You are using an older version of Internet Explorer. For the security of your Leiphone account and a better product experience, we strongly recommend using a faster and more secure browser.
This is a temporary link, for article preview only, and will expire at that time
0 
"Short sellers have died at least three times, and Palantir has risen more than 30-fold."
In the US stock market, there is a phenomenon-level stock—Palantir. It was persistently shorted and bet against, yet it kept surging against the trend, leaving every short seller in crushing defeat.
In just a few years, Palantir's stock trajectory has been explosive: from a historic low of under $6 at the end of 2022 and early 2023, it staged a desperate comeback, surging more than 30-fold, hitting an all-time high of $207.52 in 2025, with its market cap at one point approaching $500 billion—five times that of fellow software company Snowflake, and clearly higher than traditional software giants like Salesforce and SAP.
Even after a pullback from its highs, Palantir's market cap still exceeds $360 billion, and its stock has gained nearly 15-fold since going public.
But who would have thought that this soaring darling of capital was once a To G company widely shorted by Wall Street.
To G business means heavy delivery, slow turnover, and difficulty capitalizing—yet Palantir completely overturned market prejudice through AI commercialization: not only did it solidify its To G base, it also shone brightly in To B business, with To B revenue reaching 46% in 2025, becoming a new core growth engine.
A To G origin, a soaring stock price, a P/E ratio of over 200 times—sparking a wave of "Palantir-ization" in Silicon Valley and on Wall Street.
This craze quickly swept into the domestic market as well. Vendors like Fourth Paradigm, Mininglamp Technology, and Deepexi have all benchmarked and studied Palantir, accelerating their layouts in government-enterprise intelligence and data decision-making tracks, with leading big tech companies even entering the field themselves.
But the real challenge is just beginning: conservative technical routes, lagging data governance, the failure of high-gross-margin models, compounded by dual tests of strategy and resources. The path to building a "Chinese Palantir" is far harder than imagined.
Recently, Leiphone interviewed investors and analysts who have long followed Palantir, as well as executives and senior practitioners from AI vendors like Mininglamp Technology and Deepexi, to deeply dissect Palantir's comeback logic from being shorted to being embraced by capital, and to reflect on how Chinese vendors can break through their real-world predicaments.
Starting from government and defense business, how did Palantir achieve "walking on two legs" in both To G and To B? What built the company's "moat," helping it break through among AI giants? And can the domestic vendors benchmarking Palantir across the Pacific produce a "Chinese Palantir"?
To G Playbook: How to Conquer the To B Market?
A Long Dormancy in To G Business
As a software company familiar to the general public, Palantir's rise was, however, unconventional from the start.
The company's first external investment came from In-Q-Tel, the venture capital arm of the US Central Intelligence Agency (CIA). In 2005, In-Q-Tel invested about $2 million in Palantir, directly opening the door to the highest-barrier government market.
In its early days, Palantir's clients were almost uniformly from the defense, intelligence, and law enforcement systems: the Defense Intelligence Agency (DIA), the Federal Bureau of Investigation (FBI), the US military, big-city police departments, and more. The Gotham platform, oriented toward military and government scenarios, also took shape during this period.
It was precisely in these extreme scenarios—high-risk, high-stakes decisions, zero tolerance for error—that Palantir introduced a rather philosophical concept—Ontology—and turned it into its core technical methodology.
In life-or-death scenarios like battlefields and intelligence work, the thorniest problem is messy data, inconsistent standards, and the inability to integrate. Palantir's ontology is essentially a unified set of semantics and models that connects, defines, and links heterogeneous data, integrating chaotic information into a clear "battle map."
Moreover, as Li Menglin, a technical expert who has long studied Palantir and Vice President of Mininglamp Technology, pointed out, what military and government scenarios brought Palantir was the extreme refinement of top-tier technology and data security capabilities. On one hand, its data scenarios are highly closed and not publicly disclosed; on the other hand, military and government scenarios demand very high transparency in the entire reasoning process.
"In the intelligence field, if a conclusion is derived but the reasoning process behind it is unknown, it raises very serious concerns." Unlike the "black-box" reasoning of today's mainstream large models, Palantir's early reasoning framework built on ontology could lay the entire decision chain out on the table.
These advantages made To G business Palantir's ballast stone.
From its 2025 financial report, of the nearly $4.5 billion in full-year revenue, 54% came from government business and 46% from commercial revenue. Since going public in 2020, To G business has consistently contributed more than half of revenue, stable at around 55%.
Its deep ties with the US military, intelligence agencies, and other federal government departments spanning more than a decade have brought stable business and extremely high-value orders. Last July, Palantir signed a 10-year agreement with the US Army worth up to $10 billion, cementing its status as the federal government's "AI arms dealer."
In 2025, its US government revenue grew 55% year-over-year to $1.855 billion. Palantir has also made clear its intent to capture a larger share of the US federal government's software system spending.
To G Genes Injected into To B Business
But in the US stock market, To G business has always been a business "disdained" by investors: government orders have long cycles, slow payment collection, and limited room for imagination—far less sexy than To B business.
After establishing a foothold in government business, Palantir planted its commercialization ambitions early on.
Around 2010, Palantir entered To B business through the financial sector, custom-deploying the Metropolis platform for US financial giant JPMorgan Chase, building a bespoke anti-fraud and risk analysis system. Metropolis later evolved into the Foundry platform, which became Palantir's core product serving B-end clients. For more on Palantir's transformation to To B business, feel free to add the author's WeChat skylar_12_14 to discuss.
Today, To B business has long moved from early exploration into full-blown explosion.
Financial reports show Palantir's commercial segment revenue in 2025 was $2.073 billion, up 60% year-over-year. It is increasingly favored by large US enterprises, with related revenue surging 109% year-over-year in 2025 to $1.465 billion.
The secondary market's confidence in Palantir boils down to two points: first, its To G business base is very solid; second, it is genuinely converting that advantage into growth momentum for To B business.
The ontology that Palantir repeatedly refined and validated in real-world practice has become its "killer weapon" for entering the commercial market.
This methodology found a new battlefield in To B scenarios. Yingzhe, a VC investor with a data analytics background who once worked in BI (business intelligence) in the US, mentioned that different departments may understand the same concept differently—for example, the "inventory" referred to by procurement and by sales may not be the same concept, and the metrics for business growth may also differ. Ontology can unify them into the same definition through semantic models, laying the first cornerstone for intelligent decision-making.
Moreover, Palantir's To B expansion has continued the heavy-duty genes from its government service days.
The B-end product Foundry is a software platform "no less than a cloud operating system." "Ontology-based data processing, models, simulation—the complexity of this system is basically on par with Alibaba Cloud and Baidu AI Cloud," Lu Cheng, an IT veteran who has long studied Palantir and has experience at multiple cloud vendors, told Leiphone.
Palantir built Foundry to be so "heavy" because it targeted the very top of the pyramid—large enterprise customers—from the very beginning.
"High revenue, few customers, and extremely high deal sizes mean you have to go deep and thorough with every customer." The results of this strategic choice are clearly visible: top-tier large customers still contribute the vast majority of revenue to this day.
Financial reports show that in 2025, Palantir's top 20 customers generated an average revenue of $93.9 million, totaling $1.878 billion—about 42% of total revenue, up 45% year-over-year; the top three customers alone accounted for as much as 16% of revenue.
This is precisely the core difference between Palantir and the many large model vendors pursuing a lightweight approach. Lu Cheng believes that if Palantir went lightweight, its stock price might climb even higher. Although there is plenty of room in the lightweight space, there is no need to take that path. "The heavy model is its true advantage; going lightweight could actually be a dead end."
A Stunning Comeback: How Did It Achieve a Massive To B Breakout?
AIP "Ignites"
What truly made the market recognize the value of Palantir's To B business was AIP (Artificial Intelligence Platform), released in April 2023.
Before that, Palantir suffered repeated blows in the public markets in 2022: under the triple pressure of rate hikes, lockup expiration selling, and doubts about its commercialization, the market's patience with Palantir hit rock bottom—the stock fell more than 64% for the year, dropping below $6 at the end of 2022 and early 2023, hitting an all-time low.
"Many investors are hyping Palantir now precisely because they believe it was once undervalued."
In Ying Zhe's view, Palantir's early valuation didn't rise because the imagination space for a business serving only military and government clients was limited, and its highly confidential business was hard for market-oriented investors to perceive.
It was AIP that allowed Palantir to shed its "defense stock" label and leap into becoming one of the most watched AI stars in the US stock market.
Palantir staged a stunning comeback from its historical low, unleashing an epic rally: the stock surged 167% in 2023, soared another 340% in 2024, and continued climbing 135% in 2025.
AIP ignited Palantir's stock price for the following reasons:
First, it activated existing capabilities, combining Gotham and Foundry with generative AI (including large language models), upgrading the original products from traditional data analysis tools into an enterprise AI operating system.
Second, it maximized deployment efficiency, compressing deployment cycles that once took months or even years down to just a few days, dramatically reducing time costs and solving the industry pain points of slow large model deployment and poor reusability.
This is thanks to AIP's generalization capability, which opened up the possibility of large-scale "replication." Ying Zhe pointed out that unlike traditional machine learning approaches, AIP is not confined to a single industry by its training data, solving the past pain point of capabilities being hard to reuse.
What truly pushed market sentiment to a climax was Palantir's partnership with NVIDIA, the "king of AI compute":
At the end of October last year, the two parties announced the integration of NVIDIA's GPU-accelerated computing, CUDA-X libraries, and NEMOTRON models with Palantir's ontology framework and AIP platform.
This partnership was seen as a powerful alliance between a top AI decision-making platform and top-tier compute. Combined with Palantir's strong Q3 earnings, the stock hit an intraday high of $207.52 on November 3, setting a new all-time high.
AIP's boost to commercialization is clearly reflected in the financial data: Palantir's revenue growth curve in recent years shows a pattern of high growth—slowdown—re-acceleration.
It maintained growth above 40% during the post-IPO boom of 2020-2021, slowed to 24% and 17% in 2022-2023, rebounded to 29% in 2024, and climbed sharply to 56% in 2025.
Without a doubt, AIP accelerated the breakout of Palantir's To B business and opened a second growth curve for this company once seen as purely To G.
Today, Palantir's AI commercialization is still accelerating. Looking at Q4 2025 alone, its revenue momentum shows no sign of slowing: overall revenue grew 70% year-over-year, while US commercial revenue surged 137%.
The "Internal Cultivation" of the FDE Model
Beyond AIP's market ignition, Palantir's commercialization success and speed owe much to the internal cultivation of the FDE model (Forward Deployed Engineer model).
The FDE model also grew out of Palantir's early To G business: in military and government scenarios, data was siloed and environments were complex, so conventional software development approaches "broke down." Palantir had no choice but to send people to the front lines to understand the business, then bring the experience back to be distilled and abstracted into products. This approach was later systematized into the FDE model—closing the loop of customized deployment first, followed by productization.
The FDE model may seem mysterious, but at its core, it exists to solve the problem of closing the loop from software to customer scenarios. Lu Cheng offered a vivid metaphor: "If software is the highway, then the FDE model is the 'last three kilometers' of national road from the highway to your doorstep."
According to a former employee, Palantir's engineering system is divided into two types: one is the Forward Deployed Engineers (FDEs) stationed at the customer front lines; the other is Product Development Engineers (PD engineers) focused on core product development.
In terms of collaboration, FDEs, as the front-line combat troops, spend 3-4 days per week on-site, deeply understanding business processes, then use their accumulated know-how to design software that genuinely solves problems. PD engineers, as the rear support troops, are responsible for distilling these battle-tested solutions into standardized, scalable product capabilities.
This heavy delivery model means Palantir must invest extremely high upfront costs when first serving a customer. But once an "ontology model" is established—whether the customer later expands to new scenarios, or the Palantir platform serves similar customers—it can directly call on existing data models without repeatedly investing large amounts of FDE resources.
In terms of marginal effects, Palantir leverages a one-time high fixed cost to drive long-term marginal costs toward zero. This explains why its gross margin has remained high for years.
Financial reports show Palantir's gross margin in 2025 reached a high of 82%. Over the past five years, its gross margin has trended steadily upward: rising from 78% to above 80%, with Q4 2025 gross margin exceeding 84%.
It's easy to see that Palantir's business secret is "selling the platform," not "selling headcount."
"If it weren't a company that had already proven its capabilities, many B2B customers wouldn't accept this model." Lu Cheng believes Palantir's core advantage lies in "doing more than IT consulting firms."
He believes the FDE model is essentially a combination of consulting and delivery, and in the US, companies with both capabilities are rare. "Americans have a certain inherent disdain for delivery," Lu Cheng admitted. "They prefer pure consulting or pure standardized products and are unwilling to get their hands dirty with the hard work."
It is precisely the hard work put into delivery that allowed Palantir to break through in the fiercely competitive US software industry.
Yang Lei, co-founder of Deepexi, who has long studied Palantir, pointed out that over the past decades, companies like IBM and Accenture in the US software services industry have gone deep into project management, business consulting, and research. But their weakness is that they lack core products of their own—after the consulting ends, they still need to invest substantial manpower to complete subsequent development work.
Gartner analysts once analyzed that the core reason customers choose Palantir over Snowflake lies in how much To B business values the demonstration effect of lighthouse projects. Snowflake lacks practical experience in such complex scenarios, while Palantir has accumulated deep "battle-tested" experience in high-barrier industries such as defense, government, and finance.
Palantir's industry know-how was incubated and shaped precisely through the repeated refinement of the FDE model.
To this end, Palantir also played a move of "using your spear against your shield": it brought in large numbers of industry experts who came from its customers, enlisting this group of "strategists" who understand business scenarios and customer pain points best.
"What Palantir provides is first and foremost business experts, and only then IT personnel." Zhao Yang, an executive at a domestic software vendor, believes Palantir is one of the few software companies that can achieve an excellent match between industry know-how and IT technology. Many domestic software vendors have already piloted "FDE-like models"; to learn how they are faring, add the author's WeChat skylar_12_14.
How does the "combination punch" pay off?
The "combination punch" of AIP and FDE has built Palantir a barrier that rivals find hard to cross: extremely high customer stickiness.
According to the Q4 2025 earnings call, Palantir's Q4 net dollar retention rate (NDR) reached 139%, meaning existing customers increased their purchases by 39% annually on top of their original baseline.
In other words, existing customers not only renewed their contracts but also increased their annual spending by 39% year over year—clear evidence of extremely high retention, extremely low churn, and a continuously deepening dependence on Palantir.
The combination of AI implementation capability, FDE hands-on service, and industry know-how translates into the most tangible thing: customers cannot leave Palantir.
When a customer's core business processes, data models, and decision logic all run on the Palantir platform, with FDE teams deeply involved in co-building them, the customer's switching cost becomes extremely high.
This is also a trump card Palantir plays to the market: the best salesperson is the product itself, with conversion achieved entirely through real-world results and customer word of mouth.
By the end of 2025, Palantir's customer count grew from 711 the previous year to 954, a 34% year-over-year increase.
Interestingly, early Palantir long practiced an "anti-sales" culture. CEO Alex Karp even publicly expressed disdain for building a sales force. Only on the eve of its IPO, in order to promote the Foundry product aimed at the enterprise market and cope with listing pressures, did Palantir begin recruiting sales talent on a large scale and gradually build a formal sales system.
The Los Angeles Times once revealed the process of building Palantir's sales force: early members included top salespeople who had worked at major American software companies such as Oracle and IBM; it replicated its earlier European sales model—featuring 3-person teams consisting of one engineer and two salespeople.
One sales employee revealed that two months after starting, all training focused on the Foundry software itself, with no sales guidance. Clearly, even after building a sales team, Palantir's focus remained on industry know-how.
In its 2025 financial report, Palantir summarized the core secret to winning major government and enterprise customers: "The bigger and more complex the problem, the higher the technical difficulty, the greater our odds of winning." For Palantir, such projects—with high upfront investment, high implementation risk, complex data environments, and long sales cycles—precisely constitute a natural industry barrier that keeps many competitors out.
Can China produce a Palantir?
After Palantir took off, not only did Silicon Valley AI startups rush to imitate it, but the wave also spread to China.
Unlike Silicon Valley, which long ago established a mature business model for "platform software," China's software industry has long been mired in the quagmire of difficulty turning a profit and small scale—cutthroat competition in the low-end market, while the high-end market is firmly held by overseas giants. To this day, China has yet to produce a software giant with global influence like Palantir or Salesforce.
"It's like a long-distance race—Palantir is a great pacesetter. We follow behind, learn its rhythm, and then find our own stride."
Many domestic vendors are dissecting Palantir's product technology and business model, hoping to find a path to breakthrough.
Amid this wave, industry insiders believe three types of players have the potential to become "China's Palantir":
- The first type is vendors whose business logic resembles Palantir's. These vendors focus on data intelligence and AI decision-making, such as 4Paradigm, MiningLamp Technology, Deepexi, Baioo, Xunce Technology, and TRS.
- The second type is vendors benchmarking against Palantir's core defense business. These vendors focus on the military industry + AI track and do To G business, such as Jing'an Technology, Yuanting Technology, and Zhongke Shitong Hengqi. Lin Yu, an investor who has studied Jing'an Technology, told Leiphone that Jing'an wants to combine the models of Palantir and Anduril, another American defense tech company, to build drone swarm flight control systems, attempting to tell a story similar to Palantir helping find bin Laden.
- The third type is big companies with more resources and organizational capabilities. In Lu Cheng's view, to become a Palantir you still need to be a big company; it's hard for small ones. Giants like Huawei and Baidu are more likely to succeed, not only because they also have many government customers, but also because of organizational culture. Palantir's approach is a heavyweight one—heavy product deployment and heavy FDE delivery, not a small-and-beautiful path—only big companies can concentrate forces to accomplish big things and invest massive funds and resources. It is said that a certain domestic big company already has a team benchmarking against Palantir to develop related products; for more insider details, add the author's WeChat skylar_12_14 to discuss.
After Palantir became a hit, many domestic vendors have proactively aligned themselves with its model in their external communications.
In Ying Zhe's view, the reason 4Paradigm is often called "China's Palantir" is that it is a listed company with a public voice, and compared with China's "AI Four Little Dragons," it has done more solid work in this field, has a large number of B-end and G-end customers, and has launched the Sage platform similar to Palantir AIP, so it seems more alike.
"4Paradigm's stock price rose well for a time because in its quarterly reports and investor Q&A it aligns itself with Palantir's model, for example emphasizing that not every project is customized, but that, like Palantir, it highly abstracts a methodology."
But what the market cares more about is: in this wave, what exactly can domestic vendors learn from Palantir? Can they really produce their own Palantir? And what challenges do they currently face?
Challenge one: a conservative and pragmatic technology route
"It is very difficult for Chinese vendors to be completely like Palantir." Ying Zhe reached this conclusion last year after studying Agent-oriented projects in the domestic To B space.
But he believes it is not entirely impossible. Whether a Palantir-like company can emerge depends on whether the technology can eventually become strong enough to support such a narrative.
For now, domestic vendors have long lacked scenarios for honing software technology.
"When domestic B-end companies make purchases, they generally believe software isn't worth much, usually requiring integrated software-and-hardware offerings dominated by hardware, so the value of software goes unrecognized."
Ying Zhe believes that no domestic vendor has been able to hone top-tier technology with high-quality, high-value data in its early days. Although many similar companies grew up 10 years ago in the AI 1.0 machine learning era, they still remain on the periphery of business and have not entered the core domain.
According to his observation, the technology publicly available in the domestic market is still that machine learning set, and customers prefer stability—they don't want AI to freely generalize, make intelligent decisions, or take over in business scenarios, preferring AI to play an auxiliary role.
Yingzhe added that machine learning methodology is common in the financial risk control field, but it can only output probability thresholds and is hard to generalize across industries—still a gap from today's large models' ability to reason, make decisions, and master the full business chain.
This reflects the fact that domestic vendors prefer a more conservative and pragmatic technical route.
Xu Yan, formerly at IBM and now Chief Consulting Expert at Baiwang Co., told LeiFeng Wang that large language models are essentially probabilistic models, while scenarios like healthcare and finance require explainable and precise conclusions. In such business scenarios, large language models may not necessarily outperform machine learning.
According to LeiFeng Wang, several domestic vendors, including Baiwang Co., have begun combining machine learning with large model technical paradigms.
Xu Yan revealed that Baiwang currently adopts a "dual-track" technical route, with its data technology team attempting to shift from the traditional posterior mode represented by machine learning to a large-model-driven prior mode.
In Lu Cheng's view, compared with Palantir's already-hardened capability in landing large model To B applications, most domestic large model projects are still at the PoC (proof of concept) stage.
The main reason is still that domestic large models lack accuracy (i.e., the "hallucination" problem), and there are big differences in investment philosophy between China and the US, with domestic players being relatively cautious. "Everyone is desperately waiting for scenarios—only when large models can truly solve problems will they invest heavily." For the latest AI developments of domestic cloud vendors, large model vendors, and software vendors, feel free to add the author's WeChat skylar_12_14 to exchange information.
Beyond that, historical technical baggage is likewise reinforcing this conservative, pragmatic technical inclination.
Yingzhe believes that to become China's Palantir, one cannot carry too much historical technical baggage, but few vendors are willing to completely abandon their existing technology systems to transform.
He mentioned one point: if a company invested heavily in machine learning early on, forming a technical school and resource core, it becomes very hard to transition to the Transformer and Diffusion technical architectures of the large model era. They can only continue to broaden scenarios on the existing system, and it's difficult to open up entirely new business space with next-generation technology. "The original team built their success on old technology; now asking them to transform or exit is doubly difficult."
But the real gap has not cooled domestic vendors' enthusiasm for technical breakthroughs; instead, it has pushed them to use Palantir as an important reference when exploring next-generation data intelligence architectures.
Yang Lei told LeiFeng Wang (WeChat public account: LeiFeng Wang) that when Deepexi was founded in 2018, it looked to companies like Snowflake and Databricks; at that time, the industry mainly focused on data platform construction, data lakehouse construction, data real-timing, and data presentation.
But he gradually realized there was a problem with this direction: data and business knowledge were isolated from each other. Business logic was embodied in process systems like ERP (Enterprise Resource Planning) and MES (Manufacturing Execution System), while data only reflected results and described processes—the two were in a separated state.
After in-depth research on Palantir, Yang Lei and his colleagues found that this data analytics company's route was highly aligned with the direction they were exploring.
"It's like everyone climbing a mountain from different directions—Palantir climbs from the A side, we climb from the B side, and later at some point we realized: 'Hey! This is exactly how we should climb the mountain.'" Later, Deepexi built platforms like FastData Foil and FastAGI, which coincided with Palantir's Data+AI model.
Challenge Two: Data Governance Is a "Tough Nut to Crack"
Technical catch-up is only the first hurdle; what truly tests domestic vendors' internal strength is the even harder barrier of data governance.
Unlike the US, where data infrastructure is well developed and Palantir accumulated extensive data governance experience through early customized projects, the core test facing domestic vendors is having to deal with all kinds of "dirty data":
Yingzhe believes data governance capability is a fundamental skill. Once Chinese vendors benchmarking against Palantir go deep into business scenarios, the first thing they must solve is data source and quality issues. The severe problem in the domestic market is that only after large enterprises have used digital systems for many years do they have relatively structured data—and even then, there may still be anomalies, missing values, erroneous values, and other issues.
It is thus clear that China's data infrastructure is weaker, and the data governance challenges it faces are more numerous and complex than Palantir's. Yingzhe mentioned that over the past 10 years, a batch of data governance vendors emerged in China during the machine learning wave. Although they have accumulated relevant technical experience, to this day data governance remains a "tough nut to crack" in the industry.
On one hand, domestic data is severely fragmented, requiring vendors to spend more effort on governance.
Li Menglin, the aforementioned technical expert and Vice President of MiningLamp Technology, noted that overseas enterprises have embraced the cloud-native concept much better than domestic ones, and their data infrastructure is far more unified and standardized than that of domestic enterprises.
In earlier years, overseas enterprises generally used common solutions like Snowflake, Databricks, and dbt Labs for data infrastructure and accumulated large amounts of business data, making it easier to get things "running" after business modeling.
Domestic enterprises' data infrastructure levels, however, vary widely:
During early digital transformation, some deployed in local IT environments, some adopted cloud architectures, some built on open-source solutions, and some built systems directly through in-house development. Li Menglin revealed that on the first day of on-site service in China, MiningLamp first had to figure out how the customer's original systems were built and what technical architecture they used.
The massive IT systems accumulated by enterprises during digital transformation are undoubtedly a huge challenge for vendors. "Today it's simply not possible to tell customers to tear everything down and start from scratch, building a data middle platform from the ground up."
Li Menglin clearly feels that in the current market environment, enterprise decision-making has become increasingly cautious, and companies are unwilling to invest tens of millions of yuan in data infrastructure. The core reason is the long ROI cycle of data infrastructure—investments of tens of millions often take two to three years to show basic results, and enterprises find it hard to maintain sufficient patience.
On the other hand, data governance is not purely a technical problem; it also involves the pressure of organizational change.
When facing customers' internal coordination issues or other demands, vendors—even when on-site doing cross-departmental coordination and communication to connect data—face many tests. For example, a middle platform project needs to break through various departmental "walls," bringing data from different departments together for joint construction and governance before the platform can truly run smoothly—undoubtedly another hard battle.
These many practical difficulties make it impossible for Chinese vendors to directly replicate or copy Palantir's mature model.
Li Menglin told LeiFeng Wang that MiningLamp Technology (formerly Miaozhen Systems) did not benchmark against Palantir when founded 20 years ago; the clear benchmarking came later in 2014 with the founding of Mininglamp Data.
"When we initially benchmarked against Palantir, we hoped to apply its technical direction to the Chinese local market to solve similar problems." Li Menglin mentioned that Mininglamp Data's early business direction was very similar to Palantir's military and government business; at the time, they applied knowledge graph technology to public security business.
But from the practical dilemma of data governance, Li Menglin believes Palantir takes a "bottom-up" route: first helping customers connect and govern all their data, then teaching them to use it, letting enterprise employees build scenarios on the platform themselves and learn how to use it. But this model is hard to implement in the domestic market:
Domestic enterprise employees have limited willingness to adopt such complex systems, especially now that large models are relatively widespread—employees are more accustomed to asking questions directly and getting results directly; the cost of learning a complex system is too high.
Combined with the aforementioned factors, customers are unwilling to bear high labor costs just for preliminary work like data governance; they'd rather know from the start how data applications can drive business growth.
Therefore, in 2025 MiningLamp Technology proactively upgraded to an Agentic AI strategy, taking a "top-down" path: rather than diving into heavy data governance work from the start, it uses its accumulated vertical-domain proprietary data plus proprietary models to first use agents to solve customers' actual problems, letting data applications in turn drive data governance.
Challenge Three: The FDE Model Is Hard to Make Work in China
Domestic vendors benchmarking against Palantir currently face another thorny problem: heavy delivery and low gross margins. A gross margin as high as Palantir's 80%-plus is almost unimaginable in the domestic market.
"The FDE logic that relies on manpower can hardly lead to a profitable business model in China." Li Menglin said that the Chinese and US market environments differ in the price points they accept for delivery-type projects:
China's B2B market lacks the high price points and high gross margin foundation seen in the US. Relying solely on large-scale manpower for delivery makes it hard to scale, and vendors face enormous pressure on cost control and profitability.
Zhang Chi, an industry insider with over 20 years of experience who has worked at multiple major domestic and international cloud companies, did the math: Palantir can land orders starting at $100 million or even $1 billion, and can dispatch dedicated teams to serve them; but in China, apart from large compute procurement orders for infrastructure, very few particularly large deals are ever seen.
Moreover, talent costs at domestic giants like Alibaba, Huawei, and Tencent are not low—a P7-level employee's total annual package is one to two million RMB (over $200,000)—yet such a person might not even count as a top business expert on a Palantir project. To sustain such a team, orders must come with a fairly good gross margin.
This is also one reason Zhang Chi believes the Palantir model cannot be replicated in China: the lack of consulting-style large deals that send industry experts to customer sites to understand requirements, define and develop projects, and then iterate products in reverse. "No enterprise would hand a vendor a $1 billion order that is primarily consulting, on-site deployment, and engineering delivery."
This is easy to understand: while the FDE model can bridge the gap between customer needs and product/R&D teams in heavy-delivery scenarios, almost no enterprise in China is willing to pay for it.
Facing the problems of heavy delivery and low gross margins, domestic vendors are also trying different solutions.
Li Menglin told LeiFeng.com that MingLue Technology is exploring a new delivery model: while drawing on the FDE model and retaining the role of "people," it is gradually letting agents take on more delivery work. The thinking behind this is, on one hand, to use AI to lower delivery costs, and more importantly, to use AI to improve delivery efficiency and quality.
In addition, the domestic software industry has always been "heavy on delivery, light on consulting."
Lu Cheng, the aforementioned expert who has long studied Palantir, said that domestic vendors look up to the FDE model because most of them lack consulting capabilities and are only good at delivery.
In terms of deep delivery capability, 4Paradigm is in the top tier in China. Ying Zhe noted that investors favor 4Paradigm precisely because it has scenarios and data, holding the key to B2B.
But 4Paradigm is often criticized for long project delivery cycles, heavy reliance on outsourced manpower, and the need to bundle hardware; its gross margin is only a little over 30%, and it has long been unprofitable. Another problem facing such vendors is the difficulty of devoting substantial resources to forward-looking research.
Ying Zhe believes that if 4Paradigm wants to break through, it needs to take a path completely different from Palantir's: Palantir built up its skills first and then expanded into scenarios, while 4Paradigm has scenarios first and needs to improve its skills afterward—that is, "the next step is to strengthen its technical capabilities."
Challenge Four: The Dual Test of Strategy and Resources
External ailments and internal pains. Beyond industry-wide challenges, domestic vendors must also face two tests of their own.
On one hand, can domestic vendors stay committed to long-termism and maintain sufficient strategic resolve?
Xu Yan told LeiFeng.com that one thing Baio.com (Baio Stock) admires about Palantir is its "long-termism." Palantir started with To-G military and government business; from a commercial return perspective, this kind of business requires long-term investment and leans toward semi-strategic investment—Palantir endured nearly 20 years of losses before achieving full profitability.
Baio benchmarks itself against Palantir partly because of its own G-to-B-to-B DNA: participating in the construction of the national fiscal and tax data foundation through To-G business, quietly doing the underlying "bridge-building and road-paving" work behind the scenes. Xu Yan mentioned that building both G-end and B-end platforms is complex and requires huge investment, with Baio accumulating nearly tens of millions of RMB in annual investment in this area. Adopting the name "China's Palantir" is also a hope to be seen by the market.
Whether they can hold fast to long-termism is now testing a host of domestic vendors. Many investors say bluntly that most domestic products benchmarking against Palantir at this stage are "form without soul," and even those with form are few and far between.
Liu Chang, a domestic software practitioner who has studied Palantir in depth, believes that the core reason Palantir rose with AI is not agents, nor data, but a "reasoning framework" that has yet to be fully deconstructed—"Dig down to the bottom and you'll find its foundation is very deep, something we can't fully crack right now, but we should understand that this is a layout it started building many years ago."
In her observation, domestic vendors tend to pursue rapid commercial monetization and do not think as far ahead as Palantir does.
"The Chinese market is huge—even building vertical models or digital humans offers opportunities for commercial monetization. But when these areas become red oceans, what determines competitiveness will be the reasoning framework, and by then it will be too late to build one."
On the other hand, the triple bottlenecks of compute, capital, and talent together test domestic vendors' ability to integrate resources.
First, on the compute front.
Zhou Jie, an analyst at a certain institution, believes that behind Palantir's success lies strong compute support from NVIDIA's mature ecosystem. The road for Chinese vendors is harder: whether they can truly collaborate with domestic chip makers to shore up the compute shortfall is a hurdle they must clear.
Second, on the capital front.
Lu Cheng believes that Chinese vendors are not unwilling to learn from Palantir—they simply can't. The reason they can't is "poverty," i.e., insufficient investment.
Small companies lack resources, while big companies' resources are scattered—too many business lines, not focused on just one thing. "For startups, survival is the most fundamental thing, and the fastest way is quick packaging; big companies may look powerful, but once resources are spread across departments, that power is gone."
In addition, China lacks a mature mechanism for cultivating know-how talent.
"Palantir is willing to send its employees directly into the US military as officers—this is almost unimaginable in China." In Li Menglin's view, the core domestic problem is whether there are enough versatile talents who both understand technology and truly grasp industry depth, to fill the know-how gap.
The founders of Deepexi, who came out of Huawei and Alibaba technical teams, likewise recognized the importance of "understanding the business" over the past 8 years.
Yang Lei candidly told LeiFeng.com that insufficient business sensitivity was once the "weak spot" of this technically-rooted team: taking the supply chain scenario as an example, without having worked at a large logistics company, even with big-tech engineering experience, one may not truly understand how supply chains operate, and can easily underestimate business complexity, thinking supply chains are just simple goods distribution.
But over the past few years, Deepexi began investing heavily in building business teams, establishing the DIC (Deep Innovation Center). The DIC team is composed mainly of business experts who go on-site at customers to provide professional services.
In his view, talents who understand both AI and business, while also possessing strong technical capabilities, are extremely scarce today. Only by solving this problem can the Chinese market possibly give birth to a true Palantir-like company in the future.
When will the "Palantir-ization" craze fade?
In this wave of enthusiasm, many vendors have already realized that simply copying the Palantir model is unrealistic. A basic consensus has formed in the industry: borrowing from the Palantir model urgently requires localized adaptation or transformation.
Liu Chang believes that while it is still too early for domestic vendors to directly benchmark against Palantir, it is like IBM in the eyes of IT practitioners back in the day—a benchmark worthy of respect. "Take Palantir apart, study it thoroughly, and then grow something of your own—perhaps one day you can run alongside it in the leading pack."
But there are also industry practitioners pouring cold water on this wave of hype. "Chinese software vendors probably all want to learn from Palantir—some want to learn its ontology, some want to learn FDE, but in the end they'll most likely end up learning it in all sorts of weird and varied ways, and then start competing on price again; there are also companies that just want to use Palantir to pump up their stock price and package a concept."
Lu Cheng believes that within about two or three years, it will become fairly clear who has learned it properly. "Palantir's stock went up, and everyone rushed in. In the first year, a wave of them will die off; in the second year, another wave will die off; by the third year, it will all be clear."
The author of this article has long followed the moves of AI giants and the AI cloud market. Feel free to add the author on WeChat at skylar_12_14 to exchange ideas.
Note: Lu Cheng, Zhao Yang, Lin Yu, Zhang Chi, Liu Chang, and Zhou Jie in this article are all pseudonyms.
This is an original article by Leiphone. Unauthorized reproduction is prohibited. For details, see the reposting guidelines.
Related Articles
Editor
For the security of your account, please verify your email
Please verify your email