AI四层架构：应用、中间层、模型与数据中心的竞争关键问题
The Four Questions of AI
为创业者提供了清晰的AI产业链拆解框架，直接指出了中间层的高毛利机会与应用层的护城河策略，适合评估自身在AI生态中的定位。
There are 4 layers in the AI stack, each with a fundamental question before it.
AI 栈有 4 层，每层面前都有一个根本性问题。
Application
应用层
Breakout hits appear overnight. The next model release can erase them just as fast. Thin wrappers die when the model absorbs their feature. To build a moat, app companies must decide whether to build a brand, disrupt incumbents through pricing, develop their own models, or expand faster than competitors.
爆款应用一夜之间涌现。下一个模型发布可以同样迅速地将其抹去。当模型吸收了其功能时，薄封装的应用就会消亡。为了建立护城河，应用公司必须决定是打造品牌、通过定价颠覆现有巨头、开发自己的模型，还是比竞争对手扩张得更快。
Agents do the work with headless software. Will a user interface survive, or do all applications collapse into middleware harnesses1?
智能体（Agents）使用无头软件执行工作。用户界面能否幸存，或者所有应用程序是否会崩溃为中间件工具带？
Harness & Context
工具带与上下文
If applications collapse into middleware, the fight moves to who owns the harness. Model makers claim their native harness wins. The benchmarks say otherwise. In the same week, both OpenAI & Anthropic models scored higher in Cursor than in their creators’ own harnesses2. Because input context is up to 98% of all tokens2, a smart harness cuts the cost of the same answer by 71% & lifts gross margin from 38% to 75%3.
如果应用程序崩溃为中间件，争夺焦点将转向谁拥有工具带。模型制造商声称他们的原生工具带会胜出。但基准测试显示并非如此。在同一周，OpenAI 和 Anthropic 的模型在 Cursor 中的得分都高于在其各自创建者的原生工具带中[2]。由于输入上下文占所有令牌的 98%[2]，一个聪明的工具带可以将相同答案的成本降低 71%，并将毛利率从 38% 提升至 75%[3]。
Beyond the model makers, enterprises are starting to lay claim to the harness to build corporate sovereign AI4. No CIO wants to ship company secrets out the door in a prompt. Owning the context layer lets a company keep its memory & permissions in house while swapping models underneath.
除了模型制造商之外，企业也开始宣称对工具带的所有权，以构建企业主权 AI[4]。没有首席信息官（CIO）愿意在提示词中将公司机密泄露出去。拥有上下文层使公司能够在内部保留其记忆和权限，同时在下层替换模型。
Model
模型层
For foundation models, the question is at what point training costs don’t pay back. Training a state-of-the-art model costs billions5. Recursive self-improvement, AI models training better AI, has not yet bailed out the big labs. AI writes 75% of Google DeepMind’s production code, yet Google sees no end-to-end speedup in AI research6.
对于基础模型而言，问题在于训练成本在何时无法收回。训练最先进的模型需要数十亿美元[5]。递归自我改进（即 AI 模型训练更优秀的 AI）尚未让大型实验室脱困。AI 编写了 Google DeepMind 75% 的生产代码，但 Google 并未看到 AI 研究中的端到端加速[6]。
On the demand side, Rich Sutton’s Bitter Lesson argues that general models scaled with massive compute always win7. Yet SOTA models are losing token share on public gateways because they cost too much to run8. About 83% of agent tasks now route to smaller or domain-specific models fine-tuned to do one job for a fraction of the bill9.
在需求侧，Rich Sutton 的“痛苦教训”（Bitter Lesson）认为，使用海量算力扩展的通用模型总是能获胜[7]。然而，由于运行成本过高，最先进（SOTA）模型在公共网关上的令牌份额正在流失[8]。目前约 83% 的智能体任务路由到更小或特定领域的模型，这些模型经过微调以完成单一任务，费用仅为前者的零头[9]。
Challengers are attacking the training bill through process & architecture. One startup used coding agents to pre-train a competitive coding model with a single researcher for under $10m10. New architectures like diffusion models11 & JEPA, where a 109m model beat 72b LLMs with 42x less data12, bypass brute-force scaling.
挑战者正通过流程和架构攻击训练账单。一家初创公司使用编码代理，仅用一名研究人员和不到 1000 万美元的成本就预训练出了一个具有竞争力的编码模型[10]。新的架构如扩散模型[11]和 JEPA，其中 1.09 亿参数的模型在使用少 42 倍数据的情况下击败了 720 亿参数的 LLM[12]，从而绕过了暴力扩展。
Still, the bull case for scaling the frontier sits in Anthropic’s S-1. Opus 5.5 cut customer cost per task by 40%, & when Opus 4.5 delivered more capability at a lower cost per task than Opus 4.1, Jevons paradox kicked in & sequential quarterly revenue growth accelerated from 48% to 132%13.
尽管如此，扩展前沿模型的看涨论点仍体现在 Anthropic 的 S-1 文件中。Opus 5.5 将客户每项任务的成本降低了 40%，而当 Opus 4.5 以低于 Opus 4.1 的单任务成本提供了更强的能力时，杰文斯悖论（Jevons paradox）显现，季度收入环比增速从 48% 加速至 132%13。
Data Center
数据中心
Below the models, the question is whether credit & power can keep up with GPU demand. With concrete prices up 3x, electrical grids strained, & backlogs on turbines, generators, & reciprocating engines, the answer for now is no14.
在模型之下，问题在于信贷和电力能否跟上 GPU 的需求。随着混凝土价格上涨 3 倍、电网承压、涡轮机、发电机和往复式发动机积压严重，目前的答案是否定的14。
Unrestrained growth in AI requires $4t in new debt over five years, making the AI credit market larger than global private credit & equal to 91% of the US municipal bond market15. Are there enough lenders left at attractive rates? Risk strains insurance costs. Five-year credit default swap spreads have hit records from Meta (108 bps) to Oracle (322 bps) & CoreWeave (1,090 bps)16.
无限制的 AI 增长需要在五年内新增 4 万亿美元债务，这使得 AI 信贷市场规模超过全球私募信贷市场，并相当于美国市政债券市场的 91%15。还有足够的贷款人以有吸引力的利率放贷吗？风险压力推高了保险成本。五年期信用违约互换（CDS）利差已创下新高，Meta 为 108 个基点，Oracle 为 322 个基点，CoreWeave 高达 1,090 个基点16。
On the ground, local politics & grid delays are blocking new sites, while Ornn’s B300 index has climbed to $12.29 per GPU-hour (~20% CMGR since June)17.
在实际层面，地方政治和电网延误正在阻碍新站点的建设，而 Ornn 的 B300 指数已攀升至每 GPU 小时 12.29 美元（自六月以来复合年增长率约为 20%）17。
Meanwhile, three open questions hang over the math : whether cash flows match the duration of twenty-year debt & leases15, whether GPUs last five years, & whether the price of a GPU-hour will increase or decrease18.
与此同时，数学层面仍有三个悬而未决的问题：现金流是否与二十年债务和租赁的期限匹配15，GPU 是否能使用五年，以及每 GPU 小时的价格是上涨还是下跌18。
How the Layers Interplay
各层级的相互作用
None of these four questions stands alone. Applications that generate tremendous profits at the top can sustain inflation in GPU prices & higher borrowing costs at the bottom. New, more efficient architectures19 can decrease the need for concrete & copper. Surging GPU prices can swing the economics of the whole stack.
这四个问题并非孤立存在。顶层产生巨额利润的应用程序可以支撑底层 GPU 价格的通胀和更高的借贷成本。新型、更高效的架构19 可以减少对混凝土和铜的需求。GPU 价格的飙升可能会改变整个技术栈的经济性。
Like a double pendulum20, the four layers form a coupled system : simple in each part, yet complex, chaotic, & unpredictable in motion.
如同双摆20，这四个层级构成了一个耦合系统：每个部分看似简单，但整体运动却复杂、混沌且不可预测。
- A harness is the software runtime around the model, such as Claude Code, Codex, or Cursor, that governs what context flows into the prompt, which tools the agent calls, & how session trajectories are cached & logged. See Tomasz Tunguz, The Harness Is the New Battleground, July 14, 2026. ↩︎
- Tomasz Tunguz, Aftermarket Harnesses, July 28, 2026, citing Endor Labs’ Agent Security League benchmark (GPT-5.5 scoring 61.5% in OpenAI’s Codex vs. 87.2% in Cursor ; Claude Opus 4.7 scoring 87.2% in Claude Code vs. 91.1% in Cursor) & OpenRouter token composition data (input tokens representing 85.7% to 98.2% of volume). ↩︎ ↩︎
- Tomasz Tunguz, The Harness Margin Opportunity, September 17, 2026, citing UC Berkeley’s HarnessTax study (71% lower cost for the same resolved SWE-bench task with no statistically significant quality difference across 42 comparisons) & AI associate contract gross-margin modeling (38% vs. 75% on a $250k contract). ↩︎
- Tomasz Tunguz, The Price of Entry to the Frontier, August 31, 2026, on enterprise data sovereignty mandates, Zero Data Retention policies, & model access segmentation. ↩︎
- Frontier training & R&D scale : Anthropic’s S-1 (filed October 26, 2026) discloses $8.1b in FY2025 operating losses & $637b in total compute commitments (Anthropic S-1 Analysis), while OpenAI projects $278b in cumulative cash burn through 2030 & Epoch AI estimates individual next-generation frontier training runs crossing $1b+ as training compute grows ~5x per year (Epoch AI). ↩︎
- Google DeepMind, Gemini 4 Argon Safety & Security Report, October 2026 : AI generates more than 75% of DeepMind production code & fully delegated pull requests rose from 7.5% to 11.9% over five months, with no significant end-to-end acceleration of AI research timelines. ↩︎
- Rich Sutton, The Bitter Lesson, March 13, 2019 : general methods that leverage computation ultimately beat methods that embed human domain knowledge by a large margin. ↩︎
- Tomasz Tunguz, The Most Important Market in AI Is the Middle, September 23, 2026, & Honestly, Who Buys SOTA?, August 14, 2026 : 84% of tokens on OpenRouter are not state of the art, Anthropic’s flagship Fable 5.1 captured only 3.7% of gateway spending in its first twelve days (while its predecessor fell from 13.2% to 4.9% within a month), & among large corporate accounts frontier models fell from 53% of token consumption in early August to 45% by September. ↩︎
- Tomasz Tunguz, The Most Important Market in AI Is the Middle, September 23, 2026, & 9 Observations Using AI Agents, 2026 : Cursor’s Composer 2 fine-tuned open-weight Kimi K2.5 to cut costs 86% vs. its prior model, Harvey fine-tuned a domain model to cut cost per cell 55% vs. Sonnet 5 while outscoring Fable 5, & an 8b Qwen 3 model fine-tuned for task classification beat zero-shot GPT-5.2 locally on a laptop. ↩︎
- Mirendil, Pretraining at Scale Without a Pretraining Team, October 21, 2026 : one researcher paired with coding agents (which wrote 99% of the repository) pre-trained Mirendil-Base (100b to 500b parameters) from scratch in 3.5 months for under $10m all-in compute & token spend, beating GLM-5.3 (0.242 bits/byte), DeepSeek-V4-Pro (0.243), & Nemotron-550B (0.247) on held-out post-cutoff code repositories (averaging 0.222 bits/byte vs. 0.193 for Kimi-K3). ↩︎
- Diffusion language models generate & refine blocks of tokens in parallel rather than one token at a time : Inception Labs’ Mercury 2.5 (launched September 8, 2026) reaches 414 to 1,107 tokens/sec on Nvidia GPUs at $0.20/$0.75 per million tokens, roughly 9x cheaper & 3x faster than comparable autoregressive models (Inception Labs). ↩︎
- Harness（控制框架）是围绕模型的软件运行时环境，例如 Claude Code、Codex 或 Cursor，它管理哪些上下文流入提示词、智能体调用哪些工具，以及如何缓存和记录会话轨迹。参见 Tomasz Tunguz，《Harness 是新战场》，2026 年 7 月 14 日。↩︎
- Tomasz Tunguz，《售后 Harness 市场》，2026 年 7 月 28 日，引用 Endor Labs 的智能体安全联盟基准测试（GPT-5.5 在 OpenAI 的 Codex 中得分为 61.5%，而在 Cursor 中为 87.2%；Claude Opus 4.7 在 Claude Code 中得分为 87.2%，而在 Cursor 中为 91.1%）以及 OpenRouter 的令牌组成数据（输入令牌占总体积的 85.7% 至 98.2%）。↩︎ ↩︎
- Tomasz Tunguz，《Harness Margin Opportunity》，2026年9月17日，引用加州大学伯克利分校的 HarnessTax 研究（在42次比较中，解决相同的 SWE-bench 任务成本降低71%，且质量无统计学显著差异）及 AI 外包合同毛利率建模（25万美元合同的毛利率为38%对比75%）。↩︎
- Tomasz Tunguz，《进入前沿的价格》，2026年8月31日，关于企业数据主权强制要求、零数据保留政策及模型访问分段。↩︎
- 前沿训练与研发规模：Anthropic 的 S-1 文件（于2026年10月26日提交）披露其2025财年运营亏损达81亿美元，总算力承诺高达6370亿美元（Anthropic S-1 分析），而 OpenAI 预计至2030年累计现金消耗为2780亿美元；Epoch AI 估计随着训练算力每年增长约5倍，单个下一代前沿训练运行的成本将超过10亿美元（Epoch AI）。↩︎
- Google DeepMind，《Gemini 4 Argon 安全与报告》，2026年10月：AI 生成的代码占 DeepMind 生产代码的75%以上，完全委托的拉取请求在五个月内从7.5%上升至11.9%，但 AI 研究时间线并未出现显著的端到端加速。↩︎
- Rich Sutton，《苦涩的教训》，2019年3月13日：利用计算能力的通用方法最终大幅优于嵌入人类领域知识的方法。↩︎
- Tomasz Tunguz，《AI 最重要的市场在中层》，2026年9月23日，以及《说实话，谁购买 SOTA？》，2026年8月14日：OpenRouter 上84%的 token 并非最先进状态，Anthropic 的旗舰模型 Fable 5.1 在其上线的前十二天内仅占据网关支出的3.7%（而其前代模型在一个月内从13.2%降至4.9%），且在大型企业客户中，前沿模型的 token 消耗占比从8月初的53%降至9月的45%。↩︎
- Tomasz Tunguz，《AI 最重要的市场在中层》，2026年9月23日，以及《使用 AI Agent 的9个观察》，2026年：Cursor 的 Composer 2 对开源权重的 Kimi K2.5 进行微调，使其成本较前一模型降低86%；Harvey 对领域模型进行微调，使其每个单元的成本较 Sonnet 5 降低55%，同时在评分上超越 Fable 5；而在笔记本电脑本地运行的一项任务分类场景中，针对该任务微调的8b Qwen 3 模型击败了零样本 GPT-5.2。↩︎
- Mirendil，无需预训练团队的大规模预训练，2026年10月21日：一名研究人员与代码代理（编写了仓库99%的代码）合作，在3.5个月内从零开始预训练了Mirendil-Base（参数规模为100b至500b），总计算和token花费低于1000万美元，在预留的截止后代码仓库上超越了GLM-5.3（0.242 bits/byte）、DeepSeek-V4-Pro（0.243）和Nemotron-550B（0.247）（平均为0.222 bits/byte，而Kimi-K3为0.193）。↩︎
- 扩散语言模型并行生成和细化token块，而非逐个生成：Inception Labs推出的Mercury 2.5（2026年9月8日发布）在Nvidia GPU上达到每秒414至1,107个token的速度，每百万token成本为0.20美元/0.75美元，比同类自回归模型便宜约9倍且速度快3倍（Inception Labs）。↩︎
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力