Alibaba's Chip Independence Pitch Is Bigger on Slides Than on Servers

Alibaba wants to buy fewer Nvidia chips, and it wants that shift to show up on its bottom line. On the company's Q1 fiscal 2027 earnings call this week, CEO Eddie Wu called self-developed silicon "a long-term and important direction," telling analysts that deploying more homegrown chips would push gross margins higher even beyond the already-strong margins Alibaba earns on the commercial chips it currently buys.

The numbers behind that ambition are real. Alibaba's cloud unit posted $7.14 billion in revenue this quarter, up 45% year over year and its best growth rate in 22 quarters, with cloud EBITDA margin climbing 4.4 percentage points to 11.6%. AI product revenue hit roughly $1.82 billion, marking a twelfth straight quarter of triple-digit annual growth. CFO Toby Xu told the call that AI servers now recoup their five-year operating costs in just three years, and that older V100- and A100-era hardware is running at "near full capacity" — a sign Alibaba needs more compute however it can get it, whether from Nvidia or from its own T-Head chip unit.

That urgency explains why cloud has become, in the company's own framing, its most certain growth lever. Group revenue rose 9% to $39.64 billion, but adjusted earnings per share fell 42% year over year as AI infrastructure spending — capex hit $10 billion this quarter, up 75% — outpaces the profit cloud and AI are generating. Meanwhile e-commerce, still Alibaba's largest business, grew only about 4%. Investors read the mixed signals cautiously: BABA shares dipped roughly 5% intraday on the EPS miss before closing up about 1.3%, and analysts covering the earnings call, per Benzinga and Seeking Alpha transcripts, stayed broadly supportive of the AI capex bet.

T-Head, Alibaba's chip design arm, has been building toward this moment since its 2019 debut with the Hanguang 800 inference chip. It followed with the Yitian 710 Arm-based CPU, then a PPU GPU line that reportedly became China's highest-shipping domestic GPU in 2025, and most recently the Yitian 810E in January 2026, a chip Alibaba positions between Nvidia's A800 and H20 in raw capability. Reports the same month pointed to a possible spinoff or IPO for the unit. But T-Head's footprint remains modest next to the market leader: roughly 650 external customers use its chips, compared with more than 120,000 for Amazon's Graviton processors — a gap that underscores how early this substitution effort still is.

Alibaba isn't just chasing hardware independence. The company has also been pushing an open-source AI software stack it calls Zhenwu, aimed at loosening the grip Nvidia's CUDA platform holds over how AI models get trained and served. That software layer matters as much as the silicon: even chips that match Nvidia on paper are of limited use if the tooling built around them can't run the workloads customers already depend on.

The backdrop for all of this is a policy environment that has whipsawed for four years. Washington imposed sweeping export controls on advanced chips to China in October 2022, tightened them through 2023 and 2024, then cut off sales of Nvidia's China-specific H20 chip in 2025 — a move that cost Nvidia a $4.5 billion charge and a guided $8 billion in lost China revenue. The Trump administration reversed course in December 2025, allowing H200 sales into China, a decision the Bureau of Industry and Security codified into rule the following month. Each swing has given Chinese cloud providers fresh reason to hedge with domestic chips rather than bet their infrastructure plans on Washington's next move.

Alibaba isn't alone in that hedge, and it isn't relying solely on its own chips either. Huawei's Ascend line is on a roadmap that includes the 950PR in the first quarter of 2026 and the 950DT by the fourth quarter, with a 960 chip slated for 2027 — though industry reports suggest both 2026 chips actually offer less processing power than Huawei's current 910C, a hint that SMIC, Huawei's domestic foundry partner, is struggling with yield and scaling at advanced nodes. That hasn't stopped the industry from betting on Huawei anyway: ByteDance is reportedly planning $5.6 billion in Huawei chip purchases in 2026, and Alibaba and Tencent are said to be placing large orders of their own. In other words, Alibaba's chip strategy is diversification across several domestic and international suppliers, not a bet on any single one, including its own.

Baidu, whose Kunlun chips have anchored China's AI-cloud market for six consecutive years, and a Beijing procurement list that now favors nine domestically certified AI chips, round out a landscape where self-sufficiency is becoming a competitive requirement as much as a cost play. Nvidia itself estimates the China AI-accelerator market at roughly $50 billion, which gives a sense of the prize domestic chipmakers are fighting to capture piece by piece.

The gap between ambition and capability is still wide, though. SMIC's yields at 7nm reportedly run between 20% and 40%, compared with more than 90% at Taiwan's TSMC — a hard constraint on how fast Chinese foundries can scale output of competitive chips. The South China Morning Post has reported that China's leading AI labs continue to train their frontier models on Nvidia hardware specifically because of CUDA's ecosystem lock-in, and that domestic chips still underperform on coding benchmarks. A Council on Foreign Relations analysis goes further, arguing Huawei's chips still can't match Nvidia's and that export controls retain real leverage over China's AI buildout.

That leaves Alibaba's "fewer Western chips" pitch to investors as more nuanced in practice than in press-release framing. Domestic silicon looks increasingly viable for inference work and cost-sensitive deployments — exactly the kind of workload growth Alibaba's cloud unit is currently riding. Training the frontier models that define the top of the AI market, though, still runs through Nvidia's chips and the CUDA software built around them, a distinction Wu's comments to investors this week didn't dwell on.