The real reason your phone is getting more expensive
The real reason your phone is getting more expensive
手机价格上涨的真正原因
Soaring RAM costs are driving up prices across the tech industry. It may take years to settle down. 内存（RAM）成本的飙升正在推高整个科技行业的定价。这种情况可能需要数年时间才能平息。
When Apple debuts the next generation of iPhones this week, they’re likely to come with an unwanted change: a higher price tag. A price hike from the supply-chain powerhouse would be the clearest sign yet that soaring memory costs have become unavoidable — with no end to the memory crunch in sight. 当苹果公司本周发布新一代 iPhone 时，它们很可能会带来一个不受欢迎的变化：更高的价格标签。作为供应链巨头，苹果如果涨价，将是最明确的信号，表明内存成本的飙升已不可避免，且这场“内存荒”短期内看不到尽头。
Call it “chipflation” or “RAMageddon.” The shortage is reversing a decades-long decline in memory costs that helped make consumer electronics more powerful without making them dramatically more expensive. The terms “memory prices” and “memory shortage” appeared in 473 company transcripts last quarter, according to data provided by AlphaSense. The entire industry is seemingly working on the problem. And yet, we’re years away from seeing it fixed. 你可以称之为“芯片通胀”或“内存末日”。这场短缺正在扭转过去几十年来内存成本持续下降的趋势——正是这种下降使得消费电子产品在性能提升的同时，价格没有大幅上涨。根据 AlphaSense 提供的数据，上个季度有 473 份公司财报电话会议记录中提到了“内存价格”和“内存短缺”这两个词。整个行业似乎都在努力解决这个问题，但我们距离彻底解决它还有数年之遥。
The common wisdom is that AI is at fault: The technology isn’t just contributing to higher energy prices and layoffs; it’s also helping drive up the cost of everything from smartphones to game consoles. 普遍的观点认为 AI 是罪魁祸首：这项技术不仅导致了能源价格上涨和裁员，还在推高从智能手机到游戏机等各类产品的成本。
But the shortage was brewing well before ChatGPT took off — and was then amplified by AI’s unending hunger for RAM. The current market is the product of a complicated and very profitable reshuffling of the memory industry within a system that was already failing to keep up. 但短缺在 ChatGPT 爆发之前就已经在酝酿，随后又被 AI 对内存永无止境的需求所放大。当前的市场是内存行业在原有系统已无法跟上需求的情况下，进行了一场复杂且利润丰厚的重组后的产物。
“We need to build more wafer capacity,” Manish Bhatia, president and COO of Micron — one of the three largest memory manufacturers — told The Verge. “[It’s] a very different challenge for the industry than it had been for many years before, where technology alone was able to keep up with the demand.” “我们需要建立更多的晶圆产能，”美光科技（全球三大内存制造商之一）总裁兼首席运营官 Manish Bhatia 告诉 The Verge，“这对行业来说是一个与过去多年截然不同的挑战，在过去，仅靠技术进步就能跟上需求。”
For years, memory manufacturers could increase production by fitting more chips onto each wafer. But those gains were shrinking and taking longer to achieve. In 2021, even as pandemic-era electronics demand surged, Micron concluded that the problem was more fundamental: Technological advances alone would no longer create enough capacity to keep pace with long-term demand. Manufacturers would have to process more wafers — and build enormous new facilities to do it. 多年来，内存制造商可以通过在每片晶圆上塞入更多芯片来提高产量。但这些收益正在缩减，且实现难度越来越大。2021 年，尽管疫情期间电子产品需求激增，美光仍得出结论：问题更为根本——仅靠技术进步已无法创造足够的产能来跟上长期需求。制造商必须加工更多的晶圆，并为此建造巨大的新工厂。
Then the memory business collapsed. Pandemic-era purchases of computers, tablets, and phones had pulled demand forward; consumer spending weakened, and manufacturers were left with excess inventory. They lost money and slowed their expansion plans. 随后，内存业务崩溃了。疫情期间对电脑、平板电脑和手机的购买透支了需求；消费者支出减弱，制造商库存过剩。他们亏损并放缓了扩张计划。
By the time the market began recovering, generative AI had unleashed a wave of demand far larger and more memory-intensive than manufacturers had ever anticipated. 当市场开始复苏时，生成式 AI 已经引发了一波远超制造商预期的、规模更大且更依赖内存的需求浪潮。
The memory business is also extraordinarily concentrated. Three manufacturers account for about 90 percent of the market, according to Counterpoint, leaving the world dependent on a handful of companies to divide limited capacity between AI infrastructure and consumer devices. Counterpoint estimates that Samsung controlled 39 percent of the memory market in the second quarter of 2026, followed by SK Hynix at 26 percent and Micron at 25 percent. 内存行业也高度集中。根据 Counterpoint 的数据，三家制造商占据了约 90% 的市场份额，这使得全球不得不依赖少数几家公司在 AI 基础设施和消费设备之间分配有限的产能。Counterpoint 估计，2026 年第二季度，三星控制了 39% 的内存市场，其次是 SK 海力士（26%）和美光（25%）。
The memory inside your phone or computer broadly falls into two buckets. DRAM, or dynamic random-access memory, temporarily holds the information a device needs while opening apps, loading webpages, or running software. NAND flash memory provides longer-term storage for things like photos and files. 手机或电脑内部的内存大致分为两类。DRAM（动态随机存取存储器）在打开应用程序、加载网页或运行软件时，临时存储设备所需的信息。NAND 闪存则为照片和文件等提供长期存储。
“It’s not as simple as saying data centers are consuming RAM. The RAM is not the same.” “不能简单地说数据中心正在消耗内存。因为内存的种类并不相同。”
Inside AI data centers, specialized processors rely heavily on a form of DRAM called high-bandwidth memory, or HBM. “It’s not as simple as saying data centers are consuming RAM,” David Naranjo, associate director at Counterpoint, told The Verge. “The RAM is not the same.” 在 AI 数据中心内部，专用处理器严重依赖一种称为高带宽内存（HBM）的 DRAM。“不能简单地说数据中心正在消耗内存，”Counterpoint 副总监 David Naranjo 告诉 The Verge，“因为内存的种类并不相同。”
By stacking memory chips together and using advanced connections and packaging, HBM can move enormous quantities of data to and from those processors much faster and more efficiently. 通过将内存芯片堆叠在一起并使用先进的连接和封装技术，HBM 能够以更快、更高效的方式在处理器之间传输海量数据。
It’s more difficult to produce, but also more lucrative to sell. Deep-pocketed AI chipmakers like Nvidia and AMD and tech giants like Meta and Microsoft have an insatiable need for it to power their AI systems — and are more than happy to pay. 它生产难度更大，但利润也更高。像英伟达和 AMD 这样财力雄厚的 AI 芯片制造商，以及 Meta 和微软这样的科技巨头，对它有着无法满足的需求，以驱动其 AI 系统——而且他们非常乐意为此买单。
HBM also consumes significantly more manufacturing capacity than conventional DRAM. Memory chips are made many at a time on large, circular silicon wafers. Because a finished HBM product stacks multiple, larger chips together, it requires considerably more silicon. Micron estimates that producing a given amount of HBM requires roughly three times as many wafers as producing the same amount of conventional DRAM. HBM 消耗的制造产能也远高于传统 DRAM。内存芯片是在大型圆形硅晶圆上批量制造的。由于成品 HBM 将多个更大的芯片堆叠在一起，它需要更多的硅材料。美光估计，生产等量的 HBM 所需的晶圆数量大约是生产同等数量传统 DRAM 的三倍。
The incentives increasingly favor AI. Rather than guessing how many phones or laptops will sell a year from now, memory makers can lock in multiyear commitments from some of the world’s richest companies. 激励机制正日益向 AI 倾斜。内存制造商无需猜测一年后能卖出多少手机或笔记本电脑，而是可以锁定全球最富有公司提供的多年期订单。
“We have been engaging in discussions with customers, prioritizing those who can guarantee committed future captive demand,” Jaejune Kim, Samsung’s executive vice president of memory, said on the company’s last earnings call. “我们一直在与客户进行讨论，优先考虑那些能够保证未来固定需求的客户，”三星内存业务执行副总裁 Jaejune Kim 在公司上一次财报电话会议上表示。
AI is also increasing demand for conventional DRAM. Phone and PC makers want to run smaller AI models directly on their devices, requiring more sophisticated conventional memory — and more of it. Future models may need substantially more memory than their predecessors to support those capabilities. AI 也在增加对传统 DRAM 的需求。手机和电脑制造商希望直接在设备上运行较小的 AI 模型，这需要更先进的传统内存，且需求量更大。未来的机型可能需要比前代产品多得多的内存来支持这些功能。
“The three big memory guys, they’re just basically allocating the capacity that they have to these companies.” “这三家内存巨头，基本上只是在将他们拥有的产能分配给这些公司。”
“As AI spreads across various services such as search, coding, and productivity tools, the scope of demand is widening from a memory perspective,” said Song Hyun-jong, president of SK Hynix. “随着 AI 扩展到搜索、编码和生产力工具等各种服务中，从内存的角度来看，需求范围正在扩大，”SK 海力士总裁 Song Hyun-jong 表示。