Key points
- Permanent 75% price cut: On May 23, DeepSeek cut V4-Pro API pricing permanently (not a limited discount): cache-hit input at $0.003625/M tokens, cache-miss input at $0.435/M, output at $0.87/M. Cache-hit pricing is ~0.7% of GPT-5.5's $5.00/M input price—a two-orders-of-magnitude gap, framed as a de facto industry-standard declaration.
- Technical lineage: V2's MLA compressed KV caches to 512-dim latent vectors; V3.2's DSA introduced selective sparse attention; V4's CSA/HCA hybrid compresses KV entries (HCA for global context, CSA for key details). At 1M-token context, V4-Pro reportedly needs only 27% of V3.2's per-token inference FLOPs and 10% of its KV cache memory.
- Engram conditional memory: A second sparsity axis storing static knowledge (entity names, formulas) in lookup tables with O(1) retrieval. Its deterministic index (unlike MoE routing) allows offloading a 100B-parameter table to cheap CPU memory with asynchronous PCIe prefetch; H800 inference throughput drops less than 3%. A reported U-shaped scaling law suggests 75-80% of sparse parameters in MoE and 20-25% in Engram is optimal.
- TileLang: DeepSeek's open-source Python-style DSL for GPU kernels, compiling to CUDA, CANN, and OpenCL. FlashAttention shrank from 500+ lines of CUDA to 80 lines of TileLang with equal or better performance. All V4 operators were rewritten in TileLang, and eight Chinese chip vendors (Huawei Ascend, Cambricon, Moore Threads, Hygon DCU, MetaX, Kunlunxin, T-Head, Iluvatar) achieved Day-0 adaptation—NVIDIA and AMD did not receive preview access.
- Hardware economics: Huawei's official benchmarks show Ascend 950PR single-card decode throughput of 4700 TPS for V4-Pro (~20ms TPOT) and 1600 TPS for V4-Flash (~10ms). A 100-card Ascend cluster (~RMB 10M+) costs about one-third less than an equivalent H20 setup (~RMB 15M); with lower per-FLOP power draw, total savings may reach 60-70%.
- Ecosystem inversion: The author argues DeepSeek is flipping 'chips define models' to 'models define chips'—chip vendors now compete on how efficiently they run open-weight DeepSeek-V4 rather than on CUDA compatibility.
Three scenarios
1. Optimistic — China AI's 'Android moment': CSA/HCA + Engram + TileLang become the de facto standard for the non-CUDA camp, driving chip sales, cloud adoption, and cheaper inference. Key validation: Ascend 950 supernode pricing and pricing after H2 2026 volume shipments, plus V4 trainability on purely domestic clusters. 2. Neutral — dual-track 'cold war': CUDA and non-CUDA ecosystems coexist. NVIDIA remains hard to replace for large-scale RL training; DeepSeek's value is leverage—a second option that caps NVIDIA's pricing power. 3. Risk — engineering overreach: Domestic adaptation was described by insiders as 'crossing snow mountains and grasslands.' V4 arrived after a 15-month gap (Jan 2025 R1 to Apr 2026 V4) while OpenAI and Anthropic shipped multiple generations. Liang Wenfeng's stated answer: DeepSeek's primary goal is AGI; hardware efficiency is infrastructure for large-scale RL and recursive self-improvement requiring massive long-context trajectory generation.
Conclusion
The post argues DeepSeek is not fighting for market share but for the power to define what long-context, agentic inference *should cost*—anchoring the floor at $0.003625/M tokens and proving a reproducible path off CUDA. Whether this becomes a standardized route or a one-off 'moonshot' will be answered by TileLang adoption, open V4 weights, and Ascend supernode scale-out in the second half of the year.
*Sources cited in the original post: DeepSeek-V4 technical report (CSA/HCA, Engram, TileLang), TileLang DSL documentation, Engram paper (DeepSeek-AI & Peking University), Huawei Ascend 950PR official benchmarks, and research notes from Huaxi Securities and Shanghai Securities.*