The spec was signed off on a Tuesday. By Friday she had three quotes from contract manufacturers, two of which differed by a factor of two on the same drawings, and a question she had not expected to be the difficult one: who decides when this thing is finished enough to cut steel?
Writing the requirements is the part founders anticipate. We have argued elsewhere that the document is the cheapest prototype you will ever build. What surprises people is the job that starts the day it is signed, when the spec stops being an argument about the product and becomes a set of instructions that strangers on another continent will execute for money. Someone has to hold that. In a five-person hardware startup it is usually the founder who is worst at saying no.
Nobody calls it programme management at first. It arrives as a WhatsApp group with a factory and a spreadsheet that has stopped agreeing with itself.
The decisions that do not come back
A software product manager makes a wrong call on Monday and reverses it on Thursday. The cost of the mistake is a few days of someone’s attention, and the organisation has built its whole rhythm around that fact: ship, measure, undo.
Instrumental, which sells build-data tooling to electronics manufacturers, puts the hardware iteration loop at roughly a hundred times longer than the software one. A change you can define on a whiteboard in an hour takes weeks to build, days to a fortnight to get test results back, and months if a high-complexity part sits on the critical path. Their handbook makes the harder point too: because iterations are so expensive, hardware teams have to start backup plans B, C and D before plan A has failed, where a software team would simply try the next thing on Tuesday.
That changes what the job is. Feature prioritisation barely enters into it. You spend your thinking time deciding which of your handful of remaining iterations goes to which uncertainty, knowing you will run out of iterations before you run out of doubts, and you spend your talking time convincing people that a decision made now cannot be revisited in six weeks.
What the three gates are actually for
EVT, DVT and PVT get explained as if they were paperwork. They are really three different questions, asked in an order that stops you spending money on the wrong answer.
EVT asks whether the design works at all. Titoma, a Shenzhen and Taiwan based design house, puts a realistic EVT at 20 to 50 units and four to five weeks minimum, with production tooling deliberately incomplete: parts are machined or printed, plastics come off soft tools, and nobody has committed to steel. You are hunting for the flaws that would make the product impossible rather than merely annoying.
DVT asks whether this design can be built repeatedly to a standard someone would pay for. Titoma’s numbers here are 50 to 200 units and eight weeks minimum, which they call the hardest stage, and it is where certification lands. Cosmetics, drop and climatic testing, EMC, and the emissions and cyber-security work that now gates a CE mark all belong to this window rather than to some tidy phase at the end.
PVT asks whether the factory can do it, at rate, with the people who will actually be on the line. Titoma suggests 5 to 10 per cent of the first production run over four weeks minimum. The design is frozen. What is under test is the process, the fixtures, the operators and the yield.
Running the gates is the easy part. The discipline is refusing to overlap them for the reasons that always sound sensible at the time. Instrumental’s delay analysis is blunt about the cost: kicking off mass-production tooling before EVT reliability results confirm the design, and finding a major issue during reliability testing, buys you six to nine weeks of extra build time plus two to four weeks of tool revision. Simply building continuously instead of stopping to validate adds at least two weeks on its own. Even a mismatch between incoming and outgoing quality checks at the factory costs four days to a week.
Pebble’s second round of watches is a live illustration of what gate discipline looks like when it is expensive. Shipping had already moved from December 2025 into spring, then to May 2026. In June, Eric Migicovsky’s team stopped again: the latest pre-production sample had a cosmetic mark near the lugs, caused by a CNC milling step close to where the strap meets the case. They chose to modify the metal-injection-moulding tooling rather than accept a visible blemish, and moved first shipments to July 2026 with all pre-orders promised by September. There was no electrical fault and no supplier failure, only a finish defect, a tool change and three more months.
Most late-programme slips look like that. Nothing dramatic, just somebody holding a sample and deciding that shipping it would be worse than being late.
The cost line moves every week
Your BOM should carry a revision date and a name against every line, and somebody should be reading it weekly rather than producing it for the investor deck.
Two habits make the difference. The first is the cost walk: every fortnight, take last month’s landed unit cost and this month’s, and account for the delta line by line. This connector moved because we changed the mating half. This price rose because the distributor’s stock ran out and we bought on the broker market. This one fell because we finally hit a 5,000 piece price break. The total matters less than the rule that no movement stays mysterious. Teams that cannot explain their own cost drift discover in PVT that the product costs 30 per cent more than the model, and by then the model has already been used to set the price.
The second is should-cost. Before you accept a quote, build your own bottom-up estimate of what the part ought to cost: material weight, cycle time, tooling amortisation, scrap, the factory’s overhead and margin. You will be wrong, and you will also be able to ask a supplier why their number sits 40 per cent above yours. That is a different conversation from asking for a discount, and it produces better answers.
The industry likes to say that 80 per cent of a product’s cost is locked in during design. Siemens repeats the figure in its writing on electronics cost control without attributing it to any study, so treat it as directional. Two specifics in the same piece are more useful: materials run to roughly 60 per cent of electronics product cost, and late-stage changes cost ten to a hundred times more than the same change made early. The expensive change is never the one you make in EVT.
Tooling is where the abstraction becomes a cheque. Chinese moulding suppliers in 2026 quote roughly $1,000 to $3,000 and 15 to 25 days for a prototype single-cavity tool, and $5,000 to $15,000 and upwards, over 25 to 45 days, for a production multi-cavity tool, with tight-tolerance parts adding five to ten days of extra T-sample iterations. Get first articles to a European dock and you are at 10 to 16 weeks for a production tool. Those figures come from a sourcing platform’s account managers rather than a measured dataset, so treat them as orders of magnitude rather than quotes. They are enough to make the point: the tooling decision is the moment your programme stops being reversible, and it deserves the long meeting nobody wants to schedule.
Running a factory you have never stood in
The remote working habits that hardware teams were forced into during COVID did not go away, and mostly that is fine. A weekly video call with the factory’s NPI engineer, a shared build-issue tracker, photographs of every failed unit, and daily line output shared as a number rather than a feeling will carry an ordinary consumer product through EVT and most of DVT.
Third-party inspection fills part of the remaining gap. Professional inspection in China runs around $250 to $300 per man-day for a standard pre-shipment check, with regional firms as low as $150 and specialist compliance auditors above $500, and an electromechanical product often needs two or three man-days rather than one. Set against a flight and a week of your own time, booking an inspector for a during-production check is the cheapest insurance in the programme. Most first-time founders skip it until the first shipment arrives wrong.
You still fly out for three things: choosing the factory in the first place, the PVT build, and any problem that has survived two video calls. Being physically present when the line runs tells you things the camera never will, including which stations the operators are struggling with and whether your fixtures survive contact with a real shift. Eric Migicovsky, restarting Pebble a decade after the first one, described walking multiple ODMs that now build smartwatches in mass production as an experience he simply could not have had remotely on round one, when nobody made them at all.
For European founders the practical friction has dropped. China’s unilateral visa exemption for Dutch ordinary passport holders covers business visits of up to 30 days and has been extended to 31 December 2026. Book the trip around the build, not around the calendar quarter.
The calendar belongs to somebody else
Two dates dominate every schedule that touches China, and neither of them cares about your funding runway.
Chinese New Year falls on 6 February 2027, with the official public holiday running 5 to 12 February. The real disruption is four to six weeks. Output tapers from about 80 per cent two weeks out to under 20 per cent on the eve, factories close for the holiday plus three to five days, and recovery runs at 40 to 60 per cent capacity in the first week back with quality problems attached. Between 10 and 20 per cent of workers do not return at all, which means the people who ran your PVT line in January may not be the people running it in March. If you want stock in a European warehouse by the end of February 2027, production has to finish by around 10 January and start in mid-November 2026.
Golden Week is the smaller sibling and still eats a fortnight of slack. In 2026 the National Day holiday runs 1 to 7 October, with 20 September and 10 October worked as make-up days, and Mid-Autumn sits separately on 25 to 27 September. The practical working window before the shutdown is 28 to 30 September, and ocean freight books out four to six weeks ahead of it.
Buffer discipline follows from this. Put the buffer at the end of a phase where it is visible and defensible, not sprinkled inside individual tasks where it evaporates into everyone’s private optimism. Name it as buffer in the plan. Report it as a number that goes down, because a schedule where the buffer has silently reached zero is a schedule that is already late and has not told you yet.
Spreadsheet, or something with a login
Most startups should start in a spreadsheet and stop using one at a specific, identifiable moment: the first time you hand a BOM to a contract manufacturer and cannot say with certainty which revision they built from.
Once you cross that line the options in 2026 are narrower than they look. OpenBOM restructured its pricing for 2026 to $30 per seat per month on the Team tier and $90 on Company, billed annually, with CAD add-ins at $25 per seat and 2,000 data records included before you pay for scale. Arena remains the grown-up option at grown-up money, quoted in comparisons at roughly $1,200 to $2,500 per user per year. Duro, which was the obvious startup answer for years, was acquired by Altium at the end of 2025 and now routes through a sales conversation rather than self-serve pricing.
There is a lesson in that acquisition. Startup-friendly PLM tools get bought by larger EDA companies with enterprise pricing models, so choose one you could leave. Ask whether you can export every BOM, revision and approval record as CSV without asking permission first. If the answer is unclear, that is your answer.
One person, and the things they should hand away
At seed stage the programme manager is one person doing it alongside three other jobs, and the failure mode is always the same: they keep the work that feels urgent and outsource the work that feels boring, when it should be the reverse.
Keep the BOM, the schedule and the gate decisions. Those three are the programme. They cannot be delegated without delegating the product, because every one of them is a judgement about what you are willing to accept.
Hand away inspection, customs and freight brokerage, test-lab scheduling, and the grind of chasing component lead times across ten distributors. All of it is available on a day rate, and none of it requires knowing why the product exists.
The middle case is factory selection and the first NPI cycle, where a fractional programme manager or a development partner who has run the loop before earns their fee by knowing which questions to ask a factory before you sign rather than after. We do this work at RMBG, and the honest version of the pitch is that the value is mostly in the mistakes we have already paid for.
What AI does for a programme, and what it does not
Founders reach for AI to cover the gap, and it covers less than the marketing suggests. Gartner’s widely quoted forecast that AI will handle 80 per cent of routine project management tasks by 2030 is a prediction, not a finding, and the professional literature around it is mostly aspiration. The Association for Project Management’s own 2026 trends piece leads with a trust deficit, citing a consultancy that had to repay part of its fees to the Australian government after AI-generated errors turned up in a delivered report.
What does help is unglamorous. Language models are genuinely good at translating a factory’s Chinese-language DFM report at two in the morning, compressing a night shift’s build notes into the five things you need to read, and drafting the deviation request you have written forty times before. Where machine learning earns real money is on build data rather than on your Gantt chart: Instrumental publishes figures of over 900 engineering weeks saved at Meta across seven annual programmes, up to 14 days off final system builds at NVIDIA, and 5.3 times return at Toast. Those are vendor numbers at a scale you do not have yet, but they point at the right thing. The value is in the images and measurements coming off the line, not in the summary of your meeting.
If you are planning a 2027 launch, the date to write on the wall today is 10 January 2027. That is roughly when production has to be finished for stock to reach a European warehouse before Chinese New Year closes the country, and every gate in your programme has to fit behind it.
Sources
- Instrumental: Hardware engineers get more dates
- Instrumental: Hardware schedules are driven by iteration
- Instrumental: Why hardware products get delayed, EVT/DVT/PVT decoded part 2
- Instrumental: Manufacturing Acceleration Platform, published customer results
- Titoma: EVT, DVT and PVT hardware testing phases explained
- Gadgets & Wearables: Pebble Round 2 delay pushes shipping into July
- Android Authority: Pebble Round 2 shipment delay
- Siemens: The 80% rule, why electronics cost control starts in design
- Haizol: China injection moulding industry report 2026
- Minden Sourcing: China quality inspection service cost 2026
- Embassy of China in the Netherlands: notice on extending the unilateral visa exemption policy
- New Buying Agent: Chinese New Year 2027 production schedule survival guide
- SEKO Logistics: China Golden Week 2026 supply chain guide
- OpenBOM: 2026 pricing model
- oroForge: Duro PLM alternatives after Altium acquired Duro
- Duro Labs: company page confirming Altium acquisition
- Association for Project Management: five AI trends for 2026
- Zalando PMO: AI in project management
- Hardware FYI: 120 trips to the world’s factories