A founder walked us through a working prototype three weeks after his first sketch. Custom PCB, ESP32 module, BLE and Wi-Fi, printed housing, an app on his phone pulling live sensor data. Then he asked the question everyone asks: so why does everyone tell me this takes eighteen months?
He had done the hard part of 2013 in three weeks, on a laptop, for a few hundred euros. What he had not touched was any of the parts that are harder now than they were then.
Hardware is still hard, just not in the same places. The difficulty has moved off the workbench and into customs brokerage, memory allocation contracts, and the eleven months between paying a factory and being paid by a customer. Budget your risk against the old map and you will be surprised in the expensive direction.
Prototyping stopped being the bottleneck
JLCPCB advertises bare boards from $2 on 1–8 layer stackups with 24-hour production turnaround and global delivery in as little as 1–4 days. Their assembly service starts at $8 setup plus $0.0016 per solder joint, will build from two boards, and completes about 90% of assembly orders within 24 hours. A populated, tested, four-layer prototype for less than a nice dinner in Amsterdam, in your hands inside a week.
Radio certification, once a real obstacle, is now a line item. Using a pre-certified module such as an ESP32 or nRF52840 typically cuts certification cost by 60–80% and roughly halves the timeline, because the module vendor’s grant already covers intentional-radiator testing. Design your own RF front end instead and you buy that testing back at roughly two to three times the price and twice the calendar. It is the cheapest decision on the whole board, and you should almost always take it.
Soft tooling followed the same curve. Aluminium prototype tooling runs roughly $1,500–$8,000 with 2–4 week lead times, so a mechanical design can now go through two or three real injection-moulded iterations for less than a single steel tool used to cost.
AI has helped, though less than the marketing suggests and in unglamorous places. The honest picture from engineers is that language models earn their keep turning messy requirements into register maps and interface definitions, feeding synthesis and simulation errors back in to shorten debug loops, and generating testbenches and verification stimulus. That is the tedium nobody was doing thoroughly anyway. They are unreliable on timing-critical datapaths, deeply optimised pipelines and corner-case logic, where they produce plausible and wrong with total confidence. Nobody publishing on this has credible productivity numbers. Treat it as a good junior engineer with no fear and no memory.
Add it up and the first working version of a connected product is a weeks-and-thousands problem rather than a months-and-hundreds-of-thousands problem. That should make you more ambitious about what you try, but not more optimistic about your ship date.
Your BOM is now a trade policy document
On 20 February 2026 the US Supreme Court held that IEEPA’s power to “regulate … importation” does not include the power to impose tariffs, striking down both the trafficking tariffs on Canada, Mexico and China and the global reciprocal tariffs. Duties collected under those programmes became theoretically refundable, and the refund process is now a messy live litigation problem.
Five days later the administration replaced them with a 10% surcharge under Section 122 of the Trade Act, an authority statutorily capped at 15% and limited to 150 days. The Court of International Trade struck that down on 7 May 2026, and the tariff expired on schedule on 24 July anyway.
On that same date, the USTR replaced it with something far more durable: Section 301 tariffs against 60 economies found to have failed to enforce prohibitions on goods made with forced labour, covering over 99% of US imports. Seventeen partners, including Canada, Mexico, India, Malaysia, Indonesia and the UK, got 10%. Thirty-eight others, including China, Vietnam, Thailand, Turkey and Brazil, got 12.5%. The EU and Taiwan are charged net of MFN to reach 10% combined; Japan, Korea and Switzerland net to 12.5%. For China, that 12.5% stacks on the existing 25% Section 301 duty, bringing the aggregate Section 301 rate on affected Chinese goods to 37.5%.
Two legal theories were struck down inside six months, and the replacement is broader than what it replaced. If you sourced in early 2025 on the assumption that moving assembly to Vietnam would keep your landed cost flat, Vietnam is now at 12.5% and your arbitrage is 2.5 points wide instead of forty. The country-shopping game that dominated 2024 and 2025 planning has largely closed, because the current instrument is deliberately near-universal.
There is one piece of good news in the noise: the Section 232 semiconductor action that everyone feared landed narrowly. The 25% duty effective 15 January 2026 targets specific high-performance logic parts defined by processing-performance and DRAM-bandwidth thresholds (H200 and MI325X-class accelerators), not your STM32 or your buck converter. The chip tariff that would have wrecked embedded BOMs did not happen.
The cheap-parcel era ended on both sides of the Atlantic
Both major markets closed their low-value import loopholes within a year of each other, and it broke a fulfilment model a lot of hardware companies depended on. The US suspended duty-free de minimis treatment on 29 August 2025, and on 24 June 2026 CBP issued an interim final rule indefinitely suspending the $800 exemption for merchandise arriving by every mode except the international postal network. Shipments that previously cleared on a manifest now require informal entry at $2,500 or less, formal entry above that, and full classification, valuation and recipient data in every case.
The EU moved on 1 July 2026, replacing its €150 duty-free threshold with a flat €3 charge, levied per tariff classification rather than per parcel, so a shipment with three distinct product types incurs three charges. It runs until July 2028, when the EU Customs Data Hub goes live and ordinary duties resume; mandatory product identifier declaration and a further handling fee arrive in November 2026. The scale explains the politics: the Commission counted 5.9 billion low-value items entering duty-free in 2025, over 90% of them from China.
For a hardware startup this is an operations problem. Shipping engineering samples from your contract manufacturer to your desk is now a customs event with paperwork and a broker. Direct-from-factory fulfilment to end customers, which flattered a lot of early margin models, now carries per-shipment duty and clearance overhead. Plan on consolidated freight into a bonded warehouse in your own market, and price the customs work as a real function.
The shortage moved from microcontrollers to memory
If your mental model of component risk is still 2021, update it. Sourceability’s Q2 2026 reporting describes logic devices, embedded processors, sensors and optoelectronics as showing balanced fundamentals. Microcontrollers, the villain of the last cycle, are broadly fine.
Memory is not. It is allocation-only. Lead times on the most sought-after AI-adjacent parts reached roughly 40 weeks, a 67% month-over-month jump; Micron quoted 52-week lead times on DRAM and DDR4 under allocation; polymer capacitors stretched to 50 weeks at some suppliers; AI- and automotive-focused passive prices climbed 15–35% during 2026.
The price data is worse. TrendForce revised its Q1 2026 conventional DRAM contract price forecast upward from 55–60% to 90–95% quarter-over-quarter, a record. J.P. Morgan Global Research puts the total DRAM rise at more than 400% from the start of 2024 through the end of 2026, with the US import price index for computers, peripherals and parts up 37% since end-2024. Micron exited its Crucial consumer business in December 2025 to concentrate on large strategic customers, a signal about who gets served when supply is short, and it is not you.
The design consequence belongs in your architecture review, not your procurement meeting. Specify the smallest memory footprint your product can really live with, then hold that line in firmware. Choose parts with second sources on the same footprint. Assume anything with DDR in the name has a twelve-month horizon and a price you cannot forecast. An architecture that assumed 512 MB because it was nearly free in 2023 now carries a BOM risk larger than its entire mechanical cost.
China+1 is real, and it is mostly China+0.5
The diversification story is true at the top of the market and mostly irrelevant at the bottom.
India assembled about 25% of global iPhone output in 2025 (55 million units, up 53% from 36 million in 2024), exporting roughly $23 billion in devices, with Tata Electronics taking 37–40% of India’s iPhone exports, up from 13% the year before. That is an enormous industrial shift. It is also incomplete: China still produces well over 70% of the world’s iPhones, and the components, tooling and process engineering behind the Indian lines remain concentrated in China. A board assembled in Hanoi is frequently a Chinese BOM with a Vietnamese label on the box.
At startup volumes, none of this is available to you anyway. Apple can relocate a supply chain because it can guarantee a factory ten million units. You are asking someone to run 5,000, and the Indian and Vietnamese suppliers who could actually serve you are not the ones absorbing Foxconn’s overflow.
Nearshoring does eventually pencil out, but only at scale. One 2026 landed-cost analysis at 250,000 units per year puts China at $2.97 per unit against $1.63 in Mexico and $1.68 in Florida once tariffs, freight, inventory carry and quality risk are included. That is a compelling table at 250,000 units per year. At 5,000, the tooling amortisation alone inverts it: production tooling in hardened steel runs $15,000–$120,000+ with 6–12 week lead times, 8–14 weeks for complex multi-cavity moulds. Know which regime you are in before you take relocation advice written for the other one.
Nobody redesigned the money math
Everything above is solvable with attention. The cash cycle is not, and it is the reason competent hardware companies die.
You pay a tooling deposit before you have a sellable unit, a materials deposit before the line runs, the balance before the container leaves, then freight, then duty at the border, then warehousing. Only then does a retailer or distributor take possession on 60- or 90-day terms. Money leaves months before it arrives, and every unit of growth widens the gap. Sell more and you fund a bigger hole.
Then the channel takes its cut. The best-documented margin waterfalls come from packaged goods and the structure transfers directly: retailers take 22–30% gross, distributors around 13% gross on net sales, and the brand ends up with roughly 20% of the shelf price. The cost-to-shelf multiple lands near 3.3x. If your landed cost is €30, your realistic shelf price is around €99 and you see about €50 of it, before support, returns, warranty and the app you have to keep running for years.
Meanwhile the listed consumer electronics sector reports around 49% gross margin. That is an aggregate flattered by companies with brand pricing power, a services attach rate and eight-figure volume commitments. It is not available to a company buying 5,000 units against a 100,000-unit price break.
There is capital in the market: robotics startups raised $18.8 billion globally by mid-June 2026, already past 2025’s $15 billion full-year total and well past 2021’s $14.1 billion peak. But that money is chasing embodied AI and defence autonomy. It is not obviously available to fund the inventory of a well-designed €89 consumer sensor, and inventory is precisely what you need funded.
Compliance arrived with a date on it
For European teams there is one more clock, and it is the one most first-time founders discover too late.
The EU Cyber Resilience Act entered into force on 10 December 2024. Vulnerability and incident reporting obligations bite on 11 September 2026, and the full essential requirements apply from 11 December 2027, after which non-compliant products with digital elements cannot be placed on the EU market. Penalties reach €15 million or 2.5% of worldwide turnover.
Practically: a machine-generated SBOM in SPDX or CycloneDX on every firmware release, a signed and authenticated update mechanism with tested rollback, and free security updates for at least five years, or for the product’s expected lifetime if that is shorter. That last clause constrains the business model, not just the engineering. If you ship in 2027, you owe firmware engineering through 2032, and it has to be in your unit economics.
Start early and it costs less. A team that designs secure boot, update infrastructure and SBOM tooling in from the first architecture review pays a fraction of what a team retrofitting it in 2027 will pay. At RMBG it goes on the agenda before anyone draws a schematic.
What an honest 2026 schedule looks like
Prototype in weeks. Then EVT at roughly 20–50 units, DVT at 50–200, PVT at 50–500, commonly set at 5 to 10 per cent of the first production run, and mass production starting around 1,000–2,000. Plan on PVT alone taking 3–6 months.
Layer in tooling at 6–12 weeks after design freeze, certification at 3–6 weeks with a pre-certified module, a memory part on a 40-week quote, ocean freight, and a customs process that no longer waves small shipments through. In 2026, twelve to eighteen months from validated prototype to sellable inventory is a competent plan, not a pessimistic one.
Where the hard part actually lives
The founder with the three-week prototype was not wrong to be proud. He had compressed a year of 2013 into three weeks. But the schedule he was fighting was never about whether the circuit worked.
Hardware in 2026 is hard because your bill of materials is subject to trade policy that changed legal basis twice in six months, because your memory line item is on allocation to a customer more important than you, and because both of your target markets ended duty-free small parcels within a year of each other. It is hard because the channel can take half the shelf price or more, you pay in March and get paid in December, and you are signing up to maintain firmware for five years before selling a single unit.
None of that is engineering, and all of it is answerable: earlier sourcing decisions, a memory strategy in the architecture phase, landed-cost models that include duty and clearance, financing built around inventory rather than headcount, and compliance work started before the schematic. But it has to be answered by the founders, early, and it cannot be delegated to the part of the process that got cheap.
Sources
- Supreme Court Strikes Down IEEPA Tariffs — Holland & Knight
- Trump Administration Imposes 10% Section 122 Tariff — White & Case
- Section 122 Tariffs Expire, Section 301 Forced-Labor Tariffs Replace Them — Honigman
- New U.S. Tariffs Replace Expiring Section 122 Tariffs — Fennemore
- US Trade Court Strikes Down Section 122 Tariffs — Skadden
- Narrowly Targeted 25% Section 232 Tariff on Certain Advanced Semiconductors — White & Case
- Indefinite Suspension of the De Minimis Exemption — Federal Register, 24 June 2026
- EU to End De Minimis with New Customs Fee in 2026 — SupplyChainBrain
- EU ends €150 duty exemption, charging €3 per item — PPC Land
- Q2 2026 Electronics Lead Times — Sourceability
- Constraints Squeeze the Market as AI Impacts Lead Times and Prices — Sourceability
- The AI-Driven Memory Shortage: DRAM Prices, Inflation and Market Risks — J.P. Morgan Global Research
- How AI Broke the Memory Market: Inside the 2024–2026 DRAM & NAND Crunch — Octopart
- Apple iPhone production in India hits 25% of global output — TechWire Asia
- China Plus One Reality: Has the World Reached +1 or Just +0.5? — Vietnam Sourcing
- 2026 Injection Molding Pricing Report: US Costs, Reshoring & Tooling Data — Jaycon
- FCC Certification Cost (2026) — MarkReady
- PCB Prototyping — JLCPCB
- SMT PCB Assembly Service — JLCPCB
- The EU Cyber Resilience Act: New Cybersecurity Requirements for Connected Products — Pillsbury
- EU Cyber Resilience Act Forces Firmware Teams to Rethink Security Strategy — Design News
- Retail Distribution Economics: Margin After the Middlemen — Eightx
- Consumer Electronics Industry Profitability Ratios, Q1 2026 — CSIMarket
- Robotics Startups On Fire As Venture Funding Surges To Record Numbers In 2026 — Crunchbase News
- LLM-Aided Hardware Design in 2026: What Engineers Actually Trust AI With — Promwad
- Hardware product development stages: POC – EVT – DVT – PVT explained — Encata