DeepSeek raises peak-hour cache-hit prices by 11x as long-context demand strains storage
DeepSeek raised cache-hit prices by up to 11 times during peak hours on August 17, as surging long-context demand overloaded its storage and I/O systems. The hike reflects a new cost structure for large models.
The cache-hit mechanism works like a scratchpad: once the model processes a long text, it stores the computed state as KV units on storage clusters, so subsequent requests can reuse it without recalculating. DeepSeek's aggressive caching strategy has made it a favorite among developers building code assistants, knowledge-base agents, and long-document analysis tools, because it slashes the cost of repeated computation.
But the surge in long-context usage has turned this cheap feature into a costly liability. During peak hours, massive requests with millions of tokens flood GPU memory, forcing the scheduler to evict idle cache blocks to slower NVMe storage. Frequent data movement between memory and disk consumes enormous I/O resources and causes GPU clusters to stall, effectively turning expensive compute into data-moving chores. Leiphone describes the price increase as a "storage tax" that reflects the real cost of storing and moving data.
To illustrate the math, a 500,000-token codebase with 50 rounds of queries would cost over 75 yuan in input tokens without caching. With a 90% cache hit rate, only the first round pays the full precomputation cost, cutting input costs by more than 90%. DeepSeek's V4 uses two attention mechanisms—compressed sparse attention and heavily compressed attention—to shrink KV cache by over 95%, and it stores hot and cold memory in different tiers. Even with this compression, the sheer scale of demand has hit physical limits of GPU count, HBM capacity, and bus bandwidth.
Competitors face the same storage and I/O pressure. Anthropic imposes a 1.25x or 2x premium on first-time cache writes, forcing developers to mark cache breakpoints and only cache high-frequency prefixes. OpenAI, with its higher base prices, can afford more HBM memory to reduce cache pressure. DeepSeek, having priced its cache so low, has no room to absorb the load and had to raise prices.
Despite the increase, DeepSeek remains one of the most cost-effective options in the market, with costs still a fraction of Claude's. Switching providers would mean abandoning existing cache pools, so developers are better off optimizing cache hit rates. Common mistakes that invalidate caches include putting timestamps, user IDs, or random parameters at the start of a prompt, inserting tool results in the middle of the context, or accidentally changing whitespace or line breaks. Developers must now act as data administrators, carefully arranging hot and cold data and managing cache lifecycles.
The change signals a historic shift in large-model economics: compute is becoming cheaper, while state storage and data movement are turning into the most expensive parts. For developers, mastering cache hits has become a new dividing line in an era of fine-tuned efficiency.