Key Executive Summary Points for Allocators
- Software scarcity is increasingly likely to end; code is becoming a commodity.
Vibecoding and AI agents are pushing the marginal cost of software creation toward zero. The more probable regime ahead is one where scarcity and therefore value capture shifts to power, grid infrastructure, HBM, electrical parts and components, packaging, cooling, and energized land. - Portfolio exposures are misaligned with the likely emerging regime.
MSCI World / ACWI remain heavily concentrated in long-duration software moats built on expensive, slow-to-replicate human code. Those moats appear increasingly at risk as AI accelerates competition and shortens product cycles. - The growth factor is gradually migrating from software to infrastructure.
Hyperscalers are behaving more like industrial companies, driving multi-year demand for electrical equipment, advanced semiconductors, memory, thermal systems, and power generation. The balance of probability now favors growth re-attaching to atoms rather than bits. - Horizontal SaaS faces a rising probability of structural margin pressure.
Vibecoding erodes feature moats, AI agents replace seats, switching costs fall, and inference spend enters COGS. These forces are increasingly likely to compress margins and pull multiples toward historical software ranges. - Allocation implication:
Maintain exposure to durable digital franchises, but tilt allocations toward AI infrastructure bottlenecks that benefit under most forward scenarios:
memory (HBM), advanced packaging, transformers/switchgear, liquid cooling, data-center REITs, and advantaged utilities/IPPs. - Key falsifiers to monitor:
SaaS moats remain stickier than expected; power and grid constraints ease materially; AI workload growth slows; or regulation meaningfully restricts data-center expansion.
Introduction: When Human Keystrokes Stop Being Scarce
On a cold February morning in 2025, Andrej Karpathy, the former Director of AI at Tesla and a founding member of OpenAI, published a short manifesto that landed like ChatGPT had in November 2022. He called the new paradigm vibecoding: you describe what you want in natural language, the machine writes the code, runs the tests, deploys it, and you “forget that the code even exists.”
For the first time in history, syntax is becoming a zero-marginal-cost commodity.
What ChatGPT began to do to white-collar knowledge work in 2023, vibecoding is now doing to software creation in 2025–2030. The scarcity premium that has defined global equity indices for fifteen years is collapsing.
For anyone benchmarked to MSCI World or ACWI, this is not a curiosity about developer productivity tools. It is a regime shift in what the benchmark actually owns. From 2007 to 2024, cap-weighted indices became a concentrated bet on intangible software moats: high gross margins, near-zero capital intensity, and decades-long cash-flow duration. The top-10 holdings’ weight doubled to >26%, and the IT sector share rose from the mid-teens to the high-twenties while Energy and Materials shrank to historic lows. In retrospect, a large part of what we labeled American exceptionalism and dollar dominance was, at the index level, a two-decade equity moat built on the scarcity of code
In my previous note, From Bits to Atoms, I argued that AI is moving from “thinking” to “acting” via Vision-Language-Action (VLA) models, shifting value to sensors and actuators. This paper tackles the other side of that coin: what happens to the software itself? As AI gains a physical body, it is simultaneously commoditizing its own digital mind.
This paper’s core thesis is simple and brutal:
- Intangible code is becoming abundant and commoditized.
- Tangible infrastructure: power, transformers, cooling water, high-bandwidth memory, electrical parts and components, land, permits and many more are becoming the binding constraint on intelligence at scale.
The result is the Great Software Re-Rating of 2025–2030: multiples compress for large parts of the application layer while valuation power migrates to the physical stack that cannot be vibecoded away.
The Old Regime: Bits Ate Atoms (2007–2024)
From the iPhone launch through the end of 2024, the dominant equity narrative was “bits over atoms.” The benchmark math told the story in one chart: the alligator jaws between digital and physical sectors opened wider every year.
- MSCI World IT + Communication Services weight rose from ~20% in 2007 to >35% by late 2024.
- Energy fell from ~10% to <5%, Materials from ~8% to ~4%.
- Index concentration exploded: the top-10 constituents went from 10.5% of the index in 2017 to >26% in 2024.
Investors paid 10–100× sales and 40–120× earnings for best-in-class software franchises because the underlying assumption was durable: replicating a Snowflake, ServiceNow, or CrowdStrike required armies of engineers and many years. The benchmark became implicitly long “expensive human keystrokes.”
The “Growth” Trap and the Death of Asset-Light
For fifteen years, “Growth Investing” was synonymous with “Asset-Light.” The algorithm was simple: avoid factories, avoid inventory, and buy zero-marginal-cost code. This created a dangerous blind spot: growth managers systematically ignored physical industries, assuming they were all “Value traps.”
But the “Asset-Light” era is over, even for the Hyperscalers.
The Magnificent 7 are currently engaging in the largest capital expenditure cycle in human history. Microsoft, Google, Meta, and Amazon are effectively transforming from software companies into industrial energy utilities. They are pouring hundreds of billions into copper, concrete, custom silicon, and power generation. If the kings of the asset-light era are pivoting to become “kings of concrete,” the investment framework must change.
The “Growth Factor” is decoupling from “Software” and re-attaching to “Industry.” True growth, accelerating revenue and pricing power, is now found in the assets that rust and hum, not just the ones that compute. Managers who refuse to buy physical assets will be structurally short the actual growth of the next decade.
Vibecoding: From Scarce Syntax to Cheap Intent
Vibecoding is not just better autocomplete. It is a phase-change in the abstraction layer of programming.
You now specify semantic intent (“build a warehouse inventory system with RFID integration, Stripe billing, and Slack alerts”), and multi-agent systems (Cursor, Loveable, Windsurf, Claude Code, Replit Agent, etc.) generate, test, debug, and deploy production-grade code in hours or days. The constrained resource is no longer fluency in Python or TypeScript, it is high-level architecture and product taste.
Real-world evidence as of December 2025:
- Stack Overflow 2025 survey: 78% of professional developers use AI coding tools daily.
- GitHub Copilot controlled studies: 55% task completion speedup; internal Microsoft data claims 30–50% across the board.
- Y Combinator Summer 2025 batch: >25% of companies shipped with >90% AI-generated code.
- The Weekend Founder: Stories are now routine of production CRMs + Stripe + Resend + Vercel built in <60 hours for <$3,000 in compute.
The economic consequence is the rise of disposable software and shadow IT 2.0. Marketing, HR, and operations teams are vibecoding internal tools rather than paying $20–$200 per seat per month. Software is shifting from durable capital good to cheap consumable.
How the Moats Actually Erode
The Universal Translator Effect
The single most dangerous misconception is that “data gravity” will save incumbents. It will not.
Historically, migrating off Salesforce or SAP was a multi-year, eight-figure nightmare of schema mapping and data loss. AI agents now act as a universal translator: they ingest legacy exports, infer schemas, clean data, and rebuild in a new stack in hours. When data portability becomes automated, the single largest source of lock-in evaporates.
Seat-Based Pricing → Agent-Based Consumption
- Old world: 1,000 employees = 1,000 seats.
- New world: 3–10 AI business agents orchestrate the workflows of those 1,000 employees via APIs.
Revenue arithmetic collapses overnight. Vendors are forced into usage- or outcome-based pricing, and gross margins compress as inference spend enters COGS.
The Margin Equation Changes Forever
- Classic SaaS: Gross Margin ≈ 75–85% (mostly cloud hosting).
- Vibecoded + Agentic SaaS:
Gross Margin = (Revenue – (Hosting + Inference + Agent Orchestration)) / Revenue
→ 50–65% becomes the new normal for anything that ships “smart” features.
The Great Bifurcation
Not all software dies equally.
- Most exposed: Horizontal productivity (generic CRM, project management), long-tail SaaS with no proprietary data, and thin-wrapper AI coding tools.
- Resilient / Winners: Vertical SaaS in regulated domains (life sciences, insurance), Systems of Control tied to physical processes (factory robots, grid management), and platforms with proprietary data flywheels that cannot be scraped (Palantir, Veeva).
Where software is mostly a polished UI on CRUD, vibecoding is existential. Where software is inseparable from life, safety, or regulated liability, AI deepens the moat.
The Physics of Intelligence: AI Runs Into Atoms
If code is no longer scarce, what is? Joules, wafers, water, and permits.
Jevons Paradox on Steroids
Cheaper code → 10–100× more software → exponential demand for inference and training compute. Efficiency gains are real, but total resource consumption rises faster.
The 160% Wall
Global data-center electricity consumption was ~450 TWh in 2024. The 2030 base case is 900–1,200 TWh (2–3×).
Meanwhile:
- U.S. interconnection queues exceed 2.6 TW, with median time-to-operation now 4–5 years.
- Large power transformer lead times have hit 80–130 weeks, with a 25–35% projected supply deficit through 2027.
- High-bandwidth memory (HBM3e/HBM4) and advanced packaging (CoWoS) are sold out through 2026–2027.
This energy constraint parallels the “Bandwidth Wall” I discussed in From Bits to Atoms. Just as robotics is forcing a migration to optical interconnects to handle video streams, the explosion of vibecoded software is forcing a migration to liquid cooling and behind-the-meter power. Whether it is a robot hand or a CRM agent, the bottleneck is no longer code, it is physics.
The Hidden Supercycles
- Memory: Converting standard DRAM lines to HBM reduces output per wafer, creating structural pricing power.
- Liquid cooling: Rack densities are heading to 100–250 kW, driving a multi-fold increase in spend per MW.
- Energy: Behind-the-meter power & nuclear/SMR exposure offers the only credible path to 24/7 carbon-free baseload at scale.
The Value Smile Curve (2030 Edition)
Imagine a U-shaped curve. Value capture migrates to both edges, leaving the middle hollowed out.
- Left Edge (Infrastructure): High value. Power, transformers, HBM, packaging, liquid cooling, energized land.
- Middle (Application Logic): Value collapse. 3–6× sales becomes the new normal for horizontal SaaS as code becomes a commodity.
- Right Edge (Brand, Trust, & Action): High value. Entities that take legal responsibility for AI outcomes (regulated verticals) or deliver the physical action (Robotics/VLAs).
Portfolio Construction for a Benchmarked Investor
You cannot fight the benchmark, but you can tilt intelligently.
Recommended Default: The Barbell
- Core (50–70%): Quality mega-cap platforms + defensible verticals (tracking-error anchor).
- Atoms Sleeve (20–40%): Deliberate overweight to the bottlenecks:
- Memory & packaging leaders.
- Electrical equipment / grid (transformers, switchgear, HV cable).
- Liquid cooling & thermal management.
- Data-center REITs with powered land in constrained hubs.
- Select utilities / IPPs with nuclear or gas in AI load centers.
Higher-Conviction Alternatives
- Rotation: Reduce long-duration software; overweight industrials, semis, and utilities.
- Dispersion: Stay sector-neutral but ruthlessly underweight commoditizable SaaS and overweight backlog-rich atoms names inside sectors.
Risks & Falsifiers Dashboard (Monitor Quarterly)
- Brownfield Drag: 90% of enterprise value still runs on legacy COBOL/Java/.NET spaghetti. If AI agents cannot safely refactor this “toxic waste,” the Old Regime persists far longer than greenfield enthusiasm suggests.
- Moats Prove Stickier: SaaS net dollar retention stays >110% across cohorts despite AI parity.
- Power Constraints Ease: Transformer lead times drop below 50 weeks; interconnection queues shrink dramatically; modular nuclear scales by 2027–2028.
- Efficiency Outruns Demand: Hyperscaler capex guidance flattens or declines in 2026–2027.
- Regulatory Backlash: Moratoria on new data centers or water use in key states.
Conclusion: The Benchmark Is Short Reality
From 2007 to 2024, the market paid ever-higher multiples for companies that turned scarce human keystrokes into durable monopolies. Software ate the world, and the indices followed.
Vibecoding ends that era. Code is becoming abundant, disposable, and in many cases free. The scarcity, and therefore the durable economic rent, has migrated down the stack to the things no large language model can conjure out of thin air: electrons, silicon, copper, cooling water, permits, and land.
The Great Software Re-Rating is not a crash. It is a rotation from abstract bits to concrete atoms. The vibes have shifted. The benchmark will follow slowly, then all at once. Volatility will be the price of the rotation.
Position accordingly: long the constraints, selective in pure code, and own the physical bottlenecks that now gate the future of intelligence.
