Large model companies are beginning to take on the work that was traditionally done by consulting firms.
When models themselves are increasingly commoditized, what exactly is the moat for large model companies?
Over the past half month, Moonshot AI has made continuous moves: first, it was revealed to be in revenue-sharing negotiations with Microsoft, Amazon, and Google, launching an overseas revenue collection model for open-source models. Shortly after, it launched an enterprise partner program with the FDE (Forward Deployed Engineer) model, where cooperating IT service providers and integrators are dispatched to customer sites to help enterprises deploy AI into their core businesses. Last week, Moonshot AI released a set of financial industry solutions that integrate products and capabilities tailored for the financial sector, as well as institutional-level data modeling and report delivery capabilities.
None of these moves are new on their own: large models being launched on cloud vendors' platforms was already done by OpenAI and Anthropic long ago; FDE on-site delivery has been practiced by Palantir for 20 years; industry solutions are standard practices for all B2B companies. However, the intensive rollout of the three initiatives within half a month points to the same transition: Kimi is trying to transform from a model-selling company to a delivery-selling company.
This transition is not a choice exclusive to Kimi. Almost at the same time, OpenAI also released financial industry solutions, and Anthropic has long focused on the B2B track. Leading players are unanimously pouring resources into enterprise on-site scenarios, with only one core judgment behind it: Benchmark scores and ranking lists cannot bring revenue, only deployment capabilities that can be embedded in businesses and generate measurable returns can. If these actions are simply interpreted as "Kimi is also starting to do B2B business", you will underestimate the whole strategy. The real question worth asking is: why exactly now, are large model companies collectively shifting their focus to delivery?
01 The Collective Embarrassment of PoC
Over the past two years, domestic large model technologies have achieved leapfrog development, but many vendors still follow the traditional PoC delivery model. What enterprises get is often a demo polished based on carefully selected samples, where the model performs well on screened test samples. But when faced with messy internal historical documents, incomplete work orders, and cross-system permission constraints of enterprises, the model's accuracy drops rapidly, and hallucination problems break out intensively. Many enterprises, after investing budgets to complete proof of concept, see the project stagnate and fall into the "PoC Death Trap".
PoC can only prove that the model is technically feasible, but cannot prove that the model is usable in real enterprise scenarios.
The root of the trap does not lie in the model itself, but in the misalignment of labor division. The biggest shortcoming of the traditional PoC model is that it separates the connection between business, data and intelligent systems. Most vendors only output API interfaces, focus on selling Tokens, and leave all work such as business adaptation, data processing, and system docking to customers themselves. To carry out AI transformation, enterprises need to have multiple capabilities including business understanding, large model fine-tuning, and heterogeneous system integration. The vast majority of government and enterprise customers can hardly assemble such a compound team. This leads to a strange industry phenomenon: There are more and more large model products on the market, but real deployment cases that are embedded in core businesses and generate measurable business returns are very scarce.
This reveals a fact covered up by the benchmarking competition: in the second half of the large model industry, the bottleneck is not intelligence, but delivery.
02 An Open Hand: Turn Delivery into a Standardized Business
Kimi adopts the complete PDE implementation methodology in enterprise projects, namely Process (business process decomposition), Data (enterprise private data engineering), and Engine (intelligent Agent orchestration). It takes the three as pre-project prerequisites to avoid the PoC trap from the source. Different from the idea of some vendors that prioritize making demos before adapting to businesses, the PDE framework puts business processes first: in the early stage of the project, the team works with customers to sort out the complete and real business link, clarify the rights and responsibilities boundary between AI and humans, and complete the integration with existing business systems such as ERP, OA, and CRM to form a closed business loop, rather than producing a demo prototype that can only perform copy-paste operations. The second part is enterprise private data engineering: instead of relying on screened high-quality samples for demonstration, it directly connects to the original, unprocessed real business data of enterprises, completes data desensitization, permission isolation, knowledge base and evaluation set construction, to suppress model hallucinations from the source. On the engine side, relying on the Kimi Hosted Agents base, it completes prompt version management, tool call orchestration, safety guardrails, shadow testing, and full-link observability. The entire architecture is oriented to production environments, not simple API calls. The base model can be flexibly replaced, business logic is precipitated at the engine layer, supporting shadow operation and gray-scale release, to minimize the risk of business launch.
The base model is the admission ticket, and the PDE methodology is the converter that turns technology into business value. But the methodology cannot complete on-site implementation on its own, and deploying PDE requires front-line deployment engineers to undertake front-line execution. The FDE concept originated from Palantir in the United States, and later OpenAI and Anthropic also adopted the self-built FDE model. However, the overseas self-built FDE model has structural defects: Palantir relies on self-built elite engineers to achieve in-depth delivery, but falls into the heavy asset dilemma: FDE compound talents are scarce, revenue growth is deeply bound to personnel expansion, and it is difficult to give play to the scale effect of software products. The capital market values it more as a consulting company than a software company.
It is precisely by seeing this limitation that Kimi did not copy the overseas heavy asset self-built team model, but cooperated with leading IT service providers and system integrators such as Chinasoft International, Kingsoft Cloud, Teamsun, to jointly build the FDE front-line deployment engineer team. The engineers belong to ecological partners, while the methodology and platform are owned by Kimi itself. This essentially uses China's mature IT service ecosystem to solve the leverage problem that Palantir has not solved: the heavy delivery work is handed over to ready-made integrators, and Kimi retains the base model, standardized processes and component precipitation. All FDE projects must comply with unified PDE standards, project precipitated components are recycled to the platform, and FDE engineers also act as "business translators" to drive base model iteration, so that the marginal delivery cost can continue to decline.
This is a smart move, but the biggest weakness of the co-construction model is quality control: the capabilities of partner engineers are uneven, and projects will easily degenerate into ordinary outsourcing if there is any negligence. Kimi's solution is to prioritize standardization, which is the right direction. But whether it can be implemented in place still needs to be verified by the number of projects and renewal rates -- how far this model can go is also related to the ARR increment of this company.
03 The Finance Sector is a Touchstone, Not a Terminal
The financial industry solution is actually a phased achievement of the FDE model deployed in the financial track, which is an industry capability package exported after FDE on-site projects are precipitated and standardized. The choice of the financial sector as a touchstone is very deliberate: securities firms, banks, and funds have the most stringent requirements for data compliance and result accuracy across all industries. Getting through the operation here is equivalent to obtaining a trust certificate for other industries. From position morning reports and financial report comments, to project screening, in-depth research and portfolio review, this solution connects data acquisition, professional analysis and result delivery. The value of large models in the financial industry is moving from front-end single-point efficiency improvement to in-depth closed loop of core businesses.
According to the disclosed data, China Securities' "ad hoc entrusted report generation" agent jointly built with Kimi completed the access work originally planned for 2 months in only 3 working days. The manual production time for a single report was shortened from 30 minutes to 10 minutes, and manual input decreased by 67%. But we also need to be sober: 67% efficiency improvement is the "efficiency enhancement" narrative, not the "revenue generation" narrative. Whether financial institutions are willing to pay continuously for this set of solutions and whether the payment scale can cover the cost of on-site delivery is the real criterion for judging whether commercialization is established. From "lighthouse projects" to "replicable business", there is a gap that most enterprise service companies have failed to cross.
04 The True Meaning of the Three Moves
Connecting the three events, Kimi's intention is very clear: negotiating revenue sharing with overseas cloud vendors for model capabilities takes the light asset overseas expansion route; enterprise deployment realizes leverage through ecological co-construction of FDE; financial solutions as a standardized capability package prove that this set of strategies can precipitate products. With the help of the mature domestic IT service provider ecosystem, it realizes the complementarity of technical capabilities, industry experience and customer channels. The three-pronged approach essentially answers the same question: when the model itself is increasingly commoditized, what exactly is the moat for large model companies?
Industry trends are also confirming this shift. Gartner predicts that by the end of 2026, more than 85% of technology service providers will take FDE projects as the core means of AI deployment to shorten the deployment cycle. When on-site delivery becomes the industry standard, the key to competition will further move up: who has a stronger base model, more standardized methodology, and greater ecological leverage.
As the wave of large model benchmarking competitions gradually recedes, commercial deployment has become the core criterion for testing real strength. The main battlefield of domestic large model competition has been transferred to the real business scenarios of thousands of enterprises. For Kimi, the intensive actions in half a month prove that its strategy is clear, but a clear strategy never equals victory. This "delivery business" ultimately needs to answer a question as simple as that faced by all enterprise service companies: after getting through the last mile of digital and intelligent transformation, will customers renew their contracts next year?
From demo to delivery, from efficiency improvement to revenue generation, this threshold must be crossed by all large model companies, and crossing it earlier is more proactive than later. The story of large models has finally moved from "witnessing miracles" to "daily necessities". This is not the exit of imagination, but the very beginning of technology truly integrating into reality. When the spotlight shifts from benchmark ranking lists to the business scenarios of enterprises, this road is very long, but as long as the direction is correct, you will not fear the distance.
This article is from the WeChat official account "Liujiu Finance" (ID: liujiucaijing69), author: Du Hao, published with authorization from 36Kr.