China's FDE bottleneck is not the engineers

The numbers usually quoted about China: about 60 million enterprises, a community of some 30,000 members, a few thousand people who are genuinely strong with AI, and on the order of a thousand who are actually willing to take on delivery work. All four come from a community author, self-reported, and should be read as an order-of-magnitude argument rather than a census.

The last figure is the one that matters. Awareness is not the constraint: surveys in the same briefs show 88% of organisations using AI regularly against 39% reporting enterprise-level EBIT impact (McKinsey 2025), and 5% of companies reaching AI scale value against 60% seeing almost no substantive value (BCG 2025). The constraint is that almost nobody will take the last mile as a job.

Six job families, no single job title

The whitepaper maps the current supply to six observable families:

• AI solution architect
• Large-model application engineer
• Agent engineer
• AI implementation consultant
• Digital transformation lead
• AI product solution specialist

Its methodological point is more useful than the list: searching one title answers nothing. City, industry, salary, keywords, customer-site work, systems integration and AI-tool requirements have to sit on one map before any conclusion is possible. The whitepaper's verdict is modest: demand is surfacing across several roles, but titles, salary bands and boundaries still need observation.

That modesty should be preserved. Any claim that China now has a settled FDE job category is ahead of the evidence.

Telling a real role from a renamed one

Because the category is unsettled, the practical question is not "does the title exist" but "is this role actually forward-deployed". Five signals separate the two, all checkable from a posting plus one interview:

1. The posting names customer-site work, not only office or remote work.
2. It names system integration — existing systems, existing data, existing permissions.
3. It names acceptance: who signs that the thing works, and against what.
4. The salary band is published and roughly stable across postings for the same title.
5. The responsibility boundary is distinct from presales: the role is measured on delivered outcomes, not on deal registration.

A role that scores on site presence and technical stack but not on acceptance is a solutions engineer with a new label. A role that scores on acceptance but not on site presence is a product engineer with a travel budget.

The pricing structure that makes small orders uneconomic

The most concrete market finding in the Chinese material is a band, not an average: community authors put the workable range for a small team at RMB 50,000 to 200,000 per order. The reasoning is what makes it useful.

• At RMB 300,000 an order drags a small team for one to two months, with little visible progress and uncertain final payment.
• Above RMB 200,000, work tends to enter government or state-owned-enterprise procurement, which a small team can neither run nor control for collection.
• Final-payment non-collection is estimated, by the same author's observation, at possibly above 60%.

Set against that, the buyer's alternatives are cheap and legible: an enterprise RPA licence at RMB 59,800 a year, corporate training at RMB 30,000–80,000 per session, a retainer at RMB 100,000–200,000 a year. A bespoke system at the bottom of the band has to beat a licence that costs about the same and arrives without a project.

The large money exists, but it is a different game: a RMB 61.275 million industry-model procurement win is recorded in the whitepaper, and no two-person team is bidding for it. The middle of the market — the RMB 300,000 transformation project with a fixed price and a back-loaded payment schedule — is where a good order becomes a cash trap. Avoiding it is a scoping question before a price question, and the four-week rule is the cheapest test.

Salary points the same way. A community salon dataset puts the median package at RMB 366,000, with 28.25% above RMB 500,000 and tier-one averages above RMB 40,000 a month; the whitepaper puts junior engineers above RMB 200,000 a year. Treat the salon figures as community-reported.

The bottleneck is platform, method and pricing

The common explanation — Chinese engineers are not good enough — does not survive contact with the material. One structural reading is that Chinese engineering culture is in fact the closest to what FDE work demands; three other things are missing: a platform base, a methodology, and a pricing structure.

The platform point is the most expensive. Platforms come from repeated field delivery: Foundry is reported to contribute more than half of Palantir's revenue and to have been born from it. Where every delivery starts from zero, margin stays at the level of labour — which is why the community's advice is to move from selling hours to reusable assets, the conclusion scoping and pricing reaches from the vendor side.

The supply is thinner than the counts suggest

One community author built a reserve of 500+ FDEs, reviewed more than 60 real projects, visited several firms that were actually making money, then published his decision to leave FDE work. His conclusion: a large share of enterprise demand can be met with general-purpose office agents, and deep data governance is not a business a new team should take on.

This is the necessary counterweight to the 60-million figure. Five hundred people in a reserve is not five hundred people who can deliver; sixty projects reviewed is not sixty projects that paid. Supply counts inflate faster than supply.

2027–2028: three scenarios, one lean

The whitepaper sets out three scenarios rather than a forecast.

• Cautious diffusion. Enterprises treat AI mainly as an assistant, few production actions move. FDE exists implicitly inside solution and implementation roles.
• Steady productionisation. Agents enter sales, customer service, knowledge base and ticketing. FDE becomes a core role in AI service companies.
• Deep organisation. Multi-department agent systems force process redesign. FDE becomes a key job family both inside enterprises and in external service firms.

Its judgement is that China enters steady productionisation first, in 2027–2028. The observable signals to watch are narrow: whether agent deployments reach sales, support, knowledge base and ticketing as production systems rather than demos, and whether pricing moves from per-project to per-outcome.

Ending criterion for anyone entering now: ask which of the three missing pieces — platform, method, pricing structure — your work actually reduces. If the answer is none, the 60-million-firm market is a statistic about other people.