On January 2 this year, Huo Jia, Vice President of Alibaba (NYSE: BABA) Cloud Intelligence Group, ran a half marathon. That day he posted on WeChat Moments: "This year is the final year of the first Olympic cycle; after four years of training, I finally got to race."

What he was really describing was the race to land AI applications.

On September 23, during a media group interview at the Apsara Conference, Huo Jia repeated that judgment. In the fourth year of large models truly being deployed, this year is the first year of AI applications actually landing.

Token consumption is exploding and prices are rising — these are the signals that the race has begun. On the enterprise side, the situation is far more complicated.

Huo Jia's team has engaged with a large number of enterprise projects over the past year. The failures he has seen rarely stem from models not being smart enough. The causes of death are choosing the wrong scenario, overestimating one's own data, and using the methods of IT projects to do AI.

He offers two remedies: first, say No to clients; second, push engineers onto the client's site.

Four types of customers: the first two rely on demand, the latter two on value

Huo Jia explains this year's AI price increases as the stacking of three factors: exploding demand, improved model capabilities, and enterprises genuinely feeling productivity gains.

After overseas leading model products emerged, AI moved from chat and entertainment tools to the early stage of productivity tools, and the compute shortage behind Tokens was immediately exposed.

Rewind the time window by two years: back then models were constantly discounted, yet Token volumes never took off. Huo Jia says that if you figure out why things didn't take off then, you'll understand why they have taken off today.

Huo Jia divides the market into four types of customers. The first is AI-native customers, including startup teams built around new model companies, and even short-drama companies. The second is high-tech customers — internet, automotive, mobile — with strong in-house technical capabilities. The third is enterprise-level customers. The fourth is government.

In terms of large-scale usage and Token call volumes, the earlier the category, the higher the share.

"The first and second categories are demand-driven; the third and fourth are value-driven."

The first two types of customers know what they want and can use Tokens as soon as they buy them. The latter two must first be convinced it's worth using.

Huo Jia attributes the inability of enterprise-level customers to get traction to three reasons.

First, AI talent is extremely scarce.

Second, there are too few practices to learn from. In the internet and mobile internet eras, developers could see a large volume of implementation experience. In this round, everyone started at the same time four years ago, and only this year has there begun to be something to share.

Third, the working methods of traditional IT projects can't keep up with the technology. Models are released roughly once a month, and the engineering methods built on top of models have already gone through several rounds this year, from Harness engineering to loop engineering. By the time the rhythm of "plan, execute, re-plan, re-execute" completes one cycle, the technology has already changed.

Where enterprise AI projects die

How much of enterprise AI project failure is because models aren't smart enough? This was a question thrown out directly at the group interview.

Huo Jia did not give a ratio. He first stated a fact: since taking over this team at the end of August last year, there has not been a single customer complaint, not once has he needed to step forward and apologize. The reason is that projects are selected very strictly — having money and a project doesn't guarantee acceptance.

Then he described three types of failures he has witnessed.

First, forcibly training vertical-domain models.

This was the work algorithm teams loved to take on in the past. The data isn't ready, the cluster precision isn't suitable for training, the client still demands guaranteed results, and in the end training can't proceed.

A leader with a large budget once wanted to ask Alibaba (NYSE: BABA) Cloud to train a model. The first question Huo Jia asked when they met was: will you follow our model release cadence? At the time, Qwen was released once a quarter, and no one on his staff had told him about this. The other party quickly decided not to follow.

Today models are released once a month. Huo Jia's team has persuaded many customers to abandon training vertical models: "You can't outrun the frontier labs." If accuracy must be guaranteed, the solution lies at the engineering level — how to build the evaluation set, what method to switch to when RAG doesn't work — all of which can only be determined by tying it to the client's actual work.

Second, enterprises think they have data.

Last summer Huo Jia repeatedly said that "IT Data is not equal to AI Data," referring to training. This year he changed it to "IT Data is not equal to Agentic Data," referring to applications. The large amounts of data lying in enterprise systems mostly cannot be used directly in an Agent environment, yet customers insist on results.

Third, building Agents for just a few users.

A few people raised requirements, the IT department felt that finally someone was paying attention, and together they built an agent — after launch, only a few people clicked on it each day.

Huo Jia told a real case. A large enterprise had a B2B e-commerce platform, and the executive in charge approached Alibaba (NYSE: BABA) Cloud, saying you started in e-commerce and B2B and are now doing AI, so please help us design an AI-ization plan for B2B e-commerce. The team was very excited; the budget was large.

Huo Jia asked only one question: how many transactions per day?

Just over a hundred.

"Then what do you need AI for?" Manual processing takes only minutes to a few hours.

This kind of thing is still happening today.

Huo Jia's approach is to tell clients they shouldn't do it this way, and if he can't persuade them, to wait a while.

As early as the "14th Five-Year Plan" digital transformation projects, he had already said No to clients many times. "Saying No to a client is a very difficult thing, but because you say No, later on there are actually far fewer messy issues in the collaboration."

For upcoming projects, he has two more conditions.

First, the cycle must be short. For a central state-owned enterprise project, the team said in their first report that the initial launch would be three months later. Huo Jia's reply was: then don't do it, it will most likely fail. He went to meet the client's senior leadership and changed the rhythm to a monthly release, and later to weekly.

In the past, selling man-days meant the longer the project cycle, the better. Now that technology changes every month, only by compressing the cycle can you deliver something everyone agrees on at a certain point in the technology's evolution.

Second, One Team. For projects still using a client-vendor model of collaboration, he judges failure is highly likely: "You might as well just give them a standard product."

FDE is hot, but he doesn't want to mention the term anymore

The FDE term comes from Palantir. Forward Deployed Engineer — engineers deployed forward, directly stationed at the client site to write code and change processes. Palantir divides this team into two types, Echo and Delta: the former stays close to the client's business to find problems, the latter handles engineering implementation.

Large models have made this role hot again.

Research released by the MIT NANDA project in August 2025 showed that 95% of enterprise generative AI pilots had almost no measurable impact on P&L. a16z partner Joe Schmidt argued in a June 2025 article that AI startups should trade gross margin for a moat and pursue service-driven growth. From January to September 2025, FDE-related job postings grew by over 800%.

OpenAI formed an FDE team in 2024. On May 11 this year, OpenAI established the Deployment Company, with first-phase investment exceeding $4 billion, led by TPG, with Bain Capital, Advent, and Brookfield co-leading, McKinsey, Bain, and Capgemini joining as consulting and integration partners, and approximately 150 senior FDEs acquired through the acquisition of Tomoro. Anthropic is also expanding its Applied AI team.

Domestically, starting in July, there were so many FDE videos and articles that Huo Jia himself kept seeing them. By September, a batch of companies came out announcing they were doing FDE.

Alibaba (NYSE: BABA) Cloud started down this path earlier than this round of hype.

At last year's Apsara Conference, an energy company wanted to do "AI Plus" and listed a batch of scenarios from core production processes. Huo Jia felt it would be difficult, so he had the solutions team investigate first. After visiting many sites, the team found that many of the initially envisioned scenarios could not be implemented; if the client couldn't see results quickly, it would be hard to persist; and the process required introducing a large amount of new technology, which the old method of frontline reporting, back-office development, and then version iteration could not support.

When reviewing this project at the end of November last year, service staff were already advancing on-site based on pre-sales content. Huo Jia's judgment was that this seemed to be the FDE way of working. The team validated scenarios on-site and ultimately made it into a production-grade system that went live. At the internal strategy meeting in December, this model was determined to be worth piloting.

Many customers share the same pain points. They can't pinpoint scenarios, don't know how to verify technical feasibility and ROI, and don't know how to do production-grade launches — all three things they can't do.

Compared with traditional delivery, there are two differences.

First, people are pushed one step forward.

Traditional delivery first clarifies the project Scope, signs a Statement of Work (SOW), and then starts work. FDE enters during the pre-sales stage, working and negotiating at the same time. Alibaba (NYSE: BABA) Cloud did not copy Palantir's Echo and Delta; SAs take on the Echo role, and FDEs handle implementation.

When Huo Jia first formed the team, he valued large-model technical ability the most, but later found that understanding technology wasn't enough — the key was whether you could understand the client and communicate clearly. "At the most basic level, you should tell the client the Yes and No of this matter, and its cost." Many things are technically feasible but not worth the cost. Quote 100 man-days of work, and by the time you finally deliver something, the client no longer wants it. There have been far too many such demos in the past two years.

Second, the distance between the frontline and the product has shrunk.

As Huo Jia puts it, the advanced nature of all products comes from two aspects: first, forward-looking judgment about technology; second, real feedback from the market. In the past, most vendors didn't know how customers actually used models — they were holding a hammer looking for nails.

Qwen Office is the most direct example. It has a public cloud version, and state-owned and central enterprises need private deployment versions; beyond private deployment there are also large compliance requirements. The FDE team entered Qwen Office's R&D cycle directly, doing POCs at client sites, bringing requirements back in real time, and sitting with Qwen engineers to develop. "This was unimaginable in the past."

What customers use is also changing. From only using models, cloud basic products, and Bailian, they are starting to use AI-native products like Qwen Office and Qoder.

On how overseas and domestic FDE differ, Huo Jia's view is that there is no essential difference in working model and value positioning; what differs is the business environment. Overseas model companies also don't intend to make money from FDE — their goal is to get AI used. Only when the flywheel of product and usage spins will the underlying flywheel of data and models pay off.

"Domestically, wanting to rely on this to make big money may not be very realistic. Everyone shouldn't have that expectation."

Alibaba (NYSE: BABA) Cloud itself doesn't rely on it for money either. The core metric for team assessment is whether it can drive products. Alibaba (NYSE: BABA) is not a Service Company; long-term service contracts are only signed with key strategic customers. If FDE cannot become the glue and accelerator for product capability, it will be hard to sustain.

As for whether FDE is just a resident engineer with a new name, Huo Jia has already told the team he doesn't really want to talk about the term anymore. In China, the term easily takes on a different flavor — some say it's new wine in old bottles, others say it's selling dog meat under a sheep's head. Whether FDE will ultimately solidify into a model is still a question mark today.

There are only two business models

How FDE charges points to a bigger question: in the AI era, what do enterprise software and services rely on to make money?

Huo Jia's judgment starts with the old ledger of China's software industry. The commercial return rate of the software and services industry has always been mediocre; charging by feature can't clearly explain why customers should pay, homogenization is severe, and it descends into price wars, ultimately creating a vicious cycle. He used a phrase: "the dead SaaS."

He is betting on outcome-based payment, Result as a Service. Only if you can truly deliver results is there room to discuss a premium.

He only figured out this awareness after May this year. Before that, he did not believe in this model. Companies doing outcome-based payment in the previous round had a miserable time, and Alibaba (NYSE: BABA) Cloud itself also did many painful projects. The change comes from model capabilities; everyone's confidence in capability has risen.

To traditional software companies, his words are even more direct. If the competitiveness and profit margin of core products can't rise, then you can only charge by man-days. "What exactly is your moat?"

Another test criterion is whether customers are willing to sign long-term service contracts with you.

The biggest difference between Chinese companies and several major consulting firms is that they love signing Project Contracts too much and don't like signing Service Contracts. Companies that sign service contracts are inferior to product vendors, but they live better than those doing projects, and they have gross margin. Huo Jia knows several such small teams — a group of post-90s master's and PhD graduates who worked this way back when model capabilities weren't yet strong. It has already been three years, they've made profits, and the largest has grown to nearly a hundred people.

"There are only two business models: one is long-term Service contracts, the other is the product business model." Choose one first, then decide how to position the team.

The mindset of enterprise customers is also changing. A few months ago, some customers said that to budget they first had to lock in a model. Huo Jia countered: what if that model is taken offline by then? More and more customers are finding that for many tasks, putting them on MaaS is optimal for their efficiency and cost. Tokens are still very expensive; he used an analogy: with 3G-era data, would you dare binge-watch shows every day?

China's leading customers don't care how many people a vendor puts in; they only care whether the thing can get done. Huo Jia says he can explain clearly to many leaders that this problem cannot be solved today, and why it cannot be solved.

The Token price curve and the enterprise implementation curve have not yet converged today. The former is driven by AI-native and high-tech customers; the latter must be validated scenario by scenario.

The models released at the Apsara Conference a year ago and the applications at that time — no one imagined they would grow into what they are this year. FDE, outcome-based payment, long-term service contracts — which will become the norm is equally uncertain today.

Huo Jia's reminder is: "Don't lose your imagination just because you don't know."

After four years of training, the starting gun has only just fired.