Earlier this year, the OPC startup model sparked a wave of excitement. But from a primary-market perspective, OPC doesn't fit the traditional VC/PE investment logic: it lowers the barrier to starting a business, but it accounts for individual employment, not institutional capital. For the underlying logic, see this earlier article from投中网 《“OPC里的企业基本没有投资价值”｜Frontline》.
Recently, I heard a partner at a leading market-oriented investment firm mention that OPC is just an experiment, not necessarily a trend. What he's more bullish on is FDE, which he believes is the new economic model and will appear at scale.
FDE, Forward Deployed Engineer, is not a brand-new concept that appeared out of nowhere. The role was first scaled up by Palantir in the US, whose core playbook was to station engineers on-site at government and large enterprise clients, embedding standardized data analytics products into the clients' existing systems and workflows. In today's AI context, what FDEs accomplish is the "last mile" of AI.
"The large model companies have already made it. DeepSeek, Kimi, Zhipu, ByteDance and other large model companies solve the 'height' problem, i.e., model capability itself. What's truly lacking is 'breadth'—pushing AI into thousands of industries and landing it in the real economy. That kind of deployment requires a team, and that team is FDE," he said.
Unlike OPC, where one person does everything, an FDE team generally needs three or more people, essentially a combination of three capabilities: business-savvy consulting, delivery that can be implemented, and engineering that can tune models. Tools are built from scratch on the spot, and every version can be rewritten. Small teams solve big problems, and the economics work out.
He has personally invested in two companies, and is also persuading many consultants, software engineers, and coders at traditional consulting firms impacted by AI to form teams and do FDE—he's willing to invest.
Unlike OPC's grassroots popularity, FDE became an independent track back in 2016, and has even spawned this year's hottest high-paying job. Large model companies and big tech firms have been adding FDE positions this year, with a million-yuan annual salary as the baseline.
Large model companies are snapping up FDEs with million-yuan salaries
A role that was extremely niche in 2023 is becoming the most expensive headcount of the AI era.
First, the salaries. On BOSS直聘 and Liepin, Doubao AI large model FDEs earn 35,000 to 70,000 yuan per month with 15 months' pay, annual packages up to 1.05 million yuan. Ant Digital's B2B FDEs earn 40,000 to 60,000 yuan per month with 15 months' pay, annual packages up to 900,000 yuan. Zhipu's FDE lead earns 60,000 to 80,000 yuan per month, 960,000 yuan annually. Moonshot's Kimi enterprise FDEs earn 40,000 to 65,000 yuan per month with 14 months' pay, annual packages up to 1.09 million yuan. Huawei Cloud Pangu large model FDEs earn 35,000 to 60,000 yuan per month with 14 months' pay. Alibaba Cloud large model FDEs earn 20,000 to 55,000 yuan per month with 16 months' pay.
A million-yuan annual salary is the baseline. Overseas, the bar is even higher. OpenAI's FDE job posting on its official site offers $162,000 to $280,000 a year—about 1.1 to 1.9 million yuan—plus equity. Anthropic offers $200,000 to $300,000, about 1.36 to 2.03 million yuan. One headhunter offered a $400,000 salary plus fully remote work to poach an FDE with only two years of experience. Silicon Valley practitioners say compensation for this role has no ceiling, and a $500,000 total package in a year is just the starting point.
Now the growth rate. LinkedIn's global labor market trends insight report released in January 2026 showed that from 2023 to 2025, the number of new FDE positions grew 42-fold, while AI engineer positions grew 13-fold over the same period, with at least 1.3 million new AI-related jobs added globally in two years. The Wall Street Journal directly called it the hottest job in tech.
In the domestic market, Liepin's Big Data Research Institute's FDE talent supply-demand report released in September 2026 stated: the average annual salary for domestic FDE positions is 408,000 yuan, the talent supply-demand ratio is 0.28, and every 3.5 openings compete for 1 person.
This means FDEs are among the scarcest species in today's talent market.
Money has chased into this track too, and it has already produced a listed company. 海致科技 listed on the Hong Kong Stock Exchange on February 13, 2026, surging 250% on its first day, becoming China's first FDE-concept listed company.
The primary-market list includes, but is not limited to, the following companies:
中数睿智, one of the highest-valued FDE-concept companies in the current primary market. It completed 3 rounds of financing intensively in 2026, with 5 rounds in total, and CDH VGC invested in three of them.
词元无限, founded in 2025, is the fastest-moving in terms of fundraising pace. It completed a nine-figure (several hundred million yuan) angel round in 29 days. Its direction is clear: All in B2B Coding Agent, delivering runnable, verifiable, auditable enterprise-grade results via the FDE model.
素源矩阵 is even more "extreme." Its core team is only 7 people, serving large state-owned enterprise clients, with individual projects ranging from hundreds of thousands to a million yuan, and payment can be based on results. It completed its angel round in April 2026, with 奇绩创坛 as the investor. Tiny team, high customer value, pay-for-results—its very existence is a validation of the FDE organizational form.
Abroad, it started even earlier.
The first mover was Anthropic, which partnered with Accenture in December 2025 to train about 30,000 employees to use Claude. The real explosion came in May 2026: on the 4th, Anthropic joined forces with Blackstone, H&F, and Goldman Sachs to form a $1.5 billion joint venture focused on AI deployment for PE portfolio companies; on the 11th, OpenAI founded DeployCo with an initial investment of over $4 billion, with 19 institutions participating, and also acquired the consulting firm Tomoro, bringing in 150 deployment engineers in one stroke. Two months later, Microsoft followed suit, spending $2.5 billion and dispatching 6,000 engineers into the Fortune 500.
Large model companies at home and abroad are pouring money into FDE. The core reason is that under losses and commercialization pressure, FDE can serve as the key strategic lever for them to shift from "selling models" to "selling results," achieving revenue growth and model iteration.
"Enterprises bought the API, bought the model, but can't put it to use—this delivery gap is what FDEs fill. FDEs work on-site to help enterprises connect models to their data, systems, and processes, and are accountable for post-launch results. Only when the client's business actually runs will token consumption keep rolling, and only then is revenue not one-off," investor J analyzed.
Moreover, "burning tokens" is only the first layer of FDE's business logic; what's more critical is the knowledge backflow mechanism. When FDEs help enterprises deploy AI agents on-site, they organize the business processes, domain knowledge, and operational norms originally hidden in employees' heads into Prompts, Workflows, Memory, and Skills, extracting them in a structured way. This enterprise-grade domain knowledge is then internalized into the next generation of models through fine-tuning, reinforcement learning, or as training corpus.
This also means the more FDEs do, the better the models understand the industry, the less customization the next generation of models needs, and the higher the delivery efficiency. In addition, every FDE on-site engagement is collecting training data for the next generation of models that money can't buy. This feedback loop cannot be replicated by competitors. "Once a model is fine-tuned on a client's private edge cases and embedded in their daily decision-making processes, switching away would be a painful upheaval for the client. The industry calls this data gravity. The deeper the FDE embeds, the stronger the gravity," J said.
"Traditional software companies are all old-timers now—they'll be wiped out by FDEs"
The above is the logic of large model vendors. For market-oriented investment institutions, the math is different. Back to the investor at the beginning: when he invested in FDE, he wasn't investing in a new track, but in an organizational form: the new-generation software industry.
"I believe that companies like the early ChinaSoft, Neusoft, Yonyou, and Kingdee are all old-timers now, and they'll all be wiped out by these FDEs, because the tools FDEs have at hand are too powerful."
Behind it is still the math. Traditional software companies pile headcount onto projects, with long delivery cycles, and the numbers don't add up. It used to take 10 people 3 months for one project; now 3 people do it in 2 weeks. The unit economics have changed—FDE companies can make money, traditional software companies cannot.
There's also a layer of hard necessity. Trillion-parameter open-source models are simply impossible for enterprises to deploy on their own. The latest generation of open-source models runs from one trillion to two trillion parameters per team per training run, and very few teams in the market can deploy models at that scale. This requires FDE companies to act as the construction crew, design institute, and supervisor—helping enterprises design model networks, deploy and run them, and maintain them continuously. This isn't optional; it's a necessity.
He believes China will see a wave of enterprise AI application companies, with FDE as a key link among them. He draws an analogy with the semiconductor industry chain. "Upstream, think of the large model companies as TSMC or NVIDIA. Further down, Foxconn and Luxshare-type companies will inevitably emerge—a batch of token factories. Large model companies make the chips; FDE companies do the assembly, taking on model capabilities and completing the 'last mile' of delivery."
By this analogy, the ceiling for FDE companies isn't low. Foxconn became the world's largest contract manufacturer, and an FDE cluster doing business across all of China and the world is logically sound.
Where would the cluster land? His answer is Haidian. Haidian has the densest concentration of software engineers in China, and as AI coding disrupts traditional coding, this group shifting to FDE is a proactive adjustment. His suggestion is for Haidian to form several FDE clusters that go global together: not individual companies fighting alone, but consortiums of large model companies plus FDE companies, doing business across all of China together, and even the whole world.
His vision of the endgame is even bigger: the largest FDE company would resemble an insurance company. Everyone has their own toolbox and models, and one or two people can start a company. But three-person teams aren't a return to solo operators—these small teams still need a platform behind them doing four things: training and enablement, unified negotiation, in-house product development, and interfacing with large model companies. The essence is a big platform plus small front ends: the platform provides models, tools, and brand, while the small front ends deliver on-site at customer locations.
Another investor told me, "FDE isn't just worth investing in by VCs—it's even worth being owned by VCs, becoming part of their pre-investment and post-investment capabilities."
Pre-investment, an FDE team entering the picture is like external experts doing due diligence, getting a deeper read on a target company's true substance than financial statements can. Post-investment, they provide AI enablement to portfolio companies—especially in industries like retail that compete on digitalization and supply chains, this becomes the firm's differentiated service.
Of course, dissenting voices have emerged. Many people believe this is still a headcount-based business model—not sexy enough as a business, and falling short of the ideal.
Eric Young, manager of the FDE team at Stardust Haoyu, an enterprise AI company focused on AI Workers, told me a vivid case. They once served a major manufacturing client whose initial request was a single sentence: automate after-sales customer service. After repeated research on-site, what they ultimately built wasn't an AI customer service bot, but a copilot for after-sales engineers—helping engineers quickly gather information and suggest solutions, with humans making the final confirmation.
For them, this counted as a mid-sized project, taking over two months end to end, with a contract value in the high hundreds of thousands of RMB, and a model somewhere between standardized SaaS and pure customization. The hardest part of this project was that what the client says they need is often not the same as what they actually need.
They've also stumbled on delivery standards. In the past, their contracts would specify delivering one agent or two agents. But such clauses are actually very dangerous—if the software is built and the client doesn't use it, no value is generated, and acceptance and payment both become problems. Now they break the outcome into three parts: what is delivered technically, what changes in the business process, and what metrics are ultimately produced.
Being accountable for results also requires a big prerequisite: the client has to cooperate. If they don't provide data, don't provide APIs, and the organization and employees are unwilling to change processes, no amount of FDEs can produce results. "The biggest challenge in AI implementation isn't technology—it's people, organizations, and business workflows. If even one link in implementability or business feasibility is missing, the faster you move, the faster you die. That's why upfront risk control is needed," Eric said.
He also emphasized that FDE is a means, not an end. With every project completed, the capabilities on the client side should be abstracted back into the platform, becoming components reusable across industries. The less FDE work there is, the more standardized the products become, and the more clients can use them out of the box—that's the endgame of the FDE model. "Without accumulation, you just become a fairly traditional high-end outsourcing team."