An AI workstation is a high-power desktop for running AI models locally, and its speed tests do not show Nvidia market share losses. Market share means revenue and units shipped across the whole market, while workstation tests measure only one desktop configuration. Investors see headlines when a rival card wins an inference test. That win matters to buyers of that desktop, but it does not replace filings or shipment audits for the wider market.
Table of Contents
- Where Does Nvidia Earn Its AI Revenue?
- What Do Shipment Audits Show?
- Why Are Workstation AI Tests Fragile?
- What Should Investors Check Instead?
Where Does Nvidia Earn Its AI Revenue?
NVIDIA Newsroom reported fiscal second-quarter 2027 revenue of $96.2 billion, up 106% year over year. NVIDIA's fiscal second-quarter 2027 results. Data Center was about 92.5% of that quarter, at $89.0 billion of $96.2 billion, according to NVIDIA Newsroom. A single desktop result covers only the small remainder and cannot proxy company-wide share.
What Do Shipment Audits Show?
Igor's Lab, reporting Jon Peddie Research data, counted 12.5 million desktop add-in boards shipped in Q2 2026. NVIDIA held about 90%, with AMD at about 8% and Intel at about 2%, in Igor's Lab report on Jon Peddie Research data.
That audit contradicts loss claims built from isolated retailer charts. A store ranking shows what one seller sold in one week. A shipment audit shows the market.
Why Are Workstation AI Tests Fragile?
Puget Systems found workstation AI inference results were fragile. A tokenizer error cut DeepSeek-R1-Distill-Llama-8B from 61.7 to 34.5 tok/s, where tok/s means tokens per second. Virtualization passthrough first broke dual-GPU vLLM testing before bare-metal retesting worked. vLLM is software for serving large models, according to Puget Systems in Puget Systems workstation testing.
MLCommons drew a record 30 organizations and 486 datacenter-plus-edge results for MLPerf Inference v6.1, in MLCommons Inference v6.1 results. It keeps datacenter, edge and workstation in separate categories, according to MLCommons. AMD described its MI350P PCIe card as leading selected NVIDIA RTX PRO 6000 Server and H200 NVL submissions on five closed workloads. AMD framed that as a chosen comparison, not shipment or revenue share.
What Should Investors Check Instead?
NVIDIA said 6 million developers built on CUDA at its 20-year GTC mark, according to NVIDIA Developer Forums. CUDA is the software platform for running code on NVIDIA chips. That base creates switching costs, so a workstation speed win does not quickly become displacement.
Workstation tests omit 72-GPU scaling, qualification, supply and operating cost, according to MLCommons. MLCommons showed Vera Rubin NVL72 at up to 3.7 times prior-generation throughput with 99% scaling efficiency. Judge share from filings and audits, not headlines:.
- Check Data Center revenue trend in the quarterly filing
- Check unit-shipment audits for desktop add-in boards
- Check datacenter MLPerf categories for scaling before acting on a workstation chart