Structural Research · September 2026
The Agentic
Economy.
Where does the next wave of AI
become earnings?
Choose your reading pathStart with the investment brief, explore the demand themes, or examine the company cases and their failure conditions.
Chapters Investment brief
Finding the bottlenecks and second-order beneficiaries of frontier AI
A breakthrough changes the question.
GPT-6 Astra prompted me to revisit a question that sits behind every major technology cycle: when a new capability becomes practical, where does the economic value actually accrue?
I regard the transition toward models that can operate software, use specialized tools, and carry a task through successive stages of execution as a meaningful breakthrough. It changes who can attempt work that once required a team of developers, designers, analysts, or technical operators. The investment implication begins with the projects that become economical as that coordination burden falls.
The distinction matters because technological capability, commercial adoption, and equity returns follow different paths. A model may generate more designs without producing more manufacturing orders. It may help launch more applications without expanding their revenue. It may consume more tokens while improvements in routing, caching, and hardware reduce the physical resources needed for each accepted task.
This research follows those paths. It begins with what frontier agents can do and the constraints that remain. It uses the memory cycle to examine how scarcity reaches earnings, then screens the infrastructure stack for the next constrained layer. It also maps businesses that may benefit when a wider population can create interactive experiences, commercialize products, conduct scientific work, and operate specialized workflows.
My aim is to identify the point at which an expanding activity creates a payment that a supplier can retain. Valuation, balance-sheet resilience, and evidence that would invalidate the thesis remain central. Supervised execution is the working assumption throughout; the framework does not require dependable, universal autonomy.
Investment brief
From capability to cash flow.
Frontier agents can lower the minimum economic size of a project. Value capture begins at the next paid step.
Capability, bounded
Agents can browse, code, manipulate files and operate software today. Reliability remains task-specific and below universal autonomy; supervised execution is the base case.
Scarcity, priced
Optical inputs, advanced packaging and high-voltage equipment remain plausible physical constraints. Headline equities already reflect much of that scarcity.
Participation, expanding
The differentiated demand thesis is an order-of-magnitude increase in attempted projects. Paid conversion will be far lower than creation growth.
Earnings, selective
Memory's lesson was earnings multiplication. For agentic beneficiaries, monitor paid units, cohort retention and gross profit—not prompts or generated assets.
Capability reality
Everything downstream is conditioned on what agents can complete now—not on universal autonomy.
Shipping capability set
Closed-loop screen perception, mouse and keyboard control, browser and desktop navigation. Useful for heterogeneous and legacy software, with consequential actions still requiring review.
CLI, APIs, sandboxes, repository work, data analysis and multi-step debugging. Stronger than GUI operation where deterministic interfaces are available.
Search, source reading, spreadsheets, documents, media and structured deliverables; reliability depends on source access and format complexity.
Long context, retrieval and structured notes improve continuity. They do not constitute lossless, non-degrading memory across indefinite tasks.
Supervised, multi-session work is credible. Dependable unsupervised operation across days and irreversible workflows is excluded from the base case.
Cost curve and economic implication
No headline model-cost comparison is used because the inherited basket did not specify model pairs, token mix, caching, tools, retries, or accepted outcomes. The economically relevant unit is cost per accepted task: attempts + tools + sandbox + review divided by accepted outcomes. Better routing, caching and reliability can reduce this even while the frontier token price rises.
12–24 month roadmap
- Base case: supervised workflows span more applications, tolerate course corrections and continue across sessions.
- Credible: better checkpointing, verification, permissions and specialized sub-agents.
- Speculative: reliable, unsupervised general work over days with negligible exception handling.
APIs remain faster, cheaper, more auditable and less brittle where they exist.
Retrieval errors, compaction and state inconsistency remain material failure modes.
Caching, routing, sparse inference and selective retrieval break the one-to-one mapping.
Measurement and assumptions
OpenAI reports 72.6% on the offline OSWorld 2.0 partial-score evaluation and roughly 40 minutes per task in its latency simulation. Partial credit is not a complete-task success rate or a reliability guarantee for a particular business workflow. OpenAI evaluation ↗
No numerical token-to-resource or token-to-demand multiplier is used in this edition. Those conversions require a named workload, model mix, caching and hardware baseline that this report does not establish.
The historical cost-basket figures are retained only in the archived original and correction log. The documented standard Astra rates at the evidence cutoff are $10 per million input tokens, $1 cached input, and $50 output, with separate long-context pricing. Cost per accepted task remains the relevant operating measure. Model documentation ↗
Memory episode autopsy
The 2024–26 event was several overlapping cycles, not a single feature-driven shock.
AI accelerator demand reached qualified HBM capacity. Long qualification and packaging requirements limited substitutability.
Producer restraint and inventory normalization supported pricing independently of persistent-memory product features.
Conventional DRAM prices rose approximately 93–98% QoQ, broadening the earnings event beyond HBM.
Sandisk annual non-GAAP diluted EPS expanded from FY2025 to FY2026, from $2.99 to $70.88; an approximately 23.7× increase. This establishes substantial earnings expansion. The original 10× share-price and ~24× trailing-P/E observations lacked a specified price date and return interval, so they are not used here as verified return attribution.
The six-gate signature
- Deliverable capacityDemand reaches capacity that is qualified and shippable.
- Substitution resistanceScarcity survives redesign, alternate suppliers and product substitution.
- Price captureThe constrained supplier receives the price increase.
- Operating leverageIncremental revenue reaches profit after input and expansion costs.
- Earnings durationProfits persist beyond a short inventory or procurement cycle.
- Conservative expectationsConsensus and valuation do not already capitalize the outcome.
Memory evidence: TrendForce conventional DRAM survey ↗ and Sandisk annual results and GAAP reconciliation ↗. Sandisk’s non-GAAP measure differs from GAAP earnings; these are annual figures, not quarterly annualizations.
Supply bottleneck screen
Scores are MCO ordinal judgments, not company-reported measurements: 5 indicates the greatest scarcity, lead time, pricing signal or price capture. A higher supply-inelasticity score means supply responds less readily. “Already priced” is a qualitative expectations judgment; it is not a measured percentage of future demand.
| Layer | Supply inelasticity | Lead time | Pricing signal | Price capture | Already priced | Assessment |
|---|---|---|---|---|---|---|
| Qualified optical inputslasers, EMLs, indium phosphide, test | 5 | 5 | 4 | 4 | 4 | #1 conditional Qualification and yield can protect scarcity; valuation requires current, security-specific work. |
| Advanced packaging & substratesCoWoS-class capacity, ABF, inspection | 5 | 5 | 4 | 4 | 5 | #2 conditional Complex qualification and concentrated supply; valuation requires current, security-specific work. |
| Transformers & HV equipmentlarge power transformers, switchgear | 5 | 5 | 5 | 4 | 5 | #3 conditional Multi-year lead times and grid interconnection delays; valuation requires current, security-specific work. |
| HBM / DRAM | 4 | 4 | 5 | 5 | 5 | Real scarcity and pricing power, but the memory rerating is already visible. |
| Accelerators | 4 | 4 | 4 | 5 | 5 | Strategic choke point; headline exposure is the most thoroughly arbitraged. |
| Networking systems | 3 | 3 | 4 | 4 | 5 | AI fabric demand is established; component-level qualification offers better asymmetry. |
| NAND / HDD | 3 | 4 | 3 | 3 | 4 | Agents increase temporary state and data movement more reliably than permanent storage. |
| Power generation | 3 | 5 | 4 | 3 | 5 | Demand visible; permitting, fuel and regulated returns disperse value capture. |
| Cooling | 3 | 3 | 4 | 3 | 5 | Liquid cooling scales with rack density, but supplier differentiation can erode. |
| Data-center construction | 2 | 4 | 4 | 2 | 4 | Labor and sites are constrained; project economics and customer concentration cap capture. |
| Orchestration CPUs | 2 | 2 | 2 | 2 | 3 | Necessary capacity, but substitution and system integration limit bottleneck economics. |
Qualified optical inputs
Transmission: larger clusters and scale-up/scale-out traffic increase optical lanes; qualified laser and component supply must keep pace.
Watch: lead times, laser utilization, yield, EML pricing, 800G/1.6T shipments.
Relief: new qualified capacity, silicon-photonics redesign and better yields, likely over multiple qualification cycles.
Packaging & substrates
Transmission: accelerator complexity increases advanced-package area, interposer demand and substrate difficulty.
Watch: package allocation, ABF utilization, inspection intensity and yield.
Relief: capacity expansion, package redesign, panel-level progress and customer qualification.
Transformers & HV
Transmission: dense data centers convert model demand into interconnection and power-delivery equipment.
Watch: transformer lead times, backlog price/cost, utility interconnection queues.
Relief: factory additions, standardized designs, behind-the-meter generation and grid upgrades.
Second-derivative search map
These are research lanes rather than endorsed securities. The six gates must be applied at the individual company level.
| Constraint | Less-obvious value pool | What must be proven |
|---|---|---|
| Optics | Indium-phosphide wafers, optical test, isolators, precision assembly, laser packaging | Qualification protects share; price reaches the supplier; capacity is not rapidly replicated. |
| Packaging | ABF chemistry, temporary bonding, thermal interface materials, metrology, warpage inspection | Content per package rises faster than customer-driven price pressure. |
| HV equipment | Transformer cores, electrical steel, bushings, tap changers, protection relays, regional assemblers | Supplier captures backlog repricing without working-capital or raw-material leakage. |
Valuation and score discipline
Historical forward-multiple estimates have been removed from this edition because their price date, consensus provider, and forward fiscal period were not specified. Individual-company gates are MCO qualitative research judgments; they are intentionally incomplete where evidence is incomplete and do not constitute a buy rating.
Demand-side map
The primary effect is a lower minimum project size. Select a theme to inspect its transmission mechanism and gating variable.
Interactive games and experiences
Agents join asset creation, scripting, debugging, playtesting and iteration. A teacher, community, small brand or individual can produce an experience whose audience was previously too small to support a studio.
2026–28 · supervised creation works now · research scenarioDistribution, discovery, virtual economies and selected development services.
Representative exposure- Roblox
- Unity
Attention and retention. More published experiences can dilute discovery and raise trust-and-safety cost before spending grows.
Read all seven themes together
Interactive games and experiences
Agents join asset creation, scripting, debugging, playtesting and iteration. A teacher, community, small brand or individual can produce an experience whose audience was previously too small to support a studio.
Paid completion: Distribution, discovery, virtual economies and selected development services.
Constraint: Attention and retention. More published experiences can dilute discovery and raise trust-and-safety cost before spending grows.
2026–28 · supervised creation works now · research scenario
Physical-product commercialization
Agents reduce coordination costs across market research, specifications, design changes, production files, supplier comparison, packaging and reorder administration. This lowers the revenue threshold at which a niche product is worth attempting.
Paid completion: Prototypes, production, customization, logistics and qualification.
Constraint: Manufacturability, funding and actual product demand. Uploaded designs and quote requests are not revenue.
2026–29 · adoption follows physical validation · research scenario
Experimental biology as a service
Smaller scientific teams can orchestrate computational design, bioinformatics and experimental preparation with less internal technical labor. Scientific judgment and funded programs remain necessary.
Paid completion: DNA synthesis, protein production, characterization and model-ready experimental data.
Constraint: Useful hypotheses and research funding. Better prediction can reduce experiments per successful outcome.
2027–30 · specialist-led diffusion · research scenario
Personalized investment workflows
Nonprogrammers can construct portfolio analysis, screening, risk and rebalancing workflows through natural language. The practical near-term architecture retains human review at order submission.
Paid completion: Funded brokerage accounts, custody, financing, FX and applicable transaction economics.
Constraint: Trust, retained assets and revenue per funded account. More analysis or commission-free equity orders need not produce proportional revenue.
2026–28 · approval-gated workflows · research scenario
Commercial infrastructure for new software
Agents make it easier to launch and maintain niche applications, run experiments and operate subscriptions. Successful projects still need monetization infrastructure and customer acquisition.
Paid completion: Subscription management, payments, billing optimization and distribution.
Constraint: Application survival and profitable acquisition. A sevenfold rise in launches can coexist with negligible revenue from the new cohort.
2026–28 · creation already inflecting · research scenario
Microenterprise operating services
The deeper barrier after website creation is operating a business: invoicing, collections, reconciliation, filings and exception handling. Agents can lower the minimum revenue required to justify that burden.
Paid completion: Accounting, payments, filings, registered-agent services and recurring compliance.
Constraint: Business formation and survival. Agents can substitute for elementary bookkeeping and document preparation.
2027–29 · requires durable business formation · research scenario
Testing and commercial assurance
More commercial products can require physical testing, inspection and certification. A model can prepare evidence and documentation but cannot establish physical performance by assertion.
Paid completion: Laboratory testing, inspection, certification and regulatory assurance.
Constraint: Only a fraction of generated designs reach regulated commercial production; diversified incumbents dilute the increment.
2027–30 · follows commercialization · research scenario
Creation is not the revenue denominator
RevenueCat's 2026 data illustrate the gap: monthly subscription-app launches rose from about 2,000 in January 2022 to more than 14,700 in January 2026, while apps launched in 2025 or later produced only 3% of subscription revenue. Source ↗
Business-model effects
Unavoidable completion services
Fabrication, experimental validation, payments, custody, qualification and delivery earn when a digital output becomes a real transaction.
Platforms with distribution
Creation abundance increases the value of discovery, monetization, identity and trusted marketplaces—if the platform controls the paying relationship.
Human-seat software
More output may require fewer human operators and therefore fewer seats. Usage-based monetization is structurally better aligned.
Generic production services
Routine coding, standard creative work, basic document preparation and undifferentiated content libraries face direct substitution.
Advertising
More businesses and apps increase competition for customers; agent-mediated discovery can reduce merchandising, search ads and legacy attribution value.
Travel and booking
Agents may improve conversion and lower service cost, while compressing OTA take rates, loyalty and sponsored placement. Household budgets still govern trip volume.
Thirteen expressions
Filter by research priority or demand theme. Research priorities measure exposure to the usage thesis; they are not buy ratings and do not replace valuation work.
Showing 13 of 13 companies
Roblox
Interactive worldsAgentic Studio Assistant users +20% QoQ in Q2 2026; Q2 bookings +8%.
- Why it fits
- Turns consumers into persistent creators while retaining distribution, social graph and virtual-economy capture.
- Failure condition
- Creation rises but discovery, retention and bookings do not; Q3 2026 guidance already warns of near-term contraction.
- Framework read
- Monitor paid conversion, retention, bookings and valuation expectations.
Evidence: Roblox Build · Roblox Q2 shareholder letter
Xometry
Physical productsQ2 2026 marketplace revenue +45%; active buyers +20% to 89,557.
- Why it fits
- Captures the paid handoff from agent-assisted design and sourcing to qualified manufacturing fulfillment.
- Failure condition
- Quote traffic rises while conversion, repeat orders or marketplace gross profit decline.
- Framework read
- Monitor conversion, repeat orders, marketplace gross profit and valuation.
Evidence: Siemens × Xometry · Xometry Q2 results
Twist Bioscience
Experimental biologyFiscal Q3 2026 DNA Synthesis & Protein Solutions +39% to $56.6m.
- Why it fits
- Computational exploration creates physical demand for synthesis, protein production and characterization.
- Failure condition
- Better prediction reduces experiments faster than funded programs and iteration increase them.
- Framework read
- Monitor funded-program demand, experiment intensity, margins and balance-sheet resilience.
Evidence: Twist fiscal Q3 · Twist AI discovery workflow
Interactive Brokers
Financial workflowsOpen AI connectivity introduced in July 2026; users review and submit every order.
- Why it fits
- Natural-language operation expands access to sophisticated portfolio functions while funded custody anchors economics.
- Failure condition
- Usage remains occasional free analysis with no improvement in assets, retention or monetized activity.
- Framework read
- Monitor funded-account growth, retention, paid activity and incremental economics.
Evidence: IBKR AI integration
RevenueCat
Software commerceFree through $2,500 monthly tracked revenue, then standard pricing at 1% of tracked revenue.
- Why it fits
- Business-model exposure to commercial subscription apps after agents lower the operating burden for small developers.
- Failure condition
- New apps fail to monetize or model/app platforms bundle subscription infrastructure.
- Framework read
- Monitor new-app survival, paid attachment, platform bundling and pricing power. Private company.
Evidence: RevenueCat pricing · Subscription Apps 2026
Cimpress
Physical productsFY2026 revenue +10%; organic constant-currency growth +4%.
- Why it fits
- More small organizations can manage catalogs and purchase customized physical output without a design team.
- Failure condition
- More designs divide fixed merchandise budgets and small orders raise fulfillment cost.
- Framework read
- Monitor order economics, fulfillment costs, customer demand and balance-sheet capacity.
Evidence: Cimpress FY2026
Unity
Interactive worldsQ2 2026 Create revenue +2%; Grow revenue +35%.
- Why it fits
- A larger creator population can graduate into professional tools and commercial monetization.
- Failure condition
- Free creation expands while paid conversion remains weak or competing environments capture the cohort.
- Framework read
- Monitor paid conversion, professional-tool adoption, competitor capture and valuation.
Evidence: Unity Q2 results · Unity AI pricing
Robinhood
Financial workflowsSeparate funded agent trading accounts launched in beta; initial trading scope was equities.
- Why it fits
- Consumer financial automation could raise funded engagement, product attachment and wallet share.
- Failure condition
- Novelty creates churn, poor outcomes and service burden; more equity orders add little revenue.
- Framework read
- Monitor adoption quality, service burden, wallet share and incremental revenue.
Evidence: Robinhood agent accounts
AppLovin
Software commerceQ2 2026 revenue +53% to approximately $1.92bn.
- Why it fits
- If development costs fall faster than distribution costs, viable apps redirect more budget toward acquisition.
- Failure condition
- New applications lack sufficient lifetime value to advertise profitably.
- Framework read
- Monitor app economics, advertiser demand, customer concentration and valuation.
Evidence: AppLovin Q2 results
Intuit
MicrobusinessFY2026 Global Business Solutions +16%; Q4 QBO +20% partly from price and mix.
- Why it fits
- Administrative agents may improve formalization and survival, supporting accounting and payments revenue.
- Failure condition
- Customer formation stays muted while agents commoditize bookkeeping and reduce payroll seats.
- Framework read
- Monitor new-business formation, paid attachment, mix and increment materiality.
Evidence: Intuit FY2026
Protolabs
Physical productsQ2 2026 revenue +10.6%, but customer contacts served −5.3%.
- Why it fits
- More projects can progress through paid physical iteration and small-batch production.
- Failure condition
- Current pattern persists: greater spend per existing customer without broader creator participation.
- Framework read
- Monitor customer breadth, repeat activity, paid conversion and gross profit.
Evidence: Protolabs Q2 results
SGS
AssuranceH1 2026 organic growth +5.6%; Connectivity & Products +6.8%.
- Why it fits
- Commercial product proliferation can create additional testing and qualification work.
- Failure condition
- Too few generated designs reach regulated production to affect a diversified revenue base.
- Framework read
- Monitor regulated output, testing demand, pricing and incremental materiality.
Evidence: SGS H1 results
LegalZoom
MicrobusinessQ2 2026 business formations −5%; subscription revenue +11%.
- Why it fits
- Formal businesses require recurring compliance and registered-agent services.
- Failure condition
- AI substitutes for entry-level preparation faster than new formations create demand.
- Framework read
- Monitor formations, subscription retention, substitution pressure and pricing.
Evidence: LegalZoom Q2 results
Participation-to-revenue engine
An illustrative identity for the agent-enabled cohort. It is a sensitivity tool, not a forecast.
Participation expands faster than monetization. Paid completion and repeat usage determine how much economic value survives.
Multipliers describe the agent-enabled cohort relative to its baseline. They do not estimate total company revenue, profit, valuation, or investment returns.
Assisted execution
Digital creation, research, portfolio preparation and workflow orchestration expand first. Human approval concentrates at publication, purchase and transaction points.
Supervised multi-session work
Persistent state, verification and permissions support longer commercial operations and more reliable conversion from idea to completed output.
Physical and regulated diffusion
Qualification, capital, liability and real-world testing keep adoption slower in manufacturing, biology and regulated activities.
How the framework fails
The largest risk is a productivity boom without a corresponding profit pool for the selected intermediaries.
Inference efficiency
Efficiency improvements can reduce physical-resource intensity per task. This report does not use a headline throughput or resource-demand multiplier because it has not established a comparable workload, hardware baseline, or deployment mix. Optical, power, cooling and packaging demand can still grow, but each requires security-specific evidence.
Most exposed:Supply-chain names underwritten on permanent shortage.
Reliability plateau
Agents remain good enough for demonstrations but require too much supervision for routine economic use. Exception handling restores the labor cost the model was expected to remove.
Most exposed:Manufacturing newcomers, microbusiness automation and autonomous consumer workflows.
Monetization failure
Creation explodes, attention and household budgets do not. New games, apps and products divide a stable profit pool while distribution platforms absorb support and infrastructure costs.
Most exposed:Roblox, Unity, RevenueCat, AppLovin and Cimpress.
Capex digestion
Hyperscalers pause after a large build cycle, and utilization takes time to catch installed capacity. Scarcity premiums unwind before downstream agent demand matures.
Most exposed:Optics, packaging and power equipment.
Bundling and disintermediation
Model and operating-system platforms bundle billing, identity, orchestration, design and deployment. Standalone intermediaries lose price and customer ownership.
Most exposed:Unity, LegalZoom, RevenueCat and generic SaaS tooling.
Prediction substitutes for testing
Better scientific models reduce experiments required per successful result. Digital search expands while physical synthesis and validation grow more slowly.
Most exposed:Twist and broader research-service providers.
Know when the thesis is working
Track paid conversion and earnings transmission. Generated volume is an upstream diagnostic, not the objective.
First-time manufacturers, commercial creators, funded workflow users and smaller scientific buyers expand.
Activity is concentrated in existing large customers.
Published projects, quotes and designs convert to orders without falling unit economics.
Attempts rise while paid conversion declines.
Repeat orders, renewed subscriptions, retained creators and funded accounts improve.
One-time experimentation dominates.
Scale absorbs fixed costs and service intensity stays controlled.
Rework, support, fraud or inference costs consume revenue.
Recent cohorts gain a rising share of revenue, engagement or useful scientific outcomes.
Legacy products retain nearly all economics.
Review concentrates at consequential approvals.
Troubleshooting persists across ordinary steps.
Optical, package and HV lead times persist while backlog converts at attractive margins.
Capacity arrives, utilization falls and price resets.
Beneficiaries retain the customer relationship and paid attachment.
Model platforms bundle services or redirect users.
What I would underwrite next
My conclusion is that the investment opportunity begins where easier execution changes who can participate, and where that participation becomes a payment a business can retain. More prompts, designs, applications or experiments are not enough. I want to see new paying customers, repeat activity and incremental gross profit that is material to the company receiving it.
I would pursue two research tracks. In infrastructure, I would look for constraints that survive substitution and capacity expansion, then ask how much of that scarcity the share price already assumes. In downstream businesses, I would examine whether agents bring genuinely new participants into an economic activity, rather than simply help existing customers produce more within the same budget.
Roblox, Xometry, Twist Bioscience and Interactive Brokers provide different settings in which to test that question. RevenueCat offers another business-model example, but it is private, not a listed-equity candidate. These are research priorities, not completed investment cases. The next step is to connect each operating thesis to company-wide earnings, cash requirements, dilution and valuation.
The evidence that would change my mind is equally important. If creation expands without funded demand, if new customers fail to return, or if support costs and price competition absorb the incremental revenue, I would reduce conviction. If model platforms take control of distribution or bundle the paid step, I would revisit the proposed beneficiary altogether.
I regard the capability transition as consequential. I do not regard the shareholder outcome as predetermined. My task is to establish who gets paid, how much survives as earnings, how durable those earnings may be, and what I am being asked to pay for them.
Bring this framework into your investment process.
Discuss how this research could inform your organization’s underwriting, portfolio review or investment committee agenda.
Discuss this researchPrimary sources
Selected documents supporting material company and operating claims. Forward conclusions remain analyst inference.
This research synthesizes prior work completed through 5–7 September 2026. Historical market-multiple estimates have been removed where their inputs could not be reproduced. The report is analytical research, not a recommendation to transact.
How to read the evidence
Reported company metrics establish operating context; they do not isolate AI as a cause. Scenario timing, research priorities, supply scores, and transmission mechanisms express MCO research judgment. Model evaluations are vendor-reported results under defined conditions. The correction log records retired figures whose measurement basis was not reproducible.