usable agent. Memory, tools, planners, sub-agent coordination, retry and rollback. CUDA’s analog at the agent layer. The most contested architectural layer of the next three years.

• Layer 6 — Foundation Models. The frontier and the long tail. OpenAI, Anthropic, Google DeepMind, Meta, xAI, DeepSeek, Mistral. The layer everyone watches and the one most likely to commoditize first.

• Layer 5 — Compute Capacity. Hyperscaler GPU clusters, neoclouds, custom-silicon estates. Where the capex bubble lives. Where the binding constraint sat in 2024 and where it is gradually rotating away from.

• Layer 4 — Networking & Protocols. The interconnect that turns thousands of GPUs into a single coherent compute fabric. NVLink, InfiniBand, optical interconnects. Now over 20% of rack-scale system cost; growing faster than any other line item.

• Layer 3 — Silicon. Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova). The within-layer split is what makes silicon strategically interesting again.

• Layer 2 — Foundries & Packaging. TSMC, Samsung, Intel Foundry, and the HBM oligopoly (SK Hynix, Samsung, Micron). Allocation power at the foundry is now more consequential than chip design itself. The binding constraint for the entire upper stack.

• Layer 1 — Energy & Physical. Gigawatts, transmission, cooling, land, water, rare-earth magnets. The bedrock everything else stands on. The constraint that defines the ceiling on how much intelligence can be produced in any given year.

The supercycle is not unfolding in time. It is unfolding in layers. The job of any operator, investor, or builder is to know which layer they are operating inside and which layer the binding constraint is rotating toward next.

The Ceiling Is a Control Plane

Stay Ahed of The AI Supercycle →

Layer 9 does not behave like a layer. It behaves like a control plane perpendicular to the whole stack — capable of switching off any company at any layer beneath it on a timeline measured in days rather than years.

A single government directive can darken a frontier model globally in seventy-two hours. A single procurement framework inside a Fortune 100 buyer can shift demand at the model layer overnight. A single litigation outcome can re-architect what is permissible at the harness layer. Capability gating decisions inside the labs themselves can withhold a model class from market for reasons that have nothing to do with technical readiness.

Layer 9 answers to geopolitics, not to engineering. That is the structural read, and it changes the valuation logic at every layer beneath it. Models, compute, silicon, foundries — none of them can be priced as standalone technology assets anymore. They have to be priced with an embedded discount for the political volatility of the control plane.

The supercycle has a ceiling now, and the ceiling is political.

The Floor Is a Chokepoint

Stay Ahed of The AI Supercycle →

If Layer 9 is the ceiling, the physical floor is a geopolitical chokepoint. The supercycle runs on rare earths before it runs on tokens. Every GPU cooling fan, every data-center motor, every robotic actuator, every EUV lithography machine depends on neodymium-iron-boron magnets. One country controls roughly 60% of global extraction and over 90% of processing capacity.

That concentration has been weaponized. Extraterritorial export controls — modeled on US semiconductor restrictions and applied in reverse — now reach any product containing 0.1% or more Chinese-sourced rare earths by value, regardless of where it is manufactured.

Twelve of the seventeen rare-earth elements are under restriction. The “Affiliates Rule” presumptively denies export licenses to any importer with 50%-or-greater ownership by entities on China’s control lists.

The response is a triangular competition:

• United States — multi-billion-dollar EXIM commitments, Defense Production Act invocations, Brazil partnership, domestic processing buildout.

• European Union — strategic stockpiling, Australia partnerships, processing investment.

• China — processing dominance, vertical integration from mine to magnet, regulatory leverage applied with surgical precision.

Brazil emerged as the pivotal countermove — second-largest proven reserves behind China, only 30% of the country geologically mapped, US-Brazil rapprochement driven explicitly by the shared interest in breaking the processing monopoly. The next decade of critical-minerals supply chains will be reorganized around that pivot.

The supercycle runs on dirt before it runs on tokens. Whoever controls the floor sets the ceiling on what the rest of the stack can build.

The New Stack — Software, Rebuilt

Stay Ahed of The AI Supercycle →

The supercycle’s effect on software is captured by a single sentence: the agent is the consumer, and the human is the principal. Once that shift completes, the existing software-as-a-service stack becomes structurally obsolete. A new five-layer architecture replaces it:

• Knowledge Substrate — typed graph, not relational table. Agents query by description; humans query by SQL. The substrate decides which kind of company you actually are.

• Orchestration Layer — directed graph of agents with planner, sub-agents, retry, rollback, external capability calls. Not a deterministic chain.

• Access Surface — Model Context Protocol endpoints, tool catalogs, machine-readable schema, pricing APIs, policy declarations, rate limits. The exterior surface that agents call into from outside the company.

• Metering Layer — principal delegates budget; agent consumes against a rate card. Outcome definitions, machine-readable rate cards, real-time metering, multi-principal accounting, negotiation-aware pricing.

• Trust Layer — cross-cutting band that cuts through all four horizontal layers. Agent identity-as-buyer, audit trails, attestation, capability scoping.

Three anti-patterns kill incumbents trying to retrofit:

• Database-as-Substrate — relational tables pretending to be a knowledge graph. The agent cannot find what it needs by description.

• Logic-as-Orchestration — sequential code pretending to be agent coordination. Brittle the moment an agent has to plan.

• Pricing-Page-as-Metering — a marketing page pretending to be a rate-card API. The agent cannot transact.

The master diagnostic is one line: a pricing API beats a pricing page. If your competitor exposes a machine-readable rate card and you expose a “contact sales” button, you are not in the same market anymore.

The Mutation Map — Five Archetypes

Stay Ahed of The AI Supercycle →

Every existing software company has to mutate, retreat, or die. The responses cluster into five archetypes:

• Veneer — chatbot bolted onto an unchanged stack. Theater. Dies first.

• Surface — improved interface layer on the same substrate. Buys a year of relevance.

• Reprice — moves from seat-based to outcome-based pricing without rebuilding the underlying stack. Survives medium-term but hits a margin ceiling.

• Substrate — rebuilds the knowledge layer for agent consumption while keeping the upper stack. Survives long-term.

• Rebuild — full agent-native architecture from substrate to metering. Wins the next cycle.

Only Substrate and Rebuild survive structurally. Everything to the left of them is buying time. The five-year window for incumbents to commit closes around 2028, and after that the agent-native companies founded post-2023 begin eating the bottom of the existing market on price and the top on capability simultaneously.

The Analytical Toolkit

Stay Ahed of The AI Supercycle →

Five cross-cutting tools sit above the nine-layer map and help read it:

• The Three AI Tribes — Substrate Tribe (infrastructure plays), Surface Tribe (interface plays), Sense Tribe (data and judgment plays). Tells you which tribe a company belongs to and which one is best positioned at any given layer.

• The Barbelled Distribution Economy — value concentrates at the extremes (massive-scale platforms and deep-niche specialists); the middle hollows out. Explains where value accrues across every layer of the stack.

• The Four Scaling Paradigms — pretraining, post-training, inference-time, agentic-loop. Tells you which scaling regime a model company is betting on and where the next capability gain comes from.

• Cognitive Range — the human variable. AI provides breadth at every level of resolution; humans provide navigation between levels. Full-range cognitive workers earn a structural premium that automation cannot touch.

• The FRED Test — the diagnostic for whether a workflow survives agentic automation. Frequency, Repeatability, Error tolerance, Decomposability. Score high on all four and the workflow is automated within the cycle; score low on any one and the human stays in the loop.

These are not separate frameworks. They are different cross-sections of the same map.

What’s Next

Stay Ahed of The AI Supercycle →

Three structural moves to watch over the next twelve to twenty-four months:

• The control plane hardens. Expect more single-directive shutdowns, more procurement-driven model selection, more litigation deployed as competitive weapon. The companies that price governance into their architecture from day one will outperform the ones that treat it as compliance overhead.

• The floor shifts. Brazil’s reserves come online slowly — five to ten years to scale processing. The interim is where the leverage gets used most aggressively. Watch processing capacity and packaging allocation, not extraction headlines.

• Mutation pressure on the upper stack accelerates. Reprice archetypes hit their margin ceiling. Substrate archetypes start shipping rebuilt stacks. Agent-native companies founded post-2023 start eating the existing market from both ends.

The supercycle does not move evenly. It moves layer by layer, and the binding constraint rotates. The job is not to predict which layer is next. The job is to read the map well enough to recognize the rotation when it happens.

Key Takeaways & Mental Models

Stay Ahed of The AI Supercycle →

Five named models govern the supercycle as a whole:

• The Nested Clock — three cycles (5–10, 10–20, 30–50 years) run simultaneously inside the supercycle. Bubble and revolution are both true, on different clocks. Conflating them is the most common analytical error of the moment.

• The Layered Stack — competition does not happen in “the AI industry”; it happens at specific layers within a nine-layer industrial stack. Read columns for posture, rows for dominance, diagonals for profit pools.

• The Control Plane — Layer 9 is not a layer. It is a plane perpendicular to the stack. It answers to geopolitics, not to engineering, and it discounts valuation at every layer beneath it.

• The Geopolitical Floor — the supercycle runs on dirt before it runs on tokens. Whoever controls extraction and processing sets the ceiling on what the rest of the stack can build.

• The Agent Is the Consumer — once that shift completes, the existing software stack becomes structurally obsolete. Only Substrate and Rebuild archetypes survive the transition.

With massive ♥️ Gennaro Cuofano, The Business Engineer

Somewhere in the last few months, the shape of how I produce has changed. I stopped typing prompts and reading answers one at a time, and started setting a frame in the morning and watching a swarm of agents carry it through the day — drafting, optimizing, scheduling, measuring, adjusting — while I did the one thing none of them can do for themselves: decide what was worth doing, and define what good meant before any of it ran.

And I’m rarely at a desk when it happens. I set the frame from a phone — in transit, between other things, walking — because the work no longer runs where I am. It runs in a runtime somewhere else, and the thing in my hand is just the surface I steer it through. The computer I spent thirty years sitting in front of has quietly left the room.

This isn’t a piece about tools. It’s about what’s left for a person to do once the tools can do almost everything. The answer is smaller and sharper than most people expect — and the whole map of it fits in six moves. Here they are, one at a time.

Access My Full Harness in The Exec Plan

The edge has moved to framing

The scarce, valuable skill is no longer operating the model. It’s framing the work before the model runs.

Here is the rule that governs every technology shift: when a capability becomes universal, it stops being an advantage, and the edge moves up the stack to whatever is still scarce.

Access My Full Harness in The Exec Plan

Trace it. Hardware was once the edge — until hardware became cheap and abundant, and the advantage moved up to software. Software commoditized, and the edge moved to distribution. Then to data. At each step, once everyone had the layer, it stopped being where the advantage lived, and the advantage climbed one rung higher.

AI capability has now reached that point. Everyone has the same frontier models. The compute is cheap, the interfaces are frictionless, and the gap between a well-funded team and a solo operator with a credit card has all but closed. So the edge has already left the model itself. For a short while it sat one rung up — in operating the model: the skill of working a chat window well, structuring the problem, reading the output, re-prompting toward something sharp.

That rung is commoditizing too, and fast — because of agents. An agent doesn’t wait for you to prompt it. It prompts itself, reads its own output, notices its own errors, and re-runs. The skill of working the chat window is being absorbed into the system. So operating, the thing that was briefly the edge, is becoming something the machine does for itself.

Which moves the edge up one final rung — to the one input the agent cannot generate on its own: the frame. Framing is everything you decide before the agent runs. It has four parts:

• Problem definition — what is actually being solved, stated as a mechanism, not a wish.

• Constraints — the boundaries any acceptable answer has to stay inside.

• Evaluation criteria — what “good” looks like, specified in advance.

• Stopping conditions — when the agent should halt, escalate, or hand back.

The frame is the whole of the human contribution. Everything downstream of it is execution. That is the first move, and the rest of the piece follows from it: if framing is the edge, then everything else — why it matters, who’s exposed, how to do it, what it steers, and what it does to the industry — is a consequence.

Autonomy removed the safety net

Access My Full Harness in The Exec Plan

In the old way of working, a weak first attempt was cheap because you could fix it as you went. Autonomy removes that, so the first frame now decides the outcome.

Think about how working with a chat model used to go. You asked a rough question, got a rough answer, saw it was rough, and asked a better one. The cost of a bad first try was low, because you stayed in the loop turn by turn and caught the drift as it happened. Trial and error did a lot of the thinking for you. A careful operator and a careless one both reached a decent answer in the end — the careless one just took longer.

An autonomous agent removes that loop. It runs fifty steps without checking back. There is a frame at the start and a result at the end, and nothing in between where you correct course. If the frame was weak — wrong problem, soft constraints, no definition of “good” — the agent doesn’t notice. It executes the weak frame faithfully, confidently, and fast, and hands you a polished, internally consistent answer to the wrong question, delivered at a speed that makes it look authoritative.

That is the characteristic failure of this era, and it has a shape worth naming: confidently wrong at scale. A small error in the frame fans out into a large error in the output, because every autonomous step compounds the one before it. The old way punished a bad frame with a few wasted minutes. The new way punishes it with a finished deliverable that’s wrong in a way nobody catches until it ships.

Access My Full Harness in The Exec Plan

This sorts people into three tiers, and it’s worth being blunt about who’s where:

• Prompters — whose skill was phrasing. Obsolete; the agent phrases better than they do.

• Operators — genuinely skilled at working the loop turn by turn. Newly exposed, because the loop they were good at is the part being automated away.

• Framers — who can specify a problem, its constraints, and its success criteria before any execution starts. This is the tier that wins, and it’s thin.

The uncomfortable part is that the exposed tier is full of capable people. Skilled operators assume their fluency carries over. Some of it does — but the part that carries over is exactly the part that’s commoditizing, and the part that doesn’t carry over, framing, is the part that now decides everything. It was never the thing they practiced.

Access My Full Harness in The Exec Plan

How to actually frame

Access My Full Harness in The Exec Plan

Framing is a discipline with specific moves, not a personality trait. Three of them matter most.

Find the binding constraint before you delegate. Most problems aren’t what they look like on the surface, and an agent will optimize whatever you point it at — including the wrong thing. Solving the wrong bottleneck at agent speed is worse than solving nothing, because it produces motion that looks like progress. So the first act of framing is to ask: what is the one thing that, if left unaddressed, makes every other improvement irrelevant? You answer that before the agent opens, because the agent can’t.

Decide what “good” means before anything runs. When you worked in a chat window, you judged the output by eye, after the fact. You cannot eyeball fifty autonomous steps, and you certainly cannot eyeball a hundred parallel runs. So the judgment has to move earlier: you specify what a good result looks like in advance, in terms concrete enough that correctness is checkable without you re-reading everything. Evaluation stops being a final glance and becomes part of the frame. This is the hardest of the three, because it forces you to know what you want with a precision most work never demands.

Separate the mechanism from the story — because the agent won’t. Hand a model a compelling narrative and it will execute the narrative. It has no built-in defense against a frame that reads well and is structurally false. Knowing the difference between what is mechanically true and what is merely a good story is the difference between a system that compounds correctness and one that compounds a plausible mistake.

There’s a practical upshot. A thinking system — a repeatable way of mapping constraints, classifying what’s a real advantage versus a temporary one, separating mechanism from narrative — stops being a personal habit and becomes the thing you load into the machine. It’s the specification you hand the agent so it inherits your structure instead of improvising its own. The frame isn’t just the prompt you type. It’s the encoded judgment you front-load so the autonomous system runs inside it.

What the frame steers: the harness

Access My Full Harness in The Exec Plan

The frame is the human’s input; the harness is the system that executes it. The model is commoditized. The harness is where the real differentiation now lives.

A harness is the orchestration wrapped around the model: which agents you run, on what schedule, with what feedback loops, what approval gates, and what memory. The same model is available to everyone, so the model isn’t the edge. The harness is — and the ones that actually compound follow a handful of structural rules.

Use a swarm of specialists, not one mega-agent. A single agent asked to do everything degrades under load — the same reason companies have departments instead of one person who does all jobs. What works is a swarm of narrow specialists, each with one job, running on their own schedules. The agent producing content doesn’t need to know how a call-to-action gets tuned; the title optimizer doesn’t care about delivery. Specialization here is org design, not engineering.

Make the loop self-improving, or it’s just a cron job. An agent that runs the same playbook every day is automation that slowly decays. The agents that compound read their own results — which headline drew clicks, which copy converted — and feed that back into the next run. The durable asset isn’t the agent; it’s the learning loop.

Share one source of intelligence. If every agent independently asks “what’s working,” they drift and contradict each other. The fix is a shared module they all read from — one cached source of truth — so every decision is made from the same picture. Coherence is something you design in, not something you hope emerges.

Use gates, not guardrails. The instinct is to fence agents in with rules everywhere. What works is to let them run freely on low-stakes work — updates, optimization, scheduling — and put hard approval gates only in front of high-stakes actions: publishing, sending, spending. The ratio is the discipline: roughly 95% autonomous, 5% gated. Gate too much and you become the bottleneck, which destroys the speed that justified the system.

Treat memory as architecture. A system without persistent memory rediscovers its context every session, like a tourist. Structured memory — feedback rules, project state, preferences, reference pointers — loaded into every run is what lets the system accumulate knowledge. Mistakes become permanent rules; preferences become defaults. That accumulation is what turns a tool into an operating system.

The orchestration is the real engineering. Getting a model to write or send was never the hard part. The hard part is wiring eight or ten systems — content, analytics, delivery, social, payments — so an action in one triggers the right response in another. The plumbing between systems is the work now; the reasoning is the easy part.

Put those together and you reach the punchline: one person with judgment plus a harness produces what used to take a team. The old org chart for this kind of workload was a department — several content people, an SEO specialist, an email marketer, a social manager, an analyst, a developer, a designer. The harness collapses it to one person who supplies taste, judgment, and direction — the frame — while the harness supplies execution and pattern recognition at scale. The winners aren’t building AI. They’re building harnesses.

The computer left the desk

Access My Full Harness in The Exec Plan

Because the harness runs somewhere else, your physical relationship to computing inverts. The machine stops being a place you sit and becomes a process you steer from anywhere.

There’s a physical change underneath the cognitive one, and it took me a while to see it for what it was. I no longer sit in front of a computer to work. I set the frame from a phone, on the move, and the execution happens somewhere I’m not.

This isn’t a productivity habit. It’s the visible edge of a larger shift: the agent is becoming the computer. The old definition of a computer was a box you sat at — an app, on an operating system, on a chip, operated through a screen. The new definition, the one now being built straight into the silicon, is different: an agent is a brain (the model) plus a body (the harness) plus a runtime — wherever the work actually runs, whether that’s a cloud, a server, or eventually a robot. The runtime is the part that matters here, because by definition it isn’t your desk. Once the computer becomes the agent, it stops being a place you go and becomes a process running elsewhere that you hold a thin handle to.

This is also the end state of a longer progression. AI first ran in a sandbox, then learned to use tools, then learned to operate a whole computer, and finally converges with the computer itself — and convergence, experienced from the user’s side, feels like this: there’s no chair. The interface stopped being clicks and commands and became intelligence. You state a goal; you approve an outcome. And once the interface is intelligence, the form factor is free. A goal doesn’t need a keyboard. An approval doesn’t need a monitor. The big screen and the desk were artifacts of operating — and operating is exactly the part that just left.

What remains for the human collapses to two acts, and both are light enough for any device: set the frame, and clear the gate. A goal goes out; an approval comes back. Neither needs a workstation. A phone does it now. The next surface won’t even be a phone — it’ll be some AI-native object closer to an always-on personal assistant than a PC, something you speak goals into and tap approvals on. The screen shrinks because the work it used to hold moved off it.

Call the pattern the thin surface: when the interface becomes intelligence, the human’s surface shrinks to two interactions — frame-in and gate-out — that fit any device and untether the work from the desk. The harness runs remote; you carry the handle.

The strange part is how natural it feels. For three decades, “using a computer” meant arranging my body in front of one. Now the computer is a process I start with a sentence and check with a tap, from wherever I happen to be standing. I didn’t leave the desk on purpose. The work left, and I followed it.

Where the whole industry is heading

Access My Full Harness in The Exec Plan

The same logic that reorganized my desk is reorganizing the frontier. The entire industry is converging on a single contest — to own the harness.

Follow the harness up the stack and the macro picture falls into place.

The frontier AI labs are running deeply negative margins on heavy agentic usage. The easy reading is “the economics are broken.” The structural reading is the opposite: the harness is where the next moat sits, and the labs are paying to build it. Two facts make that rational rather than reckless. First, token costs are falling by orders of magnitude — what’s subsidized today is cheap to serve tomorrow, so the loss is really a financing cost on a future-cheap good. Second, once someone runs their whole operation through one lab’s harness, the cost of leaving is no longer the model — it’s the accumulated memory, the tuned loops, the wired-up orchestration. You can swap a model in an afternoon. You can’t swap a harness you spent a year teaching.

That’s why the competitive action is migrating to which features get withheld from subscription tiers and which capabilities stay API-only. The fight isn’t over raw intelligence anymore — that’s converging and commoditizing. It’s over the harness layer, because that’s where lock-in is built and where the margin eventually returns. And it’s migrating to the surface, too: if the computer is now a remote process plus a thin handle, then the runtime and the device that holds it become control points — which is why the hardware makers are suddenly racing to ship an always-on personal-agent device, not just a faster chip. Whoever owns where the agent runs, and the surface you steer it from, owns the next relationship with the user.

So here is the whole picture in one line. The model has commoditized. The frame is the human’s edge. The harness is the system’s edge. And the contest to own the harness — fought with the same logic from a solo operator’s phone to a trillion-dollar lab’s data center — is the real story of the AI industry right now. Everyone is arriving at the same realization: you don’t win by having the AI. You win by owning the harness around it, and by bringing a frame sharp enough to steer it.

The career consequence is direct. The old way rewarded lots of operators each working their own loop. The new way rewards a few people who can frame well, each steering many autonomous runs on a harness that compounds. Headcount in the operating tier shrinks; leverage concentrates in the framing tier. And the gap between a good frame and a bad one stops being measured in minutes — it’s measured in the output of the whole machine, correct versus confidently wrong, and invisible until it ships.

That’s the moment we’re in. Having the tools is table stakes. Operating them is automating. What’s left for a person is the part no model supplies for itself: the harness it runs on, and the frame you build before you let it run.

When people ask what I actually do now, the honest answer is: less, and more. Less operating — the harness handles that. More framing — which turns out to be the harder and rarer half. I live in the harness. But the harness doesn’t run on its own. It runs on the frame I bring it each morning, from wherever I happen to be standing. That’s the whole job now, and it may be the whole job for a while.

Key Takeaways & Mental Models — The Framing Models

The Framing Ladder — When a capability goes universal, the edge moves up the stack. It has now moved past operating the model to framing the work before it runs.

The Vanishing Loop — Autonomy removes the turn-by-turn iteration that let a weak first attempt get fixed. With no loop to recover in, the first frame decides the result.

Confidently Wrong at Scale — A bad frame isn’t caught; it’s executed. A small error in the frame becomes a large, polished, plausible error at the speed and volume of the machine.

Pre-Encoded Judgment — When you can’t watch every step, you have to define “good” before execution, not after. Evaluation moves from a final glance to part of the frame.

The Thin Surface — When the interface becomes intelligence, the human’s surface shrinks to two acts — frame-in and gate-out — that fit any device and lift the work off the desk.

The Frame Is the Moat — The one input that hasn’t commoditized is the structure you bring before the agent runs. Tools are universal; operating is automating; the frame is the edge that compounds.

The Harness Principles

Access My Full Harness in The Exec Plan

Swarm Over Mega-Agent — Narrow specialists on independent schedules beat one generalist. Specialization is org design.

The Self-Improving Loop — Agents that read their own performance and adjust compound; static automation decays. The loop is the moat.

Shared Intelligence — One cached source of truth keeps agents coherent; independent lookups produce drift and contradiction.

Gates, Not Guardrails — ~95% autonomous on low stakes, ~5% gated on high stakes. Over-gating makes the human the bottleneck.

Memory Is Architecture — Persistent structured memory turns mistakes into rules and a tool into an operating system.

Orchestration Is the Architecture — The plumbing between 8–10 systems is the real engineering. The reasoning is the easy part now.

Harness Theory — One operator with judgment plus an agent swarm replaces a department. The human brings the frame; the harness brings velocity

With massive ♥️ Gennaro Cuofano, The Business Engineer

I’ve been explaining since 2022, why AI is a sypercycle that looks way more like the semiconductor industry.

AI is not following the playbook of the consumer internet. It is following the playbook of the early semiconductor industry — capital-heavy, physical, strategically vital, and entangled with state power from the start.

On June 12, 2026, one US directive switched off Anthropic’s two most capable models for every foreign national on Earth — including the company’s own foreign staff. Anthropic complied within hours. Every customer on the planet lost access to the frontier tier.

Two things follow. First, the geopolitical layer at the top of the AI stack is not just a place where alliances get built. It is also a kill switch. Second, open-source frontier models — the supposed escape hatch — are not actually an escape, and the semiconductor analogy tells you exactly why.

What Happened

Anthropic released Fable 5 on June 9 — a model from the new Mythos class, the tier above the company’s Opus class — and made it generally available. Safety classifiers quietly route flagged queries down to the less capable Opus 4.8. Its sibling Mythos 5, the same model with the cyber safeguards lifted, went to a small group of cyber-defenders.

At 5:21pm ET on June 12, an export-control directive arrived. Under US national-security law, no foreign national could use either model — anywhere in the world, including Anthropic’s own foreign-national employees. There is no way to check the passport of every API request in real time, so the only way to comply was to switch the whole tier off for everyone. Every other Anthropic model stayed up. The frontier tier went dark.

Anthropic complied while disagreeing in public. Its view: the underlying concern was a narrow jailbreak — getting the model to read a codebase and patch its flaws — that does not give a real attacker any meaningful uplift and that other shipped models, including OpenAI’s GPT-5.5, can already do. Anthropic asked, again, for a transparent legal process for these kinds of blocks rather than ad hoc directives.

The government’s view is also coherent: a frontier model that can reason about cyber and biology is dual-use, and under a national-security frame, the safe move is to halt first and adjudicate later.

Both sides have a defensible case. The interesting question is not which one is right. It is what this episode reveals about how the AI industry is actually wired.

Why This Is the Chip Era, Not the Internet Era

For thirty years, the playbook for thinking about technology was the consumer internet. Move fast, ship globally, let regulation catch up later. The companies that won the era — Google, Amazon, Meta — won partly because the state mostly stayed out of the way at the moments that mattered.

AI does not work like that. It looks much more like the early semiconductor industry: expensive, physical, slow to build, and tangled up with state interest from day one.

In semiconductors, the state has never been absent. It funded the original research. It built the alliance system that decides who gets access to which generation of chips. It uses export controls to deny rivals what it has no commercial reason to deny. It treats the foundries themselves as strategic assets. The result is the world we live in: a handful of foundries, an oligopoly of equipment makers, capability gates between national blocs, and a layer of state authority threaded through the whole industry.

AI is being built inside that same kind of architecture. The deep capital stack of AI has six layers: geopolitical alliances at the top, then capital flows, then energy and minerals, then physical and digital infrastructure, then hardware, and finally software at the bottom. At every one of those layers, the state is a player, not a referee.

The U.S.–Saudi $600 billion partnership announced last year is the clearest example of how this works when everyone is cooperating: alliance-building at the top, joint investment vehicles, GE gas turbines and a dedicated energy fund, DataVolt’s $20 billion data-center build, 18,000 NVIDIA Blackwell GPUs and a $10 billion AMD agreement, and a Saudi national AI champion (Humain) sitting inside the US software ecosystem. Layer by layer, the two countries are locked together. Switching costs are designed in.

That is what the top layer looks like when it is building. June 12 is what it looks like when it is cutting.

Read more

Their team bought one of each Anthropic and OpenAI subscription plan, ran long-horizon coding tasks until the weekly limits were exhausted, and back-calculated what each subscription actually delivers at API-equivalent pricing.

The headline numbers — up to $14,000/month of usage on a $200 ChatGPT Pro plan and $8,000/month on Claude Max 20x — are 4× to 7× more generous than the widely-circulated ~$2,000/month rule of thumb.

They then modeled subscription gross margin under the assumption that the underlying API carries a 75% margin.

The result is the heat map circulating across timelines this week: ChatGPT Pro 20x runs -1,650% margin at full utilization, Claude Max 20x runs -900%, and every tier crosses into negative territory once a single power user pushes past roughly 10–20% utilization.

Their conclusion: openly nerfing subscriptions is off the table because public backlash is asymmetric to the savings, so the labs will instead withhold new features and models from subscription plans — with the explicit prediction that Mythos may end up being API-only.

We agree with all of this. The experimental approach is clean primary research, the heat map is the right framing of the data, and the policy prediction — vintage withholding rather than usage caps — is the correct structural answer to the subsidy problem.

This piece picks up where SemiAnalysis stopped. The same charts are already being read by a second wave of commentary as “AI subscriptions are a bubble” — confusing accounting margin with strategic position. The deeper structural read goes the other direction: the negative margin is not a leak, the subscription is doing three pieces of strategic work simultaneously, and the timeline on which this trade pays off is the timeline of token deflation, not the timeline of quarterly earnings.

Three things the bubble framing misses, and SemiAnalysis touches but does not fully unpack:

1. The negative margin is not a leak. It is a procurement budget. The labs are not pricing the product; they are buying an input that is not for sale anywhere else: agentic workload signal at the frontier of difficulty.

2. The subscription is a call option on token deflation. Lock in $200/month today against an API-equivalent value that compounds downward by an order of magnitude per year, and the same plan that prints -400% margin today prints +50% margin in 24 months — without changing a single line item.

3. The power user is the moat. A user willing to pay $200/month and burn through $14K of equivalent usage on long-horizon coding tasks is not a churn risk. They are the substrate on which the next frontier is trained, and the only seat from which the deflation curve becomes ownership.

This is not a bubble. It is the harness training auction — the most important strategic transaction in AI right now, and the one nobody on the surface reading is pricing correctly.

The Three-Asset Subsidy

The subscription is not one product. It is three assets bundled at a single price.

Asset 1: Harness training signal

The frontier moat has shifted. Model weights are increasingly fungible — Sonnet, GPT, Gemini, and Grok cluster within a few points on benchmarks that mattered last year. What separates them now is the harness — the agentic scaffold around the model: tool use, planning, memory, error recovery, multi-step orchestration, browser control, code execution.

This is what Harness Theory predicts: when the model commoditizes, the moat migrates to the scaffolding. And the scaffolding cannot be trained on benchmarks. It can only be trained on real long-horizon workloads — the kind a $200/month power user runs for 14 hours a day, four weeks a month.

Asset 2: Power user lock-in

Not the casual user. The casual user pays $20 and uses 2% of capacity — they are the cross-subsidy. The asset being bought is the high-utilization user: the developer who has wired their entire workflow around Claude Code or Codex CLI, who has internalized the harness, who pays $200/month and would scream if it disappeared. These users are non-substitutable in three directions: workflow ($200/mo plus a year of muscle memory), API equivalence (their usage maps to $14K/month, which they will not pay), and competitor switching (the harness is sticky in a way the model is not).

Asset 3: Token deflation call option

This is the asset nobody is pricing. Token prices have fallen ~10× per year for two years running. Opus-class capability that cost $75/MTok in 2024 is delivered in 2026 for roughly $7-15/MTok depending on routing. If the trend holds — and there is no structural reason to expect it not to — the $14,000 of API-equivalent value being delivered today on a $200/month plan becomes $1,400 next year. The subscription price is locked. The cost of delivering it is collapsing. The trade is geometric on the cost side, linear on the revenue side, and the labs are sitting on a multi-year barbell.

The bubble framing sees -1,650% margin and reads collapse. The structural read sees -1,650% margin and reads purchase order: the labs are spending negative margin today to acquire the three assets that determine who wins when the deflation curve crosses zero.

The Math of the Trade

The heat map exposes a structural asymmetry that deserves attention.

The break-even utilization thresholds are not symmetric across vendors:

OpenAI is running a more aggressive subsidy curve. ChatGPT Pro 20x breaks even at almost half the utilization of Claude Max 20x, and bleeds nearly 2× the margin at full utilization. This is not noise.

This is OpenAI declaring that the harness signal is worth more to them — either because they are further from the agentic frontier and need catch-up data, or because their consumer surface is the strategic priority and Anthropic’s enterprise tilt allows it to be more disciplined with consumer subsidies.

Either reading is structurally meaningful. The dollar amounts are also strategically meaningful. ChatGPT Pro 20x delivers $14K of equivalent usage to Claude Max 20x’s $8K — a 75% premium on the subsidy budget at the same sticker price. OpenAI is putting more chips on the table per power user, which means OpenAI is betting harder on the three-asset thesis than Anthropic is.

A useful frame: this is a backward auction. In a normal auction, buyers bid up. Here, the labs bid down — the more aggressive they get with subsidies, the worse their unit economics look on the surface, but the more harness signal and lock-in they win. The accounting metric (gross margin) and the strategic metric (harness density) are anti-correlated by design. Treating one as a proxy for the other is the categorical error driving the bubble narrative.

What the math actually says: at current utilization distributions — which are still mostly low-intensity casual usage with a long tail of power users — blended margins are likely positive or near zero. The heat map shows the worst case per cohort, not the blended reality. The labs are running a fat-tail trade where the casual masses cross-subsidize the power users, and the power users deliver the strategic asset.

Read more

Anthropic is not only pushing ahead on model capabilities. It is becoming the first major AI lab actively attempting to define the governance layer of the AI stack.

Many people still see AI as a race driven primarily by research breakthroughs and technical progress. That view is increasingly incomplete.

The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment. It is where decisions are made about who gets access to frontier intelligence, under what conditions, and for which purposes.

This is precisely why I have argued for years that the AI race cannot be understood as a purely technological competition. As the U.S. and China continue to disentangle across critical technologies, AI becomes one of the central arenas through which that broader geopolitical realignment plays out.

Anthropic’s latest move reflects that shift. The company is not merely building models. It is helping define the rules, permissions, and boundaries around frontier intelligence. And it is doing so in one of the most assertive ways we have seen from any AI lab so far.

As an analyst, my job is not to tell you whether this is good or bad. My job is to explain what is happening.

This is the emerging structure of the AI era.

Get used to it.

On June 9, 2026, Anthropic did something no AI lab had done before. It took one extraordinarily capable model and released it as two different products at the same time. The launch will be remembered for its benchmark scores. The part that actually matters is structural — and almost no one is talking about it.

This piece explains what happened, why the safety design is the most important business decision in the release, and how that one decision ripples out to reshape pricing, products, competition, and the shape of the AI market. No prior knowledge needed.

Read more

For three years, the AI industry has debated which layer of the stack captures value: the chip, the cloud, the model, the application. The debate has assumed the interface layer — how humans actually talk to computers — would remain roughly where it has been since 2007: tap an app, type into a box, swipe across a screen.

WWDC 2026 was Apple quietly signaling that this assumption is over.

The most structurally significant thing Apple announced last week is not a new model, a new chip, or even a new design language. It is an operating system in which the primary user of an app is no longer a human — it is an agent. Spotlight becomes a knowledge graph the agent reads from. App Intents becomes the surface the agent acts on. View annotations let the agent see what you see. Siri AI becomes the broker that calls apps on your behalf.

This is the agent-as-computer move. And Apple, the company most often described in 2025 as “losing the AI race,” is now better positioned than any competitor to execute it — not because Apple has a better model, but because Apple owns the only thing that matters when the agent becomes the computer: the place where the agent runs.

That is the whole story. Everything else is mechanics.

A map first, so the rest reads cleanly

The AI industry runs across roughly seven layers. From the bottom up:

1. Capital sources — sovereign wealth funds, NVIDIA’s equity book, private credit, big tech cash piles.

2. Funding vehicles — Project Stargate, neocloud SPVs, joint ventures designed to spread the capex risk.

3. Physical inputs — energy, real estate, high-bandwidth memory, fab capacity.

4. Compute and silicon — NVIDIA in the data center, custom silicon at the hyperscalers, Apple silicon on devices.

5. Cloud operators — AWS, Azure, Google Cloud, Oracle, and the emerging neoclouds.

6. AI labs and foundation models — OpenAI, Anthropic, Google DeepMind, xAI, Meta, Mistral, DeepSeek.

7. Application and governance — the operating systems, apps, and surfaces where end users actually meet AI.

Most analysts spend almost all their attention on Layers 4 through 6, because that is where the headlines are. The chip race, the cloud race, the model race.

The interesting question is not which layer is winning. It is which layer captures the rents over the next decade. And the structural pattern across previous platform shifts — desktop, web, mobile — is consistent: the layer that owns the interface to the human eventually captures the rents the layers below it generate.

That is the layer Apple owns. It always has.

Join Yearly Premium Membership To Get The Builder-PM

Access For Yearly Premium Member!

Read more

Two new chapters anchor the structure. Part IV anatomizes the role in operating detail. Part V is the transition — written for individual PMs, PM leaders at incumbents, and founders hiring into AI-native cells. An appendix compiles the named mental models from the full work into a single field guide.

How to read this book

The book is built to be read straight through, but each part stands alone.

• Individual PMs may want to start with Part IV (the role) and Part V (what to do Monday).

• PM leaders at incumbents will find Part III (the operating environment) and Part V the most actionable.

• Founders should read in order — the strategic frame in Part II will be the most expensive to skip.

PROLOGUE

The Discipline That Just Got Rewritten

In late 2024, a tech company posted a Senior Product Manager job. The description was standard: own roadmap, write specs, partner with engineering, run quarterly business reviews, drive adoption metrics. Standard ladder, standard expectations, standard interview loop.

Six months later, the same company — same business unit, same level — posted a role with the same title. The description had changed almost beyond recognition. Frames strategic bets on capability that is six to twelve months ahead of public release. Prototypes specifications directly using agentic development tools. Operates inside a small, autonomous unit with broad decision rights. Direct technical fluency required.

The title was identical. The job was not.

This is not a story about one company. The pattern is now visible across every AI-native employer and a growing share of incumbents — the role called “Product Manager” has fractured into two distinct disciplines that share a name and almost nothing else.

One is the role product management has been for the past fifteen years — scoped around discovery, delivery, stakeholder coordination, and the translation function between engineering and the rest of the company. The other is something new. It writes specifications by running them. It operates inside a unit with no roadmap document, no stakeholder review, no quarterly KPI envelope. It ships product against capability that does not yet exist in any public model. It reports outside the conventional product org because the conventional product org would kill it within a quarter.

The new role is what this book is about.

Three eras of product management

The cleanest way to see the change is through a historical lens. Product management, as a recognized discipline, has gone through two prior eras, each defined by the operating substrate of its time. The AI era is the third, and it is restructuring the substrate underneath the role.

The Cagan era — roughly 2008 to 2018. Marty Cagan’s Inspired codified the modern PM role as a four-risk function: value, usability, feasibility, viability. Dual-track agile separated discovery from delivery. The job was to find the right thing to build through structured customer research, validate it with prototypes, and partner with engineering to ship it.

• Unit of work: the user story.

• Primary value: judgment about what to build and the discipline to keep the team focused on it.

• Tools: Jira, Aha, Productboard.

• Substrate: a product organization growing from tens to hundreds of engineers, where coordination was the binding constraint.

The Lenny era — roughly 2018 to 2024. Lenny Rachitsky’s newsletter became the canonical reference, alongside an explosion of frameworks: growth loops, North Star metrics, retention curves, the four fits, jobs-to-be-done refinements, OKRs at scale. The job was to drive measurable outcomes — typically activation, retention, monetization — through a combination of strategic prioritization, growth experimentation, and operational coordination across function-specialized teams.

• Unit of work: the experiment.

• Primary value: framework fluency and the ability to operate inside an increasingly specialized org.

• Tools: Amplitude, Mixpanel, Statsig.

• Substrate: the consumer-software and SaaS organization, scaling toward thousands of employees, where measurement and alignment were the binding constraints.

The AI era — 2024 onward. Cat Wu’s March 2026 essay for Anthropic, Product management on the AI exponential, named the shift in operating terms that the role had begun absorbing. The Builder-PM frames strategic bets on capability ahead of release, prototypes specs directly with agentic tools, and owns continuous validation loops that replace ad-hoc project cadence.

• Unit of work: the loop.

• Primary value: bet selection and direct execution capacity — the ability to spec, build, and validate without the handoffs and translation layers that defined the prior eras.

• Tools: Claude Code, Cursor, agent orchestrators, evals harnesses.

• Substrate: the founder cell — a five-to-ten-person autonomous unit reporting outside the scaled product org.

• Binding constraint: no longer coordination or measurement. It is taste: choosing what to bet on, given that the constraint of building is collapsing.

The substrate, not the title

The temptation, when looking at the new role, is to assume what has changed is the seniority bar — that the Builder-PM is just a higher-skilled PM, the same job done better. This reading is comfortable and almost entirely wrong. What has changed is not the bar. The substrate the role sits on has changed.

The conventional PM existed because a software organization above a certain size produced a coordination problem that needed a dedicated function to absorb.

• Specs had to be written because engineers and designers spoke different languages and operated on different cycles.

• Roadmaps had to be maintained because stakeholders across the company needed visibility into a shared queue.

• Stakeholder reviews had to be run because the consequences of building the wrong thing — at the scale of a 50-person product org — were too expensive to absorb informally.

The PM job was a consequence of the org shape it lived inside. When that org shape changes, the role’s reason for existing changes with it.

The founder cell described later in this book has different structural properties than the scaled product org:

• Five to ten people instead of fifty.

• Direct reporting to the CEO or President, instead of through a five-layer hierarchy.

• Broad unilateral decision rights, instead of a stakeholder-driven roadmap process.

• Communication overhead of roughly one-tenth that of a comparably-staffed conventional unit.

Inside that envelope, the conventional PM job has nothing to do. There is no stakeholder layer to translate to. There is no roadmap document to maintain because the roadmap is the live set of bets the unit is running. There is no spec to write upstream of engineering because the engineering capacity is built into the role and the spec is a working prototype by end of day.

What replaces the conventional PM job is not a more senior or more technical version of it. It is a different job that happens to share the same name.

Who this book is for

Three readers will get the most out of it.

The individual PM — particularly the experienced one with five to fifteen years in the conventional discipline — who can sense the discipline shifting underneath them and wants a clear-eyed account of what’s happening and what to do about it. The book is not an optimistic pep talk. It is also not a doom narrative. It is a structural account of a role being rebuilt, with practical implications for what to learn, what to unlearn, and what career bets to make.

The PM leader at an incumbent company — VP of Product, CPO, Head of Product — who is responsible for an organization that is increasingly the wrong shape for the work that is now possible to do. The book is honest about the difficulty of carving founder cells inside scaled product orgs without triggering organizational immune responses, and it offers specific structural moves that have been demonstrated to work.

The founder or CEO — particularly at AI-native startups, but also at incumbents pursuing AI-native bets — who is making decisions about hiring, organizational design, and strategic frame at a moment when conventional product organization advice will produce predictable underperformance. The book is the operating manual for a unit shape that did not exist as a standard pattern five years ago and now defines the product velocity of the companies that are pulling ahead.

The argument in one paragraph

Product management is being rebuilt by a structural change in the AI capability curve. The gap between what frontier models can do and what is shipped into products has widened into a sustained product overhang, which rewards a different strategic posture (building for capability that does not yet exist) and selects for a different operating environment (the founder cell — small, dense, reporting outside the scaled product org) and a different role inside it (the Builder-PM — direct execution, agentic tools, bet selection, loop ownership).

Part II addresses the strategic frame. Part III addresses the operating environment. Part IV addresses the role itself. Part V addresses the transition, for individuals and organizations both. The appendix compiles the named mental models into a reference field guide.

The discipline is not dying. It is being rewritten. The remainder of this book is the rewrite.

Join Yearly Premium Membership To Get The Builder-PM

Access For Yearly Premium Member!

Or

Get It Separately

Read more

The real question is what arrangement of those people produces output disproportionate to their headcount, and the answer turns out to be geometric, not motivational.

Across two decades and many industries, a recurring organizational pattern shows up wherever 0-to-1 building happens at speed. Skunk Works at Lockheed in the 1940s. The original AWS team inside Amazon. Stripe’s API team in its first few years. Instagram inside Facebook for roughly four years post-acquisition. The original iPhone team. More recently, Anthropic Labs — the five-person unit that shipped Claude Code, MCP, Skills, and the Claude desktop app, and that now sits at the center of the discourse on AI-era product organizations.

These units differ in domain, era, and the specific products they ship. They share a structural signature: small, dense, reporting outside the scaled product org, with role composition that does not match the org chart anywhere else in the company. The signature is what this piece reverse-engineers.

The piece is the founder cell at unit-resolution. The headcount window. The role composition. The negative space — what is deliberately absent, and why each absence is a feature. The operating practices that follow from the composition rather than being arbitrary choices. The reason incumbents cannot replicate the unit by hiring more of the same roles into their existing product org.

The frame is anatomy, not doctrine. We are taking the unit apart to see how it works.

Join Yearly Premium Membership To Get The Builder-PM

Access For Yearly Premium Member!

Read more

That is now changing. Not collapsing — changing.

Underneath the headline number that says NVIDIA’s revenue keeps climbing, the silicon layer of the stack is fracturing in three directions at once: hyperscalers are building their own chips for their own workloads, a generation of specialty silicon startups is attacking the parts of the workload the GPU was never optimal for, and a much smaller number of foundry and packaging providers are quietly emerging as the binding constraint that bounds all of it.

The most monopolistic layer of the AI stack is becoming the most differentiated.

This piece maps that fracture. Four layers as before: the abstraction (why the monopoly was structurally finite), the market map (who is building what), the playbook (three observable shifts in how silicon strategy is converging), and what comes next (where the leverage moves once the GPU’s generalism stops being the winning hand).

For Premium Members:

Request Access To HQ Full Map Here!

Get The AI Map Full Visual Report

Get The Business Engineer Agent Here!

Try The Business Engineer Agent Here

Read more

Overhang.

For most of software’s history, the overhang was small. Your stack moved approximately as fast as your team could ship features against it. The two curves stayed close. Product roadmaps were exercises in execution, not in temporal arbitrage. You built for the capability you had; the capability you would have next quarter looked roughly like the capability you had this quarter, plus integrations.

Frontier AI broke this symmetry. Model capability now compounds on a six-to-twelve-month doubling cycle. Per METR’s 2026 benchmark, the time horizon a model can autonomously execute against has expanded roughly 41× in 16 months — from approximately 21 minutes on Sonnet 3.5 to approximately 12 hours on Opus 4.6. Inference price per equivalent unit of work has collapsed by a similar order of magnitude over the same window. The product capability built around those models has not, and structurally cannot, move at the same pace.

Most product organizations are built to ship for the curve they can see today. The frontier moves while they ship. The result is a permanent, widening gap, and that gap is now the most important strategic surface in software.

The Product Overhang Doctrine is the bet that you can build into the gap before it closes — shipping features for capability that doesn’t yet exist, knowing the capability will arrive on the next model release, and absorbing the pre-PMF months because the alternative is shipping into a frontier that has already moved past you.

This piece is the doctrine at depth. How it works. Who is executing it and who isn’t. The four ways it fails. The eight operating practices that make it survivable. The organizational form required to run it inside a company. And the structural choice every org now faces about which curve it’s competing on.

Join Yearly Premium Membership To Get The Builder-PM

Access For Yearly Premium Member!

The architecture of the overhang

The two-curve mechanic

Plot two curves on the same axes. The horizontal is time. The vertical is capability.

The first curve is model capability — what the frontier model can actually do, measured along whatever axis matters for your product (context length, reasoning depth, code generation, multi-turn coherence, agentic task horizon, multimodal fluency, tool-use reliability).

This curve compounds. New architectures, new training data, and new post-training techniques each push it upward, and the cycle now repeats every few months rather than every few years.

The underlying mechanism is scaling laws — empirical relationships among compute, data, parameters, and capability that have held with surprising stability since 2020 and for which no one has yet found a clean ceiling.

The second curve is product capability — what users of your specific product can actually do, end-to-end, in production. This curve does not compound. It moves at the rate of your release cadence, gated by PRDs, design reviews, engineering sprints, QA, and rollout. In an exceptionally well-run org, it inflects two or three times a year. In most orgs, it influences annual planning cycles.

The result is structural. Model capability is exponential; product capability is approximately linear. Once an exponential and a linear function are plotted on the same axes, the gap between them widens monotonically. There is no execution speed at which a linearly improving product can keep pace with an exponentially improving substrate. The gap is not a problem to be solved through better engineering; it is a structural feature of the era.

Three properties of overhang

Directionality. The gap is asymmetric in direction. Models leave products behind, not the other way around. There is no scenario in which product capability outruns model capability over the long term, because every product is built on top of the model. The arrow points one way, and any roadmap that assumes otherwise is mispriced.

Dimensional asymmetry. The compounding is not uniform across capability dimensions. A model release might extend the agentic task horizon by 10× while only modestly improving multilingual reasoning. Your overhang is dimension-specific. Building into a dimension that doesn’t compound on your horizon is the silent failure mode every builder underestimates. Picking the right axis matters more than executing well against the wrong one.

Decay. Any individual overhang bet decays. The gap you saw six months ago is partially closed by the next model release, partially closed by competing product teams, and partially closed by the user’s evolving expectations. The bet has a half-life. You ship into the gap, you capture the arbitrage, then the gap closes, and you need the next bet ready. Operators who run the doctrine well are not running one bet; they are running a pipeline of bets with staggered horizons.

Read more

This framework — Harness Theory in earlier pieces — sat alongside a related call about Microsoft specifically: the official Microsoft AI doctrine through most of 2025 and early 2026 (Suleyman’s “tight second” line, the argument that being three to six months behind the frontier was a feature not a bug) was strategic cope. Not because trailing the frontier is structurally bad — it isn’t — but because trailing without a credible frontier path of your own turns the product stack into a captive distribution channel for someone else’s IP. In Microsoft’s case, the someone else was increasingly Anthropic.

The Anthropic threat to Microsoft’s enterprise position became visible the moment Claude was integrated into Microsoft 365 Copilot in September 2025. Once the product stack accepted that frontier models are interchangeable inputs, the competitive geometry inverted. Anthropic was no longer a model supplier sitting underneath Microsoft’s distribution — it was a frontier provider with direct enterprise mindshare, increasingly preferred by developers, increasingly the model behind agentic coding workflows, and increasingly positioned to disintermediate the parts of Copilot that actually generate revenue. The product stack was eating itself: Microsoft was distributing the competitor that could erode its own enterprise franchise.

Build 2026 is the closure of that exposure. The seven MAI models, the Foundry hosted-agent runtime, the MXC containment layer on Windows, the IQ context stack, and — most importantly — Frontier Tuning with customer-owned reinforcement learning environments are not separate announcements. They are the operational answer to the harness-as-moat thesis. Microsoft just shipped the closed loop: customer agent traces feed RLE training, RLE training improves MAI models, improved MAI models run better on the Microsoft harness, the harness compounds, the loop tightens, the dependence on Anthropic and OpenAI moves from structural to optional.

Satya’s framing on stage was unambiguous: “the time has come for every company to just move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem.” That is the public retirement of the tight-second doctrine, and the public adoption of the harness-flywheel thesis. The internal record confirms the reversal was already in motion: Microsoft formed the Superintelligence team under Suleyman in November 2025; the March 17, 2026 reorganization narrowed Suleyman’s mandate to frontier models while Jacob Andreou took unified Copilot leadership; Nadella’s memo: “Progress at the AI model layer is more critical than ever to our success as a company over the next decade”; Suleyman to CNBC on the same day: “The model is the product” — the exact inverse of his April 2025 framing.

The architectural fact behind Build 2026 is now legible. Nadella owns the product stack and the platform narrative. Suleyman owns the frontier shop. Two leaders, two mandates, one company — and one closed-loop flywheel that the absolute-frontier model providers cannot replicate without becoming hyperscalers themselves.

This piece works through the four-layer analytical reading of the announcements, then closes with a focused competitive assessment of the MAI stack specifically — because the headline benchmarks obscure what’s actually being built.

What Structurally Changed

Read more

Don’t build for today's model. Build the model in six months.

Cherny followed it. He built Claude Code against a frontier that did not yet exist. For the first six months, the tool was barely usable — he used it for maybe 10% of his own code.

Then Opus 4 shipped in May 2025, and the curve inflected with every release thereafter. By late 2025, he had stopped writing code by hand entirely. By early 2026, he was producing up to 150 pull requests per day on his phone.

The five-person Labs team that produced this also produced MCP, Skills, and the Claude desktop app — then disbanded. In January 2026, Anthropic spun it back up, this time as a permanent function.

The Chief Product Officer — Mike Krieger, Instagram co-founder, two years at the company — gave up his executive title, became a Member of Technical Staff, and moved into Labs to build alongside Mann. They report directly to Daniela Amodei, the President.

The structure they built around the doctrine is what the rest of this analysis is about.

For Premium Members:

Request Access To HQ Full Map Here!

Get The AI Map Full Visual Report

Try The Business Engineer Agent Here

Read more

The real destination is the convergence of the agent and the computer itself. Rather than AI operating software through existing interfaces, the intelligence layer will become the interface. In that sense, agents won’t just use computers; they will effectively merge with them.

This is what makes AI a true second computing revolution. The first revolution gave us graphical interfaces that allowed humans to interact with machines. The second replaces those interfaces with intelligence, shifting computing from clicks and commands to goals and outcomes.

I’ve called this the agentic expansion cascade.

In its latest GTC, NVIDIA practically confirmed that!

NVIDIA stopped selling the fastest chip and started selling the unit of computation itself — the agent — turning a chip monopoly into a toll on every watt the world converts into tokens.

Everything below is the machinery behind that sentence, with the engineering numbers translated into what they actually mean for the money — and a deeper look at how the whole thing propagates, layer by layer, until it hits the places NVIDIA can’t control.

Read more

They separate. But the choice of posture is the surface of a deeper picture: three hidden games run beneath it. A regulatory game over who writes the rules. A geopolitical game over which bloc wins the frontier. And an open-source game where Western open-source AI is alive only because the closed frontier has been quietly subsidizing it through distillation.

The three games explain why each lab is making the choice it is — and why the discipline premium has consequences that reach far past Anthropic’s balance sheet.

For Premium Members:

Request Access To HQ Full Map Here!

Get The AI Map Full Visual Report

Get The Business Engineer Agent Here!

Try The Business Engineer Agent Here

Read more

Two new layers earned their place. The first is the agentic harness — everything that surrounds a foundation model and makes it usable in production. The second is governance — the deliberate, paced release of frontier capability. Both layers were dormant a year ago. Both are now structural.

This is the workshop, compressed into a written piece. Four layers of analysis, in the Business Engineer format: the abstraction, the market map, the playbook, and what’s coming next.

For Premium Members:

Request Access To HQ Full Map Here!

Get The AI Map Full Visual Report

Get The Business Engineer Agent Here!

Try The Business Engineer Agent Here

Read more

The first is the layer view — the stack from energy to governance, with a bottleneck on every floor. It tells you where physics binds, where capital cannot solve the problem, where time is the only variable. This is the view that dominated the discourse for most of 2024 and 2025.

The second is the geometry view — the three (and a half) patterns of rent capture across that stack. Vertical, horizontal, flywheel, and the niche-by-adjacency play that the small winners use to beat the giants on their own ground. This view tells you what kind of company you are looking at, not just which slot it occupies.

The third — and the one this piece is built around — is the matrix view. You score every relevant player across every layer of the stack, on a scale from absent to dominant. What emerges is not a list of competitors; it is a heatmap of who can actually deliver what, today. The matrix is the only view that surfaces the gaps each player must fill, the dependencies they must accept, and the partnerships they are quietly being forced into.

The layers tell you where the physics is. The geometries tell you how to capture rent. The matrix tells you which players will still be at the table in five years, and what they will be willing to trade to stay there.

Three layers of analysis, in the Business Engineer format: the abstraction (why a matrix view), the market map (the matrix itself), the playbook (the cascades it reveals), and what comes next.

For Premium Members:

Request Access To HQ Full Map Here!

Get The AI Map Full Visual Report

Get The Business Engineer Agent Here!

Try The Business Engineer Agent Here

Read more