Post-AGI First Principles

Assumptions

The planning horizon is five years.
Recursive self-improvement begins within the planning horizon.
Property rights remain broadly enforceable within the planning horizon.
For the current iteration, frontier intelligence remains centralized and tightly controlled.
For the current iteration, core robotics capability is solved soon after AGI and deployment is the main bottleneck.
Broad political rent shaving of AGI-adjacent bottlenecks occurs within the planning horizon.
The buildup toward rent shaving becomes visible before the market fully prices it.
Earth remains the main locus of human life and economic activity.
The investment problem is about the transition, not a distant steady state.

High Confidence

High-Confidence Statements

High-quality Earth land remains scarce.
Human nature remains human.
States and legal systems remain powerful.
Physical build times do not collapse to software speed.
Compute remains physical.
Energy remains a hard bottleneck.
Power access, semiconductor supply chains, and tooling remain strategic constraints.
Core human needs remain physical and social.
The transition path matters more than the eventual endpoint.
Existing owners of scarce physical assets start with an advantage.

Medium Confidence

Medium-Confidence Statements

Residential geography is sticky but not fixed.
Labs keep recursively self-improving systems private or tightly gated at first.
Trust and verification become scarcer.
White-collar task automation happens faster than full job elimination.
Labor income can fall faster than cost of living during the transition.
Robotics reaches meaningful scale within a few years of RSI.
Large robot fleets rely heavily on centralized cloud intelligence.
Fast AI timelines favor the U.S., while longer timelines favor China.
Leverage still kills during disorderly transitions.

Priority Questions

What remains scarce, ownable, and defensible when cognition becomes abundant?
Where can robots legally, socially, and economically deploy at scale by 2031?
Does labor income fall faster than cost of living during the transition?
Is the transition orderly or disorderly?
Which early signals tell us which branch we are on?
Does the state let private actors keep the rents from AGI and adjacent bottlenecks?
Does frontier intelligence stay centralized and expensive, or does it commoditize fast?

Questions

Do labs expose frontier recursive systems to the market or keep them mostly internal?
Which scarcities are physical, legal, social, or political rather than cognitive?
Which of those scarcities can still be owned without being regulated or rent-shaved away?
Which deployment domains open first for robots and under what constraints?
Who bears liability when robots fail in homes, streets, workplaces, and public spaces?
How reliable is AGI in high-stakes settings without human oversight?
Who captures the surplus?
Does labor income fall faster than cost of living?
How fast do white-collar jobs disappear at the task level and at the job level?
Does AGI stay expensive, or does it commoditize fast?
Can open-weight frontier models remain legal if misuse risk becomes extreme?
Does open source compress software margins?
Does the policy response dominate the technology story?
Which assets keep pricing power if cognition becomes cheap?
Which geographies are job-anchored versus family-anchored, amenity-anchored, or power-anchored?
Which land qualities matter most: safety, climate, water, tax, energy, or social density?
Which early signals tell us which branch we are on?
What would falsify the land thesis?

Open Questions

Discussion

Next question: scarcity map Robot deployment (deferred) Centralized vs commoditized Does software go to zero? Concrete commodity scenario State rent shaving Warning before repricing Portfolio under rent shaving Labor income vs cost of living Lab gating under RSI Open weights and legality Video claims

+

Assuming centralization and rapid robotics capability, what is the next question to tackle now?

Current focus: scarcity map

Why scarcity is the next step

The investment problem is to identify what stays scarce after central intelligence and rapid robot capability compress cognitive and physical labor.
Scarcity is more directly investable than a long robotics deployment debate because it points to assets you can own today.
The useful filter is not just scarcity in physics, but scarcity that is ownable, durable, and not fully confiscated by policy.
This also lets you postpone domain-by-domain robotics detail without losing the core investment thread.

Subquestions that actually matter

Which scarcities are physical: power, grid access, transformers, transmission, water, land, mineral inputs, fabs, logistics nodes?
Which scarcities are legal or institutional: permits, zoning, licenses, rights-of-way, grid interconnects, defense approvals, hospital approvals?
Which scarcities are social: trust, brand, verified identity, human acceptance, local legitimacy?
Which scarcities remain ownable under the rent-shaving assumption, and which become quasi-public utilities?
Which scarce assets are complements to both centralized AGI and robot deployment rather than substitutes?
Which of these can still be bought today at prices that do not already assume the AGI transition?

This is the next question I would tackle. More precise version: after centralization, fast robotics capability, and rent shaving, which scarce assets still have durable pricing power and are still worth owning?

+

Robot deployment question

Logged for later

Logged and deferred. Current formulation: by 2031, which domains permit large-scale robot deployment, under what liability regime, and at what all-in hourly cost relative to human labor?

+

Does frontier intelligence stay centralized and expensive, or does it commoditize fast?

Current lean: barbell

Why the fixed-supply and insatiable-demand frame is too strong

Supply is not fixed over five years. Power, packaging, memory, networking, and data center capacity can all expand.
Demand is not infinite at any price. It depends on marginal value, model efficiency, and what buyers can monetize.
If the buyers are a few giant labs and hyperscalers, they have bargaining power and can vertically integrate into adjacent layers.
If models get much more efficient, the same economic output may require less spend on some upstream inputs than people expect.

Why this still matters beyond model companies

Centralization means a few mega-buyers can internalize margins, sign exclusive contracts, and squeeze external suppliers.
Commoditization means many downstream firms get similar intelligence, which compresses software and service margins.
The answer changes where the surplus sits: upstream bottlenecks, model layers, or downstream distribution.
The answer also changes the size and speed of labor displacement, which feeds back into policy and asset pricing.

You are right that some upstream layers can win in both branches. You are wrong that the branch is irrelevant. It matters for pricing power, buyer concentration, duration of the capex cycle, vertical integration risk, and where the surplus ultimately lands.

+

Does AGI make software go to near zero or disappear into hardware control?

Current lean: code cheap, software not zero

What probably does go toward zero

The cost of writing routine code falls hard.
Simple SaaS wrappers and undifferentiated feature factories lose a lot of value.
Pure software labor and bespoke coding services get hit badly.
Some user-facing apps get replaced by direct agent interaction.

What probably does not go to zero

Software still coordinates workflows, permissions, data models, logging, safety, billing, reliability, and compliance.
Hardware does not become directly usable without control layers, protocols, operating constraints, and failure handling.
The code may be cheap, but trusted deployment, integration, and control planes are not automatically cheap.
The economic value shifts away from code production and toward distribution, data rights, trust, and operational control.

I think you are wrong to equate cheap code with zero software value. The code itself commoditizes. The economic value moves to orchestration, trust, distribution, and real-world control. Some software categories get annihilated; software as a whole does not disappear in five years.

+

What concretely changes for investors if AGI becomes a commodity?

Concrete example

What changes outside the labs

Merchant inference and cloud AI margins get compressed because customers can choose among many similar models and self-host more easily.
Application software and AI wrappers lose pricing power fastest because their model access stops being a moat.
Data centers, power, and semis can still see strong demand, but returns look more like infrastructure volume returns than scarcity-rent returns.
Customer ownership, workflow lock-in, compliance, and trusted distribution matter more than model quality gaps.

One concrete investment implication

Do not pay premium multiples for any business whose edge is mostly reselling intelligence.
Prefer businesses that own the customer, the workflow, the regulated trust layer, or the hard asset.
Within infrastructure, prefer businesses that work at lower unit margins with higher volume, not businesses that require permanent desperation pricing.
The commodity branch is more favorable to boring distributors and less favorable to glamour AI wrappers.
Even with the same megawatts or chip volume, returns change if the buyers are price-sensitive and competitive rather than desperate and concentrated.

Concrete difference: in the centralized branch, three desperate buyers can sign long-term take-or-pay contracts and bid up prices. In the commodity branch, many buyers arbitrage, switch, and push down margins. Same broad quantity story, different price and return story.

+

Does the state let private actors keep the rents?

Current lean: rent shaving is likely

Why nationalization is not the main issue

Literal nationalization is only one path and probably not the base case outside the most strategic frontier systems.
The more realistic path is windfall taxes, export controls, mandatory licensing, reporting requirements, antitrust pressure, labor-side transfers, and state-directed procurement.
The state can leave legal ownership intact while still pulling away a large fraction of the economic upside.
If AGI creates visible mass dislocation, political tolerance for private capture drops sharply.

Which layers could get hit

Labs are the most obvious target.
Cloud, semiconductors, power generation, utilities, data centers, and telecom can also get pulled into national security and public-utility style oversight.
Bottlenecks with local monopoly characteristics are especially exposed.
In a serious labor shock, the tax base can shift toward capital income, land, and excess returns across multiple layers, not just model labs.

I do not think broad U.S. nationalization is the base case. I do think broad political rent shaving across strategic bottlenecks is a real risk and should be treated as base-rate policy behavior.

+

If rent shaving is coming, do we get warning before the market fully prices it?

Current lean: yes

Why warning is likely

Major policy shifts usually show up first as rhetoric, hearings, investigations, agency signaling, draft rules, or procurement changes.
Public anger about visible private capture tends to build before full legislative or regulatory action lands.
Markets often underreact to slow political buildup until the cash flow impact is obvious.
Current U.S. federal politics probably delays the most explicit redistributive version of this story.

Why warning may be shorter than expected

Executive action, export controls, emergency powers, procurement mandates, and state-level regulation can move faster than investors expect.
The president is not the only actor. Congress, agencies, states, courts, and public utility regulators can all move the effective policy regime.
Once the narrative flips from innovation to extraction, the repricing can be sharp.
A visible labor shock or a high-profile scandal can compress the warning window.

Reasonable base assumption: yes, there is probably a usable warning period before full repricing. But do not anchor too hard on the White House alone. The signal to watch is rising political legitimacy for rent shaving, not just one person's ideology.

+

If broad political rent shaving happens within five years, how should the portfolio change?

Assume true

Increase weight toward

Businesses that benefit from buildout volume rather than uncapped monopoly pricing.
Contractors, equipment suppliers, and asset-base growers that get paid to expand capacity.
Regulated or quasi-regulated assets where allowed returns may be capped but capital deployment still grows.
Senior claims, debt, preferreds, and other positions higher in the capital structure when upside to common equity is politically truncated.
Hard assets with direct use value, not just financial rent-extraction value.
Politically boring enablers rather than visible public villains.

Reduce weight toward

Common equity priced for permanent winner-take-all margins in frontier bottlenecks.
Businesses whose thesis requires unrestricted pricing power in nationally important sectors.
Highly levered assets where policy can cap upside without protecting equity.
Assets that are obvious targets for windfall taxes, compulsory licensing, or utility-style regulation.
Long-duration stories where most of the value depends on terminal margins surviving political backlash.
Any thesis that works only if the state behaves like a passive observer.

The practical shift is this: stop paying for infinite upside in strategic bottlenecks. Underwrite capped returns, shorter duration, more policy friction, and higher redistribution. Prefer shovel sellers, capacity builders, and senior claims over politically exposed common equity.

+

Does labor income fall faster than cost of living?

Current lean: important outside pure infra

Why this matters even if AGI spending is organizational

Organizations spend because they have customers and revenue. Their budgets are not independent of household demand.
Most final demand in the economy still comes directly or indirectly from households.
If wages fall before prices fall, discretionary demand weakens, defaults rise, housing softens, tax revenue drops, and non-frontier enterprise spending gets cut.
A narrow AI capex boom can coexist with a broad demand recession.

What the distribution issue changes

If income shifts toward a small capital-owning group, aggregate spending usually weakens because rich households save more.
That does not kill demand for everything, but it changes which sectors hold up and which crack.
This matters less if the only thesis is long frontier labs, chips, and power.
This matters a lot if the thesis includes broad equities, real estate tied to middle-class demand, consumer businesses, credit, or tax-sensitive geographies.

Yes, the argument is partly about demand collapse and concentration. More precisely, it is about whether the productivity shock outruns the income-distribution adjustment. If it does, politics and asset repricing get ugly.

+

Do labs keep recursively self-improving systems private or tightly gated at first?

My probability: 80%

Why this might happen

The first lab with real RSI gets compounding internal advantage from keeping the strongest system on its own side of the wall.
Governments will treat such a system as a national security asset, which pushes toward restricted access.
Frontier systems will likely be compute-constrained and expensive, which makes broad release unattractive even before safety concerns.
Labs will worry about misuse, theft, cyber offense, autonomous replication, and loss of control.
If the system can improve itself, each extra internal cycle may be worth more than external API revenue.

Why this might not happen

Labs still need revenue, distribution, developer lock-in, and real-world usage data.
Competition can force partial release if one lab starts monetizing externally.
Labs may not immediately know that they have crossed into meaningful RSI.
Tightly gated API access can still be strategically useful while the deeper self-improvement loop stays internal.
Some frontier organizations are structurally built to ship products, not to disappear into secrecy.

My base case is not total secrecy. My base case is that the strongest recursive system stays private or very tightly gated, while weaker derivatives, narrower agents, or rate-limited APIs still get sold to the market.

+

Can releasing an open-weights AGI model remain legal if bio-risk is real?

Current lean: not stably

Why it could be legal at first

Software publication has speech-like protections in parts of the legal system.
Current rules are patchy, and regulators may lag the capability frontier.
Open release can happen before policymakers agree on a test for what is too dangerous to publish.
Some actors will argue that open release is necessary for competition, transparency, and national resilience.

Why it probably does not stay legal

If a model materially lowers the barrier to pathogen design, optimization, or evasion, political tolerance collapses.
Governments can use export controls, licensing, compute rules, mandatory evaluations, or direct prohibition even without a clean AGI statute.
There is no stable equilibrium where downloadable extinction-relevant capability is treated like a normal open-source release.
Once weights are released, proliferation is close to irreversible, which gives regulators strong incentive to intervene before or at release.

Today, some powerful open-weight releases are still legal. True open-weights AGI with credible bio-offense uplift probably does not remain legal for long. That does push the legal frontier toward centralization. But legal centralization and illegal leakage are not the same investment branch, because one preserves monitored choke points while the other creates uncontrolled proliferation risk.

+

Specific Claims From The Dylan Patel Video

Transcript grounded

[00:15] Dwarkesh Patel: "If you add up the big four, Amazon, Meta, Google, Microsoft, their combined forecasted capbacks ... this year is $600 billion ... that would be like close to 50 gigawatts ... that's paying for compute that is going to be coming online over the coming years."
[33:59] Dylan Patel: "Google has gotten absurdly AGI-pilled ... They bought an energy company. They're putting deposits down for turbines. They're buying a ridiculous percentage of the powered land. They're going to utilities and negotiating long-term agreements."
[34:51] Dylan Patel: "I think the biggest bottleneck is compute. And for that, the longest lead time supply chains are not power or data centers. They're actually the semiconductor supply chain themselves ... It switches back from being power and data center as a major bottleneck to chips."
[35:08] Dylan Patel: "In the chip supply chain, there's a number of different bottlenecks ... There's memory. There's logic wafers from TSMC. There's fabs themselves. Construction of the fabs takes a couple years, two to three years, versus a data center takes less than a year."
[37:03] Dylan Patel: "By 28, 29, the bottleneck falls to the lowest rung on the supply chain, which is ASML ... currently, they can make about 70. Next year, they'll get to 80. Even under very aggressive supply chain expansion, they only get to a little bit over 100 by the end of the decade."
[75:36] Dylan Patel: "It's like fast timelines, U.S. wins, long timelines, China wins."
[77:24] Dylan Patel: "Anthropic could actually release a slow mode ... They could probably reduce the price of Opus 4.6 by 4X, 5X and reduce the speed by maybe just like 2X ... And yet they don't because no one actually wants to use a slow model."
[144:33] Dylan Patel: "You don't need to have all the intelligence in the robot ... a lot of the planning and longer horizon tasks are determined by a much more capable model in the cloud ... and then it pushes those directions to the robots."
[146:06] Dylan Patel: "Any robot you deploy needs leading edge chips because the power is really bad for robots ... all of a sudden you're taking power and chips that would have been for AI data centers and you're putting them in robots."
[149:14] Dylan Patel: "Just shipping out all the engineers and blowing up the fabs means China has a stronger semiconductor supply chain than the rest of the world."
[150:10] Dylan Patel: "If something happens to Taiwan ... your incremental ability to add compute goes to almost zero ... maybe like 10 gigawatts across Intel and Samsung or 20 gigawatts."