YK Research
Research Note · Valuation & Regime

The Future of Software

If an LLM Can Rebuild Your Product in a Weekend, Your Moat Isn’t a Moat

April 2026 · YK Research · Software / Valuation

Core Thesis

For 15 years the software business model was a gift: sell a seat, lock in the contract, compound the ARR. The stock market rewarded it with 60–80% of equity value sitting in terminal value. The unexamined assumption that the moat would still be there in year 11 and beyond.

That assumption only holds if the product is hard to rebuild. AI just changed what “hard” means. If an LLM can replicate the core of your software in weeks, or a weekend, the moat isn’t a moat, and the terminal value doesn’t exist. Which means 60–80% of the equity simply ceases to exist.

2025 was the first inning. Multiples quietly compressed. The market went long silicon, short SaaS. Software isn’t dead, but the software business model is getting rewritten in real time.

The Software Split: What Survives, What Dies

Survives
  • Distribution-moated: platforms that own the user (Google, Meta, TikTok). The LLM clone doesn’t matter if nobody finds it.
  • Data-moated: proprietary data you can’t scrape (Palantir ontologies, Bloomberg terminal, private-market data).
  • Infra-moated: physical silicon, cloud regions, compute capacity. The stuff the LLM runs on.
  • Regulation-moated: compliance, audit, systems of record in healthcare, defense, finance.
  • Workflow-moated: mission-critical systems where the cost of getting it wrong is enormous (ERP for a public co., EHR for a hospital).
Dies (or at least re-prices hard)
  • Single-feature SaaS: note-taking, grammar, transcription, basic CRM. A weekend project with a decent prompt.
  • Thin vertical tools: the “Shopify for X” wrappers where the X is commodity.
  • Consumer ed-tech: Duolingo −67.5% from mid-2025 peak. ChatGPT is a free, personal, infinitely-patient tutor.
  • Creative-tool rentiers: Adobe −24.6% from 2025 peak. Generative models ate the top of the funnel.
  • PE-owned SaaS rollups: 1,900+ software cos bought for $440B (2015–25), now ~25% of the private-credit book. LBO debt serviced by cash flows AI is unbundling.

The one question every software company has to answer

“How will this not be unbundled by an LLM?”

If the honest answer is “data,” “distribution,” “regulation,” or “physics,” you might be fine. If the answer is “features” or “brand,” the terminal value is dying.

Mag 7 Share of S&P 500
17% → 36%

2018 to October 2025. Big Tech didn’t just get bigger, it doubled its slice of the entire index. Every one of those percentage points is priced on the assumption that the moat is real.

Fraction in Terminal Value
60–80%

Most DCFs forecast 5–10 years of explicit cash flows. Everything past year 10 is a terminal value assumption, the belief the business still earns money in year 11+. That assumption is the whole ballgame.

Put the two together: 36% of America’s public equity rests on a 60–80% terminal-value fiction. That is the exposure.

2025 Scoreboard: Long Silicon, Short SaaS

Silicon (Physical Infra)
LONG
Broadcom (AVGO)
SK Hynix
TSMC / ASML
Cable & Power (HWM, GEV)
Can't be prompt-engineered away. Capex-heavy, long-cycle, terminal value intact.
SaaS / App Layer
SHORT
Duolingo−67.5% from mid-25 peak
Adobe−24.6% from 2025 peak
PE-owned SaaS rollups
Vertical point solutions
If an LLM can replicate the product in weeks, the moat is rhetorical. Terminal value evaporates.

The Question Every Business Has to Answer

“How will this not be unbundled by an LLM?”

If there's no credible answer, the terminal value assumption is dead.

Three Stages of the Repricing

Stage 1TODAY

P/E compression

Market applies a larger discount factor to future cash flows. Multiples quietly compress. We're here now.

Stage 2TRANSITION

P/E → FCF multiples

Market flips from earnings-based multiples to multiples of current free cash flow. GAAP-only reporting. No more adj. EBITDA. SBC treated as the real cost it is.

Stage 3NEW REGIME

Short-duration world

Growth investing breaks. Pre-revenue unicorns de-fund. Capital rotates to physical world. M&A logic inverts. Short-termism becomes rational.

Signal the flip from Stage 1 to Stage 2: watch companies stop reporting adjusted EBITDA. When SBC finally gets treated as the real cost it always was, the regime has shifted.

Four Implications If This Plays Out

Growth equity math breaks

Pay-now-harvest-later only works if 'later' is financeable. A 12x ARR multiple presumes a moat that compounds. If LLMs unbundle the moat, the 12x is a fiction.

Private credit is the canary

PE bought 1,900+ SaaS cos for $440B (2015–25). Software is now ~25% of private credit. Cash flows repay the LBO debt. If AI replaces the product, cash flows die, loans default. Watch the default rate.

The AI capex paradox

$300–500B/yr of AI infra only pays back over 7–15 years. If markets reprice to short-duration multiples, that capex can't be financed. The disruption engine disrupts its own funding.

Likely outcome: oscillation

Not a permanent new regime. An oscillating one. Shorter cycles, higher vol, terminal values compress and recover as AI capex ebbs and flows. Fat tails, not a new mean.

Five Forms of Leverage in an AI-Divided World

The geopolitical organizing principle is no longer who protects you, but who gives you access to prosperity. If you are not the US or China, the honest answer is you have almost nothing to negotiate with. Almost. Five forms of leverage still matter. Countries that hold any of these have something to trade. Countries that hold none face an uncomfortable question: which version of intelligence do they rent, and what do they concede in return?

01

Critical Minerals

Rare earths, lithium, gallium, copper. The raw inputs to every chip and every grid upgrade.

AU, CA, CL, ID

02

High-Tech Manufacturing

Endemic fab capacity, lithography, advanced packaging. Hard to move, harder to replicate.

TW, KR, NL, JP

03

Capital

Deployable sovereign balance sheet. Can fund the next fab, the next grid, or the next deal.

SG, UAE, CH, NO

04

Energy

Gigawatt-scale power, cheap and firm. Data centers are power plants with chips attached.

SA, QA, US, AU

05

Data Center Infra

Land, cooling, fibre, permitting speed. The physical substrate AI runs on.

US, IE, SG, NL

“The organizing principle of geopolitics is no longer military alliances, but supply chain position.”

Investment read: every one of these five levers is a physical asset class. None of them are software. This is the same point the multiples are making, just in different language.

Atoms, Not Bits

The lasting bottlenecks in the AI stack are physical: critical minerals, energy storage, and actuation. The same logic applies to capital allocation.

“If your moat requires atoms, not bits, a better language model cannot unbundle it overnight. For now.”

Energy Infra
Farmland
Toll Roads
Batteries
Critical Minerals
Actuation

The two-word test before any long position: atoms or bits? If the answer is bits alone, keep the size small.

The Sovereign Megacorp

The logical endpoint of Big Tech hitting its stride is the vertically integrated megacorp: a firm so embedded in physical supply chains that it becomes too big to disrupt. Elon is assembling one in the open, with Tesla, SpaceX, and xAI converging into a single industrial entity. These companies are starting to finance themselves like sovereign states.

40-Year Bonds
MSFT · AMZN · AAPL

Issued long-dated paper that only makes sense for entities assumed to still exist in 2066.

100-Year Bond (2026)
GOOGL

Century bond, the first tech issuer to try. Priced with confidence that the franchise outlives most nations.

Order Book
10x oversubscribed

The market agreed. Credit is pricing these firms as sovereigns, not companies.

Counter-Risk

The same dominance invites trust-busting. Big Tech is the modern railroad monopoly, hoarding talent, stifling innovation, and concentrating wealth. The sovereign balance sheet is a target as much as a moat.

Portfolio read: sovereign megacorps are the closest thing to “atoms plus bits” you can own on public equity markets. Keep them long, but size around regulatory tail risk.

Positioning Takeaways

Overweight physical infra

Silicon supply chain, energy, actuation, critical minerals. Assets where terminal value is priced on real capex and real tonnage, not SaaS switching costs. Our AI Power, XLI, and ASML theses all sit in this bucket.

Underweight replaceable SaaS

Single-feature SaaS, thin vertical software, PE-owned rollups levered against cash flows the LLM can eat. Especially anything still trading on 10–15x ARR post–2025.

Sell vol, don't buy it, on the AI names

Oscillating terminal value = fat tails, not a new mean. IV on hyperscalers stays bid. Fits the option seller's playbook and harvest the premium, size for the drawdown.

Accept shorter horizons

If Stage 2 arrives, 10-year DCFs are fantasy. Reprice your own book to what you can prove in 2–3 years. Earnings, not narrative.

Bottom Line

The partial move is already underway. A full repricing would be the most consequential structural shift in capital markets since the postwar era. You don't need to bet that it happens fast. You need a book that survives whether it does or doesn't: long the stuff that has to exist regardless, short the stuff whose moat is a sentence.

Source

Social Capital, “2025 Annual Letter” (April 2026). Read the original. This write-up extracts and frames the valuation-multiple thesis for YK Research; interpretation and trade framing are ours.