The Leveraged Service Delivery Thesis — Deep Market Research v3 by Eric San

Exposed Wages (US)

$3.7T

Karpathy/BLS, Mar 2026

Actual Displacement

1.3%

316K jobs since Jan 2023

Pro Services TAM

$6.3T

global market 2026

Roll-Up Capital

$3B+

GC + Thrive + others

I. The Core Insight

Andrej Karpathy scraped all 342 US occupations from the Bureau of Labor Statistics, scored each on AI exposure (0–10), and published the data in March 2026.¹ The results inverted conventional wisdom:

Income Bracket	AI Exposure	Implication
$100K+ earners	67%	Highest-paid knowledge workers are most exposed
<$35K earners	34%	Physical work is structurally protected
Bachelor’s degree holders	Most exposed credential	The degree that was “safe” is now the most vulnerable

Average exposure across all jobs: 5.3/10. Software developers: 8–9. Medical transcriptionists: 10. Roofers: 0–1.²

China shows the same pattern. Geek mapped ~250 Chinese occupations — the right side is wall-to-wall high-risk knowledge work.³

The gap is the market. AI can do 40–60% of most knowledge work tasks today. But actual displacement is 1.3% of layoffs (316K total jobs since Jan 2023).⁴ DisplaceIndex shows a 17-point divergence between capability and adoption. The capability exists. The deployment hasn’t happened. That gap is the arbitrage window.

II. The Leverage Math

Jensen Huang set the benchmark: a $500K/year engineer should spend $250K on AI tokens — a 1:2 salary-to-token ratio.⁵ He’d be “deeply alarmed” at $5K token spend. The result: ~10x output per engineer. Every engineer orchestrating 100 AI agents.⁶

But that’s the tech company version. The service economy version is where the real opportunity lives:

Metric	Jensen’s Engineer	Typical Service Firm	AI-Native Service
Worker cost	$500K/yr	$20–40K/mo per head	Expert + token budget
Token spend	$250K/yr (50%)	~$200/mo (~1%)	30–50% of expert cost
Leverage ratio	~10x	~1.01x	5–10x
Effective cost per unit	Low (high base)	High (no leverage)	Lowest

The typical professional service firm spends ~$200/month per employee on AI tools. On a $20K/month payroll, that’s 1% leverage — negligible. They’re using AI as a productivity boost to existing humans, not as a replacement architecture.

The structural trap for incumbents. Service firms sell hours. If they 10x productivity with AI, they either fire 90% of staff (destroying delivery capacity and culture) or cut rates 90% (destroying their business). They literally cannot adopt at Jensen-level ratios. This isn’t a talent gap — it’s a business model impossibility.

The Replacement Math

Comparison isn’t “human + AI tools” vs “human + AI tools.” It’s:

	Old Model (Incumbent)	New Model (AI-Native)
Who does the work	Human does 99%, AI helps at margin	AI does 80%, expert orchestrates 20%
Cost basis	Headcount × salary	Expert + tokens (~1/5 to 1/8)
Pricing model	Hours billed	Outcomes delivered
Scalability	Linear with hiring	Exponential with compute

III. Two Plays: Bottom-Line vs Top-Line

Bottom-Line Play — “More Is NOT Better”

Client wants the same output, cheaper. You win by being structurally cheaper while matching or exceeding quality.

Industry	What Client Buys	Current Cost	AI-Native Price	Client Saves
Legal (doc review, research)	Associates at $300–500/hr	$50–100K/matter	$15–30K/matter	50–70%
Accounting (compliance, audit)	Staff at $150–300/hr	$30–60K/engagement	$10–20K	50–65%
Recruitment (screening)	Agency at 20–25% of salary	$20–40K/hire	$5–10K/hire	60–75%
Consulting (analysis, decks)	Analysts at $200–400/hr	$100–300K/project	$30–80K	60–70%

Top-Line Play — “More IS Better”

Client wants capabilities they couldn’t previously justify. You unlock new budget, not compete for existing budget.

Industry	What Was Unaffordable	Why	AI-Native Delivery	What Unlocks
Field sales ops	Per-client intelligence	Would need 3–5 analysts	Expert + Opus + capture	216 clients serviced like 20
Content & marketing	Multi-channel at quality	Agency = $20K+/mo	Expert + Opus	10x output at 1/3 cost
Market research	Continuous competitive intel	Firm = $50K+/study	Expert + Opus	Always-on, not project-based
Customer success	Full-book health monitoring	Only top 20% covered	Expert + Opus	100% account coverage

The bottom-line play is faster to close but commoditizes. The top-line play is harder to sell but creates dependency and expansion revenue. Best companies do both — cost-cut to get in the door, then expand into top-line as the client sees what’s possible.

IV. The Roll-Up Model — PE Meets AI

The biggest capital is flowing not into AI tools but into AI-enabled service roll-ups — buying existing service businesses and transforming their economics with AI.⁷

How It Works

Acquire service businesses running at 5–15% margins (standard for professional services)
Deploy AI to automate 30–70% of repetitive tasks
Transform economics from services (5–15% margin) to quasi-software (40–65% margin)
Re-rate the business from a 10x services multiple to a 30–50x software multiple

Who’s Doing It — The Capital Table

Player	Capital	Portfolio	Results
General Catalyst	$1.5B (from $8B fund)	Crescendo (contact center), Long Lake (HOA mgmt)	Crescendo: $500M valuation, 60–65% gross margins vs 25% industry avg. Long Lake: $100M EBITDA in <2 years.⁸
Thrive Holdings	$1B+ vehicle	Crete (accounting), multiple firms	Crete: $300M+ revenue, 30+ accounting firms rolled up, AT fastest-growing 2025.⁹
OpenAI / Thrive	Equity + embedded engineers	Direct engineering inside portfolio cos	OpenAI took equity stake Dec 2025, engineers embedded.¹⁰
Bezos / Prometheus	$6.2B raised, seeking $100B	Manufacturing (not services)	120 employees from OpenAI/DeepMind/Meta. First acquisition: General Agents.¹¹

Bezos’s play is the physical version of the same thesis. Buy undervalued manufacturing assets, deploy AI to replace human labor, re-rate from industrial multiple to AI-software multiple. As @YaseenKhanYous5 put it: “fire 60% of the human workforce, replace them with automation, and instantly re-rate from a 10x manufacturing multiple to a 50x AI software multiple.” Same math — different modality (physical vs knowledge work).

The Multiple Arbitrage

Business Type	Typical Multiple	After AI Transformation	Uplift
Traditional services firm	5–10x EBITDA	20–30x EBITDA	3–5x
AI-native services firm	15–25x revenue	(born at this level)	—
SaaS company	8–15x revenue	—	(benchmark)
Manufacturing (Bezos target)	8–12x EBITDA	30–50x EBITDA	3–5x

V. Does the Roll-Up Make Sense for Our Thesis?

The PE roll-up model (GC, Thrive, Bezos) and the bootstrapped AI-native model are two different games playing the same arbitrage:

	PE Roll-Up (GC/Thrive/Bezos)	AI-Native Build (Our Model)
Starting point	Buy existing business + clients	Build delivery OS + win clients
Capital required	$100M–$100B	$0–$1M (expert + tokens)
Revenue day 1	Yes (acquired)	No (must earn)
Margin transformation	5–15% → 40–65%	Born at 40–65%
Integration risk	High (culture, tech debt, people)	None (greenfield)
Speed to scale	Fast (if integration works)	Slower (organic growth)
Moat	Capital + scale	Domain expertise + delivery IP

The insight: we’re building the thing the PE firms need to buy. GC and Thrive need AI-native delivery engines to inject into their acquired firms. They’re building internally now (Crescendo, embedded OpenAI engineers). But purpose-built vertical engines — like Donna for field sales — are exactly what a roll-up acquirer wants: proven delivery + margin structure + domain expertise.

Three Paths From Here

Path	Description	Fit
1. AI-Native Service Co	Sell leveraged delivery directly to clients. Our current model. Expert + Opus, outcome-priced. Grow organically or with light capital.	Best fit now. Already running with Hopeman. No capital needed. Prove the unit economics first.
2. Vertical Roll-Up	Once delivery OS is proven, acquire small service firms in target verticals (F&B marketing, insurance brokerages, recruitment agencies). Deploy Donna/mufu into each. Transform margins.	18–24 months out. Need proven delivery + 3–5 reference clients. Then the math works: buy a 10-person firm at 5x EBITDA, transform to 40%+ margins, worth 20x.
3. Be Acquired Into a Roll-Up	Build the vertical engine (field sales, F&B, services), then sell to GC/Thrive as the delivery platform for their acquisitions in our vertical.	Exit path. GC paid $500M valuation for Crescendo. Vertical delivery OS with proven clients is exactly what they’re buying.

VI. The Displacement Index — Where the Math Works Best

Combining Karpathy’s exposure scores with the DisplaceIndex data and professional services market sizing:¹⁴

AI Exposure by Service Category (Karpathy Score × Market Size)

Legal services

7.8/10

Financial analysis

7.5/10

Consulting/advisory

7.2/10

Recruitment

7.0/10

Marketing/content

6.8/10

Accounting

6.5/10

Field sales ops

6.0/10

Customer support

5.8/10

Where Each Play Fits

Category	Exposure	Play Type	Roll-Up Capital Active?
Contact centers	5.8	Bottom-line (cost cut)	Yes — GC’s Crescendo
Accounting	6.5	Bottom-line (compliance)	Yes — Thrive’s Crete
HOA/property mgmt	5.5	Bottom-line (ops)	Yes — GC’s Long Lake
Legal doc review	7.8	Bottom-line	Not yet (regulatory friction)
Recruitment	7.0	Both	Emerging
F&B/marketing services	6.8	Top-line (more is better)	No — fragmented, low margin
Field sales ops	6.0	Top-line (more is better)	No — our space
Manufacturing	3–5	Physical automation	Yes — Bezos $100B

Field sales ops and F&B marketing services have no roll-up capital pointed at them. Contact centers, accounting, and property management are already being consolidated. Legal is high-exposure but regulatory-heavy. Our vertical — field sales + SME marketing services in Asia — is the white space between what PE is buying and what AI can disrupt. Too fragmented and regional for US PE, but perfect for an AI-native operator.

VII. The Pricing Shift — Hours to Outcomes

The professional services industry is repricing from hours to outcomes.¹² When AI completes a 12-week migration in 2 weeks, outcome-based pricing captures full value while hourly billing slashes revenue by 80%.¹³

Emerging AI Service Pricing Models

Model	Structure	Example	Who Wins
Outcome-based	Pay per result delivered	$5K/workflow saving 10+ hrs/wk	Provider (captures full value of speed)
Retainer + variable	Base fee + per-deliverable	$4K/mo + $500/report	Both (predictable + aligned)
Agent licensing	Setup + monthly ops	$20K setup + $2K/mo	Provider (recurring, low marginal cost)
Traditional hourly	Time × rate	$300/hr × 200 hrs	Client IF they switch to AI-native

The Cannibalization Dilemma.¹⁴ If incumbents price per seat, every AI improvement becomes a revenue leak — fewer people needed, fewer seats sold. If they price per outcome, they admit their old model was overpriced. Either way, the incumbent’s business model works against them. The AI-native player is born at outcome pricing and never has this problem.

VIII. Our Proof Point — Hopeman

The Hopeman trial is the first live test of this thesis:

Element	Thesis	Hopeman Reality
Expert + tokens	1 expert with AI delivers like 5–8 people	Eric + Donna delivering Malaysia franchise data, brand books, CV screening, WA intel for 216-client book
Client cost	Lower than alternatives	HK$20K/mo vs hiring 2–3 staff at HK$15K each
Leverage ratio	5–10x	1 person servicing what previously needed a team of 3–5
Outcome pricing	Pay for deliverables, not hours	Flat monthly retainer for defined service scope
Top-line unlock	Client gets what they couldn’t afford	Bob’s team never had market research, competitive intel, or automated reporting before

Kill date: June 30, 2026. If the trial doesn’t convert to ongoing engagement, the thesis needs revision — not the model itself, but the specific vertical application.

IX. Red Team

Bull Case

$3.7T in exposed wages with only 1.3% actual displacement — massive adoption gap
Incumbents structurally trapped: can’t adopt AI at scale without destroying their business model
PE firms validating the economics: Crescendo at $500M, Long Lake at $100M EBITDA in <2 years
Outcome pricing kills hourly billing — early movers capture disproportionate value
Bezos deploying $100B on the physical version of the same thesis
Our vertical (field sales, SME services, Asia) has zero roll-up capital pointed at it
Hopeman already running as live proof of the unit economics
30–40% of $6T professional services market expected to migrate in 5 years
AI-native delivery creates compounding data advantage (outcome data > usage data)

Bear Case

The 1.3% displacement is accurate — AI augments, rarely replaces, and the gap may persist
Incumbents may adapt faster than expected (Accenture acquired Faculty, $400+ AI specialists)
Roll-up economics require scale; may not work at bootstrapped levels
Outcome pricing requires trust that’s hard to build without track record
“Expert + tokens” model depends on finding domain experts who can orchestrate AI
Quality control at scale is unsolved — Eric’s taste in the loop doesn’t scale beyond managed service
Regulatory risk in high-exposure sectors (legal, finance) could slow adoption
Bezos’s $100B may be a leading indicator for manufacturing, not services
The window may be shorter than expected — 2–3 years before large firms retool

Verdict

The general model is sound and capital-validated. The roll-up path is real but premature for us. Build the engine first.

The math works. AI-native service delivery at Jensen-level token leverage (30–50% of expert cost in tokens) produces 50–70% cost savings for clients with 40–65% gross margins for the provider. PE firms (GC, Thrive) have deployed $3B+ validating this thesis across contact centers, accounting, and property management. Bezos is deploying $100B on the physical version.

The roll-up model is the fastest path to scale — but it requires capital and proven delivery first. We’re at the “prove delivery” stage with Hopeman. The correct sequence:

Now: Prove the unit economics with Hopeman. Document: cost per deliverable, client satisfaction, time savings, leverage ratio. This is the evidence packet for everything downstream.
3–6 months: Expand to 3–5 clients across 2 verticals (field sales + one other). Codify the delivery OS into repeatable workflows, not Eric-dependent judgment.
6–12 months: Decide: stay AI-native service company (organic, capital-light) or raise to roll up. By then you have: clients, margin proof, delivery IP, and a team.
12–24 months: If roll-up: acquire 2–3 small service firms, deploy the engine, transform margins. If organic: scale through referrals and vertical expansion.

The key insight from the roll-up research: GC and Thrive are proving that AI-transformed service businesses get re-rated from 5–15x to 25–50x multiples. You don’t need $1.5B to exploit this. You need a proven delivery engine that can be deployed into acquired firms. Build the engine. The capital will find you.

References

[1] karpathy/jobs — BLS Occupational AI Exposure — GitHub, Mar 2026. 342 US occupations scored 0–10. Average: 5.3.

[2] Kaito thread on Karpathy jobs release — X/Twitter, 14 Mar 2026. 12K likes. SW devs 8–9, roofers 0–1, transcriptionists 10/10.

[3] Geek — China AI job displacement map — X/Twitter, 18 Mar 2026. ~250 Chinese occupations. Right side all high-risk knowledge work.

[4] DisplaceIndex — Real-time US Labor Market AI Tracker — Mar 2026. Index: 44.7/100. 316K jobs cut total. 1.3% of layoffs. 17-point sentiment divergence.

[5] Jensen Huang: $500K Engineers Should Spend $250K on Tokens — Business Insider, Mar 2026. “Deeply alarmed” at $5K spend. 1:2 salary-to-token ratio.

[6] Every Engineer Will Have 100 AI Agents — Digit, Mar 2026. Engineers become orchestrators. Output shifts from volume to judgment.

[7] Founder’s Guide to AI-Enabled Roll-Ups — Capital Founders OS. Buy at 5–15% margins, automate 30–70%, re-rate to software multiples.

[8] Our Creation of Crescendo — General Catalyst. $500M valuation. 60–65% gross margins vs 25% industry average.

[9] AI Roll-Up Investors: GC, Thrive, Bessemer — Capital Founders OS. Crete: $300M+ revenue, 30+ firms, AT fastest-growing 2025.

[10] AI Roll-Up Technology: OpenAI, Proprietary AI — Capital Founders OS. OpenAI took equity + embedded engineers in Thrive portfolio, Dec 2025.

[11] Bezos Aims to Raise $100B for AI Manufacturing — Reuters, 19 Mar 2026. Project Prometheus. $6.2B raised. ADIA + JPMorgan in discussions.

[12] Why AI Services Will Outgrow AI Software — Valere Labs, Medium, Mar 2026. Services budget 6x software budget. Already priced on outcomes.

[13] State of Professional Services 2026 — TSIA. PS 1.0 (billable hours) → PS 2.0 (outcome-based). Not optional.

[14] State of AI for Technology Services 2026 — TSIA. Cannibalization dilemma: better AI = fewer seats = revenue leak.

[15] Ben Sigman on Karpathy jobs data — X/Twitter, 15 Mar 2026. $100K+ earners 67% exposed. Plumber safer than analyst.

[16] Bezos $100B AI Factories — Bull Theory, X/Twitter, 20 Mar 2026. Project Prometheus. Physical version of the same roll-up thesis.

[17] The $6 Trillion Shift — Agentic Private Markets. 30–40% of $6T pro services to migrate in 5 years. 1–2 year window to recalibrate.

[18] AI Agency Pricing Strategies 2026 — Digital Applied. Outcome-based: $5K/workflow. Agent licensing: $20K + $2K/mo.

[19] Adriana Sobota on exposure vs replacement — X/Twitter, 15 Mar 2026. “Exposure doesn’t mean replacement. It means the ones who learn the tools first get paid more.”