Confirmed in a primary source 22 Sept 2026 Compute

Alibaba said its model has been improving itself for a month, and published the score.

On September 22 at its Apsara Conference in Hangzhou, Alibaba said its Qwen3.8-Max model completed 33 rounds of self-improvement over one month with no person writing the changes, moving its score on the independent Artificial Analysis index from 40 to 45, and that in a separate run a model spent more than 60 hours and over 10,000 calls to chip design software cutting a chip layout's area by 42 percent. The same presentation launched a new accelerator, the Zhenwu V900, and committed Alibaba Cloud to more than 20 gigawatts of data centre capacity by 2032.

The caveat that matters is that every one of those numbers is Alibaba's own, measured inside Alibaba, with nobody outside the company given the model, the harness or the logs to check them.

Until now

Recursive self-improvement was a hazard frontier labs described in safety documents. On September 21 OpenAI proposed that progress in it be measured by shared standards, and no company had published a number.

Now

Treat recursive self-improvement as a figure a listed company now reports to investors, and treat this particular figure as unaudited until somebody outside Alibaba is allowed to run the test.

The chip is the part that costs money and the part everyone will write about. It is real: 216 gigabytes of memory, 1,200 gigabytes per second between chips, native 4-bit arithmetic, three times its predecessor by Alibaba's own measurement, mass production in the first quarter of 2027, clusters specified to half a million cards. Around it sits the rest of a stack nobody outside China is assembling in one place: server processors, network chips, 20 gigawatts of data centre capacity promised by 2032, and 53 billion dollars over three years to pay for it. None of that is a surprise. Alibaba has been saying a version of it since 2025, and the capital number was restated rather than raised.

The surprise is four sentences in the model section. Alibaba says Qwen3.8-Max ran 33 cycles of improving itself over one month, unattended, and that an independent index score moved from 40 to 45. It says a model spent sixty hours and ten thousand tool calls redesigning a chip and cut its area by 42 percent. The day before, OpenAI published an architecture for global AI standards whose first named measurement was evaluation of progress in recursive self-improvement, and whose printed sentence ruled out licences, prerelease review and approval. Twenty-four hours later a Chinese company supplied the measurement itself, in the format of a product slide, with no outside party holding the instrument. Treat recursive self-improvement as a figure a listed company now reports to investors, and treat this particular figure as unaudited until somebody outside Alibaba is allowed to run the test.

How it got here

Aug 2026
Alibaba raises 80 billion Hong Kong dollars, about 10.2 billion dollars, in a share placement to fund AI infrastructure.
Sep 20
Alibaba's Qwen team changes the licence on its image model. Carded in Edition 028.
Sep 21
OpenAI publishes its standards architecture, naming shared measurement of recursive self-improvement and ruling out licences and approval. Carded in Edition 029.
Sep 22
At Apsara, Alibaba publishes its own recursive self-improvement numbers, launches the Zhenwu V900 and sets a 20 gigawatt capacity target for 2032.
Who has to react
Nvidia, whose China business the Zhenwu line is built to replace and whose interconnect numbers now have a published Chinese comparison; the United States Bureau of Industry and Security, which has to decide whether a domestic Chinese accelerator at this specification changes what export controls are for; OpenAI and Anthropic, both of which have argued in the last two weeks that the frontier should be paced and neither of which has published a self-improvement figure of its own; the AI safety institutes named in OpenAI's September 21 architecture, none of which has a way to test a Chinese model it cannot obtain; Artificial Analysis, whose index has just been cited as evidence for a claim it did not make; Alibaba's cloud customers and the 650 T-Head chip buyers, who now have a 2027 supply date to plan against; Korean and Taiwanese memory suppliers, since 216 gigabytes per card is a high-bandwidth memory order; and the Chinese grid operators who would have to deliver 20 gigawatts.

The evidence

ObservedTuesday September 22, 2026, Hangzhou. At the 2026 Apsara Conference, Alibaba chief executive Eddie Wu presented a full-stack plan spanning chips, cloud, models and agents, and Alibaba's own newsroom published the figures.

The whole observation, as recorded

On self-improvement: Qwen3.8-Max completed 33 iterative self-improvement cycles over one month, and its Artificial Analysis score moved from 40 to 45. In a separate chip design experiment the company reports more than 60 hours of self-improvement, more than 10,000 calls to electronic design automation tools, and a 42 percent reduction in chip area. On models: Qwen 4 is in training, and Qwen 4.5 and Qwen 5 are planned at 5 to 10 trillion parameters, against Qwen 3.8 Max at about 2.4 trillion. On silicon: the Zhenwu V900 carries 216 gigabytes of memory and 1,200 gigabytes per second of chip-to-chip interconnect bandwidth, with native arithmetic down to 4-bit floating point; Wu called it the most powerful AI chip in China today and three times the performance of the Zhenwu M890 it replaces; mass production is set for the first quarter of 2027 and a single cluster is specified to hold up to 500,000 cards. Alibaba's T-Head chip unit reports more than 650 customers across automotive, finance, large language models, energy and manufacturing, and new Yitian 720 and 730 server processors arrive in 2027, the 730 on T-Head's own microarchitecture and rated 40 percent faster per gigahertz than the Yitian 710. On land and power: Alibaba Cloud will operate more than 20 gigawatts of global data centre capacity by 2032. On money: the three-year commitment of 380 billion yuan, about 53 billion dollars, was restated rather than raised, and Alibaba raised 80 billion Hong Kong dollars, about 10.2 billion dollars, in a share placement in August 2026. On products: QwenWork reached 30 million users in its first month. Wu's framing, in his own words: Machine Thinking currently represents less than 3% of all Human Thinking. If that scales to 1,000x human capacity, Machine Thinking still has enormous growth runway. The market moved little: Hong Kong-listed shares closed about 2 percent higher and the American listing traded about 3 percent higher before the New York open.

Established

Alibaba stated on September 22, 2026 at its Apsara Conference that Qwen3.8-Max completed 33 self-improvement cycles over one month and that its Artificial Analysis score moved from 40 to 45.
Alibaba launched the Zhenwu V900 with stated specifications of 216 gigabytes of memory and 1,200 gigabytes per second of interconnect bandwidth, mass production in the first quarter of 2027, and cluster support up to 500,000 cards.
Alibaba Cloud stated a target of more than 20 gigawatts of global data centre capacity by 2032, and restated rather than raised its 380 billion yuan three-year capital commitment.
Alibaba stated that Qwen 4 is in training and that Qwen 4.5 and Qwen 5 are planned at 5 to 10 trillion parameters.

Not established

That the self-improvement result is what it sounds like. No third party received the model, the loop, the harness or the logs, a five point move on a published index has more than one way to be produced, and Artificial Analysis did not report this.
That the chip performs. There is no independent benchmark of the Zhenwu V900, mass production has not begun, and a 500,000 card cluster is a specification rather than a machine anyone has run.
That new money was committed. The 380 billion yuan figure is the existing one, and no new capital number was announced on the day.
That 20 gigawatts is procurable. No site, utility, interconnection agreement or transformer order was named, and grid connection is the binding constraint on every such target anywhere in the world.

Why it weighs this much

Signal strength

CostHigh

CommitmentHigh

ReversibilityHard

PersistenceYears

Cost is high but not uniformly so, and the two halves of this announcement deserve different weight. The silicon and the capacity are expensive: a 216 gigabyte accelerator taped out and scheduled for volume in one quarter, 650 existing chip customers, 53 billion dollars committed and 80 billion Hong Kong dollars raised in August to help pay for it. The self-improvement claim cost a slide. Commitment is high on the same split: a mass production date and a 2032 gigawatt figure are quotable back at the company, and Wu made both in his own name. Reversibility is hard for the parts that involve fabs, land and grid connections, which is most of the money. Persistence is years: a chip specified today is in racks until 2031, and a data centre commitment to 2032 is a decade of procurement. What we are not doing is letting the expensive half lend its weight to the cheap half.

What to watch

strengthens An independent evaluator, Artificial Analysis included, publishes a measured score for a Qwen model produced by the self-improvement loop Alibaba described, or Alibaba publishes the loop's harness or logs, by March 31, 2027.
weakens Zhenwu V900 mass production slips past the first quarter of 2027, or no third party publishes a benchmark of the part, and Alibaba discloses no further self-improvement results, by June 30, 2027.

The words

Recursive self-improvement: A model improving itself, then using the improved version to improve itself again, with no person writing the next change. It is the mechanism safety frameworks name as the point where progress stops being predictable.
Parameters: The adjustable numbers inside a model. More of them usually means more capacity to learn and far more hardware to train and run.
Artificial Analysis: An independent firm that scores AI models on a common set of tests and publishes one index number. Alibaba quoted a move on that index; the scoring of the self-improved versions was reported by Alibaba, not by Artificial Analysis.
Electronic design automation: The software chip engineers use to lay out a processor. Counting tool calls is a way of counting how much design work happened without a person in the loop.
Gigawatt: A billion watts, roughly the output of a large nuclear reactor. Data centre capacity is now quoted in the same unit as power stations.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.