比其他公司便宜五十倍，好像不影响股票涨
现在不是号称deepseek牛的一比吗
版主： who， shepherd17
-
shepherd17
- 精英
- 帖子： 6559
- 注册时间： 2023-12-24
- Has thanked： 994 time
- Been thanked： 845 time
Re: 现在不是号称deepseek牛的一比吗
你忘性比较大。DeepSeek 刚推出来时，美国相关股票应声大跌。
下面是 AI 汇总。
DeepSeek 在早期引发过股价暴跌，但没有造成持续性下跌，因为市场很快意识到：它的极低价格伴随明显性能限制，并不足以削弱美国公司的核心竞争力，也不会显著减少高端算力的长期需求。
为什么 DeepSeek 没有引发持续的美股下跌
1. 最便宜的模型准确率极低
DeepSeek 的 V4‑Flash 虽然比美国同类便宜 百倍以上，但只有 37% 准确率、84% 幻觉率。
企业无法用它替代 OpenAI、Anthropic、Google 的高端模型。
便宜 ≠ 可用于关键业务。
这直接限制了它对美国 AI 收入的威胁。
2. 市场早在 2025 年就经历过“DeepSeek 冲击”
DeepSeek 确实导致过巨震，只是发生在更早的时间：
- 2025 年 1 月 27 日：NVIDIA 单日蒸发 6000 亿美元市值
- AI 代币与算力估值普遍暴跌 90%
到 2026 年，投资者已经对中国开源权重模型的冲击“免疫”，不再恐慌。
3. 便宜模型不会减少对昂贵算力的需求
即使是 DeepSeek 的高端模型（如 V4‑Pro Max），仍需要大量算力，并与美国前沿模型竞争，而不是替代它们。
对 NVIDIA 等公司来说：
- 模型更高效 → 单次推理成本下降
- 推理更便宜 → 总使用量上升，整体算力需求反而增加
市场看到了这种双向效应，因此没有长期下跌。
4. 企业关注的是“总成本”，不是“每 token 价格”
低价模型往往带来：
- 更多人工校对
- 更高风险
- 更低可靠性
- 更复杂的流程
因此，DeepSeek 的低价并不能真正削弱美国高端模型的商业价值。
5. 美国科技股的估值依赖“稀缺性叙事”
美国 AI 的估值建立在：
- 巨型数据中心
- 高端 GPU
- 专有研究体系
中国开源权重模型挑战了这一叙事，但并未推翻它。
因此市场出现波动，而不是崩盘。
6. DeepSeek 的最强模型并未明显超越美国前沿
DeepSeek 的便宜模型很便宜但质量低；
它的强模型（V4‑Pro、V4‑Pro Max）虽强，但并未明显优于 GPT‑5.5 或 Claude Opus 4.6。
市场关注的是性能，而不是价格。
7. 地缘政治与监管限制了 DeepSeek 的全球渗透
即使 DeepSeek 技术更强，美国企业也面临：
- 数据安全
- 合规要求
- 供应链信任
- 法规限制
这些因素阻止 DeepSeek 立即蚕食美国 AI 收入。
为什么相关美股没有持续下跌？
| 因素 | 对美国 AI 股价的影响 |
|---|---|
| DeepSeek 极便宜 | 造成短期恐慌，但企业难以采用 |
| 性能差距明显 | 保护美国高端模型的定价能力 |
| 2025 年已提前暴跌 | 冲击已被市场吸收 |
| 算力需求仍增长 | 支撑 NVIDIA、云厂商估值 |
| 监管壁垒 | 限制 DeepSeek 的海外影响 |
| 前沿模型仍昂贵 | 稀缺性叙事仍成立 |
DeepSeek did cause stock drops—just not a sustained collapse—because markets quickly realized its ultra‑low prices come with major performance limits, do not eliminate U.S. firms’ competitive advantages, and do not meaningfully reduce long‑term demand for high‑end compute.
Why DeepSeek didn’t trigger a lasting stock crash
1. Its cheapest models are extremely inaccurate
DeepSeek’s V4‑Flash is indeed 105× cheaper than Anthropic’s Claude Fable 5, but it has only 37% accuracy and an 84% hallucination rate. Forbes
This means enterprises cannot replace premium U.S. models with DeepSeek for most critical workloads. Cheap ≠ economically viable.
This sharply limits the real-world threat to OpenAI, Anthropic, Google, Microsoft, and Amazon.
2. Markets already priced in the “DeepSeek shock” in 2025
DeepSeek did cause a massive stock drop—just earlier.
- On January 27, 2025, NVIDIA lost $600 billion in market cap in a single day after DeepSeek released R1. proptechusa.ai
- Token prices and AI‑compute valuations collapsed industry‑wide by 90%. proptechusa.ai
By 2026, investors were no longer surprised by cheap Chinese open‑weight models. The shock was already absorbed.
3. Cheap models don’t eliminate demand for expensive compute
Even DeepSeek’s own high‑end models (e.g., V4‑Pro Max) still require substantial compute and compete with U.S. frontier models rather than replacing them. Crypto Briefing
For NVIDIA and other chipmakers, the effect is ambiguous:
- More efficient models reduce cost per query.
- But cheaper inference expands total AI usage, increasing overall compute demand. Yahoo Finance
Markets recognized this nuance, preventing a prolonged crash.
4. Cheap AI often shifts costs elsewhere
Enterprises care about total cost, not token price.
Low‑accuracy models create:
- rework
- delays
- higher risk
- more human oversight
Forbes notes that DeepSeek’s low price “does not automatically translate to low cost overall.” Forbes
Thus, DeepSeek’s pricing doesn’t meaningfully undermine the business case for premium U.S. models.
5. U.S. tech valuations rely on scarcity, not just price
American AI stocks are priced on the assumption that frontier AI requires:
- massive data centers
- premium GPUs
- proprietary research pipelines
Chinese open‑weight releases (DeepSeek, Kimi K3) challenge this assumption, but they haven’t disproven it.
As Abacus notes, these releases “expose how sensitive U.S. tech stocks have become,” but they do not show U.S. spending is wasted. Abacus
Hence: volatility, not collapse.
6. DeepSeek’s strongest models are not fully superior
DeepSeek’s cheapest models are extremely cheap but low‑quality.
Its strongest models (V4‑Pro, V4‑Pro Max) are competitive but not clearly better than GPT‑5.5 or Claude Opus 4.6. Crypto Briefing
Markets respond to performance, not just price.
7. Geopolitical and regulatory factors limit adoption
Even if DeepSeek were technically superior, U.S. enterprises face:
- security concerns
- regulatory constraints
- data‑sovereignty rules
- trust and supply‑chain issues
These factors prevent DeepSeek from immediately cannibalizing U.S. AI revenue.
Synthesis: Why no sustained stock drop?
| Factor | Impact on U.S. AI Stocks |
|---|---|
| DeepSeek is extremely cheap | Short-term fear, but limited enterprise adoption due to accuracy issues |
| Performance tradeoffs | Protect premium U.S. models’ pricing power |
| Shock already priced in (2025) | Markets no longer surprised |
| Compute demand still grows | Supports NVIDIA, cloud providers |
| Regulatory barriers | Limits DeepSeek’s U.S. penetration |
| Frontier models still expensive | Scarcity narrative remains intact |
