Power Generation as a Side Business: Taiwan Routes, Costs, and Software and Edge AI | Private AI Cluster Log, Day 6

Before reading: I asked AI to research this using public sources and to map possible routes. This article contains no firsthand record of building a power system, running an energy company, or completing customer interviews, and it does not verify that any route is profitable. The cost and revenue examples are teaching assumptions. They show what to ask and calculate next, not actual quotations, regulatory approvals, or investment promises.

If one person in Taiwan wants to make electricity generation a side business, should they sell electricity, sell equipment that generates it, or manage energy for someone else? These are different businesses. This survey examines each route for an individual with software or AI skills, roughly 10–15 hours a week, and no existing power engineering team. It asks what site rights, capital, qualifications, customers, and maintenance obligations come first, when revenue might arrive, and what would allow the business to grow into a company.

Day 6 adds a question: can software AI and AI running on the equipment improve the full development-to-operations workflow? They can help organize requirements, analyze measurements, find defects, predict faults, and reduce inspections. They cannot replace an energy source, qualified protection, field tests, product inspection, or accountable engineers. The article counts their added electricity use, costs, and errors alongside any benefits.

Here, “building a complete power product” means integrating generation, storage, controls, and protection equipment that can lawfully be used for the intended application. It does not require a person to manufacture solar cells, inverters, or a novel reactor alone. Equipment sales are product revenue; self-consumption saves an electricity bill; electricity sales earn revenue from delivered power. Feed-in sales mean selling electricity under an applicable purchase scheme and contract.

The conclusion depends on the starting position. With no usable roof or other energy site and limited capital, the more workable entry is paid measurement and maintenance work for a small commercial site, followed by integration of existing solar and storage equipment. Someone with a lawful roof and funds can also study self-consumption or electricity sales. A proprietary complete product becomes worth evaluating after repeat orders, an electrical engineering partner, and a clear compliance route. This is a ranking of requirements, not proof that customer demand has already been validated in Taiwan.

The survey also examines the more distant prospect of fusion R&D. Fusion is not a later version of the small solar product. If the ultimate goal is fusion electricity supplied at one-tenth to one-hundredth of the full delivered cost of a competitor providing equivalent service, this remains an unproven long-term research target. The later section works backward through physics, engineering, capital, and regulatory evidence. Revenue from research services cannot be counted as fusion electricity profit.

Sources were checked on October 3, 2026. The article distinguishes official historical data, forecasts, illustrative calculations, and market questions that only customer interviews can answer. This is a desk survey with scenario calculations, not a record of an operating startup or a construction drawing.

What one person can do first

If you have software skills but no team, start by finding a problem someone pays to solve and offering a data service that does not control electrical equipment. One person can begin these tasks and produce something concrete from each:

1. Find the problem: interview small commercial sites or equipment owners about the real costs of outages, electricity bills, and inspections, and identify who can approve a purchase. Produce an anonymized problem list and a list of people willing to continue the conversation. These interviews are proposed work, not interviews already completed for this article.
2. Establish a baseline: with the owner's permission, use bills and data exports allowed by the equipment manufacturer to assemble hourly load, missing-data, outage, and service-call records. Start with a spreadsheet and fixed rules; no AI chip or wiring is needed.
3. Build a software prototype: use authorized or public data for a one-page dashboard with anomaly summaries, cost estimates, or organized maintenance tickets. Let users trace conclusions back to source data. AI offers suggestions first; it does not control equipment.
4. Test willingness to pay: offer one prospective customer a measurement or reporting pilot with a defined deliverable. Track setup time, cloud fees, false alarms, and whether the customer will pay. A service fee is software or analysis revenue, not proof that you generate and sell electricity.
5. Prepare evidence for a complete product: if a customer really needs off-grid power, collect the load, environment, candidate manufacturer's equipment, quotations, and possible product classification. Have qualified engineering and inspection partners confirm the design, protection, installation, and market-entry requirements. One person can organize requirements and software; AI cannot replace that professional verification.

The stop condition is equally practical: without lawfully usable data, a paying problem, or a gap left by standard products, do not invest in a proprietary power system yet. If you already control a lawful roof, ask qualified providers to assess self-consumption and electricity sales separately; that is a different path requiring a site and capital.

Survey findings: five entry points for one person

This table compares entry requirements and types of revenue; it does not claim that one route is most profitable at every site. If “generation as a side business” strictly means producing electricity oneself, the first three rows apply. The final two routes enter the energy industry and build customer and delivery capability, but service fees are not electricity-sales income.

Three common starting points divide the route. A lawful roof but no electrical team: check site rights and grid capacity, then ask qualified installers to assess self-consumption or electricity sales. No roof but software skills: find customers willing to share load and maintenance data, then test a paid service. An electrical engineering partner and a defined customer: assess a complete power product, beginning with classification, installation, and after-sales responsibilities. In every case, actual site, supplier, and customer evidence must replace the examples in this article.

Before the first payment, a service using already accessible data usually has the smallest capital burden, but first needs a paying customer. Own-roof generation requires installation funds, approvals, and grid-connection time before a slow payback. A complete product needs development, inspection, inventory, and warranty funds before shipment, potentially offset by customer deposits. Wind, hydro, and fuel-based routes depend first on resources, sites, and permits. These are relative thresholds, not reliable payback durations without a real site and contract. In the first 90 days, a part-time entrant can interview customers, obtain lawful load data, and deliver one paid measurement service; the detailed plan and stop conditions appear later.

How this survey compares the routes

Every route faces the same questions: Where does the energy or site come from? Who has installation rights? Who pays? How much cash is needed before the first payment? Who handles safety and maintenance? Can the work be repeated for a second customer? Equipment power ratings are separated from energy actually delivered; product inspection is separated from site permission; and per-unit contribution is separated from business profit and the timing of cash receipts.

Evidence is kept in four categories. Official historical data and current rules describe the background and scope of regulation. Official forecasts show possible directions only. Illustrative calculations demonstrate how to make a decision, without pretending to be market quotations. Unvalidated market hypotheses require customer interviews, site records, supplier quotations, and formal guidance for the specific project. The comparison above and the later 90-day plan give a quick route through the article; the technical and legal sections provide detail.

A possible progression from side business to company

The following sections explain electricity and AI trends, then the Taiwan rules and economics for power products and electricity sales. Fusion comes last for readers studying long-term R&D after an energy business has been established.

Before reading the figures, remember: power is the rate of electricity supply; energy is the amount supplied over time. Delivering 1 kW (kilowatt) for one hour supplies 1 kWh of energy. MW and GW are larger power units, equal to one thousand and one million kW respectively; 1 TWh is one billion kWh. The article uses these same units for products and power plants so that a machine’s power rating is not confused with its annual output.

What is changing in the electricity industry?

An electricity system has five parts: generate it, deliver it, store it, use it, and coordinate everyone’s timing. Better generation technology addresses only the first part.

The IEA’s 2026 review estimates that renewables provided around 34% of global electricity generation in 2025, or around 43% when nuclear is included; wind and solar together provided around 17%. Supply is changing, but global electricity is still far from entirely free of fossil fuels. IEA global electricity supply review

Storage is expanding too: the IEA estimates that about 108 GW of battery storage power capacity was added in 2025, up around 40% year on year. This is GW of power, not GWh of energy. How much power a system can deliver and how long it can keep delivering it are different questions. A battery that lasts a few hours can help shift peaks; that does not mean it solves seasonal energy supply. IEA battery storage review

The grid is another constraint. The IEA’s Electricity 2026 reports that more than 2,500 GW of renewable generation, large loads, and storage projects worldwide are waiting for grid connections. Not all will necessarily be built, but the queue illustrates the need for transmission and distribution construction to keep pace with new generation and demand. IEA grid analysis

Solar cells convert light directly into electricity; wind, hydro, and fuel-based generation usually involve mechanical or thermal processes. All require an external energy source. Batteries merely store energy, and hydrogen is usually an energy carrier whose production first consumes energy. A “magnetic generator” that continuously produces net electricity without fuel, light, wind, flowing water, or another energy input cannot provide the physical foundation for a business plan.

A complete solar PV system includes more than panels: it also needs mounting, power conversion, distribution, and protection. Storage becomes an additional consideration when power is needed at night. The US Department of Energy’s introductions to PV system design and solar plus storage are useful starting points.

How does AI change electricity demand?

Global growth is rapid, but check the year and scope

In its April 2026 report Key Questions on Energy and AI, the IEA estimates that global data center electricity consumption increased by around 17% from 2024 to 2025, reaching about 485 TWh in 2025. Its central projection is approximately 950 TWh in 2030, close to 3% of global electricity demand. This covers all data centers, so their entire consumption should not be attributed to generative AI. The 2030 figure is also a forecast, not a confirmed level of demand. IEA: Key Questions on Energy and AI

Several changes deserve closer attention.

First, loads are concentrated. A few percentage points of additional global electricity demand does not mean every city experiences the same increase. Data centers cluster in particular places, where local substations, lines, and available supply capacity may become constraints first. A country’s annual generation total cannot by itself answer whether a particular plot can connect a large new data center.

Second, computing chips are not the only electricity consumers. GPUs are chips commonly used for AI computation. Alongside servers are networking, storage, power conversion, and cooling. PUE means “total facility electricity use divided by information technology equipment electricity use.” A PUE of 1.2 means that for every kWh used by that equipment, the facility uses 1.2 kWh in total. It helps track overhead consumption, but does not tell you whether each AI task is worth computing.

Third, training and inference offer different kinds of flexibility. Interruptible training, batch inference, or data processing can potentially move to more suitable times. Real-time voice, customer support, and search are constrained by latency and service commitments. An entire data center cannot be treated as a load that can be switched off at any moment.

Fourth, efficiency gains and higher total demand can happen together. Representing models with more compact numbers, choosing smaller models, reusing previous results, and using better chips may reduce energy per task, while lower prices may encourage more usage. Lower energy use per answer does not by itself prove that total electricity demand will fall, and it is equally wrong to assume that additional usage will inevitably consume all the savings.

The IEA’s analysis of data center energy demand provides background on these effects and uncertainties. This survey infers that electricity software opportunities will arise where outcomes can be measured, operations can be scheduled, and benefits can be demonstrated, rather than only from generating more electricity.

AI can also improve energy systems

Examples include forecasting tomorrow’s sunlight and load, detecting faults earlier, adjusting charge and discharge schedules, or helping technicians find maintenance documentation. But a model needs reliable sensor data and controls that can actually carry out its recommendations.

“Generation was lower this month” could mean worse weather, shaded panels, faulty equipment, or missing data. An AI alert that cannot distinguish among these causes merely adds to the customer support workload. The IEA’s analysis of AI applications in the energy sector provides broader industry context.

Taiwan needs its own assessment

The Energy Administration’s 2025 National Electricity Supply and Demand Report, published in 2026, records Taiwan’s 2025 electricity consumption at 283.01 billion kWh, down 0.30% year on year. National gross generation was around 289.71 billion kWh, including self-use generation equipment, so its coverage differs from statistics for the Taipower system alone. Gas, coal, renewables, and nuclear accounted for approximately 47.7%, 35.3%, 13.3%, and 1.1% of generation, respectively, with the remainder from other sources. Electricity consumption and gross generation are not interchangeable figures either. Energy Administration supply and demand report, pages 2 and 5

The same report forecasts average annual electricity demand growth of 2.5% over 2026–2035, and growth of 2.7% in nighttime peak load. These figures support the importance of medium- and long-term electricity infrastructure, but do not support a claim that AI has already caused total electricity demand to surge across Taiwan. Industrial structure, the economy, efficiency measures, weather, and AI all affect the outcome.

Dates also matter when describing nuclear power. On September 24, 2026, the Nuclear Safety Commission announced that it had completed its review of Maanshan Nuclear Power Plant’s “restart plan.” Implementation, submission of results, and further required reviews remained. That announcement does not mean the plant had resumed generating electricity. Nuclear Safety Commission review update

For an individual side business, four questions are more practical: Is the site legally usable? Can it connect to the local grid? When does the customer need electricity? Who can repair the equipment when it fails? An AI company signing a long-term green electricity purchase agreement and an engineer developing a 5 kW rooftop project operate at very different transaction scales.

Nor should electricity prices be assumed to keep rising. On September 18, 2026, the Ministry of Economic Affairs announced that it would defer a decision on the current electricity price adjustment until further discussion in December. The investment examples below therefore do not depend on prices “definitely going up.” MOEA electricity tariff review announcement

What are the options for making your own power equipment?

Rated power is particularly misleading for small wind systems. A 1 kW rating does not imply an average output of 1 kW throughout the year. Turbulence around buildings can also affect generation and service life. The US Department of Energy specifically highlights the turbulence, vibration, and cost issues with rooftop wind turbines. DOE small wind FAQ

For hydro, examine flow and effective head before generator size. Flowing water does not automatically mean sufficient energy or a right to use it. DOE hydropower basics

For a software engineer starting from zero, integrating solar power systems is a plausible first technical route given component maturity and access. The parts you truly own can be requirements, monitoring, mechanical layout, the user experience, and repair processes. You do not need to start by making solar cells, inverters, or lithium cells yourself.

When each of the three revenue models fits

EMS stands for energy management system: it reads generation, consumption, and battery status, then provides analysis or scheduling. It does not necessarily need AI. At sites with few devices, dependable rule-based control may be easier to maintain.

For someone without a site and with limited capital, the survey ranks service first, integration second, and an own-brand complete product last. If the customer simply needs a portable power station that already exists, acknowledge that a standard product solves the problem. A side business does not have to sell its own hardware on every project.

Someone with lawful access to a roof can assess electricity sales in parallel. Without customers, demand should be tested before holding panel and battery inventory.

The service route also needs workable unit economics

“Start with a service” can sound nearly cost-free. In practice, alerts, false positives, site visits, and customer support consume time. This purely illustrative monthly model assumes ten small remote sites have signed and each pays NT$1,200 a month, with lawful access to data from existing equipment and no new meter or gateway installation. It is not a completed sale. NT$800 per hour values the founder’s time as an economic cost; it is not necessarily cash wages paid that month. Actual contracts must replace every assumed price and hour.

That NT$1,300 has not paid for customer acquisition, insurance, taxes, software development, or other administration; it is not net profit. One additional visit with the same labor and travel costs adds NT$6,000 and turns the contribution into a NT$4,700 loss. The contract therefore needs clear terms for remote support, included visits, service hours, and additional callouts. Ten sites could also belong to one customer, so they do not automatically diversify customer risk.

Onboarding has a cost too. Suppose connecting, cleaning, and testing data across the ten sites takes 20 hours at the same NT$800 per hour, plus NT$4,000 in travel and incidentals: NT$20,000 in one-time onboarding costs. If customers pay no setup fee and monthly contribution stays at NT$1,300, recovering just that cost takes about 15.4 months, before development and customer acquisition. Alternatively, an upfront NT$20,000 setup fee needs a defined deliverable in the contract. Any case requiring new meters or gateways needs separate hardware, installation, and integration quotations; this existing-data example cannot be reused unchanged.

A person working 10–15 hours a week cannot independently promise around-the-clock site response. If a customer needs 24-hour incident coverage, arrange qualified maintenance partners and include their fees in the price. The customer’s arithmetic matters too: NT$1,200 per site per month must be justified by verified savings on inspections, downtime, or data problems. Both sides need a workable case for the service to endure.

Who might buy the first power product?

A customer group worth testing first is commercial sites where connecting power is difficult, loads are small, and owners need to know remotely whether equipment is still working. Examples include environmental sensors, agricultural monitoring, and outdoor data gateways installed with the owner’s consent. The first version should focus on applications that are not safety-critical, avoiding an initial promise of uninterrupted power for life support, firefighting, or disaster response.

The later product example assumes a NT$110,000 off-grid system serving an average 40 W load. That load uses about 350 kWh a year; even at an assumed NT$5 per kWh, the electricity alone is worth only about NT$1,750 annually. A buyer would have to save on power-connection work, routine visits, outages, or another verified expense. Actual bills and maintenance records, rather than a claim about cheap electricity, must establish that value.

This is a product hypothesis to validate, not a claim that Taiwan already has a sufficiently large market. In interviews with ten potential customers, check:

1. How is the equipment powered today? What do a power connection, battery changes, and site visits each cost?
2. What are the actual average load, startup load, and daily operating hours?
3. How long can the site tolerate losing power? What happens during consecutive rainy days?
4. Who holds the site usage rights, and who controls procurement and maintenance budgets?
5. Would an existing product plus manual inspections already be cheap enough?

A first specification saying “how long the specified equipment runs under defined conditions, when an alert is sent, and how quickly faults are handled” is more useful than “AI smart generator.”

Get watts and kilowatt-hours right first

kW measures power; kWh measures energy. One kWh is one unit of electricity on a Taiwan electricity bill. A continuous 1 kW load uses 24 kWh over 24 hours. kWp usually means a PV system’s rated peak output under standard test conditions; multiplying it by every hour of the year does not give actual generation.

“Equivalent peak sun hours” converts a day of changing sunlight into the number of hours the panel would need to run at its rated peak output to collect the same solar energy. The 3.5 hours below are an example input, not a forecast for a Taiwan site.

The following assumptions are for illustration. They are neither a sunlight forecast for locations across Taiwan nor a design ready for production.

Illustrative daily usable solar energy = 0.6 × 3.5 × 0.8 = 1.68 kWh
Daily load energy                     = 0.04 × 24       = 0.96 kWh
Energy the battery can deliver        = 2 × 0.8 × 0.9   = 1.44 kWh
Runtime from full charge, no sunlight = 1.44 ÷ 0.04     = 36 hours
Capacity for 72 hours without sunlight = 0.04 × 72 ÷ (0.8 × 0.9) = 4 kWh

Solar yield and battery runtime are two preliminary checks here. They cannot simply be added to produce a daily energy supply figure. The actual energy balance depends on when generation, charging, and discharging occur, and on losses at each stage. The 4 kWh figure also excludes aging, temperature effects, and design margins.

This example is initially defined as a DC power system for small loads, without a grid connection. The 40 W figure is average consumption, not maximum output capability. The finished product still needs specified output voltage, continuous rated power, short-duration peak power, supported loads, and environmental conditions. Adding AC output requires separate confirmation of the inverter, standby consumption, protection, and applicable inspections.

Enough energy on an average day does not guarantee continuous operation through several overcast or rainy days. When sunshine returns, the system must power the load and replenish the battery at the same time. Simulate using the site’s hourly sunlight, shading, worst season, and load records, then test it.

PVWatts can estimate solar output, but it is a generation estimation tool for grid-connected PV, not a guarantee of off-grid supply reliability. Batteries, control logic, and hours of unmet demand need separate modeling. PVWatts guidance and model inputs

Under the same conditions, an AI workstation continuously drawing 1 kW needs 24 kWh per day. It would require at least 24 ÷ (3.5 × 0.8) ≈ 8.6 kWp of solar capacity merely to balance this average assumption. Overnight operation and consecutive days of bad weather require a separate battery calculation. A “2 kWh power supply” in a small box cannot make a heavily loaded data center energy-independent over the long term.

Technical architecture from prototype to product

Start the prototype by reading data through communications interfaces permitted by the supplier. At a minimum, record time, generation power, load power, battery charging and discharging, estimated charge level, temperature, alarms, and connection status. Estimated charge level is commonly called SOC; it is an estimate, so understand its error and calibration method.

If equipment supports an industrial communications protocol such as Modbus, you can follow its documentation to read data fields, such as power or temperature. These fields are usually called “registers.” Which fields can be read, and which allow settings to be written, still depend on the model and firmware documentation. This manufacturer’s Modbus TCP documentation is an interface example, not evidence that the brand or model suits every application in Taiwan.

Use easy-to-audit rules for the first control version: alert on low charge, then reduce nonessential loads; if data becomes invalid, stop optimization and fall back to a conservative local mode. AI may suggest schedules or maintenance actions, but must not override voltage, current, temperature, or protection limits. Remote updates need source authentication and recovery after failure, and devices must not all share a single set of credentials.

Do not simply copy a battery lifetime claim of “several thousand cycles.” Lifetime depends on depth of discharge, temperature, current, and the remaining capacity that defines end of life. Manufacturer battery specifications show why these conditions must be read together. Choosing lithium iron phosphate does not eliminate thermal, short-circuit, or fire risks.

Where software and on-device AI fit from development to operations

First separate where the models run. Software AI on a computer or in the cloud reads documents, historical data, and work orders to help people compare designs, write code, forecast, and schedule work. On-device AI puts a validated small model on a site gateway, microcontroller, or chip with an AI accelerator so it can interpret sensor data locally, even while offline. An accelerator is often called an NPU, a processor optimized for model computation. Neither produces electricity: both depend on measurements and power from elsewhere. A US Department of Energy microgrid research plan identifies edge inference chips and distributed sensors as enablers. The IEA finds that forecasting, inspection, and maintenance are entering practice sooner than safety-critical real-time control. DOE microgrid AI research plan, IEA grid AI applications

This table is a list of testable uses, not evidence that one person can now complete all five stages with AI. A language model can draft documents and code, but manufacturer manuals, customer tickets, and webpages are data, not a source of new control instructions. It cannot sign off an electrical design, certify a product, or rewrite safety limits. As the IEA notes, moving AI closer to real-time control raises requirements for validation, explainability, cybersecurity, fallback, and accountability. IEA grid AI applications

Applied to the five business routes above, personal rooftop self-use or feed-in sales first need site, installation, and generation evidence; AI may not beat simple monitoring. Energy services can test software AI for data review and dispatch without buying much hardware. Complete power products need field evidence that offline detection or low latency matters before adding on-device AI. Wind, hydro, and fuel-based projects still depend first on resources and sites. For a part-time entrant, the order is measurement and rules, then a software-AI trial, and only then a decision about adding a chip.

AI on the equipment also uses electricity

Take the earlier 40 W average load and 2 kWh battery as a purely illustrative case. Adding an on-device AI module that continuously draws 2 W makes the average load 42 W. AI adds 2 × 24 = 48 Wh a day, 5% of the original 0.96 kWh daily load. With the same battery delivering 1.44 kWh, full-charge runtime without sunlight falls from 36 hours to about 34.3 hours. For 72 hours without sunlight, nominal capacity rises from 4 kWh to 4.2 kWh, before aging or design margins. The 2 W assumption is not a specification for any particular chip. Measure candidate hardware, including connectivity, standby, cooling, and update energy. Cloud AI also incurs subscription, transfer, human error-checking, and data-governance costs.

Whether a chip pays for itself depends on avoided dispatches and downtime covering the hardware, energy, maintenance, and false-alarm costs. In the earlier service example, an extra visit carries NT$6,000 in valued labor and travel. Avoiding one such visit would be a possible measurable benefit; causing one false visit could reverse it. Model accuracy cannot be booked directly as net profit. NIST research on industrial condition monitoring likewise calls for evaluating the model within actual risk and maintenance processes. NIST industrial AI condition-monitoring study

Move from advice to automation through evidence gates

At a minimum, compare false alarms per site-month, missed real faults, dispatches, downtime, electricity actually delivered, model and communications energy, and total monthly costs. The earlier four to eight weeks of data can expose workflow problems, but generally cannot establish rare-fault, typhoon-season, or annual performance. Continue validation across sites, seasons, and equipment versions. If connectivity, data quality, or model confidence fails, stop AI suggestions and return to engineer-approved local rules and manufacturer protection. NIST’s AI risk framework likewise emphasizes monitoring during operation, human intervention, and graceful degradation. NIST AI risk framework

How software and AI should prove their value

First save four to eight weeks of baseline data, then compare performance under similar seasons, sunlight, and workloads. A short pilot cannot establish annual performance. Report downtime, site dispatches, false-alarm rates, labor hours, and actual costs. If a model is more complex than simple rules but delivers no verifiable improvement, hold off on deploying it.

For a small system, avoiding one inspection trip to a remote site may be worth more than generating a few extra kWh. This is a commercial hypothesis to validate through interviews and paid pilots, not a basis for guaranteeing a customer’s payback.

What must a complete product clear before being sold in Taiwan?

Define the product first: is it a portable power supply, stationary storage, an inverter, a complete solar power system, or an on-site engineering project? Products with the same battery capacity may face different requirements because their applications, installation methods, inputs, outputs, and functions differ.

Choose the first inquiry based on what will be delivered. Read-only measurement or analysis: confirm data access, system-access rights, and the service contract with the customer; consult qualified engineers if site equipment will be altered. An off-grid complete product: bring model and functional details to BSMI or a recognized body to confirm classification; check NCC where wireless functions are included, and separately check installation conditions. Grid-connected rooftop PV with electricity sales: first establish roof rights, possible Taipower grid capacity, and the applicable energy-authority route for applications, metering, and contracts. Permission for one route does not replace the requirements of another.

Product inspection and site permits are different matters

BSMI is Taiwan’s Bureau of Standards, Metrology and Inspection. Its June 2026 announcement lists new product categories subject to inspection from July 1, 2026: power conversion systems of 20 kW or less, stationary lithium storage devices of 20 kWh or less, and lithium battery packs for energy storage from 1 kWh to 20 kWh. kW measures output power; kWh measures stored energy. These are potentially relevant categories, not proof that the proposed 2 kWh complete product belongs to all three. Classification depends on function, model, and form of sale. “I will not connect it to the grid” does not by itself establish exemption. BSMI June 2026 inspection announcement

For a proposed 600 Wp/2 kWh product, the right step is not to declare it exempt yourself. Bring product specifications, a functional block diagram, installation method, model number, and intended form of sale to BSMI or a recognized certification body to confirm the product category and inspection route. For wireless communications, separately confirm National Communications Commission (NCC) rules and module integration conditions. A certified radio module may still entail registration and labeling requirements for the final product. Article 17 of the NCC equipment approval regulations

Buying compliant batteries, chargers, or wireless modules can reduce development costs. But arbitrary reassembly, a new enclosure, or rebranding does not automatically let the complete product inherit the original certificates. Before shipping, obtain written confirmation of the model, certificate holder, labeling, and permitted changes. CE, FCC, or overseas test reports also do not by themselves establish access to Taiwan’s market. BSMI explanation of household storage conformity

This table is a product development checklist to close out item by item, not a legal conclusion that every product needs the same permits. Storage fire requirements depend on building use, capacity, installation location, and effective dates. Amendments not yet in force and proposed drafts should not be treated as today’s applicable thresholds. Start with the National Fire Agency’s storage fire safety management guidance announcement and the amendment and effective-date information for consumer electrical equipment inspection rules.

Passing inspections and having a manufacturer’s warranty do not automatically remove your product liability to customers. Where a consumer relationship exists, designers, manufacturers, importers, and distributors may bear responsibility according to their roles. Discovering that a product may endanger consumers also triggers obligations such as recalls. Prepare warnings, operating instructions, and incident procedures according to applicable laws and product risks. For commercial procurement, also clarify contractual and other legal responsibilities. Consumer Protection Act, Articles 7–10

Before shipping, establish serial-number traceability, assess product liability insurance, and check whether the customer’s contract requires coverage. This is a risk-management judgment, not a claim that every product has the same statutory insurance obligation.

Which parts can you start making yourself?

You can handle requirements and energy simulations, telemetry, the frontend, alerts, maintenance tools, mechanical requirements, and testing workflows. Have partners with the appropriate capabilities and qualifications design, install, and inspect power circuits, battery packs, protection coordination, building wiring, and grid connections.

Even in low-voltage systems, high-current batteries can cause serious short circuits and fires. Do not backfeed a prototype through a household outlet, modify the manufacturer’s battery protection, or treat “it lights up” as validation for sale. Fuel generators also pose a carbon monoxide hazard and must not operate indoors or anywhere else with unsafe exhaust conditions.

If the goal is to sell electricity

Self-consumption, selling all output, and selling surplus

Self-consumption offsets your own electricity costs; it is not revenue from selling electricity. The value depends on the time period and usage tier being displaced. Selling all output under a feed-in arrangement and selling surplus after self-consumption must follow approved wiring and metering arrangements and the electricity purchase agreement. You cannot connect equipment to the grid yourself and expect Taipower to pay based on a meter running backward.

A completely off-grid system does not sell electricity to Taipower. A grid-connected system configured for zero export still requires confirmation of grid connection and consumer equipment requirements. Conversely, ordinary grid-connected solar may not continue powering a home during an outage. Backup requires suitable equipment, isolation, and switching design; panels on the roof are not enough.

For a typical small fixed PV installation, first confirm that the site is legally usable and a grid connection is feasible, then address approvals, contracts, construction, and registration. Under the current installation management rules, Type 3 means renewable generation equipment for self-use with capacity below 2,000 kW. This is an equipment classification, not a blanket exemption from applications below 2,000 kW, and not a universal tariff or inspection threshold. Renewable generation equipment installation management rules, Articles 3, 4, and 7

This approval framework covers fixed equipment with total installed capacity of at least 1 kW. The earlier 600 Wp off-grid product is an example of a power supply product. The electricity sales calculation below instead uses a 5 kWp rooftop project. The same application and feed-in purchase conditions cannot simply be applied to the 600 Wp product.

The 2026 renewable energy feed-in schedule lists NT$5.6279 per kWh as the published ceiling for rooftop PV of at least 1 kW and less than 10 kW, the same in both periods. Whether it applies, and whether adjustments are added or deducted, depends on the installation, relevant rules, and contract. It is not a universal purchase price for every generator or a guarantee for a new project in a later year. Official 2026 rate schedule

You can first check Taipower’s available grid connection capacity, then obtain a formal assessment for the project. Capacity shown online is not a reservation for your project. Solar projects receiving preliminary approval from July 1, 2026 onward also need proof of full payment of module recycling fees for equipment registration. Include that requirement in the schedule and budget. Installation management rules, Articles 12 and 13

The feed-in purchase period is twenty years where the announcement and contract conditions are met. Rules cover completion periods, equipment subsidies, additional rates, and changes, among other matters. “Twenty years” should not be treated as a guarantee of twenty years of net earnings. 2026 tariffs and calculation formulas

Direct corporate sales and operating as an electricity retailer

A corporate power purchase agreement is usually called a PPA. Direct supply means electricity delivered directly under the applicable approvals; wheeling involves delivery and settlement through the grid. A PPA involves the parties’ eligibility, power sources, those delivery methods, metering, grid charges, contracts, and potentially certificates. Owning a generator or completing ordinary company registration does not automatically qualify you to operate an electricity retail business. Electricity Act, Articles 2, 15, 45, and 69

Lawfully installed Type 2 and Type 3 renewable generation equipment for self-use may explore selling electricity to a renewable electricity retailer, which can then resell it to corporate customers. Every small project does not need to establish its own electricity retail company. Energy Administration explanation of Type 2 and Type 3 sales to renewable electricity retailers

Ordinary equipment sales, lawful feed-in sales, and operating a renewable electricity retail business should not be conflated. Taiwan Renewable Energy Certificates (T-RECs) document the renewable attribute of particular electricity output; the environmental benefits of the same kWh cannot be promised more than once. Electricity sold under the feed-in tariff scheme cannot also receive T-RECs. Where self-use equipment sells surplus under that scheme, the qualifying self-consumed portion may receive certificates, but the feed-in portion does not. Corporate PPAs should specify ownership, transfer, and usage claims for both electricity and certificates. National Renewable Energy Certification Center FAQ

For a part-time engineer without a site or an electricity buyer, this is generally worth pursuing later than energy software services or equipment integration.

Is a generation project financially worthwhile?

The following is an illustrative model you can recalculate, not a quotation, investment advice, or a promise of returns. Rooftop construction, structural reinforcement, grid connection, rent, tax, financing, insurance, declining output, replacement equipment, and removal can all change the result.

Assume a 5 kWp rooftop project that qualifies for the tariff category above and whose eligibility for that ceiling rate has been formally confirmed, with total investment of NT$300,000, first-year net generation of 1,200 kWh per kWp, and maintenance and other recurring costs provisionally set at NT$3,000 per year. All three figures are assumptions used in this article. The electricity sales calculation temporarily uses NT$5.6279 per kWh without other adjustments.

In plain terms, the assumed NT$300,000 investment brings in about NT$33,800 in year-one electricity sales. After NT$3,000 in recurring costs, about NT$30,800 remains. Dividing the initial investment by that one-year figure gives a simple payback of about 9.8 years. Later replacements, site rent, financing, and taxes are still missing, so this is not a guarantee of profitability in year 9.8.

Annual generation         = 5 × 1,200             = 6,000 kWh
Gross revenue, all output sold = 6,000 × 5.6279    = NT$33,767.4
After illustrative recurring costs = 33,767.4 − 3,000 = NT$30,767.4/year
Simple payback            = 300,000 ÷ 30,767.4    ≈ 9.8 years

With the same investment but different generation assumptions:

These figures exclude the time value of money and later major repairs, so payback years should not be mistaken for a rate of return. For an actual decision, list cash received and paid each year, then translate future amounts into today’s value. Subtracting the investment gives net present value. Internal rate of return is the discount rate that makes that net present value zero; it still needs to be assessed alongside cash flows and risks. Test at least low generation, lease termination, inverter replacement, and delayed grid connection.

If instead 70% is self-consumed and 30% is lawfully sold as surplus, and the displaced electricity price is assumed to be NT$3.5 per kWh:

Self-consumption savings = 6,000 × 70% × 3.5    = NT$14,700
Surplus sales revenue   = 6,000 × 30% × 5.6279 = NT$10,130.22
Combined benefit        = NT$24,830.22/year, before costs

This is not a quotation of Taiwan’s current residential tariff. It simply illustrates that a higher self-consumption rate does not necessarily produce a greater financial benefit: compare the actual electricity price displaced with the available sales terms. Self-consumption, backup, and decarbonization may also serve different goals, whose values should be considered separately.

Battery economics require more than tariff spread times capacity

Arbitrage means charging when electricity is cheaper and discharging when it is more expensive to save money. Calculate it using kWh actually delivered to the equipment. “Round-trip efficiency” is the share of charging energy that can be recovered; “cycle degradation” is the battery aging cost caused by use.

Net benefit per kWh discharged
= avoided peak tariff − off-peak charging tariff ÷ round-trip efficiency − cycle degradation cost per kWh

Assume a peak tariff of NT$6, an off-peak tariff of NT$2, round-trip efficiency of 85%, cycle degradation cost of NT$1 per kWh, and 600 kWh delivered per year. The illustrative benefit is (6 − 2 ÷ 0.85 − 1) × 600 ≈ NT$1,588. All inputs are illustrative, not current tariffs or battery warranty figures. If some charge is reserved for outage backup, less energy remains available for arbitrage.

When calculating the full project, avoid counting both battery replacement cash flows and the degradation allowance above for the same cost. A demand charge is based on the highest power drawn during a defined period. Savings depend on whether it applies to the customer and whether that maximum demand actually falls. Nameplate output power cannot simply be converted into guaranteed savings. Actual customers should first check Taipower’s time-of-use tariff calculator.

Calculating the cost of selling a complete system

Again using illustrative assumptions, the table below estimates the earlier 600 Wp/2 kWh system for small remote loads. Supplier quotations must replace these provisional costs. Amounts are in New Taiwan dollars before tax. The sensors themselves and specialized civil works are excluded.

At an assumed pre-tax sale price of NT$110,000, the contribution per unit is NT$34,000, or about 30.9% of the selling price. This is the per-unit contribution after the delivery costs above, not company net profit or gross profit under a strict accounting definition. Development, certification, customer acquisition, administration, insurance, and inventory financing still need to be covered.

If a provisional one-time budget of NT$600,000 is set aside for development, testing, inspection, and fixtures, recovering just that amount requires 600,000 ÷ 34,000, rounded up to 18 units. NT$600,000 is not a certification quote and may not cover all actual requirements. It should not be used as a financing basis before a classification decision and formal laboratory quotations are obtained.

If per-unit costs rise by 15%, contribution falls to NT$22,600, and recovering the same fixed investment requires 27 units. Meanwhile, the original delivery cost of 20 units already totals NT$1.52 million. A profit on paper does not mean there is cash available to pay suppliers.

Now assume all 20 units receive 50% deposits, bringing in NT$1.10 million upfront. If you must first secure the NT$1.52 million delivery budget plus NT$600,000 of upfront development funding, there is still a NT$1.02 million funding gap, before any buffer for operations, taxes, or delayed acceptance. A warranty reserve may not require an immediate cash payment. Recalculate actual funding needs month by month using the timing of deposits, supplier payments, installation, final customer payments, and warranty spending.

During small-volume production, use clear specifications, staged payments, and supplier lead-time management to avoid buying large inventories upfront. For service subscriptions, specify communications, cloud hosting, alert handling, whether site visits are included, and the service term. Do not promise a one-time purchase with free on-site service forever.

Why would a customer buy a NT$110,000 system? Under this article’s assumptions, a 40 W load consumes about 350 kWh per year. Even at an assumed NT$5 per kWh, the electricity is worth only around NT$1,750. The purchasing rationale must come from avoiding power connection work, site inspections, downtime, or another value confirmed by the customer. “It saves money on electricity” is not a viable justification for this example by itself.

What to do in the first ninety days

The plan below assumes around ten to fifteen hours per week. If certification, engineering work, or site permits remain incomplete, postpone dependent activities. A calendar does not replace qualification requirements.

First revenue can come from site assessments, data analysis, or maintenance of manufacturer-supplied equipment that you can already deliver lawfully. It need not wait for your own complete product. A prototype lent out free of charge is not automatically exempt from safety or regulatory requirements. Before entering a customer site, confirm permitted testing arrangements and responsibilities.

Success at ninety days means finding a use case with a willing payer, repeatable requirements, and a workable technical and regulatory route. It does not mean promising certification and volume production within three months.

A twelve-month route to repeatable shipments

Two manufacturing terms matter here: BOM means bill of materials, the list of component models, quantities, and versions. ODM means original design manufacturer, a supplier engaged to help design and manufacture a product. This can reduce the investment needed to build your own factory, but design rights, quality, and certification responsibilities still need agreement.

Design validation asks whether power, energy, thermal performance, environmental performance, and fault behavior meet the specification. Production validation asks whether another batch of components or another operator can produce the same quality. “The prototype ran for a week” cannot answer both questions.

Leave destructive battery tests and electrical safety tests to suitable laboratories rather than creating hazardous conditions at home. Outdoor testing is not equivalent to structural certification for typhoons. Claims about waterproofing, service life, or uninterrupted supply require corresponding evidence.

When signing a manufacturing contract, address more than unit price and minimum order quantity: specify ownership of designs and firmware, the certification applicant, a ban on unauthorized component substitutions, discontinuation notices, responsibility for defective batches, spare-parts supply, and maintenance documentation. The manufacturer’s warranty and the warranty you give the customer are separate commitments.

Which skills should you build, and which partners should you find?

The most useful skills for a software engineer to add first are DC/AC fundamentals, power versus energy, measurement error, reading solar and battery specifications, communications protocols, time-series data, system failures, and testing methods. Learn forecasting and optimization after those foundations, rather than reversing the order.

Find electrical engineering professionals who understand power and protection, installation partners qualified to carry out the on-site work, and testing/certification bodies familiar with product classification early. Fixed outdoor installations may also require structural, building, and fire-safety expertise, depending on the project.

You do not have to become a battery chemist, power electronics designer, and electricity retailer at once. Your core role can be turning customer needs into verifiable specifications, equipment status into trustworthy data, faults into actionable work orders, and finally one-off engineering into a repeatable product.

When to stop or change direction

If any of the following occurs, resolve it before expanding:

• After ten interviews, you still cannot identify the decision-maker or reason to pay; people are interested but have no budget.
• The panel area, battery capacity, or installation conditions exceed what customers will accept.
• An existing product plus a simple service is cheaper, and your version offers no measurable advantage.
• Moving forward requires a major order before the certification, site, or installation route is clear.
• Once warranties, dispatches, and repairs are included, contribution per unit is insufficient to sustain operations.
• The finances only look attractive if electricity prices surge, every day is sunny, batteries never degrade, or there is no after-sales work.
• A key supplier will not provide necessary testing, traceability, maintenance documentation, or change notifications.

Taking site access, capital, and compliance together, a plausible starting sequence is to obtain data from a real site and build a useful energy service first. If the customer also needs a power supply, deliver an integrated system using maintainable equipment whose compliance can be confirmed. Invest in your own brand and complete system after the same requirement keeps recurring.

AI draws more attention to electricity, but the opportunities a small team can capture are usually concrete: one fewer equipment outage, one fewer site visit for an owner, or a clearer understanding of supply costs. Do those things well before scaling the side business.

From a power equipment business to fusion: a conditional research route

If the long-term goal is to use fusion to provide reliable electricity at a total supply cost one-tenth, or even one-hundredth, of competitors’ costs, that remains a long-term research goal requiring step-by-step evidence. As of October 3, 2026, this review found no commercial fusion operating record that supports such a cost advantage.

A conditional route based on funding and evidence is to make your energy equipment business self-sustaining first, then sell research tools and engineering services, and next co-develop key technologies. Form a dedicated team to develop a fusion power system only when independent validation indicates a credible opportunity to meet your cost target. This may allow the equipment and research-service businesses to earn revenue while breakthroughs remain pending. Becoming a fusion electricity supplier would require separate proof.

The solar system discussed earlier and a fusion system are not two versions of the same machine. Measurement, data, equipment management, customer delivery, and quality control are transferable capabilities. Reactor engineering, materials, fuel cycles, and radiation protection require additional expertise. Replacing part of a solar system with a “fusion core” would not bypass those engineering challenges.

This survey defines “lowest-investment R&D” as spending as little as possible on things you cannot recover before the next major investment, to investigate the issues most likely to make the plan fail. The evidence is not yet sufficient to identify a commercial fusion reactor design that is definitively the cheapest.

Agree on what ten to one hundred times cheaper means

This survey expresses “ten times cheaper” as a cost reduced to one-tenth, or a 90% reduction, and “one hundred times cheaper” as a cost reduced to one-hundredth, or a 99% reduction. This avoids ambiguity.

The comparable metric is the full lifecycle cost of each kWh actually delivered under equivalent service conditions. “Lifecycle” means the entire period from research and construction through maintenance and component replacement to final shutdown and equipment disposal.

If you cannot obtain a competitor’s internal costs, use public research or available quotations for a complete electricity supply service as a proxy, and state the limitations. If you can verify only a lower selling price, claim only the price difference, not proof of the competitor’s costs.

Looking only at water, fuel, or electricity bills during operation can omit the most expensive equipment and financing. A restaurant still pays rent, kitchen equipment costs, and wages even if its ingredients are free. If fuel hypothetically accounts for 5% of a system’s total cost, eliminating fuel costs leaves total costs at 95% of the original amount. This example is not an actual fusion cost breakdown.

Work backward from your goal into a cost table you cannot skip

Start with a purely illustrative assumption to see how demanding the goal is: a competitor’s full cost is NT$3 per kWh delivered. This is neither Taiwan’s current electricity tariff nor an estimate of a fusion electricity price.

No reactor design can yet be identified as achieving this. The cost table instead asks: How much construction cost does this permit? How much maintenance? How much time without generation?

To understand the constraints, first express the system per “1 kW of net rated capacity available for external delivery.” Net capacity already deducts the electricity required by the machine itself. It cannot be replaced with the thermal power released by the reaction.

Cost per kWh ≈ (capital investment per kW × annual capital recovery factor + annual fixed costs per kW)
               ÷ annual kWh actually delivered per kW
               + variable and delivery costs per kWh

For now, think of the “annual capital recovery factor” as the proportion used to spread upfront investment and financing costs across each year. It is not a rate of investment return. A formal model must use a discount rate and economic life. The discount rate reflects debt, equity, and risk; it cannot simply be the bank lending rate. Put plainly, borrowing has a cost, and shareholders also require a return on their investment. NLR cost calculation guidance

Capital investment must also include engineering, construction-period financing, and a reasonable allocation of R&D, rather than just the reactor itself. Construction-period costs, subsequent capital recovery, replacements, and residual value should sit within a consistent cash-flow model to avoid double-counting. The equation above is only for initial screening; a formal investment analysis requires annual calculations.

Next, deliberately use favorable assumptions: annual delivery equivalent to 90% of rated net capacity, no additional transmission or distribution losses, and an annual capital recovery factor of 0.10. The 90% is an optimistic assumption used to test the target, not a demonstrated fusion power plant result. Each 1 kW would therefore deliver 8,760 × 0.9 = 7,884 kWh per year.

If maintenance, fuel, power delivery, and all other costs are temporarily set to zero, the maximum allowable total capital investment is:

These are optimistic ceilings you must stay below, not market quotations or construction costs already achievable. Adding any other cost lowers the ceilings. For example, if other delivery costs are assumed to be NT$0.2/kWh, the tenfold target leaves only NT$0.1 for capital costs, reducing allowable capital investment to NT$7,884/kW. The hundredfold target would not even cover those non-capital costs.

The purpose of this table is to reject unworkable ideas early. If a lower bound on power delivery, maintenance, or financing within the same service scope already exceeds the target, removing it from the spreadsheet does not establish success. You can change the technology or business arrangement, or acknowledge that the cost target is not yet supported. If you change the comparison market, you must disclose the new baseline too.

A 2026 original research paper in Nature Energy questions whether applying other industries’ rapid cost reductions directly to fusion plants is overly optimistic. Its analysis primarily addresses magnetic confinement and laser fusion. It does not prove that every future design is impossible, but it is a useful reminder: “Mass production will make it cheaper later” needs evidence; it cannot fill today’s cost gap. Research on fusion cost learning rates

Understand in plain language what fusion still needs

Fusion combines lighter atomic nuclei and releases energy. The plasma discussed in research can initially be understood as a gas-like state containing charged particles. Confinement means keeping those particles under conditions suitable for the reaction, so they do not disperse or lose energy too quickly.

But a reaction is only the starting point. You ultimately want to sell electricity that customers can use, so five kinds of achievement must be distinguished.

The commonly used Q is an experimental energy gain, but different devices may use different measurement boundaries. LLNL’s 2025 annual report records 8.6 MJ of fusion energy from an NIF experiment on April 7, 2025, with 2.08 MJ of laser energy delivered to the target. MJ means one million joules, a unit of energy; the key here is the ratio between the two amounts. This is an important target-gain result, but the denominator does not include all electricity used by the facility. NIF official annual report

ITER explicitly states that it does not produce electricity, and its planned Q metric also has a specific definition. Experimental gain should therefore not be converted directly into a margin on electricity sales. ITER official FAQ

For you, the most important R&D gaps include how much electricity the machine consumes, how heat or other energy becomes electricity, how long components last, how long replacements take the plant offline, and whether fuel can be supplied continuously. Routes using deuterium-tritium fuel must also address radioactive tritium supply and cycling. Materials exposed to particle bombardment have lifetime limits too. Adding an AI model does not let you skip these engineering problems.

The US Department of Energy’s 2026 roadmap still identifies materials, fuel cycles, and integrated facilities among the challenges to solve, with a pilot power plant in the mid-2030s as a development goal. This is a national R&D program’s target, not a guaranteed schedule for an individual startup. DOE fusion R&D roadmap

Can the lowest-cost fusion reactor be identified today?

For each approach, investigate which cost is hardest to reduce, then choose collaborations based on access to reliable data, experts, and experimental evidence.

Start with the IAEA’s overview of fusion research to compare approaches, then read the teams’ original results. W7-X research progress illustrates stellarator research, while the University of Wisconsin’s IEC research goals help distinguish particle-source research from electricity supply.

Low-neutron fuels should not be simplified into “no radiation, therefore definitely cheaper.” For example, a 2026 original study on proton-boron inertial fusion analyzes radiation energy-loss constraints under particular conditions. Its model has a defined scope and should not be expanded into a claim that every design must succeed or fail. Study of proton-boron fusion energy losses

The research entry route does not imply a reactor choice. Without a core team with expertise in a particular approach, legally accessible experimental data, and a complete energy and cost model, keeping options open is more defensible than committing early to market “own fusion technology.”

Which capabilities can transfer from your energy equipment business?

For a software-led entrant, tools for experimental data and equipment maintenance are a possible first research-adjacent product: help researchers identify the settings and analysis-code version behind a result, and trace it back to the original measurement. This is closer to your software skills and can also serve general industrial and academic customers, reducing dependence on the pace of fusion development alone.

Your first project can be narrow: organize historical measurement data that the customer already has the right to use, so selected cases can be reanalyzed using the original data, settings, and code versions. Deliver a data inventory, rerunnable analyses, a list of missing data, and a comparison report. Agree with the customer in advance on acceptable differences when reproducing results, time spent organizing data manually, and the scope of support. This is a service you can quote and assess against acceptance criteria, and you can initially deliver it offline.

First ask where the customer’s existing tools actually fall short. If they already have good free tools, provide the necessary integration and validation services. Do not rebuild a general-purpose platform without a problem someone will pay to solve. Nor should ordinary cloud software be connected directly to protection systems or real-time plasma control; that involves a different scope of expertise and responsibility.

Open-source tools can save time spent rebuilding basic functions, but they do not automatically provide a credible power plant design:

Read the official documentation for PlasmaPy, TORAX, and PROCESS first. Before commercial use, pin versions and check the licenses for code, bundled data, and dependencies individually. Permission to use open-source code does not grant rights to someone else’s research data, patents, or trademarks.

An evidence sequence for low-investment fusion R&D

Write a one-page plan for each research effort first: the problem, existing evidence, the new data needed, who evaluates the results, the spending limit, and the results that would trigger a stop. This is cheaper than buying equipment and then searching for a research question.

AI can help organize literature, write data-processing code, check units, detect anomalies, and propose hypotheses to test. Its outputs must be checked against original papers, test cases, and experimental data. Having a model make a graph smoother does not demonstrate extra energy. Making a simulation run one hundred times faster does not mean electricity supply costs fall one hundredfold.

More advanced on-device AI can enter fusion experiment control at specialist facilities. In 2026, the US Princeton Plasma Physics Laboratory reported that its PACMAN framework had been tested in five real-device experiments. Multiple models combined diagnostics and control at millisecond timescales, while independent hardware safety limits constrained final outputs. This is research with a team and experimental facility, not a controller that can be transplanted by adding an AI chip to a small solar product. It is not evidence of commercial fusion generation or one-tenth electricity cost. PPPL PACMAN research report

Every result should distinguish three columns: measured, model-estimated, and not yet demonstrated. This helps nonspecialists understand what is reliable and practitioners quickly identify the research gaps.

By “independent validation,” I mean having qualified experts who were not involved in the model’s main development review the accounting boundaries, original data, units, uncertainties, cost assumptions, and failure scenarios, and document their findings and unresolved questions. This is more than seeking an endorsement, and it cannot replace permits. Passing software tests does not establish that the physical assumptions are valid.

Research partnerships and regulation in Taiwan

A lower-barrier research entry is data and software work that does not involve radiation sources. You can approach NCKU’s FIRST team or NTHU’s industry collaboration office with a specific research question, or find relevant expertise through NIAR’s industry–research matching service. These are institutions you can contact, not evidence that they have agreed to take your project, provide equipment, or perform research for free.

Your first collaboration proposal should explain your existing verifiable capabilities, the specific problem to solve, the data you need, the budget and schedule, acceptance criteria, ownership of results, and what may be published. Agree on data and intellectual property rights early, rather than discovering after completion that you cannot commercialize the work.

Taiwan’s Nuclear Reactor Facilities Regulation Act defines a reactor in relation to a self-sustaining nuclear fission chain reaction. This does not mean every fusion device follows the same procedures, nor does it imply that fusion is unregulated. Nuclear Reactor Facilities Regulation Act

Equipment that generates ionizing radiation and radioactive materials are separately subject to the Ionizing Radiation Protection Act and other rules. Ionizing radiation is radiation capable of removing electrons from atoms or molecules. Research use, not yet generating electricity, or not using tritium does not establish a blanket exemption. Confirm applicable permits before manufacturing equipment; sites, personnel, modifications, and transport each have requirements too. Ask the competent authority to confirm whether a permit or registration is required for the equipment, materials, and applicable thresholds. Ionizing Radiation Protection Act, Regulations governing radiation operations

I therefore would not include building your own neutron source, purchasing fusion fuel, or setting up a residential laboratory as cost-saving shortcuts. Any necessary research should be performed by institutions with the appropriate permits, facilities, and personnel. Signing a university partnership does not automatically provide authorization.

This review did not find a complete Taiwan commercial fusion power plant application template that can be directly applied to your concept. If you eventually move into electricity supply, first have the Nuclear Safety Commission and energy authorities confirm the procedures for the design, then address the site, environmental requirements, electricity business requirements, grid connection, and sales contracts. Research authorization and power plant authorization are separate stages.

Account separately for three kinds of profit

Grants and equity financing can support R&D, but they are not revenue from customers buying products. Profitable research tools do not mean profitable fusion generation. Financial reporting must preserve that distinction.

Formal supply-chain entry points do exist. For example, UKAEA procurement information and the CFS supplier portal can help an entrant understand requirements and qualifications. Registration is not an order, and a list of needs does not guarantee a contract. An initial proposal can offer one narrow, clearly defined capability; a full factory requires stronger evidence of orders and funding.

Developing your own reactor changes the scale of funding. In August 2025, CFS announced that it had raised nearly US$3 billion in total. This is one company’s financing history, not a construction quotation for a power plant or a minimum funding threshold for every approach. It is a reminder not to assume that a small equipment business’s gross profit can independently fund whole-plant R&D. CFS official funding announcement

Funding limits for a first research budget

First reserve funds for salaries, taxes, after-sales support, inventory, and reasonable operating reserves in your equipment business. Only use the remaining cash available for allocation for new research. A ceiling of 10%–20% of annual cash available for allocation is one scenario for discussion. This is management advice, not a statutory percentage. Before each allocation, confirm that cash remains above a predefined minimum reserve for operations and contractual obligations; funds already allocated cannot be counted again. If the available amount is insufficient to complete a research package with clear acceptance criteria, narrow the question, find a paying customer, or raise separate funding.

The following is an example budget for a one-year data and research tools project, not a quotation for a fusion experiment. It deliberately includes the founder’s labor to avoid creating false profits through unpaid work. Actual contracts must replace all assumed unit prices and sales volumes.

This is not a commitment you should make as soon as equipment shipments become reliable. If the core business pays the entire NT$900,000 in cash, the 10%–20% ceiling above requires NT$4.5–9 million in annual cash available for allocation. The value assigned to the founder’s time may not require an immediate cash payment, and customer installments also change the funding gap. Prepare three separate schedules for economic costs, actual cash outflows, and the timing of receipts. If funding is insufficient, first approve a research package of public-case reproduction and customer interviews that can be completed in two to four weeks, then decide whether to expand. Do not omit your own time and claim that the research is free.

Assume two engineering service contracts at NT$250,000 each, plus four annual licenses at NT$120,000 each, with all six contracts fully performed and eligible for revenue recognition within the illustrative year. Revenue would be NT$980,000. Further assume that the budget above covers all delivery work and agreed support. After subtracting the budget, including the NT$140,000 contingency, only NT$80,000 of budget headroom remains. This is neither actual pre-tax profit nor cash available for allocation; shared costs, taxes, and future maintenance commitments still need checking. With one fewer license and the rest of the budget unchanged, there would be a NT$40,000 shortfall. These sales volumes illustrate sensitivity; they are not my forecast of the orders you will win.

Contracts should tie payments to deliverables that can be accepted and addressing data provision, revision limits, acceptance deadlines, intellectual property ownership, and ongoing maintenance fees. When cash arrives matters too; do not look only at whether total year-end revenue is sufficient.

Taiwan’s SBIR program can be assessed separately. Under the 2026 central program, an individual Phase 1 project lasts six months, with a grant ceiling of NT$1.5 million and funding of no more than 50% of total project costs, subject to business eligibility, matching funds, and review requirements. Do not treat the maximum grant as guaranteed cash or assume that personal research interests establish eligibility. 2026 SBIR application guidelines

The first twenty-four months after entering fusion-related work

This schedule begins after your equipment business can deliver reliably and you have an approved R&D budget. It does not mean building a fusion plant within two years from today.

You can build fusion-related revenue and research capabilities during this period. The schedule cannot substitute for the scientific evidence, team, and capital needed for your own pilot power plant. A precise commercial operation year for later stages would be unsupported without further evidence.

Evidence needed before developing a fusion power system

Preserving the end goal does not mean owning every step yourself. You can co-develop with an existing team, obtain a lawful license, or form a joint venture. But before any capital commitment, meet at least the following conditions.

First rank the costs that most need a breakthrough. If the largest cost is expensive components that need frequent replacement, validate lifetime and maintenance. If it is the machine’s own electricity consumption, validate the full energy balance. If it is complex construction, investigate designs that reduce on-site engineering. These are research directions, not grounds for entering tenfold improvements into a spreadsheet without data.

Each quarter, list the three largest uncertainties in electricity supply costs: their estimated shares of total costs, how much an improvement could reduce total costs, and what it would cost to obtain the next credible piece of evidence. Prioritize questions that can be resolved at lower cost and whose answers would change the investment decision. Research tools that are easy to sell and technologies that most help reduce electricity supply costs may be different items; rank them separately.

Reducing size, component costs, and internal electricity consumption while extending lifetime may not all be compatible. Greater compactness, for example, can worsen cooling and maintenance demands. You cannot multiply five independently optimistic twofold improvements and claim a thirty-twofold cost advantage.

How to demonstrate that low-cost electricity supply has actually been achieved

Final acceptance criteria should be written into the research plan rather than choosing favorable calculations after the machine has been built.

1. A third party must be able to recalculate the whole-plant energy account. Include startup, standby, heating, cooling, controls, fuel processing, and other electricity use, with clear measurement boundaries and uncertainties.
2. Measure complete operating cycles. Deduct energy charged into batteries or capacitors before startup, and match or correct for initial and final stored-energy states. Recovering previously supplied electricity is not newly generated fusion electricity.
3. Collect representative long-term data. Record net delivery, reduced output, faults, planned maintenance, component replacements, and downtime. Short tests cannot simply be extrapolated into years of service life.
4. Demonstrate sustainable safety, fuel, and maintenance arrangements. Provide evidence for the supply and cycling required by the chosen fuel route, material lifetimes, waste handling, and lawful operation.
5. Use real engineering and contract data for costs. Include quotations, construction, financing, labor, spares, insurance, decommissioning, and power delivery. Estimate the first unit separately from future mass-produced units.
6. Compare against the competitor baseline agreed in advance. Examine adverse scenarios and uncertainty ranges alongside the central estimate. If only a highly optimistic scenario meets the target, it remains a research goal.

The cost test must show both the 1/10 and 1/100 targets. Reaching one-tenth does not mean another tenfold increase in scale will naturally deliver one-hundredth. If some costs have reached a lower bound, a new technical or system-level breakthrough is needed.

If the credible range of full costs remains above the target, continue a valuable supply-chain business, research the actual bottlenecks, or stop pursuing a particular reactor approach. Do not present profits from research tools as completion of the generation goal, or use large irreversible investments to sustain an unsupported story.

Survey findings: the first three things to validate

First, make the equipment business generate cash that is truly available for reinvestment. Judge that from payments received, full costs, and after-sales records, rather than shipment counts alone.

Second, build an electricity supply cost model with a transparent method that you can recalculate yourself. Fix the scope of the competitor comparison and retain the 1/10 and 1/100 targets. Label every input as a measurement, quotation, literature value, or assumption.

Third, approach research partners and paying customers with one narrow data or engineering problem that has clear acceptance criteria. First demonstrate that you can save researchers time, reduce errors, and improve data reliability, then decide which hardware capabilities are worth owning.

This conditional route may limit early capital exposure while preserving fusion electricity supply as a long-term goal. Each step can produce something deliverable. A tenfold to hundredfold reduction in full electricity supply costs, however, must be earned through subsequent physical, engineering, and commercial evidence. It cannot be treated as an established promise today.

Sources and future updates

Relevant paragraphs link directly to official reports, regulations, manufacturer documentation, and original research. Years, scopes, and forecasts are distinguished for global and Taiwan statistics. The architecture, business sequence, timelines, and calculations are analysis in this survey, not official recommendations or supplier quotations. Fusion experimental gain, whole-plant net electricity, and commercial costs are checked separately; company fundraising is not treated as a plant construction price.

Before starting an actual project, reconfirm that year’s feed-in tariffs, applicable electricity rates, BSMI effective dates, NCC integration conditions, local building and fire requirements, and Taipower’s formal position on the site. The most important inputs to update are the compliance determination for specific models, quotations from three suppliers, and the customer’s hourly load. They will affect whether the project is worth doing more directly than macroeconomic forecasts.

For fusion, keep updating complete energy balances, component life and maintenance data, regulators’ guidance on the specific design, and competing costs for equivalent electricity service. Recalculate research priorities and the 1/10 and 1/100 targets whenever these change. This review is not a guarantee of future commercialization.