Where a Fellow-Level Technical Leader Can Have the Most Leverage: A 2026 Ranking of Venture Frontiers

The venture market in 2026 looks generous and narrow at the same time. According to the KPMG Venture Pulse snapshot, $227.4 billion was invested globally in the second quarter, while deal count fell to a level not seen since 2017. Large late-stage AI rounds heavily shape the total. For a new founder, this is not a promise of easy money but a warning: capital exists, yet it concentrates where scale is already visible or the company has access to scarce data, infrastructure, or distribution.

The technology transition is far from complete. The Stanford AI Index 2026 reports that 88% of surveyed organizations used AI in 2025, while agent deployment remained in the single digits across almost every function. Models are useful enough to create demand, but the production layer of trust, authority, verification, and recovery has not yet become ordinary infrastructure.010101I covered the same gap on my channel through McKinsey’s survey: roughly two thirds of companies have not yet scaled AI enterprise-wide, yet 62% already at least experiment with agents. The distance between interest and operation is exactly the opening worth entering.Knizhny kub · McKinsey’s state-of-AI report (in Russian)01I covered the same gap on my channel through McKinsey’s survey: roughly two thirds of companies have not yet scaled AI enterprise-wide, yet 62% already at least experiment with agents. The distance between interest and operation is exactly the opening worth entering.Knizhny kub · McKinsey’s state-of-AI report (in Russian)

This leads to the ranking’s central thesis. For a Fellow-level leader, the best territory is not inside the next general-purpose model but between the model and an expensive real-world outcome. That gap demands simultaneous domain understanding, reliable system design, negotiated risk boundaries, workflow redesign, and production operations. One excellent engineer cannot cover it; neither can one excellent manager.

rare leverage = technical depth × the right to enter the workflow × the ability to build operations × willingness to own the consequences

Each frontier is rated from zero to ten on six criteria. The weights are specific to the stated profile: one quarter of the score asks how much the combination of technical depth, management, and hands-on agentic building changes the odds. Market size matters, but without a fast wedge and a compounding advantage it becomes a decorative investor-deck number. The nine themselves were assembled by expert judgment: there was no long formal list with a cutoff, and the scoring compares already-selected frontiers plus the four nearest candidates from section six.

The resulting scorecard is not an expected-return calculation. It compares the starting position of someone without access to classified defense projects, a clinical network, an energy asset, or an exclusive manufacturer agreement. If that access already exists, adding 0.5–1.0 points to the relevant frontier and rerunning the choice is reasonable. When total scores tie, the frontier with the faster first signal ranks higher; if speed also ties, profile leverage decides.

score = 25% profile leverage + 20% willingness to pay + 15% wedge speed + 15% compounding advantage + 15% venture scale + 10% portability

The table should be read by tier, not by ordinal. The ranking has only two meaningful boundaries: after second place, where the gap reaches 0.4 points, and before ninth, where it is 0.5. The top two differ by just 0.2—inside the noise band—yet both pull clearly away from the middle. The middle, third through eighth, sits entirely within 8.2–8.5: pairs 03–04 and 07–08 are exact ties, and no adjacent scores diverge by more than 0.2. The faster-first-signal rule above settles those ties, but it only explains how the numbers came out — not how much they mean: the order inside tier B is nominal, and the numbers exist as section navigation, not as a ranking.

The second honest caveat concerns the weights. An equal-weighted sum of the same six scores produces the same order—the ranking is held up by the scores themselves, not by the weight profile. Weights start to matter when the reader changes. For a strong engineer without management experience, I would drop profile leverage from 25% to 15% and raise wedge speed from 15% to 25%: energy then moves into third place, healthcare catches up with industrial autonomy, autonomous R&D slips to sixth, and both tier boundaries stay where they are. That is what makes the method portable: the score table is shared, the weight profile is yours.

Finally, gates operate on top of the score—the same four that close the ninety-day expedition at the end of this article. The score measures the strength of a position after entry; it does not substitute for the right to enter. For two frontiers that right is binary, and their scores are marked as conditional in the table: healthcare without a clinical and regulatory partner, and dual-use systems without a chosen jurisdiction and clearance, do not slide down a few places—they drop out of the comparison entirely. The gate first, the score second.

The map exposes an important tradeoff. A fast signal does not mean an easy entry: healthcare can deliver a verifiable software outcome quickly, yet sits high on the map because of clinical responsibility; energy and industry move more slowly but generate hard operating data. Dual-use systems and space ventures can grow into very large businesses, yet they require accepting long cycles, jurisdictional limits, and costly verification from the outset. The map axes deliberately do not repeat the table’s criteria: the horizontal is close to wedge speed, while “entry burden” folds together the regulatory, capital, and field barriers that the scorecard spreads across portability, willingness to pay, and compounding advantage. The positions are placed by judgment, not recomputed from the scores.

01 · Vertical agentic systems of action — 9.1

The strongest general bet is not an assistant for a profession. It is a system that completes one expensive workflow and is accountable for a verifiable outcome: handling an insurance claim, executing procurement, closing a freight operation, preparing an engineering change or audit evidence, or performing a compliance check. Model-generated text is only an intermediate artifact; value appears when the case is closed in the system of record.

Bessemer’s work on vertical AI shows rapid growth among its own portfolio companies and a path to enter alongside an incumbent system instead of replacing it wholesale. That is a useful market signal, but not a complete picture of the market: the investor sees a selected portfolio and has an interest in the category. The more durable mechanism is that structured work can be independently accepted, while the product performing it can be paid for from an operations or services budget—not only a software budget.020202It helps to keep the opposite bet in view. I covered the Manus AI story on my channel: the same “brain plus hands” formula, but a general product rather than one expensive workflow. Both strategies are alive; they simply defend differently — a narrow vertical compounds access and data, a general agent has to compound speed.Knizhny kub · the story of Manus AI (in Russian)02It helps to keep the opposite bet in view. I covered the Manus AI story on my channel: the same “brain plus hands” formula, but a general product rather than one expensive workflow. Both strategies are alive; they simply defend differently — a narrow vertical compounds access and data, a general agent has to compound speed.Knizhny kub · the story of Manus AI (in Russian)

• —Entry point: one frequent case class with known input, permitted actions, completion criteria, and cost of error.
• —Path to a large company: case → system of action → primary system of record → transactions, financing, or an operator network.
• —Compounding advantage: exceptions, human corrections, system-of-record integrations, and authority over a growing share of the workflow.
• —Stop condition: the buyer praises the demo but withholds real cases, cannot define acceptance, or continues to pay only for operator hours.

The rare profile is particularly valuable here: the job is not merely to assemble an agent, but to design accountability, build a team of domain experts and engineers, make difficult exceptions visible, and control the cost of human review. The primary trap is quietly becoming a consultancy with an AI interface.

02 · Security and control planes for agents — 8.9

A conventional service receives a predefined role and makes a known set of calls. An agent develops a plan at runtime, reads untrusted context, invokes tools, delegates subtasks, and can travel far beyond a person’s original intent. It needs a distinct identity, short-lived authority, constraints on combinations of actions, goal validation, a delegation trail, and the ability to replay an episode after an incident.

In May 2026, NIST recorded broad market agreement that agents introduce new threats, security already impedes adoption, and conventional practices need adaptation. A separate NIST agent standards initiative highlights identity, authorization, interoperability, and security evaluation. OWASP has already translated the problem class into an applied risk map.030303The market says the same thing in its own words. I covered Palo Alto Networks’ outlook on my channel: an unguarded agent is called the new insider, identity the new perimeter, and accountability for an incident is drifting up to the executive. Demand for this layer is not driven by fashion.Knizhny kub · Palo Alto Networks’ 2026 outlook (in Russian)03The market says the same thing in its own words. I covered Palo Alto Networks’ outlook on my channel: an unguarded agent is called the new insider, identity the new perimeter, and accountability for an incident is drifting up to the executive. Demand for this layer is not driven by fashion.Knizhny kub · Palo Alto Networks’ 2026 outlook (in Russian)

• —Entry point: grant narrow authority for actions in GitHub, cloud, CRM, ERP, or financial systems, with mandatory approval for dangerous steps.
• —Path to a large company: machine-actor inventory → runtime policy enforcement → compliance evidence → risk management for the entire autonomous workforce.
• —Compounding advantage: a graph of who did what on whose behalf, a library of real incidents, and policies that reduce manual review without increasing missed threats.
• —Stop condition: the product sees only model prompts rather than real authority and consequences, and one feature from a cloud or model provider can replace it.

The strong version is not a universal “bad-prompt shield.” It becomes a neutral point for issuing and controlling authority across models, clouds, and customer systems. The opportunity is enormous, but IAM, PAM, cloud, and security vendors are moving fast. The company needs a new execution primitive, not one more observability dashboard.

03 · Industrial autonomy deployment — 8.5

The compelling target is not a universal humanoid but one repeatable workflow on existing equipment: machine tending, visual inspection, sorting, packaging, inventory, professional cleaning, or work in a hazardous zone. The software layer connects perception, planning, safe shutdown, remote intervention, and fleet data. Hardware can come from several manufacturers.

The IFR counted 542,000 industrial robot installations in 2024—more than twice the number ten years earlier. Professional service-robot sales grew 9%, while the robot-as-a-service fleet grew 31%; at the same time, the IFR identifies 944 suppliers, most of them small and medium-sized companies. The market is real but fragmented: deployment and operations can connect models to a heterogeneous installed base.040404I used the same 542,000 installations on my channel when explaining why robotics reached an interesting point: the industrial base already exists, vision-language-action models have arrived on top of it, and Open X-Embodiment is trying to make skills portable across platforms. The deployment layer profits precisely from that fragmentation.Knizhny kub · why robotics is worth doing now (in Russian)04I used the same 542,000 installations on my channel when explaining why robotics reached an interesting point: the industrial base already exists, vision-language-action models have arrived on top of it, and Open X-Embodiment is trying to make skills portable across platforms. The deployment layer profits precisely from that fragmentation.Knizhny kub · why robotics is worth doing now (in Russian)

• —Entry point: one operation in a controlled environment where operating hours, human interventions, downtime, and payback can be counted.
• —Path to a large company: software overlay → heterogeneous robot management → failure data and simulation → subscription operations for an entire class of work.
• —Compounding advantage: rare edge cases, a validated safety envelope, process fixtures, field service, and an update loop.
• —Stop condition: an impressive demo cannot survive a full shift, and the second deployment remains as manual a project as the first.

The IFR names IT/OT convergence among the top robotics trends for 2026, and the deployment layer lives exactly on that seam. The leader must build two worlds at once: an IT team that learns quickly from data, and an operating organization that values predictability, spare parts, and safe recovery. This is difficult work, but unusually well matched to someone who can operate across architecture and organization. How robotics arrived at this point—from cybernetics and Shakey to VLA models—is the subject of the companion longread on Physical AI, published on the same day as this text.

04 · Energy orchestration for the AI era — 8.5

AI has made electric capacity, cooling, transformers, and speed to power part of the compute stack. The IEA estimates that data-center electricity use grew 17% in 2025 and could nearly double—from 485 to 950 TWh—by 2030. AI-focused facilities grow even faster, while rack density tests the limits of power and cooling infrastructure.050505What the buyers do shows how serious this became. I covered on my channel how Meta set up a separate subsidiary and won regulatory approval to trade electricity wholesale; Google, Apple, and Amazon had taken the same route before it. Energy stopped being a budget line and became a strategy of its own.Knizhny kub · Meta enters wholesale power trading (in Russian)05What the buyers do shows how serious this became. I covered on my channel how Meta set up a separate subsidiary and won regulatory approval to trade electricity wholesale; Google, Apple, and Amazon had taken the same route before it. Energy stopped being a budget line and became a strategy of its own.Knizhny kub · Meta enters wholesale power trading (in Russian)

The software does not create new energy. Its value is turning compute load into a controllable energy resource. A scheduler can move eligible training and inference in time or across sites while accounting for SLAs, price, carbon intensity, and grid constraints. Other wedges include site and interconnection selection; joint dispatch of grid, batteries, generation, and cooling; and flexible-load participation in energy markets.

• —Entry point: a behind-the-meter site—a data center, industrial campus, or energy storage site—where fewer parties are involved and released megawatts can be measured.
• —Path to a large company: one-site control → asset portfolio → market dispatch → capacity and flexibility contracts.
• —Compounding advantage: real-load telemetry, constraint models, certified control, and a track record of meeting commitments.
• —Stop condition: the product displays an attractive consumption chart but does not change physical operation or share measurable savings with the customer.

Locality is the primary risk. Interconnection, trading, and licensing rules differ, and selling to a utility can take years. The technical core travels internationally; the contractual shell does not. The first buyer should therefore value speed to power or available capacity now—not merely promise a long joint program.

05 · Autonomous R&D and engineering — 8.4

A self-driving lab connects literature search, hypothesis, plan, instrument, measurement, analysis, and the next iteration. In July 2026, the NSF announced a $380 million investment in twenty teams building a network of AI-programmable laboratories, with a philanthropic partner adding upwards of $20 million. It signals emerging common infrastructure—not a proven mass market.

The scientific frontier remains genuinely difficult. An agent experiment in atomic force microscopy found that models which answer domain questions well may coordinate a real instrument poorly, deviate from instructions, and remain sensitive to phrasing. The authors of the Nature Communications study explicitly call for testing and safety protocols before autonomous use.060606I covered the optimistic half of this picture on my channel through John Jumper’s talk: AlphaFold 2 trained on 1% of the available data matched the previous version trained on all of it. His conclusion is practical — an idea can be worth many times more than extra data, and that is exactly the lever a small team can pull.Knizhny kub · John Jumper on AI for science (in Russian)06I covered the optimistic half of this picture on my channel through John Jumper’s talk: AlphaFold 2 trained on 1% of the available data matched the previous version trained on all of it. His conclusion is practical — an idea can be worth many times more than extra data, and that is exactly the lever a small team can pull.Knizhny kub · John Jumper on AI for science (in Russian)

• —Entry point: one instrument or a “plan → fabricate → measure → select the next experiment” loop in materials, chemistry, biology, or engineering simulation.
• —Path to a large company: adapters and data provenance → lab orchestration → an installation network → a proprietary stream of negative and positive results.
• —Compounding advantage: protocols, calibrations, failed experiments, safety constraints, and knowledge transfer across installations.
• —Stop condition: scientists like the system, but there is no budget owner, domain cofounder, or right to experimental results.

A sensible start is a software layer and a partner laboratory, not an expensive facility of your own. A Communications Materials review tracks the same trajectory: the frontier is shifting from isolated self-driving instruments to orchestrating entire laboratories. Full automation is optional too: a business exists if the system reduces wasted experiments, improves reproducibility, and leaves scientists the decisions where domain intuition truly matters.

06 · An operating system for healthcare work — 8.4

The most rational healthcare wedge is not an autonomous doctor. It is end-to-end work with a human at the clinical decision point: chart preparation, referral, payment authorization, coding, scheduling, discharge, and longitudinal-care coordination. The outcome can be verified without immediately handing clinical liability to a model.

The global workforce shortage remains severe: the WHO projects a nursing and midwifery shortage of about 4.8 million by 2030. In the United States, the CMS rule requires affected payers to implement APIs for authorization and data exchange largely from 2027; operating metrics began appearing in 2026. These APIs are creating digital infrastructure for a workflow that once relied on calls, faxes, and manual reconciliation.

• —Entry point: one workflow with a clear provider–payer pair, a document package that can be structured, and expert confirmation.
• —Path to a large company: administrative case completion → multi-party coordination → longitudinal operating record → outcome-based payment.
• —Compounding advantage: healthcare-system integrations, local rules, validated outcomes, professional trust, and proven reductions in time or denials.
• —Stop condition: data rights, regulatory class, and budget owner are unknown, while success is measured by answer quality rather than completed work.

The FDA list shows many already authorized AI-enabled devices, but authorization of one product does not prove readiness of the whole category or transfer to another country. Healthcare scores highly on pain and scale, and poorly on portability. Without a strong clinical and regulatory partner, it is better left alone: sixth place holds only on the far side of that gate.

07 · Protecting assets from climate risk — 8.2

This means protecting assets and infrastructure from the effects of a changing climate, rather than reducing emissions. The frontier becomes interesting where the product does not end at a risk map. The complete loop is observation → priority → work order → physical change → independent verification. The target might be water leakage, wildfire or flood protection, power-line resilience, urban cooling, forest maintenance, or supply-chain protection.

UNEP estimates that developing countries will need $310–365 billion per year in adaptation finance by 2035, versus $26 billion in international public flows in 2023. The enormous gap proves the importance of the problem, but also reveals market failure: the beneficiary, payer, and party avoiding the loss often differ.

• —Entry point: one asset owner and one loss mechanism already visible in insurance premium, downtime, repair, water, or capital expenditure.
• —Path to a large company: detection → verified risk reduction → work procurement and financing → an evidence standard for insurer and regulator.
• —Compounding advantage: a history of “signal → cause → action → verified effect,” an asset graph, and a local vulnerability model.
• —Stop condition: everyone agrees on the social value, but no one will pay from an existing budget for a specific avoided loss.

Swiss Re reports that secondary perils—including wildfires, severe convective storms, and floods—accounted for 92% of insured natural-catastrophe losses in 2025. This is a reinsurer’s estimate, not a universal climate model, but it points to a buyer with a direct financial interest. A strong company sells verified risk reduction, not one more analytics layer.

08 · Dual-use autonomous systems — 8.2

Here, high budgets meet a high cost of error. Compelling layers include navigation without satellite signals, resilient communications, sensor fusion, counter-drone protection, logistics, maintenance, energy resilience, simulation, and open control architectures for heterogeneous platforms. Many have a civilian first market in maritime operations, ports, rescue, wildfire response, and critical infrastructure.

The European Defence Fund allocated €1 billion to its 2026 program, with one quarter directed to critical technologies and advanced capabilities. NATO DIANA is seeking work in autonomy, sensing, resilience, and logistics. Funding and test channels exist, but a program award is not the same thing as repeatable commercial revenue.

• —Entry point: a capability useful to both a civilian operator and a security buyer, testable in a difficult but accessible environment.
• —Path to a large company: communications, perception, or planning module → multi-platform interoperability → trusted mission architecture → serial delivery.
• —Compounding advantage: field data, tests on real equipment, resistance to interference, customer trust, and the ability to pass procurement.
• —Stop condition: one closed contract, no second market, and export, ownership, or citizenship constraints discovered after development.

Portability is lowest here for someone not tied to one country: jurisdiction, investor composition, team citizenship, and permitted export destinations must be chosen early. The ethical boundary is also an entry condition, not a line in a future policy. With an active clearance and a strong procurement channel, this frontier rises into the top five; without them, eighth place is more honest than impressive budget slides.

09 · Autonomous space infrastructure operations — 7.7

Space is the ranking’s purest adventure, but a software product can be the starting point without owning a rocket or constellation: ground planning, fleet operations, conjunction-risk analysis, maneuver coordination, fuel optimization, onboard autonomy, mission simulation, and evidence for regulators or insurers.

ESA estimates the spacecraft manufacturing and launch market at roughly €75 billion in 2025, and the downstream market for satellite data and services at about €490 billion. No single startup can claim that whole sum as its addressable market, yet it shows that value has long extended beyond rockets. Meanwhile, the ESA Space Environment Report records about 40,000 tracked objects and, in the agency’s model, more than 1.2 million debris objects larger than one centimeter.070707The engineering side of this shift was described well in a podcast I covered on my channel: old space is a mainframe, one craft for ten to fifteen years, while new space is a cluster where a satellite is a node, reliability comes from constellation redundancy, and updates ship almost continuously. The software entry point exists because of that change of model.Knizhny kub · a podcast on the new space industry (in Russian)07The engineering side of this shift was described well in a podcast I covered on my channel: old space is a mainframe, one craft for ten to fifteen years, while new space is a cluster where a satellite is a node, reliability comes from constellation redundancy, and updates ship almost continuously. The software entry point exists because of that change of model.Knizhny kub · a podcast on the new space industry (in Russian)

• —Entry point: ground-operations or autonomy software on top of existing data and spacecraft; revenue before an owned launch.
• —Path to a large company: one-mission planning → fleet management → operator coordination → servicing, insurance, and orbital infrastructure.
• —Compounding advantage: precise spacecraft states, verified maneuver outcomes, failure models, and an intent-sharing operator network.
• —Stop condition: the product merely repeats free government alerts, the buyer pool is small, and the team is already financing its own spacecraft.

NASA demonstrated a lightweight, explainable agent for event-driven onboard operations on its MMS mission, with the in-flight demonstration completed in 2025. It is a useful model for the wedge: not a general intelligence for space, but bounded autonomy that conserves scarce communication and operator time. The hardware version of the same thesis carries an entirely different risk profile and horizon.

Several frontiers missed the ranking not because they cannot support a large company. Some are entry patterns that cut across industries; others require a special advantage absent from the starting profile. This is precisely why personal history can lift them above a universal list.080808The right of agents to spend money has a research line of its own. I covered a paper on agent economies on my channel: if such markets emerge spontaneously and stay permeable to the human economy, settlement, limits, and reconciliation stop being optional. That is the argument that moves programmable payments out of the second list.Knizhny kub · virtual agent economies (in Russian)08The right of agents to spend money has a research line of its own. I covered a paper on agent economies on my channel: if such markets emerge spontaneously and stay permeable to the human economy, settlement, limits, and reconciliation stop being optional. That is the argument that moves programmable payments out of the second list.Knizhny kub · virtual agent economies (in Russian)

The discriminating power of the criteria is easy to stress-test on the candidates themselves, run through the same six: profile leverage, willingness to pay, wedge speed, compounding, scale, portability. The numbers below are an authorial estimate made for this check, not a full scoring pass. Programmable payments without banking partners lose on wedge speed and portability: 8 · 7 · 4 · 9 · 9 · 5 yields 7.2. Compute optimization sinks on compounding, since clouds absorb the gains: 9 · 9 · 5 · 4 · 8 · 9 yields 7.5. Digital public services run into procurement and locality: 8 · 6 · 3 · 7 · 7 · 4 yields 6.6. All three land below ninth place. The one exception is verifiable modernization: a strong profile and a live budget give it 10 · 9 · 7 · 4 · 6 · 9, or 7.8—level with space, inside the noise band. What keeps it out of the ranking is structure, not score: it is not a separate industry but an entry pattern for frontier number one.

Verifiable modernization deserves a separate note

An agentic factory that recovers a legacy system’s behavior, creates a test oracle, migrates one slice, and produces verifiable proof of equivalence is an ideal use of a Fellow profile. Yet it is not a separate industry; it is a repeatable vertical system of action. The team must choose a specific stack and migration class—for example, legacy ERP forms, a batch calculation, or a data pipeline. Without that focus, the company quickly becomes ordinary custom development.

Name the attractive traps in advance

• —Training another frontier model without a unique scientific result, compute resource, or distribution channel.
• —A universal “agents for everyone” platform whose core features are rapidly absorbed by models, clouds, and open source.
• —A full humanoid built from scratch before proving the economics of one job and a service network.
• —A climate dashboard without an operator, work order, and party receiving measurable savings.
• —An owned drug, rocket, or chip fab pursued only because the market is large: the starting profile alone does not confer the required scientific or physical advantage.

The author’s home turf is not in this ranking

An attentive reader will notice one more missing frontier: AI tools for software development. It is the area where my personal access is greatest—and my current job—which is exactly why it stays outside the frame: scoring it with the same rubric would mean a conflict with my present role, and this bet was deliberately sought as diversification, away from the day-job specialty. That is a boundary of the ranking, not a hidden verdict: the frontier was never scored, and its absence from the table says nothing about its strength.

Fellow level does not oblige anyone to become a solo founder. In a hard domain, technical honesty can demand the opposite: find a partner who already has the right to speak with the clinician, factory, utility, laboratory, or government buyer. The form of participation should match the largest unknown.

A simple test helps. If the primary unknown is “can this be built and operated,” the technical-cofounder role is genuinely central. If it is “who will buy and let us into the workflow,” a domain partner or the founder’s direct market work comes first. If it is “which of three adjacent industries should we choose,” an entrepreneur-in-residence structure is useful only until evidence appears and full focus becomes necessary.090909Brian Halligan, formerly of HubSpot, puts the price of the scaling half well in a conversation I covered on my channel: starting a company is easier than ever and scaling it is harder, and mature leaders spend up to half their time on hiring and team design. The form of participation deserves to be chosen with that work in mind too.Knizhny kub · Brian Halligan on the CEO job (in Russian)09Brian Halligan, formerly of HubSpot, puts the price of the scaling half well in a conversation I covered on my channel: starting a company is easier than ever and scaling it is harder, and mature leaders spend up to half their time on hiring and team design. The form of participation deserves to be chosen with that work in mind too.Knizhny kub · Brian Halligan on the CEO job (in Russian)

Agentic development can produce a persuasive interface quickly—and thereby increase the risk of self-deception. The first three months should buy four rights rather than product breadth: the right to see real work, verify the outcome, receive payment, and retain data for the next cycle. What follows is an authorial exploration protocol, not a required fundraising sequence. In slow-signal frontiers—energy, dual-use, and space—ninety days can realistically clear only the first two gates; the honest equivalent of payment there is a signed letter of intent and test access with a predefined conversion to money.

Days 0–15: choose three hypotheses and gain access to a real workflow

For each hypothesis, describe one event rather than a market: who starts the work, which systems it touches, what completion means, where exceptions occur, who owns an error, and which budget currently pays for the workaround. Hold enough conversations to see repeated exceptions, and observe the full workflow at least once. Fifteen substantive interviews are a practical guide, but one real observation is worth more than ten general conversations about the future of AI.

Days 16–30: price the unit of outcome and the cost of error

Establish a baseline for time, cost, completion rate, human review, and difficult tails. Classify errors as reversible, detectable, or irreversible. Decide which action an agent may take in shadow mode, which needs approval, and which remains forbidden. The budget owner and lawful path to data should already be visible.

Days 31–60: close one case with independent acceptance

Build a narrow prototype, using agents and behind-the-scenes manual work where automation has not yet earned trust. Separate execution from verification: hidden tests, an expert, or a physical-outcome measurement should not belong to the same loop. Count the full episode—retries, human intervention, delay, and the cost of an incorrect action—not one model call.

Days 61–90: secure paid commitment or stop

The objective is a paid pilot, a test contract, or at least a letter of intent naming the owner, scope, success criterion, and next budget decision. At the same time, negotiate the right to use de-identified traces and outcomes to improve the product. A free pilot is acceptable only when it provides rare access that cannot be bought more cheaply and has a predefined conversion to payment.

a large round cannot repair missing access, acceptance, buyer, or compounding advantage—it only makes the mistake more expensive

A universal ranking narrows the field, but a company begins with unfair advantage. An active clearance raises dual-use systems. A clinical network and strong medical cofounder raise healthcare. Factory data and a manufacturer partnership raise industrial autonomy. A funded mission raises space. Without access like this, one more degree of technical strength usually matters less than one person able to bring the team inside a real workflow.101010An honest counterexample is worth holding too. I covered an interview with Higgsfield AI’s cofounder on my channel: their unfair advantage was not industry access but speed — the team took someone else’s model through an API and closed the missing control faster than anyone. That path exists, but it lives in a market where the model refresh cycle is measured in months.Knizhny kub · an interview with Higgsfield AI’s cofounder (in Russian)10An honest counterexample is worth holding too. I covered an interview with Higgsfield AI’s cofounder on my channel: their unfair advantage was not industry access but speed — the team took someone else’s model through an API and closed the missing control faster than anyone. That path exists, but it lives in a market where the model refresh cycle is measured in months.Knizhny kub · an interview with Higgsfield AI’s cofounder (in Russian)

Before the big bet, I would require a positive answer to seven questions:

• —Is there a current budget at least an order of magnitude larger than the expected price of the first product?
• —Does the system perform a verifiable action rather than only produce advice or text?
• —Can objective feedback arrive before the next funding round?
• —Can an error be bounded, detected, and—where possible—reversed?
• —Does every completed case produce data or authority that cannot be bought from a model provider?
• —Will the second deployment be mostly the same product rather than a new project?
• —Is there a person who will open the data and environment, then actually authorize payment?

The strongest general strategy sounds less exciting than a futuristic pitch: a narrow software product first, hard domain advantage second. An agent moves the idea into contact with reality quickly. The Fellow designs the boundaries, verification, and architecture. The leader builds sales, operations, and an organization that can own the outcome. Only together do those three roles turn an interesting adventure into a venture company.