Founders are paying thousands of dollars to walk through Chinese factories. That is the headline fact from a Reuters report dated September 3. US startup founders are roaming Shenzhen's hardware markets to source components. European executives are reportedly booking the same tours out of fear of falling behind in AI and robotics.
The GPU is no longer the scarce thing for most of these founders. Cloud rental for an Nvidia H100 fell from an early 2024 hyperscaler peak of roughly $8 to $9 an hour to as low as $1.8012 an hour by Q2 2026, according to a July 2026 venture analysis. Compute got cheap. Cameras, batteries, enclosures, and a factory that turns a sample around in days did not.
Damaging admission first. Nobody has a hard count of these pilgrims. That is a trend, not a census, but the direction points to a real change in where AI hardware moats live.
The Iteration Radius
AI hardware moats used to be argued in one dimension: who has the best model. That is the wrong dimension for a device company. Call the right one the Iteration Radius, the number of days between a design change and a physical sample in your hand. The smaller the radius, the more shots you take before the money runs out.
Four numbers that explain why founders fly instead of renting more compute.
Scarcity in AI hardware now lives in three layers, and each behaves differently. Layer one is the foundry. TSMC's CoWoS packaging is "very tight and remains sold out through 2025 and into 2026," in the words of CEO C.C. Wei. SK Hynix and Micron have confirmed their HBM3E capacity is fully booked through 2026.
You do not get into that line by flying to Shenzhen. You get in by being Nvidia. The contrarian analysts are right that the four largest AI chip designers consumed roughly 90%20 of global CoWoS and HBM capacity by value in 2025, according to Epoch AI estimates. That layer is closed to startups, wherever they sit.
Layer two is compute, and it is loosening. More than 30021 new cloud providers reportedly entered the market in 2025, per the same venture analysis. Cast AI's 2026 telemetry across 23,00022 production clusters puts average GPU utilization at 5%. The chips people already own are sitting idle.
Layer three is assembly. Shenzhen's Huaqiangbei district and the Pearl River Delta supply chain let a team go from prototype to sample in days. US hardware teams sourcing overseas measure the same loop in weeks or months. A 2025 XIN Summit recap counts more than 2,000 AI companies in Shenzhen and reports a robotics cluster of roughly 74,00024 firms.
The founders on the tours are betting on layer three. For smart glasses, wearable recorders, robot vacuums, and humanoid components, that bet is correct. AI hardware is the fastest-growing category of consumer AI products. In that category, bill-of-materials cost and iteration speed beat model quality.
Weights Copy Fast, Factories Compound Slowly
A model weight is a file. A factory is a relationship. Files copy in minutes. Relationships compound over decades, and that asymmetry is why founders are flying east.
The tour works as a lesson in shoshin, beginner's mind. Western founders arrive believing hardware is software with a shipping delay. They leave understanding that the enclosure tolerance and the camera module lead time decide the product more than the model does. The model is a commodity input you can rent by the hour.
Two companies already proved this. DJI built its drone business from Shenzhen with its supply chain inside driving distance. Apple owns none of the factories that build its phones, yet its Guangdong supply chain took roughly a decade to assemble. Neither moat was a chip. Both were proximity turned into trust.
Now hold that against US export policy. Washington's chip controls were designed to slow Chinese AI compute, and Nvidia now ships China-specific accelerator variants as a result. The dependency running the other way remains largely unaddressed. Western AI device startups need Chinese contract manufacturing, batteries, cameras, and enclosures.
Put it plainly. Washington restricted the chip. It did not restrict the chip's body. Every founder walking Huaqiangbei is exploiting that gap.
Impermanence cuts both ways here. TrendForce data shows DRAM supplier inventories fell from 13 to 17 weeks in late 2024 to 2 to 4 weeks by October 2025. Prices in some memory segments more than doubled between February and October 2025. The founder embedded in Shenzhen re-specced around what sat on the shelf that week. The founder in Austin waited for a revised quote.
The 70% rule applies. You will never have full information about a supplier, a part, or a tariff. Decide at 70% confidence, fly, and correct on the factory floor. Waiting for certainty in hardware is how you spend two years and ship nothing.
Manufacturing veterans push back, and they should. Shenzhen incubators and contract manufacturers say sustained presence and guanxi, the local relationship network, drive real advantage. They see two-week tours as tourism. My read on this: they are right about the moat and wrong about the ticket. The tour is the entry fee. The moat belongs to whoever stays.
The bottleneck moved down the stack
The foundry tier is not open to startups.
Epoch AI estimates the four largest AI chip designers consumed roughly 90% of global CoWoS and HBM capacity by value in 2025. SK Hynix and Micron have confirmed HBM3E capacity is booked through 2026. No flight to Shenzhen changes that queue.
The chips people already own are idle.
Cast AI's 2026 telemetry across 23,00022 production clusters puts average GPU utilization at 5%, and more than 30021 new cloud providers reportedly entered the market in 2025. Compute is a rentable commodity input, which is exactly why it stopped being a moat.
Downstream controls would flip proximity into liability.
Washington restricted the chip, not the chip's body, and every founder walking Huaqiangbei is exploiting that gap. If CHIPS Act logic extends to contract manufacturing for consumer devices, China proximity becomes an asymmetric liability overnight. A barbell with a second qualified assembly line hedges both futures.
2031. Pull back five years. China has stated its goal of global AI leadership by 2030, as documented in RAND's June 2025 "Full Stack" report. That timeline overlaps almost exactly with the product cycles the tech pilgrims are planning right now.
Two futures split from here. In the first, export controls stay focused on advanced compute. Western device startups keep iterating in the Pearl River Delta and ship consumer AI hardware at BOM costs their domestic rivals cannot touch. Only cash is real, and the cash sits with whoever has the lowest cost per unit and the fastest fix.
In the second, controls creep downstream. The US CHIPS Act and parallel European and Japanese programs already aim to reduce dependence on Chinese manufacturing for critical infrastructure. It is unclear whether Washington extends that logic to contract manufacturing for consumer devices. If it does, China proximity flips from asymmetric advantage to asymmetric liability overnight.
Enterprise and defense buyers are already hedging. Many decline hardware with heavy China exposure in critical components. A founder selling smart glasses to consumers can ignore that. A founder selling edge inference boxes to a utility cannot.
I think the durable position is a barbell, and it amounts to counterpositioning against both camps. Iterate in Shenzhen where the Iteration Radius is measured in days. Qualify a second assembly line outside China for the customers who require it. Domestic-only rivals cannot match your speed, and China-only rivals cannot pass your buyers' compliance review.
The contrast pair for the decade: GPUs buy a demo. Supply chains buy a company. The founders who understood that in 2026 will own categories in 2031. The founders still optimizing model quality will be licensing someone else's weights and shipping someone else's enclosure.
Map Your BOM Before You Book Flights
You do not need a factory visa this weekend. You need a spreadsheet and three small experiments. Here is the order.
First, write out your bill of materials. A BOM is the full list of every physical part in your device, down to the screws. Tag each line by where it is manufactured. If more than half the lines trace to the Pearl River Delta, your Iteration Radius is currently measured in flights, not days.
Second, read before you fly. Point Iris.ai Researcher Workspace at the export control filings and supply-chain analyses instead of a search bar. It analyzes and summarizes research documents in addition to returning content-based document lists. Feed it the September 3 Reuters piece and the 2026 packaging reports and ask it where your parts sit in the three layers.
Third, build the software half now. Rocket v1.0 generates one-prompt apps with one-click GitHub sync and Netlify deploy built in, with a free tier and paid plans from $2513 a month. Ship the companion app for your device before the device exists. Show, don't tell: a working app makes supplier conversations concrete.
Fourth, track suppliers with guardrails. Google has announced that Gemini for Google Workspace will shift from suggesting edits to editing your Google Docs based on comment feedback. Set write limits before you connect it to anything. Let it read your supplier sheets and draft comparisons. Do not let it send emails to factories.
Fifth, get your reps in. Order the same sample from three Huaqiangbei vendors through an intermediary. Two will disappoint you on tolerance, lead time, or communication. That is not failure, that is the data you flew to collect, and you collected it without leaving your desk.
None of this requires a CS degree or a mechanical engineering background. It requires a list, a few tools, and a willingness to look foolish for a month. Take a deep breath and take it step by step. The pilgrims paying thousands of dollars started with the same spreadsheet.
Map your bill of materials before you book a flight.
- Tag every BOM line by origin. Write out every physical part in your device down to the screws, then mark where each one is manufactured. If more than half the lines trace to the Pearl River Delta, your Iteration Radius is measured in flights, not days.
- Read the filings before you fly. Point Iris.ai Researcher Workspace at the September 3 Reuters piece, the 2026 packaging reports and the export control filings, then ask where your parts sit across the foundry, compute and assembly layers.
- Ship the companion app first. Rocket v1.0 generates one-prompt apps with GitHub sync and Netlify deploy on a free tier and paid plans from $2513 a month. Then order the same sample from three Huaqiangbei vendors through an intermediary and compare tolerance, lead time and communication.
The moat is proximity turned into trust, and it does not copy in minutes.
A model weight is a file and a factory is a relationship, which is why DJI built inside driving distance of its suppliers and why Apple's Guangdong supply chain took roughly a decade to assemble. Compute at $1.8012 an hour is available to anyone, and 5% average utilization proves nobody is short of it. What remains scarce is the number of days between a design change and a sample in your hand, and the founders shortening that number are doing it in Shenzhen. The manufacturing veterans who dismiss two-week tours as tourism are right about the moat and wrong about the ticket: the tour is the entry fee, and the advantage belongs to whoever stays while also qualifying a second line for buyers who require one.