MiniMax's H3 Is a Quotation Mark for Alibaba — China's Video-Gen Race Just Reset the AI Investment Cycle

The brief frames this as a Sora/Veo chase, but the live competition is domestic. MiniMax's H3 lands at $0.13/sec for 2K; Kuaishou's Kling 3.0 starts at $0.084/sec standard and ByteDance's Seedance 2.5 is closed-source at unspecified pricing. H3's value proposition is not raw quality — Artificial Analysis ranks it behind Gemini Omni Flash on text-to-video — but the combination of price, open weights, and Chinese-chip compatibility. That bundle is aimed at enterprise and developer adoption inside China, where the buy decision is increasingly local.

TechTimes and MiniMax's own blog confirm H3 was designed from inception to run on Chinese-made semiconductors alongside Nvidia GPUs. The implication is concrete: as MiniMax's $300M ARR base scales via the Hailuo consumer app and the MiniMax-H3 API, marginal inference loads can land on Cambricon or Huawei Ascend 910C silicon rather than H20s. With Nvidia recording zero H20 sales to China in Q2 FY26 and Bernstein projecting Nvidia's China AI-chip share falling from 54% in 2025 to 8% in 2026, H3's chip-agnostic posture accelerates that substitution.

Downstream, the consumer Hailuo app already generates >370M videos globally (MiniMax IR, June 2025) and the API is live on Fireworks AI. Enterprise demand routes through the open weights once shipped — a structural deflationary force on inference compute pricing that benefits customers and hurts any vendor locked into Nvidia-only stacks.

Alibaba sits in three seats at once. It led the $600M Series A into MiniMax at a $2.5B valuation in March 2024 (Wikipedia/Alibaba-led 600M funding); it is the cloud provider hosting MiniMax's training and inference (Alibaba Cloud case study, Jan 2026); and it independently ships HappyHorse 1.0/1.1, which sits at the top of Artificial Analysis' video leaderboard. H3's $0.13/sec 2K pricing undercuts HappyHorse's $0.14/sec on Atlas Cloud — within the same Alibaba ecosystem.

The honest read: Alibaba's Cloud Intelligence Group just posted 38% YoY growth to RMB 41.6B in the March quarter and AI-related cloud revenue has printed triple-digit growth for 11 straight quarters (Alibaba IR). That is the engine. If H3's open weights cannibalize HappyHorse's enterprise pipeline, the equity gain on the MiniMax stake does not offset the cloud and model-revenue loss. That asymmetry — own the lab at one price, or accelerate the in-house model at another — is why Alibaba is the named beneficiary-or-victim of this release.

OpenAI discontinued its Sora video platform; Chinese models now dominate Artificial Analysis' leaderboard after Alibaba's HappyHorse, ByteDance's Seedance and Kuaishou's Kling took the top slots (WSJ, Apr 2026). Google's Veo 3.1 and Gemini Omni Flash remain the only Western models in the mix — and the brief's 'closing on Sora/Veo' framing misses that the live race is internal. H3's actual benchmark is HappyHorse (same Alibaba-linked ecosystem) and Seedance 2.5 (same day, same tier). MiniMax ranks #1 on Artificial Analysis' video-editing arena but trails on text-to-video — meaning HappyHorse still owns the leaderboard crown Alibaba cares about.

MiniMax is the messenger; the message is that Chinese frontier video-gen runs on Chinese silicon at one-third the price and ships open weights within a week. The lab itself is uninvestable outside the Hong Kong float, where the stock trades at ~116x sales and ~155x EV/Sales on $300M ARR — priced for the H3 optionality already. The asymmetric trade is upstream: Cambricon, which already doubled revenue to RMB 2.88B in Q1 2026 with profit +185% YoY, absorbs H3 inference loads that would otherwise have gone to Nvidia. Nvidia's China-share thesis continues to deteriorate on every open-weights release from a chip-agnostic lab.