What Happened
On September 22, Alibaba used its annual Apsara Conference in Hangzhou to lay out what it calls a full-stack artificial intelligence strategy, one that runs from the chips inside its data centers to the AI agents embedded in enterprise software. Alibaba Cloud CTO Li Feifei introduced the centerpiece of the event: AgentCore, an enterprise-grade platform for building, running, and governing AI agents that Alibaba says can raise task completion rates by 99% and cut the total cost of ownership of running agents by 70%.
The company paired that software announcement with new hardware. T-Head, Alibaba's chip design unit, unveiled the Zhenwu V900, a processor the company describes as the most powerful AI chip built in China to date. The V900 triples the performance of its predecessor, the M890, carries 216 GB of memory, and can be linked into clusters of up to 500,000 chips. Alibaba also moved up mass production of the chip from the third quarter of 2027 to the first — a six-month acceleration that suggests the company wants supply on hand before rivals close the gap.
Alibaba Group chairman Joe Tsai and CEO Eddie Wu both spoke at the event, framing the announcements as chapters in a longer strategy rather than a one-off product launch. "Machine Thinking represents less than 3% of Human Thinking, with enormous growth runway ahead," Wu told the audience. Tsai said intelligence "goes beyond guiding AI from technological breakthroughs toward value creation" — language that points to Alibaba's push to turn its AI spending into revenue, not just research headlines.
Why It Matters
AgentCore is Alibaba's answer to a problem that has slowed corporate adoption of AI agents: they tend to be expensive and unreliable to run at scale. The platform adds long-running task support, automatic retries, and checkpoint recovery — the infrastructure plumbing that turns a demo into something a bank or logistics company can depend on. Alibaba also introduced Agent Context, which it says reduces token usage by up to 67% in knowledge-heavy applications, and a caching system called Tair KVCM that claims a 99% cache hit rate and cuts the cost per token in half.
Those figures matter because token costs are what decide whether an AI agent stays a novelty or becomes a business tool. If Alibaba's numbers hold up under real workloads, AgentCore would undercut rival platforms from Microsoft, Google, and Amazon on price — a familiar Alibaba Cloud strategy of competing on cost while the underlying models converge in quality. The push comes as Alibaba's capital spending on AI infrastructure jumped 75% to 67.7 billion yuan in the June quarter, even as profit fell — a bet that only pays off if enterprise customers actually show up.
Reaction
Investors gave the announcement a mixed but net-positive reception. Alibaba's Hong Kong-listed shares rose about 3% in the session that followed the Apsara keynote, even as the stock remained down roughly 25% for the year on its U.S. listing. Analysts described the market's response as pleased but cautious, reflecting how much of Alibaba's AI narrative now rests on execution rather than announcements. On the hardware side, Alibaba said its Zhenwu chips have already found more than 650 external customers, up from around 400 in May — evidence that demand for a domestic alternative to Nvidia hardware is real regardless of how AgentCore performs.
What's Next
The roadmap Alibaba laid out stretches years, not months. Qwen 4 is already in training, and later versions are expected to scale to between 5 and 10 trillion parameters — up to four times the size of today's flagship Qwen 3.8-Max. T-Head's next CPUs, the Yitian 720 and 730, are slated for 2027, the same year the Zhenwu V900 is due to reach mass production. Alibaba Cloud also said it plans to grow its global data center capacity to more than 20 gigawatts by 2032, a target that would put it in the same league as the largest U.S. hyperscalers. Competition is not standing still either: Huawei has pulled forward its own Ascend 960DT chip to early 2027, setting up a direct collision with the V900 on Alibaba's home turf.
Closing Thoughts
What stood out about this year's Apsara Conference wasn't any single number — the 70% cost cut, the 500,000-chip cluster, the 20-gigawatt target — but how deliberately Alibaba tied them together into one pitch: cheaper agents, run on cheaper chips, trained on models that keep getting bigger. It is a strategy that only works if all three layers hold up under pressure at the same time.
Eddie Wu's framing of machine thinking as still in its early innings is a useful way to read the whole event. Alibaba is not claiming to have solved enterprise AI adoption; it is betting that whoever makes agents cheap and reliable first will own the layer everything else gets built on. Whether AgentCore's numbers survive contact with messy, real-world enterprise data is a question Alibaba will have to answer at its next earnings call, not at a conference keynote.
9월 22일 알리바바가 항저우에서 열린 연례 컨퍼런스 '항저우 클라우드 서밋(Apsara Conference)'에서 반도체부터 AI 에이전트까지 아우르는 풀스택 AI 전략을 공개했다. 알리바바 클라우드 CTO 리페이페이는 기업용 AI 에이전트 플랫폼 '에이전트코어(AgentCore)'를 소개하며, 작업 완료율을 99% 끌어올리고 총소유비용(TCO)을 70% 줄일 수 있다고 밝혔다. 동시에 반도체 설계 자회사 T-Head는 이전 모델 대비 성능이 3배, 최대 50만 개 칩까지 클러스터 연결이 가능한 신형 AI 칩 '전우(Zhenwu) V900'을 공개했으며, 양산 시점도 2027년 3분기에서 1분기로 앞당겼다.
조 차이 알리바바 그룹 회장과 에디 우 CEO는 이번 발표를 단발성 신제품 공개가 아닌 장기 전략의 일환으로 강조했다. 에디 우는 "머신 씽킹은 아직 인간 사고력의 3%에도 못 미치며, 성장 여력이 막대하다"고 말했다. 발표 이후 알리바바의 홍콩 상장 주가는 약 3% 상승했지만, 올해 들어 미국 상장 주가는 약 25% 하락한 상태다. 2분기 자본지출은 전년 대비 75% 급증한 677억 위안을 기록했고, 전우 칩의 외부 고객사는 5월 약 400곳에서 9월 650곳 이상으로 늘었다.
알리바바는 향후 몇 년에 걸친 로드맵도 함께 제시했다. 차세대 모델 '큐원(Qwen) 4'는 이미 학습 중이며, 이후 버전은 최대 5조~10조 파라미터 규모로 커질 전망이다. 2032년까지 전 세계 데이터센터 용량을 20기가와트 이상으로 늘리겠다는 목표도 내놨다. 다만 화웨이가 경쟁 칩 '어센드 960DT'의 출시 시점을 2027년 초로 앞당기는 등 경쟁도 만만치 않아, 이번 전략이 실제 기업 고객 확보로 이어질지는 다음 실적 발표에서 가늠할 수 있을 전망이다.
참고: TechAfrica News, KuCoin, TheStreet