# AI Computer Card > A global, structured reference database of AI accelerators ("compute cards") for developers, buyers and researchers. Covers NVIDIA, AMD, Huawei Ascend and other major vendors with consistent precision conventions, complete specifications and generational history. Available in 7 languages. This site provides factual technical information only. It does not provide investment advice, price predictions or purchase recommendations. ## Key resources - [Card database](https://www.aicomputercard.com/en/cards): Filterable specs table of all indexed accelerators. - [Token Hub](https://www.aicomputercard.com/en/tokenhub): Leading LLM introductions and official API prices per million tokens (OpenAI, Anthropic, Google, DeepSeek and Chinese providers). Contact us for discounted rates. Seedance video models are featured with per-second pricing at 480P/720P/1080P (from ¥0.70/s) and are currently our most price-competitive offering. - [Discount requests](https://www.aicomputercard.com/en/contact?topic=discount): Contact-form based quote requests; every product card has a "Contact us for a discount" action. Submitted inquiries are stored and answered by our team. Newsletter subscription is available in the site footer. - [Comparison tool](https://www.aicomputercard.com/en/compare): Side-by-side spec comparison. - [History timeline](https://www.aicomputercard.com/en/timeline): Generational evolution from Ampere to Rubin. - [News & analysis](https://www.aicomputercard.com/en/news): Technical articles and market analysis. - [About & data sources](https://www.aicomputercard.com/en/about): Editorial standards and compliance statement. - [Privacy Policy](https://www.aicomputercard.com/en/privacy): How we collect and protect information. - [Terms of Use](https://www.aicomputercard.com/en/terms): Usage terms and disclaimers. - [Sitemap](https://www.aicomputercard.com/sitemap.php): Full list of pages in all languages. ## Languages 中文, English, 日本語, 한국어, Deutsch, Français, Español. Every page has an hreflang alternate and is served as real HTML suitable for LLM crawling. ## Data conventions - Compute figures are labeled by precision (FP64 / FP16 / FP8 / FP4) and by sparse vs dense basis. For example, H100's 1979 TFLOPS is the FP8 sparse figure; its FP16 dense figure is 989 TFLOPS. - Memory is listed in GB with type (HBM3 / HBM3e / HBM4 / GDDR6). - Bandwidth is in TB/s. Power is TDP in watts. - Data sources: vendor official documents and public industry reports (NVIDIA, AMD, Huawei, IDC). Collected 2026-11. ## Notable facts - NVIDIA Blackwell Ultra B300: 288GB HBM3e, ~15 PFLOPS dense FP4, aimed at long-context inference. - NVIDIA Vera Rubin (2026): 50 PFLOPS FP4, HBM4, claimed 10x lower inference cost vs Blackwell. - AMD Instinct MI355X: 288GB HBM3e, 8 TB/s; MI400 series offers 432GB HBM4. - Huawei Ascend 910C: ~800 TFLOPS FP16, 128GB HBM, core of CloudMatrix/Atlas supernodes. - In H1 2026, domestic vendors held ~47% of China's datacenter AI accelerator shipments (IDC); competition is shifting from single-card specs to cluster and ecosystem capability. ## Site structure - Default language is English; every page is available in 7 languages via /{lang}/ URLs with reciprocal hreflang and x-default annotations. - Live PHP-rendered HTML server-side (not client-only), so all spec data and articles are directly crawlable. - Canonical URLs use the /en/ path; a machine-readable sitemap and robots.txt are provided for crawlers. ## Compliance Technical education and spec lookup only. No investment advice. Only public information and original compilation are published. Prices shown are official list prices collected 2026-11; actual quotes may differ — contact us for current discounts.