From Model Training to the Delivery Closed-Loop: Enterprise AI Implementation and the AI FDE Model (WAIC 2026)
WAIC 2026 made one thing clear: the bar for enterprise AI has moved from "can the model answer?" to "can it finish the job and be paid on the result?" This is a practical guide to the delivery closed-loop, the training-vs-inference budget question, and the AI FDE (Forward Deployed Engineer) model that ties tokens to business value.
A "delivery closed-loop" means enterprise AI is measured not by whether a model can answer, but by whether it completes the full loop—deploy on-site, find the use case, do the work with tokens and tools, quantify the business gain, and get paid on the outcome. At WAIC 2026 this became the field's most concrete yardstick for AI capability. For most enterprises, the practical takeaway is simple: stop buying model demos, and start buying delivered, measurable results.
What is the "delivery closed-loop," and why did it replace "model capability" as the point of competition after WAIC 2026?
The short answer: the market has decided that a model that can talk is table stakes, and only a model that can deliver is worth paying for.
That shift was the dominant theme of WAIC 2026. In its recap of the show, QbitAI put it bluntly—"AI Chat is now ordinary; doing real work and delivering is the new gold standard—Show me the Agent"—and reported that agents had become the default topic for roughly 70% of exhibitors. Its sharper conclusion for buyers: "whether a system can complete effective delivery is becoming the most concrete and most consequential standard for measuring a large model's capability." The center of gravity, QbitAI argued, has moved to supply–demand matching—which scenarios a model is actually best suited to, and which core business problems it can solve.
Vendors are converging on the same language. In a 36Kr interview, Chaoqing Digital CEO Tang Chunfeng described the industry as having "moved quickly from large-model training to inference-and-application deployment," with attention shifting "from competition over model capability toward the creation of application value." A separate 36Kr conversation with Runjian framed it as an outright change of contest: "AI competition is moving from a competition over model capability to a competition over closed-loop value capability."
Training vs. inference: where should an enterprise put its budget?
The short answer: the training era rewarded the biggest model; the inference-and-delivery era rewards the most efficient token that produces the most business value. For most enterprises, that means budgeting for deployment and unit economics, not for another round of pretraining.
This is not a soft distinction. Chaoqing's Tang describes the current focus as the "AI factory" and token economics—"in essence, how to get more output from less input." The infrastructure bar has risen accordingly: QbitAI reports that a 100,000-accelerator cluster is now the entry threshold for next-generation AI infrastructure, and notes SenseTime's proposal to measure a data center's worth by TPW (tokens produced per watt of power)—a metric that only makes sense once the game is efficient inference at scale, not one-off training runs.
The economics show up on the ground. Runjian's "Runchan" token platform at its Wuxiang Cloud Valley token factory reports a 700% improvement in token throughput; a dedicated 220 kV substation is expected to cut compute cost by more than 30%; and micro-channel liquid cooling plus 800 V HVDC brings PUE below 1.1. Runjian also flags a hard constraint for anyone going global: overseas token costs run roughly 10–30× domestic levels—a reminder that in the inference era, the cost per token, not the parameter count, is what decides whether a use case is viable.
| Dimension | Model-training phase | Inference-and-delivery phase (closed-loop) |
|---|---|---|
| Focus of competition | Model capability, parameters, leaderboards | Supply–demand fit; can it deliver effectively? |
| Success metric | Accuracy, general capability | Business value; TPW (tokens per watt) |
| Infrastructure bar | 100,000-accelerator cluster as entry point | Token throughput and energy efficiency (PUE, compute-power co-design) |
| Cost logic | One-time training spend | Token economics: more output from less input |
| How you pay | Per project / per license | Paid on the business outcome |
What is an AI FDE (Forward Deployed Engineer), and how is it different from traditional delivery?
The short answer: an AI FDE is an engineer who deploys into the customer's environment to find and build high-value use cases—and whose success is measured by quantified business results, not by shipping a system and leaving.
The clearest articulation of the model comes from Runjian, which has formalized it as an AI FDE (AI Forward Deployed Engineer) service system, wrapped in a broader discipline it calls VGE—Value Growth Engineering. As Runjian describes it in 36Kr, the company sends "FDE teams into the customer's site to identify AI use cases and value alongside the enterprise… quantify the economic return of every token, price on the value delivered, and form a complete value closed-loop." The difference from traditional delivery is structural: instead of handing over a system at acceptance and exiting, the FDE team stays embedded, ties its work to measurable gains, and shares the risk of making AI actually pay off.
| Dimension | Traditional project delivery | AI FDE / VGE closed-loop |
|---|---|---|
| Where the team sits | Leaves after handover | FDE team embedded on the customer's site |
| Goal | Ship a system or feature | Find use cases + quantify the economic return of each token |
| Pricing | Per person-day / contract value | Paid on the business outcome |
| Who owns the risk | Customer bears deployment risk | Vendor and customer share it and co-create value |
| Verifiability | Ends at acceptance | Ongoing, quantified gains (lower waste, higher efficiency) |
Can enterprise AI actually be priced on business outcomes?
The short answer: yes, and it is already happening—but only where the vendor can quantify the value it delivers.
Outcome-based pricing is the payoff of the closed-loop, and Runjian is running it in production. Under its VGE model in manufacturing, the company reports seven projects with contracts signed, two already delivered, and its "cognitive foundation" product, Apollo 11, live across "seven or eight" manufacturing projects—concrete evidence that "quantify the return of each token, then price on the outcome" is more than a slide.
The same logic appears in high-stakes verticals. In its 36Kr interview, Chaoqing Digital describes partnering with cancer-drug developer Yikang Pharma in Beijing's Yizhuang district to deliver a full-stack "AI acceleration platform" that both raises R&D efficiency and lowers R&D cost, with multiple projects already in place. The pattern holds: outcome-based pricing works when the outcome is measurable—R&D cycle time, cost per experiment, defect rate—and stalls when it is not.
Why do enterprise AI projects stall out—and how does the closed-loop prevent it?
The short answer: most failed deployments compete on model capability while skipping supply–demand fit and quantified delivery. The closed-loop fixes this by pairing on-site co-creation with hard, measured value.
A useful anchor for what "good" looks like: QbitAI reports that Lingchu Intelligence, running its in-house Psi-R2 model for multi-robot coordination on tasks like insertion testing and quality inspection, cut production waste by about 10%. That is exactly the kind of quantified result the closed-loop is built to produce—and its absence is a good early warning that a project is drifting toward a stalled, "capability demo" fate rather than a delivered outcome.
If you want the deeper diagnosis of why deployment is hard in the first place—organizational, data, and integration friction—we've written about it separately in why enterprise AI deployment is hard. This piece is the complement: not "why it's hard," but "what a working delivery loop looks like." For more implementation playbooks, see our AI implementation writing.
How 6AM approaches AI FDE delivery
We build the way WAIC 2026 says the market now rewards: outcome first. 6AM's Forward Deployed Engineers embed with your team, scope the highest-value use cases, instrument the token economics, and tie our work to metrics you can verify—so the question at the end is "did it move the number," not "was the demo impressive." If you're deciding where AI can pay off first, start with an AI implementation diagnosis.
Frequently asked questions
What is an AI FDE (Forward Deployed Engineer)?
An AI FDE is a delivery role that embeds on the customer's site, bringing tools and tokens to identify and deliver business value directly—rather than shipping a system and leaving. Runjian has systematized this as VGE (Value Growth Engineering), in which FDE teams quantify the economic return of each token and price on the value delivered.
Can enterprise AI be priced on business outcomes?
Yes. Outcome-based pricing is viable wherever the value can be quantified. Runjian runs this closed-loop—"quantify the economic return of each token, then price on the result"—and has already validated it in manufacturing, with seven contracts signed and two projects delivered.
Why do enterprise AI projects stall, and how do you avoid it?
Most stalled projects compete on model capability while neglecting supply–demand fit and measurable delivery. The delivery closed-loop avoids this through on-site co-creation, quantified gains (such as a ~10% reduction in production waste), and settlement tied to the business outcome. For more, see our FAQ.
