Summary
Chinese AI startup DeepSeek will make permanent a 75% price cut on its flagship V4-Pro model, keeping API prices at one-quarter of original level. New pricing: 0.025–6 yuan per million tokens (~$0.0035–$0.83), depending on usage type; down from 0.1–24 yuan previously. DeepSeek did not disclose whether the permanent cut was tied to increased supply of Huawei Ascend 950 chips, which V4 was developed to use. Huawei’s AI chip sales have benefited from U.S. export controls preventing Nvidia from selling its most advanced semiconductors in China; separate U.S. curbs on chipmaking equipment exports have limited Huawei’s ability to scale Ascend production. When DeepSeek launched V4 in April 2026, it said Pro pricing would fall sharply once Huawei Ascend 950 supernodes launched in volume H2 2026. The permanent cut suggests either Ascend supply has increased or DeepSeek is pricing for share against a U.S. competitive set.
Key Points
- The cut: 75% permanent reduction in V4-Pro API pricing. Old: 0.1–24 yuan/million tokens. New: 0.025–6 yuan/million tokens (~$0.0035–$0.83/million tokens).
- The competitive context: V4-Pro is DeepSeek’s flagship model. The cut is presented as a permanent pricing decision, not a promotional one.
- The Huawei Ascend dependency: V4 was adapted to use Huawei Ascend 950 chips because U.S. export controls cut off Nvidia’s most advanced semiconductors from China. DeepSeek did not say whether the price cut was caused by improved Ascend supply — but the V4 launch (April 2026) explicitly told customers Pro pricing would fall once Ascend 950 supernodes launched in volume H2 2026.
- The structural constraint on Huawei: U.S. curbs on chipmaking equipment exports have limited Huawei’s ability to scale Ascend production. The supply side is not unconstrained.
- The Pro vs. Flash pricing structure: When V4 launched in April 2026, Pro was priced up to 12x more than Flash “due to constraints in high-end compute capacity, limiting availability.” The 75% Pro cut now compresses that ratio significantly.
- The April 2026 Big Chinese Tech Firms Scramble for Huawei Chips Reuters piece (April 29, 2026, linked in the source) provides the demand-side context: Chinese cloud and tech firms were competing for limited Ascend supply.
Newsletter Angles
- The pricing signal is the export-controls-feedback-loop receipt: U.S. export controls (intended to constrain Chinese AI compute) drove Chinese AI firms to develop and adopt Huawei Ascend chips. The Ascend supply curve is now demonstrably elastic enough that DeepSeek can do a permanent 75% price cut on its flagship model. This is the Export Controls-as-industrial-policy-for-the-target pattern — a recurring feature of the wiki’s Tech-State Conflict cluster. Worth noting alongside the U.S. AI EO postponement (see prior ingest’s Trump Postpones AI Executive Order — CBS - 2026-05-21) — the U.S. is hesitating on its AI policy framework while the Chinese alternative compute stack is delivering price competition.
- The 75% cut is the pricing-power-shift signal: If DeepSeek can sustain a permanent 75% cut on its flagship API, the marginal cost of high-end Chinese model inference has dropped to roughly $0.83/million tokens at the high end (or $0.0035 at the low end). For comparison, OpenAI’s GPT-4.1 charged ~$2/M tokens input + $8/M tokens output (as of late 2025). DeepSeek is now priced 2–3x below the U.S. competitive set on the high end and orders-of-magnitude below on the low end. The wiki’s AI Cost Incidence concept should track this as a competitive-floor pressure point for U.S. labs.
- The “did not disclose” framing is the analytical anchor: DeepSeek explicitly did not say whether the cut was supply-driven or competition-driven. The two readings have very different implications: supply-driven means Huawei is hitting scale and the U.S. export-controls regime has lost analytical traction; competition-driven means DeepSeek is buying share against U.S. labs and accepting margin compression to do it. The wiki should track Q3/Q4 2026 financial signals for which reading holds — if Huawei Ascend 950 supernode launches deliver the H2 2026 volume that DeepSeek’s April 2026 announcement promised, the supply-driven reading wins.
- The buried lede is the Huawei Ascend 950 supernode launch timeline: H2 2026 — explicitly named in the April 2026 V4 launch. If supernodes ship in volume in Q3, the DeepSeek pricing decision was an early-confidence signal. The wiki should add the Ascend 950 supernode launch to the watch list as a leading indicator for whether the U.S. AI competitive cost structure resets.
Entities Mentioned
- DeepSeek — Chinese AI startup; the actor making the pricing change
- Huawei — chip supplier; the Ascend 950 supply curve is the load-bearing constraint
- Nvidia — barred from selling most advanced semiconductors to China by U.S. export controls
- OpenAI — implicit competitive comparison
Concepts Mentioned
- AI Cost Incidence — DeepSeek pricing as a competitive-floor pressure point
- Tech-State Conflict — the U.S.–China AI / export-controls dynamic
- AI Sovereignty — Chinese alternative-stack performance under U.S. export pressure
- Frontier AI — competitive set context
Quotes
DeepSeek will make permanent a 75% price cut on its flagship V4-Pro AI model, keeping prices at a quarter of their original level. — Reuters
The company cut V4-Pro API costs to between 0.025 and 6 yuan per million tokens (about $0.0035 to $0.83) depending on usage type, from 0.1 to 24 yuan previously. — Reuters
Pro pricing was expected to fall sharply once Huawei Ascend 950 supernodes are launched in large quantities in the second half of the year. — DeepSeek, April 2026 V4 launch (paraphrased in this Reuters piece)
Notes
- Source tier: Reuters, byline Abu Sultan in Bengaluru and Eduardo Baptista. Direct citation of DeepSeek’s Saturday statement and the V4 launch communications. No Huawei or DeepSeek financial disclosure cited; the “did not disclose” framing on the supply-vs-competition question is honestly flagged.
- Pricing math caveat: The 0.0035-$0.83 USD range translation is based on the Reuters yuan-to-USD conversion at time of publication. The wide range reflects DeepSeek’s tiered usage pricing (different input/output, caching, off-peak pricing all factor in).
- Open follow-ups: (1) any Huawei announcement on Ascend 950 supernode shipment timeline; (2) competitive responses from OpenAI / Anthropic / Google on API pricing; (3) third-party benchmark comparisons of V4-Pro vs. U.S. frontier-model performance at the new price points; (4) whether the Chinese MIIT or related agencies are subsidizing the Huawei–DeepSeek pricing dynamic.
- Wiki gap: Neither Huawei nor Nvidia currently has an entity page despite both being recurring subjects across the AI / chips / export-controls coverage. Worth elevating to a near-term stub priority — listed in deferred-stubs for this ingest.