$100 billion. That’s the combined cloud and AI revenue target Alibaba has set for 2031, and it’s the number that explains everything else the company has been doing for the past year. You don’t chase a figure like that by renting someone else’s silicon.
Alibaba has built its own AI stack from the bottom up. Proprietary CPUs. High-performance AI accelerators. A rebuilt cloud layer. And sitting on top, the Qwen family, with Qwen 3.7-Max explicitly designed for agentic workloads. The company was named to the 2026 TIME100 Most Influential Companies list, and the framing there was telling: Alibaba is trying to convert its lead in open models into something much bigger, a full-stack operation where it owns every layer it depends on.
What full-stack actually means when you’re the one shipping bots
I build agents for a living, so my interest in vertical integration is entirely selfish. I care about three things: what it costs per run, how fast the loop closes, and whether the thing I wrote last quarter still works next quarter.
Vertical integration touches all three. On May 20, 2026, Alibaba unveiled a homegrown AI chip that triples the performance of its predecessor, alongside a new flagship model and that rebuilt cloud. When the same company designs the accelerator, the serving stack, and the model architecture, it can tune them against each other in ways a company assembling parts from three vendors cannot. That’s not marketing. That’s just what happens when the feedback loop between hardware and model design is short.
Agentic workloads are a different animal
The detail I keep coming back to is that Qwen 3.7-Max was designed for agentic workloads specifically. If you’ve built anything beyond a single-shot chat wrapper, you know why that matters.
A chatbot is one request in, one response out. An agent is a loop. It reads a tool schema, decides, calls, parses the result, decides again, and repeats until a condition is met or your patience runs out. A single user task might fire twenty inference calls. The economics and failure modes are completely different from chat:
- Cost compounds per step. A model that’s slightly cheaper per token is dramatically cheaper per completed task when the task takes fifteen turns.
- Latency compounds too. Two hundred milliseconds of overhead is invisible in chat and painful across a twenty-step loop.
- Reliability compounds worst of all. A model that emits malformed tool calls two percent of the time will fail a twenty-step chain roughly a third of the time. Agent builders live and die by structured output discipline.
- Context handling determines what’s even possible. Long-running agents accumulate state, and how the model handles that accumulation sets the ceiling on task complexity.
Designing a model for that profile means optimizing for different things than leaderboard chat scores. If the hardware underneath is being tuned for the same profile, the gains stack.
The competitive picture got noisy fast
Alibaba hasn’t been slow about releases. Qwen 3.8 Max arrived in early August 2026, with the company positioning it as a rival to Anthropic’s best. By late September, Alibaba was out with further details on both the Qwen models and the full-stack infrastructure underneath.
I’d treat vendor comparison claims the way I treat all vendor comparison claims, which is to say I’d run my own evals. But the release cadence itself is the signal worth reading. Shipping that frequently across model and silicon at once suggests the internal pipeline is genuinely working, not just the press release pipeline.
What I’d actually do about it
My advice hasn’t changed, it’s just gotten more urgent. Build your agents against an abstraction layer. Not a heavy framework, just a thin interface between your orchestration logic and whatever model is answering. One adapter per provider, a shared tool-calling format, and your business logic sitting behind both.
Do that and swapping in a Qwen model to benchmark against your current setup is an afternoon of work. Skip it, and you’ve written your infrastructure decisions into your application code permanently.
Then build a real eval suite for your specific tasks. Not general benchmarks, your tasks. Twenty to fifty scenarios covering your actual tool schemas, your actual edge cases, your actual definition of success. When a vendor tells you their model excels at agentic work, that suite is how you find out whether they mean your kind of agentic work.
Alibaba’s open-model track record is the part I find most practically interesting here. Open weights mean you can evaluate without committing, run locally, and understand what you’re getting before you route production traffic through it. Whether the full-stack strategy hits that 2031 revenue target is a question for investors. For builders, the near-term question is simpler and better: does it make your agents cheaper, faster, or more reliable? That one you can answer yourself, this week, with a decent eval use and a couple of hours.
🕒 Published: