Exclusive-Huawei’s new AI chip find favour with ByteDance, Alibaba which plan to place orders, sources say

Huawei plans to ship ~750,000 units of its new 950PR AI chip this year, with samples sent in January and mass production slated to start next month ahead of full shipments in H2. The 950PR is priced at ~50,000 yuan (~$6,900) per card (70,000 yuan for an HBM premium) and is reportedly attracting orders from major Chinese tech firms including ByteDance and Alibaba due to improved CUDA compatibility and stronger inference performance. The launch comes as US export controls have restricted many Nvidia chips in China, making the 950PR a potentially sector-moving domestic alternative for AI inference deployments.

Analysis

A credible domestic alternative that meaningfully lowers software migration friction will re-route procurement cycles inside China: CIOs and cloud teams optimize for cost, latency and vendor risk, not headline FLOPS. Expect procurement mix to shift toward inference-optimized hardware and away from top-end training accelerators in China first; that reallocation should shave incremental GPU demand growth inside China by a few percentage points over 6-12 months, while increasing server and systems spending in the same vendors that integrate the new stacks.

The memory and server supply chain sees asymmetric impacts. Inference-focused designs favor cheaper DDR configurations over premium HBM in many deployments, pressuring HBM pricing and capex plans for HBM-focused suppliers, while lifting volumes for mainstream DRAM and PCIe-connected subsystem vendors. Server OEMs tightly coupled to the incumbent accelerator architecture face margin compression in China; conversely, local OEMs or cloud-native builders that standardize on the alternative stack can lock in multi-year service contracts with sticky revenue and higher integration fees.

Near-term catalysts that could accelerate or reverse these flows are policy (export approvals or fresh restrictions), incumbent product and pricing responses, and developer tool maturation. Regulatory signals can move market share in days; enterprise procurement cycles and mass-deployment economics move in quarters. A durable shift requires both cost parity at scale and an ecosystem of production-ready tooling and support — absent that, adoption will plateau at proof-of-concept and pilot scale.

The consensus risk is binary framing: either ‘incumbent toppled’ or ‘no impact’. Reality is a multi-year, regionally segmented market share transfer with concentrated winners among Chinese cloud/platform integrators and second-order losers in premium HBM and Western server OEM exposure to China. Positioning should be tactical and hedged around regulatory newsflow and product roadmap milestones rather than outright permanent conviction shifts.

Key Decisions for Investors

• Overweight BABA (12-month horizon): increase exposure by 2-3% of portfolio via equity or a 9-12 month call spread to capture upside from accelerated cloud/integration wins in China; target +25-35% upside with a tactical stop at -12-15% if adoption stalls or regulatory headwinds re-intensify.
• Short NVDA China-exposure tactically (3-month horizon): buy a limited-risk put spread (e.g., 10-15% OTM) sized to 25-40% of your delta exposure to NVDA to hedge downside from Chinese share loss and pricing pressure; risk equals premium paid, potential pay-off ~2-4x if China revenue guidance materially disappoints.
• Long APP (6-12 months): buy 6-12 month calls or add small equity stake to play increased monetization from cheaper inference enabling richer product features; asymmetry: expect +20-30% on positive adoption vs -18-22% downside if user monetization fails to accelerate.
• Pair trade: long BABA / short SMCI (6-12 months): express regional divergence by overweighting Chinese cloud/platform execution (BABA) and reducing exposure to US-centric server OEM sales into China (SMCI); target net positive carry and aim for a 2:1 upside/downside payoff where a 20% BABA move offsets a 10-12% SMCI adverse move.