Why DeepSeek’s Cache Costs Jumped 11‑Fold: Long‑Context Surge and the New “Storage Tax”

DeepSeek, once known for its rock‑bottom pricing, increased the cache‑hit fee from 0.025 ¥/M tokens to 0.3 ¥/M tokens—a 5‑fold rise in idle periods and an 11‑fold surge during peak hours. The article explains that “cache hit” means reusing previously computed KV (key‑value) states stored as a compressed “draft paper” for long‑context prompts.

The low cost previously relied on DeepSeek’s KV compression techniques—Compressed Sparse Attention (CSA) and Heavy Compression Attention (HCA)—which merge 4 to 128 tokens into a single memory block, shrinking the KV memory by over 95 %. Combined with a hot‑cold tiered storage architecture, hot caches stay in fast GPU HBM while cold caches are off‑loaded to inexpensive enterprise‑grade NVMe SSDs, making a 1 M‑token cache occupy only ~10 GB and cost just a few cents per million tokens.

For a practical cost illustration, the author assumes a code‑base agent built on DeepSeek V4 Pro with 500 k tokens and 50 daily queries. Without caching, each query recomputes the entire 500 k tokens, totaling 25 M tokens and >75 ¥ in input cost. With a 90 % cache‑hit rate, only the first round incurs full cost, reducing input expenses by over 90 %.

During high‑traffic periods, massive long‑context requests fill GPU HBM, forcing the scheduler to evict less‑used cache blocks to NVMe. Each eviction moves hundreds of megabytes to gigabytes of data, and subsequent requests must reload these blocks, turning the GPU from a “brain” into a “muscle” that wastes compute cycles on I/O. This phenomenon is dubbed the “storage tax”.

The article notes that OpenAI and Anthropic face similar constraints. Anthropic mitigates pressure by requiring developers to mark cache breakpoints and charging 1.25‑2× a premium on first‑write cache, while OpenAI’s higher base prices allow larger HBM allocations, reducing cache stress.

To keep costs down, developers are urged to maximize cache‑hit rates: ensure prompt prefixes are stable (avoid timestamps, user IDs, random values at the start), prevent extra whitespace or line‑break changes, and avoid inserting tool results that break the cache boundary. Understanding which data is hot (kept in GPU memory) versus cold (stored on SSD) and managing cache lifecycles are essential as inference systems evolve into large distributed storage platforms.