Bear
$20B
AI remains a strong copilot; open models, bundling and limited non-software adoption cap spend.
Bottom-up thematic research
When every engineer becomes an AI pilot
A bottom-up estimate of the mature annual market for LLMs and agent harnesses across software, hardware, industrial, mechanical, civil and adjacent technical engineering work.
Base
$85B
Professional engineers use managed agents daily, with software workflows monetizing most deeply.
Bull
$300B
About 140M engineers and technicians supervise persistent agents at roughly $180 per month.
The answer
The base case is an $85B annual market
The decisive variable is not whether engineers use AI. Software adoption is already broad. It is whether agent task volume grows faster than model prices fall—and whether CAD, EDA, simulation and industrial tools support the same depth of delegation as code.
Investment conclusion
Harnesses should be a large market, but the model layer does not automatically capture all of it. The durable tollbooth is the product that combines model intelligence with proprietary engineering context, execution rights, verification and auditability. In software that may be an agentic IDE or repository platform; in chip, mechanical and industrial design it is more likely the incumbent system of record.
What this TAM is—and is not
This is mature annual end-customer spend on engineering LLM inference, harnesses, orchestration, tool execution, evaluation, security and governance. It is not current revenue, productivity value, GPU infrastructure, consulting, or the existing price of CAD/EDA/PLM software. Accounting, legal and all other knowledge-work markets are excluded.
Eligible engineering workers
×
Paid AI penetration
×
Annual AI wallet
Counted once
- Model inference and metered agent runtime
- Engineering-agent and harness subscriptions
- Context, orchestration, evaluation and audit controls
- Incremental AI premiums in CAD, EDA, BIM and PLM
Excluded
- Upstream model COGS already embedded in a harness bill
- Servers, accelerators and general cloud infrastructure
- Existing non-AI engineering software and services
- Legal, finance, medicine, sales and other knowledge work
Current evidence
Software proves willingness to pay; physical engineering does not—yet
Heavy agent use already costs $100–$250 per developer per month
OpenAI guides to roughly $100–$200 and Anthropic to $150–$250 for active enterprise usage, far above a legacy $10–$40 copilot seat.S5S6
Distribution has reached millions of paid software seats
Microsoft reported 4.7M paid GitHub Copilot subscribers in FY26 Q2, up 75% year over year.S8
Step 1 · Workforce
Start with occupations, not “developers”
Millions of employed people, 2025. Codes are mutually exclusive main occupations. Source: ILOSTAT; Pharoi calculations.
How the global count was built
- Take each country’s latest complete ILO level-two occupation mix for codes 21–26 and 31–35.
- Calculate codes 21, 25, 31 and 35 as shares of their respective major groups.
- Reweight 121 covered countries to ILO’s 2025 modelled employment for professional and technical major groups.
- Gross up to world totals; covered countries represent 66.3% and 74.5% of the two major groups.
Biggest caveat: ILO lacks a complete global three-digit occupation panel. Code 21 contains some scientists, while code 31 contains some supervisors and controllers. China is the largest missing detailed-country mix and the main extrapolation risk.
Accessible workforce data table
| ISCO-08 group | Workers (M) | Role in scenarios |
|---|---|---|
| 25 · ICT professionals | 35.09 | All cases |
| 21 · Science & engineering professionals | 50.55 | Strict subset in bear; all in base/bull |
| 31 · Science & engineering associates | 42.33 | Bull |
| 35 · ICT technicians | 10.99 | Bull |
| Broad technical workforce | 138.96 | Rounded to 140M |
Steps 2–3 · Penetration and wallet
Three internally consistent end states
The bear resembles a managed copilot seat. The base resembles one daily agent per professional engineer. The bull resembles a software power user today—then extends that economic intensity to the broad technical workforce.
Bear · Copilot
$20B
36M paid seats · $47/month blended
- 60M strict eligible workers
- 70% ICT / 45% engineering paid adoption
- $600 / $480 annual segment wallet
- Bundling and open models suppress pricing
Base · Daily agent
$85B
72M paid seats · $99/month blended
- 86M professional eligible workers
- 90% ICT / 80% engineering paid adoption
- $1,800 / $700 annual segment wallet
- Managed agents plus inference and controls
Bull · AI pilot
$300B
139M paid seats · $180/month blended
- 100% of broad professional + technical workforce
- Multiple persistent agents and tool execution
- $75–$300 monthly wallet by occupation group
- Agent task growth outruns inference deflation
Annual end-customer vendor revenue, 2026 USD billions. “Strict engineering” in bear is an analyst subset of code 21; the bull adds codes 31 and 35.
Full scenario arithmetic
| Case / segment | Eligible workers | Paid adoption | Annual wallet | TAM |
|---|---|---|---|---|
| Bear · ICT professionals | 35.09M | 70% | $600 | $14.7B |
| Bear · strict engineering subset | 24.91M | 45% | $480 | $5.4B |
| Bear total | 60.00M | 60% blended | $562 / paid seat | $20.1B |
| Base · ICT professionals | 35.09M | 90% | $1,800 | $56.8B |
| Base · science & engineering professionals | 50.55M | 80% | $700 | $28.3B |
| Base total | 85.64M | 84% blended | $1,182 / paid seat | $85.2B |
| Bull · ICT professionals | 35.09M | 100% | $3,600 | $126.3B |
| Bull · science & engineering professionals | 50.55M | 100% | $2,400 | $121.3B |
| Bull · engineering associates | 42.33M | 100% | $1,000 | $42.3B |
| Bull · ICT technicians | 10.99M | 100% | $900 | $9.9B |
| Bull total | 138.96M | 100% | $2,158 / paid seat | $299.9B |
The “AI pilot” mechanism
Engineering shifts from producing to directing and verifying
The harness becomes more valuable than chat when it can observe the project, plan across tools, change artifacts, run tests or simulations, and return evidence for a human decision. Human attention moves toward specification, exception handling, trade-offs and sign-off.
01 · INTENT
Engineer sets the mission
Requirements, constraints, interfaces, risk budget and acceptance criteria.
02 · CONTEXT
Harness builds the state
Repositories, drawings, models, standards, telemetry and prior decisions.
03 · EXECUTION
Agents operate tools
Code, EDA, CAD, BIM, PLC, simulation, test benches and documentation.
04 · EVIDENCE
System verifies work
Tests, simulations, design rules, traceability, security and audit logs.
05 · JUDGMENT
Human handles exceptions
Trade-offs, novel failure modes, physical liability and final approval.
Software
Plan, implement, test, review, migrate, deploy and monitor across repositories and cloud systems.
Chips & hardware
Specification to RTL, verification, synthesis, layout, board design and lab-debug loops.
Industrial & mechanical
Parametric design, simulation, controls, manufacturing planning, quality and maintenance.
Civil & AEC
BIM authoring, code checking, quantity takeoffs, clash resolution and construction sequencing.
Sensitivity
A small wallet change moves tens of billions
Because the seat pool is enormous, monetization intensity matters more than false precision in the workforce count. These sensitivities are directional and hold other inputs constant.
+$12B
for each additional $10/month across 100M paid seats
+$10B
for 10 points of penetration on 110M seats at $75/month
−$17B
if the base engineering seat pool contracts by 20%
Flat
if agent task volume doubles while price per task halves
Value capture
The $85B is shared across a stack
This allocation is analytical—not a second market to add to TAM. Model vendors may sell harnesses; systems-of-record vendors may internalize model cost. Count the customer bill once.
55%
Models, inference & runtimeReasoning, multimodal understanding, parallel agent compute and tool execution
≈ $47B
30%
Harness & workflowAgent UX, planning, orchestration, repository/design-system integration
≈ $26B
15%
Context, evals & controlsEnterprise knowledge, verification, observability, security and governance
≈ $13B
| Research queue | Why it matters |
|---|---|
| MSFT / GitHub | Distribution, harness and Azure/model economics |
| GOOGL / AMZN | Models, cloud runtime and enterprise orchestration |
| CDNS / SNPS | Chip-design context, verification and executable workflows |
| ADSK / Siemens / PTC / Dassault | CAD, BIM, industrial and PLM systems of record |
| OpenAI / Anthropic / Cursor | Private leaders in model-plus-harness monetization |
Representative research queue, not recommendations. Infrastructure suppliers are second-order beneficiaries and intentionally outside the defined TAM.
Falsifiers
Why $300B may never materialize
The strongest counterargument is that intelligence commoditizes before vendors can monetize the extra work. Adoption can be universal while revenue remains modest.
Price falls faster than task volume rises
Open-weight models, distillation, caching and better chips can collapse inference cost and bargaining power.
Engineering seats shrink materially
If one pilot replaces several roles and new engineering demand does not offset it, per-seat ARPU rises against a smaller base.
Verification remains the bottleneck
“Almost right” work can increase review, simulation and debugging; liability keeps agents advisory in physical systems.
Incumbents bundle AI for free
CAD, EDA and PLM vendors may use AI to defend core subscriptions rather than expose an incremental AI line item.
Enterprise context stays fragmented
Proprietary formats, undocumented plant state and weak data lineage prevent agents from operating end-to-end.
Generated output does not become shipped output
Code, drawings and simulations can explode while releases, qualified designs and installed capacity improve far less.
What to monitor next
Paid agent ARPU; non-software paid attach rates; hours of autonomous tool execution; verified releases or qualified designs per engineer; gross margin after inference; AI-specific pricing at CAD/EDA/PLM vendors; safety and IP incidents; and whether engineering employment expands or contracts as the cost of producing engineering work falls.
Sources & method
Evidence ledger
Workforce and pricing inputs use current primary sources. Productivity evidence separates independent studies from company claims. Data cut-off: 25 July 2026, 13:00 BRT.
S1ILO detailed level-two occupation data, updated 24 Jul 2026. Primary statistical data.
S2ILO 2025 modelled major-group employment, Nov 2025 model vintage. Primary statistical data.
S3ILO ISCO-08 structure and definitions. Primary methodology.
S4ILO modelled-estimates methodology. Primary methodology.
S5OpenAI Codex rate card. Company-reported usage-cost anchor.
S6Anthropic Claude Code costs. Company-reported usage-cost anchor.
S8Microsoft FY26 Q2 earnings. 4.7M paid Copilot subscribers; primary company disclosure.
S9JetBrains 2026 developer survey. Survey of more than 10,000 professional developers.
S10Cui, Demirer et al., Management Science. Three field experiments, 4,867 developers; independent research.
S12Google DORA 2025. Technology-professional adoption and delivery evidence.
S13Anthropic Claude Code expertise study. Company research on planning versus execution behavior.
S14Cadence autonomous virtual engineer. June 2026 company announcement; early access.
S16Synopsys agentic engineering. March 2026 company announcement and performance claims.
Calculation conventions. Dollars are annual, nominal 2026 USD. “Wallet” is the customer’s total incremental
engineering-AI bill, including separately metered usage; it is not the provider’s gross revenue plus its upstream model bill.
Occupation counts are employed people in a main occupation, not current paid AI users. Scenario values are:
bear $20.1B, base $85.2B and bull $299.9B before rounding. The base value-pool split is a Pharoi analytical allocation,
not reported revenue. No probability weighting or target year is assigned.