AI 还不够。杰文斯悖论一直是这个时代的标志,就像芯片领域的摩尔定律一样。它会过早失效吗?
我们仍然受制于供给,这是势头强劲与快速普及的信号。——Sundar Pichai,Alphabet 2026 年第二季度。1
到 2026 年,我们仍然没有足够的容量来满足我们所有的需求,而且我相信这种态势在 2027 年也会同样如此。我们目前已经收到的 2028 年需求令人震惊。——Andy Jassy,Amazon 2026 年第二季度。2
Blackwell 的销售火爆异常,云 GPU 已经售罄。——Jensen Huang,NVIDIA 2027 财年第一季度。3
那么,如果 AI 的价格翻倍会发生什么?
它们已经在上涨了。Anthropic 于 7 月 24 日发布 Fable 5,定价为每百万输出 token 50 美元,是 Opus 5 的两倍,并树立了新的前沿价格上限。Google 的 Gemini 旗舰模型在四代产品中从 1.50 美元攀升至 12 美元。4
或者至少大多数都在上涨。OpenAI 在 Fable 5 发布五天后将 GPT-5.6 Luna 的价格下调了 80%,这要么是抢占市场的举措,要么是根本性的成本突破。5
到目前为止,AI 给出的说辞是:杰文斯悖论意味着价格下降会推动更多消费。但价格正在上涨。
AI 实验室正在押注分层。只要价值层和中间市场前沿能吸收那些因定价过高而被挤出高端层的负载,消耗的 GPU 小时总量就会持续增长。分层支撑了杰文斯悖论。
高端层的成本是价值层的 13 倍,而智能只多出五分之一。中间市场前沿是新来者:GPT-5.6 Sol 和 Kimi K3,以 40% 的成本提供前沿智能的 96%。67 价值前沿则从 GLM-5.2 一路延伸到 DeepSeek V4 Flash,以高端层成本的 1% 到 5% 提供 84% 的智能。
随着分层成为常态,路由器成为买方与模型之间的战略层,是让应用把查询转交给合适模型的基础管道。路由器将内置于模型之中,外置于客户的软件之中,也存在于各类 harness 之中。
各大实验室都需要在三个层级中都有布局,否则就会在路由器竞价中因价格而落败。初创公司则通过占据前沿上的某个单点来取胜,而这是大型实验室在经济上无法匹敌的。DeepSeek V4 Flash 以 $0.03 的价格就是这一点的存在性证明。
这一切都是为了进一步推动杰文斯悖论。
-
Alphabet 2026 年第二季度财报电话会议,Sundar Pichai 发言,2026 年 7 月 22 日。↩︎
-
Andy Jassy 表示,Amazon 今年将支出 2200 亿美元,但仍将没有足够产能满足需求,Fortune,2026 年 7 月 30 日。↩︎
-
NVIDIA 2027 财年第一季度财报:营收、每股收益与展望,2026 年 5 月 20 日。↩︎
-
GPT-5.6 定价自 2026 年 7 月 30 日起生效:Sol 为每百万 tokens 5 美元/30 美元,Luna 为每百万 tokens 0.20 美元/1.20 美元。OpenAI GPT-5.6 性价比前沿公告;VentureBeat:OpenAI 将 GPT-5.6 Luna 价格下调 80%。↩︎
-
Artificial Analysis Intelligence Index v4.1,访问于 2026-08-04。↩︎
-
中端市场 GPT-5.6 Sol 与 Kimi K3 智能指数平均值(58)对比高端 Opus 5 与 Fable 5 平均值(60.5)= 96%。成本比 1.05 美元 / 2.75 美元 = 38%。↩︎
There’s not enough AI. Jevons’ Paradox has been a hallmark of this era like Moore’s Law in chips. Will it fail early?
We continue to be supply constrained, a sign of momentum & rapid adoption. — Sundar Pichai, Alphabet Q2 2026.1
We will still not have enough capacity to meet all the demand we have in 2026, & I believe this dynamic will also be true in 2027 too. The demand we already have for 2028 is striking. — Andy Jassy, Amazon Q2 2026.2
Blackwell sales are off the charts, & cloud GPUs are sold out. — Jensen Huang, NVIDIA Q1 FY2027.3
So what happens if the price of AI doubles?
They’re already increasing. Anthropic launched Fable 5 on July 24 at $50 per million output tokens, doubling Opus 5 & setting a new frontier ceiling. Google’s Gemini flagship climbed from $1.50 to $12 across four generations.4
Or at least most are increasing. OpenAI cut GPT-5.6 Luna prices by 80% five days after Fable 5’s launch, either a market-capture push or a fundamental cost breakthrough.5
The pitch so far from AI : Jevons’ paradox drives more consumption as prices fall. But prices are increasing.
AI labs are betting on segmentation. As long as the value & mid-market frontiers absorb the workloads priced out of premium, total GPU-hours consumed keeps growing. Segmentation sustains Jevons.
The premium tier costs 13x the value tier for a fifth more intelligence. The mid-market frontier is the new arrival : GPT-5.6 Sol & Kimi K3, delivering 96% of frontier intelligence at 40% of the cost.67 The value frontier runs from GLM-5.2 down to DeepSeek V4 Flash, offering 84% of the intelligence at 1 to 5% of the premium cost.
As segmentation becomes the norm, routers become the strategic layer between buyer & model, the plumbing that lets an application shift a query to the right model. Routers will be internal to models, external in customers’ software, & in harnesses.
The major labs each need entries in all three tiers, or they lose the router auction on price. Startups win by owning a single point on the frontier that a big lab cannot economically match. DeepSeek V4 Flash at $0.03 is the existence proof.
All in furtherance of Jevons.
-
Alphabet Q2 2026 earnings call, Sundar Pichai remarks, July 22, 2026. ↩︎
-
Andy Jassy said Amazon will spend $220 billion this year and still won’t have enough capacity to meet demand, Fortune, July 30, 2026. ↩︎
-
NVIDIA Q1 FY2027 earnings: revenue, EPS & outlook, May 20, 2026. ↩︎
-
AI Model Inflation. Vendors are pivoting from share-taking subsidies to margin-taking pricing as capex hits records. ↩︎
-
GPT-5.6 pricing effective July 30, 2026 : Sol at $5/$30, Luna at $0.20/$1.20 per million tokens. OpenAI GPT-5.6 price-performance frontier announcement; VentureBeat : OpenAI cuts GPT-5.6 Luna prices by 80%. ↩︎
-
Artificial Analysis Intelligence Index v4.1 accessed 2026-08-04. ↩︎
-
Mid-market average of GPT-5.6 Sol & Kimi K3 intelligence (58) versus premium average of Opus 5 & Fable 5 (60.5) = 96%. Cost ratio $1.05 / $2.75 = 38%. ↩︎