NextFin News — Qoder, Alibaba Cloud’s AI coding product, has become a real business. At the Yunqi Conference in September, the team said annualized recurring revenue had reached $100 million. Enterprise accounts already supplied about 70% of revenue by May. The commercial story is clear. The product story is still moving.

A public farewell from a team member, Huang Fengda (also known as Phodal), put a sharper edge on that gap. He thanked colleagues who had built the product with him, then wrote that shipping a good product inside Alibaba is hard—and that many of the obstacles sit outside pure engineering. The remark is one person’s view. It lands against a year of feature launches, retirements, and migration notices that force buyers to relearn where their work lives.

Qoder’s early pitch was straightforward: read a code repository, explain the project, hand a full development task to an agent, then review what comes back. That remains the core for many engineers. Usage did not stay there. At Yunqi, Qoder lead Zhang Liaoyuan said more than 30% of users were not professional developers, and more than half of tasks were not directly about writing code. The same stack that drafts functions can build a web page, clean a spreadsheet, or draft a briefing.

Those jobs look similar in a chat window and diverge in what “done” means. A developer may need tests, a clean merge, and code that someone else can maintain. An operations user may need an editable table and nothing else. Supporting the second path means tracking versions, human edits, and state that a chat log cannot fully store. Huang had argued in an August essay that agents must work on living artifacts—slides, documents, models—not only on a transcript. Peers are stretching in the same direction. Cursor has added planning steps before execution and project layers that keep knowledge in sync across runs. The fight is no longer only about autocomplete. It is about who owns the work after the first draft.

In September, Alibaba Cloud said QoderWork would wind down, with a final shutdown date of September 5, 2027. The notice set out two recommended homes. Development, testing, design, and product work should move to the Qoder desktop client, with accounts, package benefits, task history, skills, memory, and organization settings largely carried over. Document generation, data analysis, research, and web-page work should move to Tongyi office tools (Qianwen Office). That path uses a separate account and benefits system: new registration, no shared credits, refunds of old benefits under a stated process, then a fresh purchase, plus a separate move for organization data.

The split is not clean on features alone. Days after the migration note, a Qoder desktop update added conversation-based site building and direct handling of presentation files—capabilities that sit next to the “office” list pointed at Qianwen Office. Users choosing the office route pay switching cost even when the underlying job feels familiar. Quest, once pushed as a standalone task window, is also being retired as an independent surface, with task-oriented users steered to the new Qoder under existing rights after data import.

Other vendors are consolidating rather than splitting purchase relationships. Anthropic has moved to merge chat and Cowork so work started in one place can continue in one thread, while keeping a code-oriented entry. ByteDance’s TRAE line has been described as folding coding and work agents into a single TRAE aimed at engineers’ wider cycle—requirements, tests, technical writing—while keeping agent and IDE modes. The industry is still searching for a stable shape. Alibaba’s choice puts development continuity on one commercial track and general office work on another.

That choice echoes an older tension inside the group. DingTalk was once expected to drive consumption of Alibaba Cloud resources under a tightly coupled “cloud plus DingTalk” story, then was set on a more independent path, then remained linked to the cloud as channel and infrastructure. Selling infrastructure and selling subscription software pull on different metrics. Shared runtimes and a cloud marketplace can help an application grow. They do not decide which product name owns a customer’s renewal after a reorg.

Qoder’s leadership has stressed infrastructure—models, context, and the harness that lets agents call tools and finish jobs—as the main investment. That strength can improve reliability. It does not remove the need to decide who inherits a paid seat, a permission tree, and a year of project memory when labels on the storefront change. QoderWork’s dual migration makes the trade-off explicit: both routes offer a product; only one keeps the old commercial relationship intact.

Enterprise buyers will configure access, train teams, and embed tools in process. They will also notice when “the same job” requires a refund and a second checkout. Qoder has proven it can sell. The harder test for Alibaba Cloud is whether application software can keep a stable customer promise while the company redraws which App sits on which side of the desk.