AI 先吃掉的是穷人的手机AI Eats the Cheap Phone First
AI 数据中心的算力扩张,代价落在了一个还没买过第一台智能机的人身上。这是暴力三角里最典型的一笔:本可以上网的人,因为内存颗粒被挪去喂 AI,上不了网。差额就在这里,而且它是被精确地制造出来的。
机制很清楚。三星、SK 海力士、美光三家占了全球内存芯片九成以上,2025 年底它们把绝大部分供应转向 AI 数据中心,消费电子的货就没了(来源:Rest of World)。内存涨价,涨价被转嫁给消费者。三家厂和美国的 AI 巨头签了合作,S&P Global 估算美国超大规模厂商 2027 年的资本开支会超过 1.3 万亿美元。这不是市场意外,是资源分配的一次主动选择:谁的需求被优先满足,谁的被挤掉。
挤掉的正好是 150 美元以下的机器。2025 年全球出货的智能机里超过四分之一低于 150 美元,这个市场在萎缩。低于 100 美元的机型第二季度同比掉了近六成。中国厂商占全球出货约六成,今年大幅削减了入门级项目,转去做利润更高的高端机。小米在印度把 Redmi 15C 从 12499 卢比提到 16999 卢比,涨了 36%;Oppo 在东南亚的百元以下出货暴跌 96%;非洲去年 81% 的智能机低于 200 美元,百元以下的出货第二季度同比降了 34%。分析师说,原来低于 150 美元的,以后可能变成低于 250 甚至 300 美元,而且不指望回到 2025 年以前的水平。
结构暴力不需要有人动手。它只需要一条供应链按利润率排序。涨价的幅度按地区分:印度 21%,亚太 19%,中东和非洲 18%,美国只有 5%。同一个芯片短缺,落在不同地方轻重完全不同——最买不起的地方涨得最狠。GSMA 的数据是,入门设备要花掉最穷的 20% 人口月收入的 44%,撒哈拉以南非洲是 76%。收入涨不过手机价,结果不是晚点买,是根本没有联网的设备。
这里没有反派的脸,所以更容易被当成天气。三家内存厂没有做错什么,它们只是把货卖给了出价更高的一方;中国手机厂商也没有做错什么,它们只是不再做不赚钱的生意。所有人都在做对自己最优的选择,合起来就是把几亿人挡在网外。这正是共谋者理论要说的:不需要谁有恶意,只需要所有人按既定秩序各取所需,秩序本身的暴力就自动执行了。
被武器化的是“AI”这个词。它现在是万能的正当代词,任何资源向它倾斜都自动获得合理性。数据中心要内存,所以内存归它;入门机不赚钱,所以入门机消失。没人需要为这个决定负责,因为每一步都合乎理性。理性在这里就是元暴力的伪装——它把一次分配选择说成技术进步的必然。
真正被削掉的东西也很具体:GSMA 原本预计到 2030 年还有近 8 亿人会用上移动互联网,现在这个预测因为手机涨价而变得可疑。这不是抽象的“数字鸿沟”,是一个个本来今年能买第一台手机的人,明年也买不起。
报道里没有出现的是那三家内存厂的名字被追问:它们和 AI 巨头签的合同里,有没有给消费电子留哪怕一条产线?也没有出现任何政府或监管机构在问这个问题。最穷的那批消费者在整条供应链里没有议价权,也没有代言人,他们的需求在利润排序里排在最后,然后被写成“市场调整”。
AI 的账单不会寄给用 AI 的人。它寄给那个还没上网的人。
The buildout of AI data centers is being paid for by people who have not yet bought their first smartphone. This is the Violence Triangle at its most literal: people who could have gotten online will not, because memory chips were redirected to feed AI. The gap is right there, and it was manufactured on purpose.
The mechanism is plain. Samsung, SK Hynix, and Micron make up more than 90% of the global memory chip market, and in late 2025 they shifted the vast majority of their supply toward AI data centers instead of consumer electronics (source: Rest of World). Memory got scarce, memory got expensive, and the increase was passed to the consumer. All three have partnerships with U.S. AI giants, and S&P Global estimates U.S. hyperscaler capex will exceed $1.3 trillion in 2027. This is not a market accident. It is a choice about whose demand gets served and whose gets squeezed.
What gets squeezed is precisely the sub-$150 phone. More than one in four smartphones shipped globally in 2025 cost less than $150; that market is shrinking. Sub-$100 shipments fell almost 60% year over year in Q2 2026. Chinese makers, about 60% of global shipments, drastically cut entry-level projects this year and pivoted to premium devices. Xiaomi raised the Redmi 15C in India from 12,499 to 16,999 rupees, a 36% jump; Oppo's sub-$100 shipments in Southeast Asia plunged 96%; in Africa, where 81% of phones shipped last year cost under $200, sub-$100 shipments fell 34%. What used to be under $150 may become under $250 or $300, and the analyst does not expect costs to return to pre-2025 levels.
Structural violence needs no villain. It only needs a supply chain sorted by margin. The price hikes land unevenly: 21% in India, 19% across Asia-Pacific, 18% in the Middle East and Africa, just 5% in the U.S. The same chip shortage, wildly different weight—heaviest where people can least afford it. GSMA finds an entry-level device costs the poorest 20% the equivalent of 44% of monthly income, and 76% in sub-Saharan Africa. Incomes will not keep pace. The result is not a delayed purchase; it is no internet-enabled device at all.
There is no face to blame here, which is exactly why it gets treated as weather. The memory makers did nothing wrong—they sold to the higher bidder. The phone makers did nothing wrong—they stopped making unprofitable products. Everyone made their optimal choice, and together they shut hundreds of millions out of the network. This is what the co-conspirators theory describes: no malice required, only everyone acting in their interest under the given order, and the order's violence executes itself.
What has been weaponized is the word "AI." It now functions as a universal justification: any resource tilted toward it becomes reasonable by default. Data centers need memory, so memory is theirs. Entry-level phones are unprofitable, so entry-level phones disappear. No one has to answer for the decision, because every step is rational. Rationality here is the disguise of meta-violence—it recasts a distributional choice as the inevitable march of technology.
What is actually being erased is concrete: GSMA had projected nearly 800 million more people would use mobile internet by 2030. That projection is now in doubt because phones got more expensive. This is not an abstract "digital divide." It is a person who could have bought a first phone this year and will not next year either.
What the report does not show is anyone asking the three memory makers whether their contracts with AI giants left a single line for consumer electronics. Nor does any government or regulator appear to be asking. The poorest consumers have no bargaining power anywhere in this chain and no advocate. Their demand ranks last in the margin sort, and then gets written up as a "market adjustment."
The AI bill is not sent to the people using AI. It is sent to the person who has not yet gotten online.