You are using an older version of the IE browser. 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, used only for article preview, and will expire at that time
0 
"Are the cloud giants about to get their hands dirty with heavy lifting again?"
Recently, Amazon Web Services (AWS) spent $1 billion to build a team of "AI resident engineers." This seemingly "backward-looking," asset-heavy move has sparked considerable controversy in the tech world.
You see, for a long time, the international cloud giants' favorite story was "extreme standardization": open up public cloud APIs, then lie back and count the money. Heavy on-site presence? Basically nonexistent.
But in 2026, this model of lying back and selling standardized products may be about to change.
AWS is spending $1 billion to build a team of thousands of "embedded AI engineers" — that is, FDEs (Forward Deployed Engineers) — who will go straight into customers' offices and hand-hold enterprises through deploying AI agents and building AI production systems.
This is not an isolated case.
Over the past six months, AI and cloud giants like Google, OpenAI, Anthropic, and Meta have all been aggressively setting up FDE positions or dedicated companies, with total investment exceeding tens of billions of RMB; domestically, Alibaba, ByteDance, Tencent, and other LLM-plus-cloud giants have also been intensively assembling their own "on-site iron armies" in the first half of this year.
"Deep on-site presence," "high-end outsourcing," "IT construction crews"... these somewhat negative, outdated labels have now reappeared in the narratives of the AI cloud giants.
Why are the cloud giants, who once pursued making money effortlessly, suddenly so keen on on-site customer engagement in the AGI era? Will they repeat the "unlimited customization" mistakes that domestic cloud giants once made? Behind this dramatic reversal — from collecting rent while lying down to on-site construction — what industry changes are actually hidden?
The work AWS's embedded AI engineers have to do is not easy.
First, they need high-level skills. LLM fine-tuning, multi-layer agent orchestration, advanced algorithm optimization like RAG, and fine-grained token cost reduction... these AI engineering capabilities cannot slip up.
Second, they must not only provide technical solutions, but also roll up their sleeves to help customers write low-level code, do data cleaning, and even restructure business processes.
No wonder that when the FDE role comes up, many people's minds conjure shadowy labels like "unlimited customization," "led around by the nose by customers," and "high-end outsourcing experts."
But the question is: will the cloud giants' on-site customer engagement in the AI era be the same as their private deployment work for government and enterprise customers back in the cloud computing era?
At the very least, AWS has offered a different attempt.
Leiphone has observed that AWS has imposed at least three restrictions on the "embedded AI engineer" service.
First, the nature of the deliverable is fundamentally different.
To avoid sinking into the quagmire of "unlimited customization," AWS has defined the core of FDE on-site delivery this time not as application-layer custom code, but as the "semantic layer" and "knowledge graph."
In AWS's plan, when engineers go on-site at a customer, their main task is to deploy a standardized semantic layer within the customer's AWS account.
"The so-called semantic layer is like a translator. It connects the customer's scattered enterprise data and internal rules and regulations, generating an AI-driven, governed knowledge graph, so that after the engineers leave, the customer's own AI agents can reason and operate on this graph," explained Zhang Xiaofeng, a partner at an IT consulting firm in Shanghai.
When the deliverable is clear and specific, there is no falling into unfocused, unlimited customization.
Second, AWS has played another big card: breaking the old tradition of "per-head billing" and instead billing by results or fixed increments.
In the past, especially in on-site services for government and enterprise customers, cloud giants were often passively dragged into the quagmire of traditional system integrators, with projects firmly locked into per-head or per-hour billing arrangements.
This model has an inherent paradox: customers want to spend the least money to get the most done, squeezing on-site staff to the limit. As a result, project timelines stretch longer and longer, the cloud giants' gross margin is gradually eaten away by high labor costs, and ultimately every deal loses money.
This time, when AWS launched its $1 billion FDE program, it explicitly stipulated: FDEs will be paid based on "fixed-price outcomes," not per-head billing.
Before entering the site, the giant and the enterprise customer align on one or two very specific goals — for example, using an AI agent to shorten the invoice audit process from 3 days to 45 minutes. Once the goal is achieved, the project is declared delivered.
This "results-oriented" approach forces both sides to focus on core business pain points, preventing customers from endlessly demanding "while you're at it, add a small feature."
Meanwhile, AWS has imposed strict limits on the on-site cycle, shortening it from "months" to "weeks" — the "AI-45 rule."
To reduce on-site costs, AWS has internally set an extremely strict "AI-45 delivery rule" for its FDE teams: 45 minutes to go from a customer pain point to an AI idea; 45 hours to produce a runnable prototype; 45 days to complete on-site embedding and final delivery.
"OpenAI and Anthropic in Silicon Valley are now using a similar 'blitzkrieg' model for their on-site work," Zhang Xiaofeng added.
"Keeping on-site time within a few weeks means the giant's engineering costs are locked in and controllable, avoiding a long-drawn-out 'on-site quagmire.'"
The answer is: not necessarily.
"The root cause is that rapidly evolving AI technology, immature and non-standardized AI products, and scarce hybrid talent, when combined with customers' complex AI expectations, fill projects with a great deal of uncertainty," said Fei Lianghong, former chief architect at AWS.
In his view, the FDE is a currently scarce hybrid talent who needs to serve on-site as project manager, developer, architect, and AI implementer all at once, capable of leading projects and guiding customers to ultimately land AI products such as agents.
As for FDEs who lack experience and skills and rush into the role after hasty transitions, they may well, due to a lack of confidence and capability, end up being led around by the nose by customers like outsourced engineers, ultimately ruining the project.
Moreover, because the agent industry is so new and developing extremely fast, even experienced traditional PMs or consultants overseas face a "dilemma" of insufficient experience and inability to prove themselves to customers when it comes to actually landing the combination of AI and business processes.
"In the past, consultants could still rely on information asymmetry — whether in IT or so-called advanced experience — to impress customers. But now AI technology changes by the day, everyone is on the same starting line, and it's not easy to withstand customers' doubts and convince them how things should be done," Fei Lianghong said.
In particular, today's AI products are not yet a mature standardized product.
Traditional cloud computing is a clear "standardized product"; customers have a very clear path for using it, and technical account managers or Solutions Architects only need to provide support on details like selection, configuration, and cost-effectiveness.
But in the era of agents and tokens, what AI delivers is merely raw, unfinished material like "rebar and cement" — there is no standard answer at all for how to use it or how to use it well.
Moreover, the API model that the giants once believed in was only a story told to technical people; it couldn't communicate with business and process managers. Now, with agents, anyone from any business side can get involved and raise requirements, which further increases the difficulty of the FDE role.
"Everyone is still exploring," Fei Lianghong admitted. Palantir, whose stock price soared and fame surged last year, is arguably the one that has done FDE work the longest and best, but Palantir mainly serves limited customer scenarios like the U.S. government and the Department of Defense, not a fully open, competitive market, so its example lacks the characteristics for large-scale replication.
Meanwhile, large model vendors like OpenAI and Anthropic have only just begun accumulating experience and practice in FDE, and the lessons they have distilled are still insufficient.
In China, this situation is even more complicated.
At present, Lei Feng Wang has learned from multiple sources that major domestic cloud providers are also forming FDE teams, but with varying organizational structures.
Some are still mixed in with the SA track or broader categories, while others select talent from across the company to form standalone teams that report directly to higher-level management. Their work focuses mainly on helping customers sort out their processes and increase token usage.
In the view of Wang Jian, a domestic cloud sales professional, the Silicon Valley giants are being somewhat "too precious" about it. On-site customer service is a standard offering in China. "It's just the work that service managers and service experts did in the traditional cloud era—only now it's been renamed 'large model expert services.'"
"But the awkward thing is, domestic FDE positions can't command a premium." Liu Gang, an IT analyst at a securities firm, told Lei Feng Wang that although FDE work involves AI and is more technically demanding, customers are still unwilling to pay for it, because in their eyes it's no different from the IT construction crew model. "Customers still see you as headcount, not experts, so it falls back into the old rut of customization."
Moreover, in this field in China, there are too many hyper-competitive companies, all of them former software consulting and delivery firms. "It's really just implementation. There was already a lot of algorithm-related delivery before, and now with all kinds of tools available, they lean toward people with solution backgrounds. In China, the FDE role itself is almost a dead end..." said Liu Ning, head of a headhunting firm, helplessly.
For more details and firsthand experiences of FDE roles at major domestic AI companies, feel free to add the author's WeChat xf123a to discuss.
Amid the broader trend of AI and cloud giants going on-site with customers, some say the cloud giants are moving from MaaS (Model as a Service) toward CaaS (Consultant as a Service).
In the past, everyone bet on whose model had more parameters or whose API was cheaper. Now people have discovered that the AI competition has ultimately become a service industry. Whoever can help customers actually put AI to use will lock in future compute orders.
AWS's $1 billion investment is really about "buying the road" to future tens of billions in cloud compute consumption.
And this encroaches on the turf of the traditional consulting and outsourcing giants.
Many IT industry insiders believe that this move by AWS and its peers directly carries the battle into the heartland of traditional IT consulting and outsourcing giants like Accenture, McKinsey, IBM, and companies such as Softstone and ChinaSoft International.
In the past, cloud vendors mainly sold technology while outsourcers did the dirty work. But now, cloud vendors with their own models and chips are not only entering the field themselves—their on-site teams, backed by first-hand models and chips from the original vendor, are so much more efficient at solving large model debugging and optimization problems that a generational technology gap has opened up with traditional outsourcers.
Tech giants are leveraging their underlying technology advantages and FDE roles to nibble away at the upstream profits of the traditional outsourcing industry.
But at the same time, cloud giants spending heavily to help customers deploy AI may give another type of player a free ride, which in turn could threaten the cloud giants themselves.
"Today's AI-driven cloud computing market is more complex than ever, and the traditional cloud computing camp is facing the impact of 'NeoClouds,'" Fei Lianghong told Lei Feng Wang (public account: Lei Feng Wang).
Overseas, NeoCloud vendors represented by Lambda Labs and CoreWeave are primarily AI-oriented, offering compute and a Token Factory model, while domestically there are comparable companies like SiliconFlow.
In his view, as cloud giants' FDE roles accelerate the deployment of customer AI agents, enterprises with mature AI production systems in the future will mainly consume tokens and flexibly deploy applications, becoming less dependent on traditional cloud storage, compute, and networking, and instead relying mainly on APIs and token generation.
This will give the new generation of NeoCloud players the advantage, while traditional cloud computing giants may gradually lose ground. Just as in the US market, the fastest-growing sellers of GPUs and tokens are no longer the traditional established cloud giants.
Back to the original question: do cloud giants really want to do the heavy lifting? The answer is clearly no. They grit their teeth and send AI engineers to the grueling on-site front lines in order to help customers bridge the last mile of AI deployment in the early days of the AI era.
Whether this carefully calculated "new on-site business playbook" will ultimately pay off depends on complex factors. Whether the cloud giants' move will fatten their own token revenue or end up making wedding clothes for the watching "NeoCloud" upstarts remains to be seen.
The author has long tracked the capital moves, cutting-edge technologies, and behind-the-scenes stories of overseas AI giants. Feel free to add the author's WeChat xf123a to exchange ideas.
This article is original content by Lei Feng Wang. Unauthorized reproduction is prohibited. See republishing guidelines for details.
Related Articles
Editor
To secure your account, please verify your email
Please verify your email