6 rivals tracked across 2 segments — direct datacenter GPU competitors and the cloud providers designing their own silicon. Threat level reflects displacement risk to NVIDIA's datacenter position, not share price.
Two kinds of competition sit in this list. Direct rivals sell an accelerator you could buy instead — the question there is software: whether your models run on their stack without a rewrite. Custom silicon is different: hyperscalers building their own chips don't sell them to anyone, they simply stop buying, which removes demand rather than winning a deal.
Announced 27 October at Oracle AI World. Oracle will host Google's TPUs on Oracle Cloud Infrastructure, with availability in calendar 2027 and a multi-year ramp. Financial terms were not disclosed. Oracle continues to offer NVIDIA systems alongside; this adds a second merchant-cloud route to non-NVIDIA accelerators.
On 16 October China's Ministry of Commerce and Ministry of Science and Technology published a draft revision of the technology export catalogue. It covers 40 technologies: 23 in AI (including large models and agents; models of 2 trillion or more parameters could be restricted), several chip-design technologies for AI processing and memory chips, 12 solar and 5 biomedicine items. It is a draft and not yet in effect; no effective date was given.
The Wall Street Journal reported, and Reuters relayed on 2026-10-12, that Amazon is in talks to acquire Israeli AI startup Decart for about $7 billion, against a valuation of nearly $4 billion in May per PitchBook. Decart builds real-time video models (Lucy) and simulated environments for robotics and autonomous-driving training (Oasis). NVIDIA joined Decart's $300 million May round as a new investor. Anthropic had explored a deal of about $6 billion in August and decided against it. Amazon declined to comment, and Decart did not respond. Coverage frames the deal as support for Amazon's Trainium chips, which compete with NVIDIA's GPUs, but that rationale comes from reporting, not from Amazon.
At the Apsara Conference in Hangzhou on 22 September, Alibaba's T-Head unit presented the Zhenwu V900: 216GB of memory, 1,200GB/s of inter-chip bandwidth, native FP8 and FP4, and a claimed three times the performance of the Zhenwu M890 it replaces. More than 1,000 V900s can operate as one supernode over T-Head's ICN Switch interconnect, and Alibaba says a cluster can scale to about 500,000 accelerators. The part is not yet in production. T-Head targets mass production and sale in Q1 2027. Alibaba tied the chip to a goal of 20GW of Alibaba Cloud data-centre capacity by 2032.
At Huawei Connect 2026 in Shanghai on 17 September 2026, Huawei launched the Atlas 960E SuperPoD, which it describes as the industry's first SuperPoD built on near-packaged optics. It scales to 4,096 NPUs under unified memory addressing, is rated at 8 EFLOPS FP8 and 16 EFLOPS FP4, integrates 5,500 Hi-ONE optical units, is fully liquid cooled, and is claimed at 99.8% system availability with over 550kW less power than the conventional approach. In the same keynote, rotating chairman David Wang moved the Ascend 960DT to Q1 2027, three quarters earlier than previously planned, with the 960PR in Q3 2027, the Ascend 970 in 2028 and the Ascend 980 in 2029 — a one-generation-a-year cadence. The roadmap dates are reported from the keynote rather than from a Huawei product page, and none of the Ascend 960 parts are available yet. The competitive point is not the per-chip specification but the scale-up domain: Huawei is answering rack-scale NVLink systems with optically interconnected pods measured in thousands of NPUs, which is the level at which large training and inference jobs are actually bought.
NVIDIA announced on 2026-09-03 an agreement to acquire Hugging Face, the open model and dataset hub, for $12,930,300,000. The deal had not closed as of this record, and NVIDIA's announcement did not state a closing date. NVIDIA states Hugging Face will keep supporting open-source and open-weight models from across the ecosystem with multi-cloud and multi-accelerator deployment, and that NVIDIA compute will not be required to build on or deploy through the platform. A software and distribution acquisition; no silicon or supply change.
The Information reported, as carried by Investing.com on 2026-08-31, that ChangXin Memory Technologies (CXMT) has begun low-volume HBM3E production, with broader commercial availability targeted for 2027. Alibaba's T-Head and Cambricon are reportedly testing the memory for processors as early as next year. The report cites low yields and restricted access to advanced equipment, and places CXMT's HBM three to five years behind SK Hynix, Samsung and Micron. CXMT has not confirmed it.
0 of 6 threat assessments carry citations; refresh them with the /nvda-research pass.