I wrote earlier that frontier reasoning token prices will drop 10-100x over 3 years. A reader asked:

“What does it mean for white collar roles? If can’t be the same thing isn’t it? 22th June 2026 and I don’t have enough end point to comprehend this — extrapolate with high probability what will happen, just like the iPhone moment.”

Fair. Let me try.

The iPhone Moment — A Pattern for Inflection

The iPhone (2007) wasn’t the first smartphone. It was the first smartphone where the economics crossed a threshold:

Before iPhone (2005):
  Smartphone cost: $500-800
  Data speed: 100-200 Kbps (EDGE)
  Apps: Pre-installed, carrier-approved
  Touch: Resistive stylus
  Market: Business executives

After iPhone (2010):
  Smartphone cost: $200 (subsidized)
  Data speed: 1-7 Mbps (3G)
  Apps: 300,000+ in App Store
  Touch: Capacitive, multi-touch
  Market: Everyone

The threshold that mattered was not a single feature. It was the intersection of price, capability, and distribution reaching a point where new behaviours became economically rational.

Uber launched in 2009 — it needed GPS + 3G + app store + payment API. None of these existed at scale before 2007-2009. The iPhone didn’t create ride-sharing. It created the conditions for ride-sharing to be viable.

The token price collapse will do the same for cognitive labor.

The Token Price Trajectory

Architecture Diagram

This is not linear. Each generation of models is more capable than the last, so the price per unit of “reasoning quality” drops faster than the raw token price. A 2029 frontier model at $0.50/M tokens will be as capable as a 2026 model that doesn’t exist yet.

The Cognitive Cost Curve

The critical frame: what does it cost to do a unit of cognitive work?

Architecture Diagram

At $0.001-0.10 per hour of cognitive output, the marginal cost of “thinking” approaches the marginal cost of compute — essentially zero.

This is the threshold. When thinking costs near-zero, the question is no longer “should I use AI for this?” but “why wouldn’t I?”

Phase 1 (2024-2026): Augmentation — The Copilot Era

Architecture Diagram

Today. The human does the work; the LLM accelerates it. A lawyer writes a brief with AI drafting. A consultant builds a model with AI generating code. An analyst reviews AI-generated insights.

Impact on roles: - Junior roles compressed (less need for “ground work”) - Senior roles amplified (they direct AI instead of juniors) - Headcount may stay flat but output increases 2-5x

Who loses: Entry-level knowledge workers who traditionally spent 1-3 years learning by doing grunt work. That grunt work is now done by AI, so the training pipeline is disrupted.

Signs: - “We’re not hiring juniors this year. We’ll use AI.” - “Every analyst now produces director-level output.” - Billable hours models under strain.

Phase 2 (2026-2028): Agentification — The Delegation Era

Architecture Diagram

2026-2028. Token prices drop enough that running multiple agents for hours on a single problem is economical. The human shifts from “doing” to “directing.”

A single senior analyst now manages 3-5 AI agents that: - Monitor industry news across 100+ sources - Generate weekly reports in natural language - Surface anomalies requiring human attention - Draft responses, recommendations, and briefs

Impact on roles: - Middle management layer thins (AI tracks project status, generates reports, schedules) - Specialist roles (research analyst, compliance reviewer, data analyst) compress - Generalist roles that coordinate specialists become more valuable - The “10x engineer” becomes “100x engineer with 5 agents”

Who loses: - Mid-level managers whose primary job is information aggregation and reporting - Specialists whose expertise is narrow and learnable - Any role where 80% of the job is “read, summarize, write”

Signs: - “We run 50 agents per team. Humans handle exceptions.” - “my job went from writing reports to writing prompts and verifying outputs.” - Companies announce “AI-native” org structures with flat hierarchies.

Phase 3 (2029+): Structural Shift — The Thinking-as-Commodity Era

Architecture Diagram

By 2029, when frontier reasoning costs 10-100x less than today, the organizational structure of knowledge work inverts:

The old stack (pyramid): - Few executives at top - Layers of managers, analysts, associates - Each layer adds cost and delay - Information flows up, decisions flow down

The new stack (thin): - Strategists define goals and constraints (small, high-value) - Builders create the agent systems, prompts, and guardrails (medium) - Verifiers sample-check AI output for quality (small) - AI agents do the actual analysis, writing, coordination, execution (massive)

The “middle” of the pyramid — the analysts, associates, coordinators, reviewers — compresses because the AI agent layer absorbs their function.

The iPhone Moment for Knowledge Work

The iPhone didn’t kill the mobile phone industry. It killed: - Standalone GPS devices (TomTom) - Standalone MP3 players (iPod) - Point-and-shoot cameras - Physical keyboards (Blackberry) - Printed maps

But it also created: - App economy (iOS developer, QA, designer) - Gig economy (Uber driver, Deliveroo) - Creator economy (Instagram influencer, YouTuber) - Mobile-first banking, payments, health, dating

The common pattern: specific devices/roles got absorbed into a general platform. The iPhone didn’t replace “having a phone.” It replaced “having a separate device for each function.”

The token price collapse does the same for cognitive labor:

Roles that get absorbed: - Research analyst - Compliance reviewer - Data analyst (basic) - Legal associate (document review) - Accounting clerk - Customer support (tier 1-2) - Report writer - Content moderator

Roles that bifurcate: - Software engineer → Strategist + Verifier + Builder - Lawyer → Strategist + Verifier + Builder - Consultant → Strategist + Verifier + Builder - Accountant → Strategist + Verifier + Builder

Roles that strengthen: - The person who defines what to do (domain expert + strategist) - The person who validates AI output (domain expert + verifier) - The person who builds the agent system (engineer + domain expert) - The person who owns the relationship (sales, BD, therapy, negotiation)

Demand Elasticity — The Missing Piece

Here’s the counter-argument everyone misses:

When token prices drop 10-100x, demand for cognitive output explodes.

Think about bandwidth:

1995: 56k modem → people browsed text
2005: 10 Mbps → people watched YouTube
2015: 100 Mbps → people streamed 4K
2025: 1 Gbps → people run 50 devices

Bandwidth didn’t get cheaper and people stopped using the internet. People used more internet because new use-cases became viable.

Same for tokens:

Architecture Diagram

At today’s prices, you use AI for important tasks. At 10x cheaper, you use AI for every task. At 100x cheaper, you use AI for tasks you didn’t even know existed because they weren’t worth doing manually.

This means:

The Three Scenarios

Architecture Diagram

My assessment: between Base and Bull.

The analogies that give me confidence:

  1. Excel didn’t kill accountants. It killed bookkeepers and created financial analysts. The number of accounting jobs increased after spreadsheets because cheap calculation made analysis economical.
  2. Google didn’t kill researchers. It killed librarians and created SEO specialists, data journalists, and information architects.
  3. AWS didn’t kill sysadmins. It killed server rackers and created cloud architects, SREs, and DevOps engineers.

Each wave of automation expanded the scope of the domain while compressing the execution layer.

What I’d Tell Someone Starting Their Career Today

Architecture Diagram

The winning human in 2029 is not the one who competes with AI on execution. It’s the one who:

  1. Knows what to do (domain expertise + strategy)
  2. Can build or direct the AI system (prompts, agents, tools)
  3. Can verify the output (deep enough to catch hallucinations)
  4. Owns the relationship (the client/stakeholder trusts them, not the AI)

These skills compound. They cannot be automated because they require: - Context that exists outside the prompt (relationship, trust, history) - Judgment that cannot be reduced to a training set (what should we do?) - Accountability that cannot be assigned to software (who gets blamed?)

The Real iPhone Moment

The iPhone moment for white-collar work is not “AI replaces your job.”

The iPhone moment is: the marginal cost of high-quality reasoning approaches zero, and everything built on expensive reasoning gets rebuilt.

Just as no one in 2005 predicted Instagram (a photo-sharing app that exists entirely because iPhone had a good camera + 3G + always-on), no one today can predict what happens when:

These create new categories of work that don’t exist today: - Agent deployment engineer - AI output quality auditor - Prompt systems architect - Human-AI interaction designer - AI risk modeler

Summary

Today (2026) 2029
Frontier token price $2-5/M $0.02-0.50/M
Cost of 1hr cognitive work $0.10-1.00 $0.001-0.10
Primary AI role Copilot Agent
Human role Doer + Reviewer Director + Verifier
Org structure Pyramid (many layers) Thin (strategists + builders + verifiers + agents)
Total white-collar employment Baseline -20% to +30% (composition shift)
Best skill to have Domain expertise Domain expertise + AI orchestration + verification

The juice is not in predicting doom. It’s in understanding that when the marginal cost of reasoning drops 100x, the demand for cognitive output grows more than 100x. Total human employment in the cognitive sector may not shrink — but which humans and what they do will change completely.

This is the iPhone moment. Not the death of the old industry. The birth of categories we can’t name yet.

This is extrapolation, not prophecy. I’ve been wrong before. But the economic forces (token price dropping 100x, demand elasticity >1, platform economics) are structural, not speculative. History suggests structural shifts in input costs produce structural shifts in output organization.