Why Hardware Startups Die Between Prototype and Production

The most dangerous moment in a hardware company is often not when the product fails. It is when the product works—once. A convincing prototype unlocks customers, hiring, purchase orders, tooling, and a larger financing round. Those commitments begin compounding before the company has proved that it can build the same product repeatedly, inside specification, at a cost and cadence the balance sheet can survive.

The usual diagnosis is that hardware requires more capital than software. That is true but incomplete. Hardware companies more often break because commercial commitments grow faster than manufacturing knowledge. The prototype proves possibility. Production has to prove control.

This report examines the gap between those two proofs: why yield is a finance variable, why revenue can consume cash during a ramp, why outsourcing production does not outsource manufacturing risk, and what founders and investors should require before paying for scale.

A prototype answers, “Can this design be made to work?” Production answers a harsher set of questions: Can different operators, tools, suppliers, lots, and environmental conditions produce units that behave the same way? Can the company detect drift before the customer does? Can it repair or replace failures without erasing gross margin? Can it finance the time between paying suppliers and collecting from customers?

That distinction is not semantic. A prototype is normally optimized for learning speed. Engineers can hand-select components, tune individual units, accept long cycle times, and debug failures with the people who designed the system standing nearby. A production process is optimized for a distribution: not the best unit, but the range of outcomes across every unit.

In aerospace, NASA makes representativeness explicit: prototype-qualification hardware is expected to use the same drawings, materials, tooling, manufacturing processes, inspection methods, and personnel competency planned for flight hardware. Most startups do not need aerospace qualification, but the test is useful. If the unit was built through a different system than the product that will ship, its performance is weak evidence about production. Source: NASA-STD-7002B with Change 1.

The US Government Accountability Office has studied this transition for decades in complex acquisition programs. Its framework identifies three knowledge points before major investment: technology maturity, design stability, and mature production processes. In a 2002 review, the GAO found that leading commercial companies stabilized design and demonstrated reliable, controlled manufacturing before committing to production. Programs that entered later phases with only about one-quarter of drawings complete and fewer than half of manufacturing processes in control experienced poor cost and schedule outcomes. The lesson is not that startups should inherit defense procurement. It is that a functioning technology, a stable design, and a capable process are separate assets. Source: GAO, 2002.

Startups regularly finance them as though they were one.

“Scaling manufacturing” sounds like one workstream. In practice, it is a system of risks that amplify one another.

The coupling is what makes the gap lethal. A late design change can obsolete long-lead inventory. The replacement part can require a new supplier process. That process can alter qualification evidence. Requalification can delay acceptance. Delayed acceptance can defer collection while payroll, rent, and purchase commitments continue.

A software team can often ship a correction to the installed base. A hardware team may have to identify affected serial numbers, stop the line, quarantine inventory, rework finished units, update test fixtures, negotiate with a supplier, and decide whether fielded units need service. The failure is not one bad part. It is the number of financial and operational promises attached to that part.

Design stability is not design perfection

Freezing a design too early preserves mistakes. Freezing it too late converts learning into scrap.

The right question is not whether engineering changes have stopped. It is whether the company understands the cost and blast radius of a change. A controlled design has named requirements, interfaces, tolerances, approved parts, test methods, and revision ownership. Changes can still happen, but they move through a system that exposes which tools, suppliers, tests, inventory, certifications, and customers are affected.

The GAO often uses completed engineering drawings as one tangible indicator of design stability in major programs. A startup does not need a government documentation burden, but it does need a production baseline: the minimum complete description from which a supplier and an internal team can independently build and test the same revision.

Without that baseline, “manufacturing problems” are often unresolved product decisions arriving on the factory floor.

Process stability comes before process capability

Yield is the most visible production metric, but a single yield percentage can hide the risk that matters.

First-pass yield asks how many units clear a process without rework. Final yield can look healthier because repaired units eventually pass. Neither number alone shows whether defects are random, concentrated in one supplier lot, created by one station, or escaping into the field. A team can improve final yield by adding technicians and inspection while making no improvement to the underlying process.

The National Institute of Standards and Technology separates stability from capability. A stable process has a constant mean and variance over time; only after stability is demonstrated does it make sense to compare the process distribution with specification limits. A process can therefore have a good average and still be uncontrolled, or be stable but consistently centered too close to a limit. Source: NIST Engineering Statistics Handbook.

For a startup, the practical sequence is:

• Define the critical-to-quality characteristics that predict customer performance.
• Instrument the stations that create those characteristics.
• Preserve traceability across component lot, operator, tool, software version, and finished serial number.
• Separate first-pass yield, rework recovery, scrap, and field failure.
• Establish that the process is stable before extrapolating a pilot average into a production forecast.

This is not quality theater. It is the difference between a forecast built on a process and a forecast built on a lucky batch.

Founders often model gross margin as a future reward for buying components at volume. During the ramp, margin is governed at least as much by the conversion of purchased material into accepted units.

The simple model below is illustrative, not an industry benchmark. Assume a product sells for $100, purchased material costs $45 per build attempt, and labor plus other conversion cost is $15 per shipped unit. To isolate yield, assume material in a failed attempt is lost and ignore warranty, freight, overhead, and recoverable rework.

The arithmetic is purchased material divided by yield, plus conversion cost. Real factories are messier: some material can be recovered, rework consumes variable labor, failures can occur at different points in the routing, and capacity lost to rework can be more expensive than the scrapped component. The model is useful because it shows the shape of the problem. Moving from 50% to 75% yield does not merely add 25 points of output. Under these assumptions, it changes a loss-making unit into a unit with $25 of contribution.

Yield also changes capacity. If the bottleneck process can attempt 100 units per week, 60% yield produces 60 good units before recovery; 90% produces 90. A sales plan built around nominal machine capacity silently assumes a yield plan.

That is why an investor should distrust a margin bridge that begins with supplier price reductions but does not show first-pass yield, scrap, rework hours, throughput at the constraint, warranty reserve, and the mix of units that have passed customer acceptance.

Production businesses usually pay before they are paid. Suppliers may require deposits for tooling, minimum orders for custom components, or payment before shipment. The company then holds raw material, work in process, and finished goods. The customer may pay only after delivery, installation, commissioning, or acceptance.

Growth increases the amount of cash traveling through that interval.

The financing problem is best understood as three clocks. The technical clock runs through tooling, qualification, yield stabilization, and customer acceptance. The cash clock starts earlier, when supplier deposits, inventory, labor, and freight are paid, and ends later, when the customer pays for an accepted unit. The financing clock is imposed by runway, grant periods, debt maturities, and the next fundraise. The company fails when the cash and financing clocks expire before the technical clock produces repeatable, collectible delivery.

Consider another illustrative model. Assume each accepted unit consumes $60,000 of cash cost and there are no customer deposits or supplier credit. The table shows cash tied up solely by the delay between production spend and customer collection. It excludes fixed operating burn, tooling, scrap, taxes, and capital equipment.

At 50 units per month, adding 30 days to the cash gap requires another $3 million before considering any additional factory burn. A forecast can therefore be correct about demand, revenue, and even eventual gross margin while being fatal to liquidity.

Public-company filings show the mechanism clearly because the damage becomes visible in inventory charges, purchase commitments, expedited freight, and negative gross profit.

In the first quarter of 2022, Rivian said low volumes on lines designed for higher volumes created significant labor and overhead pressure. It also recorded $188 million of charges related to writing inventory down to net realizable value and losses on firm purchase commitments, while noting added expedited-freight costs. By the third quarter, those inventory and purchase-commitment effects were $696 million for the quarter. Sources: Rivian Q1 2022 filing and Q3 2022 shareholder letter.

This is the production-ramp paradox: volume is needed to absorb labor, depreciation, and overhead, but increasing volume before process and demand are synchronized can create more inventory, more purchase exposure, and more negative cash flow.

“We will fix margin with scale” is not a plan unless the company can identify which costs decline because of learning, which decline because of purchasing leverage, which are fixed and absorbed over volume, and which increase because the process is not ready.

No case study proves a universal law. Public records overrepresent large companies, distressed companies, regulated products, and businesses that eventually filed with a court or securities regulator. They remain useful when the evidence is treated as a mechanism—not a morality play.

Tesla: automate the stable work, not the uncertainty

Tesla’s Model 3 ramp is remembered as “production hell,” but the company’s own filings are more useful than the slogan.

In October 2017, Tesla reported that only 260 Model 3 vehicles had been produced in the third quarter and attributed the shortfall to a handful of bottlenecks, even though most manufacturing subsystems could operate at high rate. Before the ramp, Tesla had correctly described production as an S-curve constrained by the least successful element in the entire supply chain and process. Sources: Tesla Q3 2017 update and Q2 2017 update.

The 2017 annual report described higher automation in material conveyance and battery-module production, interim semi-automated lines that would add labor cost, new suppliers that had to ramp to Tesla’s schedule, and significant cash and management resources required by the plan. The 2018 filing later said Tesla reduced automation in challenged areas and introduced semi-automated or manual processes at additional labor cost. Sources: Tesla 2017 10-K and 2018 10-K.

The lesson is not that automation is bad. Automation replicates a defined process quickly. When product interfaces, material behavior, or station logic are still changing, rigid automation can industrialize assumptions before they are true. Manual or semi-automated work can be rational during learning because it exposes failure modes and preserves flexibility. Automation becomes leverage after the work is understood.

Tesla survived the learning bill. Many startups finance the factory as if that bill will not arrive.

Rivian: utilization helps, but the income statement can hide the factory

Rivian demonstrates both sides of the ramp. In 2022, low utilization, inventory adjustments, purchase commitments, and logistics costs produced severe negative gross profit. By 2025, the company reported $144 million of consolidated gross profit, an improvement of more than $1.3 billion from 2024, citing higher average selling prices, software and services performance, and lower cost per vehicle. Its automotive segment, however, still reported a $432 million gross loss for the year; software and services generated $576 million of gross profit. Source: Rivian 2025 results.

That distinction matters. Consolidated gross profit can turn positive without the manufactured product itself crossing the same threshold. Investors underwriting a factory should separate hardware contribution, regulatory credits, software or development payments, service revenue, and one-time commercial arrangements.

Rivian is not evidence that scale always repairs economics. It is evidence that production cost can improve materially—and that segment mix must be read before declaring the ramp complete.

Peloton: capacity is a demand bet with cancellation penalties

Peloton shows the opposite synchronization error: capacity and inventory committed against demand that changed.

In May 2022, management wrote that excess inventory had consumed “an enormous amount of cash” and left the company thinly capitalized for its scale. By the fiscal fourth quarter, Peloton described inventory commitments as an existential threat, reported $1.1 billion of ending inventory, and recorded a $182.3 million increase in inventory reserves that contributed to a deeply negative connected-fitness gross margin. The company exited owned manufacturing and shifted toward third parties. Sources: Peloton Q3 FY2022 letter and Q4 FY2022 letter.

The production failure was not an inability to make the product. It was the conversion of a demand forecast into fixed facilities, supplier commitments, logistics capacity, and inventory before the forecast had earned that rigidity.

Outsourcing afterward made more cost variable, but it could not retroactively cancel the commitments already made.

Proterra: customization turns backlog into working capital

Proterra’s 2023 bankruptcy declaration described a different trap. Its transit business built highly customized electric buses, requiring extensive and often unique inventory and labor-intensive manufacturing. Supply disruptions created production inefficiency; long-term contracts became economically unfavorable; and minimum commitments to a bus-body supplier created liabilities when Proterra could not take the agreed volume. Management concluded that merely shrinking the transit business would not solve its working-capital and cost-structure problems. Source: Proterra first-day declaration, August 2023.

Backlog is not liquidity. A large order for a customized product can be a claim on cash if pricing is fixed, input costs can move, configurations are unique, delivery dates carry penalties, and payments arrive after acceptance.

This is why contract quality belongs in manufacturing diligence. Two companies with the same backlog can have opposite financing profiles depending on deposits, milestones, cancellation rights, price adjustment, acceptance, and whether inventory can be redirected to another customer.

Northvolt: demand does not rescue an unfinished process

Northvolt entered Chapter 11 in November 2024 after raising substantial capital and securing major customer commitments. Its first-day declaration said the company had secured $55 billion in orders but had difficulty scaling production to meet that demand. The filing described bridge financings used to support the ramp and a deteriorating liquidity position. At filing, Northvolt reported roughly $30 million of available cash—about one week of operations—and $5.8 billion of debt. Sources: Northvolt first-day declaration and Reuters, November 21, 2024.

It would be too simple to attribute Northvolt to yield alone. The company also faced strategy, execution, supply-chain, market, and financing pressures. The defensible conclusion is narrower: customer demand and strategic importance did not make production maturity or liquidity optional.

Orders can validate the market while the factory remains unvalidated.

Manufacturing maturity should not be reduced to one “production ready” label. The Department of Defense’s Manufacturing Readiness Level framework uses ten levels to create a shared vocabulary for manufacturing risk. The exact government process is too heavy for most startups, but the principle is useful: evidence should become more representative before capital commitments become less reversible. Source: GAO review of Manufacturing Readiness Levels.

Flux’s simplified evidence ladder is below.

The order matters. Factory capacity is a solution to insufficient throughput. It is not a solution to an unstable design, an unqualified product, or uncertain demand.

Qualification and acceptance are different

Complex hardware companies often use “testing” to describe several economically different activities.

NASA’s systems-engineering guidance distinguishes qualification from acceptance. Qualification demonstrates that the design meets functional and performance requirements under anticipated environmental extremes; it is generally performed once while the design remains unchanged. Acceptance is a smaller verification set performed on each flight unit to show that manufacturing and workmanship conform to the qualified design. Source: NASA Systems Engineering Handbook, Product Realization.

The distinction translates beyond aerospace:

• Development testing discovers how the design behaves.
• Qualification proves that a representative design can survive the required environment.
• Acceptance testing determines whether a specific produced unit conforms.
• Field reliability tests whether the assumptions remain true over time and use.

A pilot can succeed while production acceptance is undefined. That is not a small omission. If the customer and company have not agreed what constitutes an accepted unit, the company cannot reliably forecast delivery, revenue recognition, collection, or support cost.

The test plan is part of the commercial contract.

Equity is unusually tolerant capital. It can finance technical uncertainty, design iteration, team formation, and a pilot process with no collateral or predictable repayment schedule. That does not mean it should finance every machine, inventory cycle, or project asset forever.

The cheapest capital usually arrives only after uncertainty has been converted into evidence.

The sequencing principle is straightforward: use flexible capital to retire uncertainty, then use cheaper and more structured capital for repeatable assets and cash flows.

Government energy lending makes the boundary visible. The Department of Energy’s financing office says that for certain innovative energy and supply-chain projects its engineers generally look for 1,000–2,000 hours of successful demonstration or pilot operation, preferably including operating and maintenance cycles. It also describes independent engineering review, completed front-end engineering work, and technical, schedule, cost, and repayment diligence. These are program-specific expectations, not a universal readiness benchmark. They illustrate why project lenders arrive after technical evidence, not in place of it. Source: US Department of Energy applicant guidance.

Founders get into trouble when they raise expensive, flexible equity to construct assets before the process is known, then discover that the supposedly cheaper follow-on capital still requires the evidence the factory was meant to create.

A contract manufacturer can provide equipment, trained labor, purchasing leverage, certifications, and operating systems that would be irrational for a startup to recreate. It can also make fixed costs more variable. Those are meaningful advantages.

But the manufacturer cannot own an ambiguous requirement on the startup’s behalf.

If drawings are incomplete, tolerances are unnecessarily tight, test coverage is weak, demand mix changes weekly, or key components are single-sourced, an external factory inherits the instability and prices it through NRE, minimum commitments, longer lead times, change charges, buffers, or margin. The startup still owns architecture, supplier choices, forecast quality, acceptance, field performance, and the economics of every revision.

The make-versus-buy decision should be made by learning loop:

• Keep work close when it contains proprietary process knowledge, changes frequently, determines core performance, or produces data that improves the product.
• Outsource work when the process is understood, suppliers have demonstrated capability, interfaces are stable, and external scale or specialization is genuinely advantaged.
• Avoid vertical integration justified only by fear. Owning a weak process does not strengthen it.
• Avoid outsourcing justified only by accounting. Moving equipment off the balance sheet does not remove deposits, inventory exposure, quality escapes, or forecast liability.

The goal is not maximum ownership. It is minimum distance between a failure and the team capable of learning from it.

A production dashboard should expose the causal system, not reward units shipped at any cost.

At minimum, founders and boards should see:

• Revision stability: engineering changes by week, their cause, and inventory or tooling affected.
• First-pass and final yield: by station, supplier lot, product revision, and time—not one blended percentage.
• Rework: hours, queue age, recovery rate, and whether rework consumes the bottleneck.
• Throughput at the constraint: demonstrated output over sustained runs, not instantaneous nameplate speed.
• Supplier performance: on-time delivery, incoming defects, lead-time movement, capacity evidence, and alternate-source status.
• Qualification: requirements passed, open anomalies, retest triggers, and whether units were production-representative.
• Reliability: failure distribution by operating hours or cycles, field conditions, and revision.
• Unit economics: purchased material, direct labor, burden, scrap, rework, freight, warranty, installation, and service.
• Cash conversion: deposits, inventory days, acceptance timing, receivable days, and purchase commitments.
• Demand quality: funded orders, cancellation rights, configuration concentration, deposits, and gross margin by contract.

The important board conversation is not “Did yield improve?” It is “Which mechanism improved yield, did it persist across lots, and what cash commitment is now justified because of that evidence?”

The strongest manufacturing diligence begins with reconciliation.

The bill of materials should reconcile to the unit-cost model. The unit-cost model should reconcile to purchase orders and supplier quotes. The throughput forecast should reconcile to station cycle times, uptime, staffing, and yield. The delivery forecast should reconcile to qualification and customer acceptance. The revenue plan should reconcile to the cash calendar.

Useful diligence questions include:

• Which production metric would invalidate the current financing plan first?
• What is the lowest demonstrated yield across multiple lots, not the best pilot yield?
• Where is the current constraint, and where does it move if that station improves?
• Which design changes would obsolete the most inventory or require requalification?
• Which supplier commitment becomes payable before the corresponding customer payment?
• What percentage of backlog includes deposits, cancellation protection, or price adjustment?
• Does the quoted gross margin include scrap, rework, warranty, freight, installation, and field service?
• Which processes are stable, which are merely improving, and what data supports the distinction?
• What part of capacity can be delayed until demand or process evidence is stronger?
• If the next financing closes six months late, which commitments can actually be stopped?

The data room should contain bad builds. A company that shows only the golden unit is presenting product possibility, not production knowledge.

Hardware moats are often described as patents, supply chains, factories, or vertical integration. Those can matter. The deeper advantage is a closed learning system that converts every build and field hour into faster control over design, process, cost, and reliability.

Manufacturing data becomes strategically valuable when it is traceable enough to answer:

• Which input changed?
• Which process step created the variation?
• Which serial numbers are exposed?
• Which test could have detected it earlier?
• Which design or supplier action prevents recurrence?
• What did the corrective action do to cost, throughput, and reliability?

That system compounds. It lowers the cost of the current product and shortens the path to the next one. A factory without this loop is capacity. A factory with it can become product-development infrastructure.

This changes how capital efficiency should be judged. The right numerator is not simply prototypes built or revenue booked. It is durable uncertainty retired: stable specifications, controlled processes, qualified suppliers, accepted units, predictable service burden, and a cash cycle that improves with scale rather than deteriorates.

The last prototype may be the best-looking unit the company has ever built and the worst predictor of what comes next.

Production begins when performance is no longer dependent on the people who invented the product being in the room. It requires a stable enough design, a measured process, suppliers that can repeat, tests tied to customer acceptance, and capital whose duration matches the learning cycle.

Hardware companies do need more money. But money cannot substitute for missing manufacturing knowledge; it only increases the size and speed of the commitments made without it.

The companies that cross the gap do not eliminate uncertainty before scaling. They make each commitment reversible until the evidence earns irreversibility.

This report was prepared from public information available through September 7, 2026. Company cases rely primarily on SEC filings, bankruptcy declarations, and company disclosures; government frameworks are used for manufacturing principles, not as a claim that startup production should copy government acquisition. The two financial tables are illustrative Flux models with stated assumptions and are not industry benchmarks.

• GAO: Capturing Design and Manufacturing Knowledge Early Improves Acquisition Outcomes
• GAO: DOD Can Achieve Better Outcomes by Standardizing Manufacturing Risk
• NIST Engineering Statistics Handbook: Assessing Process Stability
• NASA Systems Engineering Handbook: Product Realization
• NASA-STD-7002B with Change 1: Payload Test Requirements
• Tesla Q2 2017 update
• Tesla Q3 2017 update
• Tesla 2017 Form 10-K
• Tesla 2018 Form 10-K
• Rivian Q1 2022 filing
• Rivian Q3 2022 shareholder letter
• Rivian full-year 2025 results
• Peloton Q3 FY2022 shareholder letter
• Peloton Q4 FY2022 shareholder letter
• Proterra first-day bankruptcy declaration
• Northvolt first-day bankruptcy declaration
• Reuters: Northvolt files for Chapter 11
• US Department of Energy: Questions Applicants Should Ask Before Applying