日益严峻的算力短缺
The Growing Compute Shortage

原始链接: https://www.apollo.com/wealth/insights-news/insights/2026/06/growing-compute-shortage

人工智能革命已从纯粹的软件竞赛,转变为一场针对物理基础设施的工业级竞争。先进制程半导体产能(特别是台积电的 N3 节点)、高带宽内存(HBM)以及电力供应的严重短缺,正造成重大的瓶颈。据预计,到 2030 年,数据中心的电力需求将达到 200 至 300 吉瓦,这导致变压器和燃气轮机等关键组件早已被预订一空,未来几年的产能已供不应求。 这种稀缺性正在从根本上改变市场估值。曾经被视为过时的资产,如拥有现有电力和电网接入的前加密货币挖矿场地,正作为“拉撒路资产”(Lazarus assets)重新焕发活力。与此同时,计算能力的获取已成为一种战略性的“竞争护城河”;尽早确保供应的企业占据了主导地位,而其他企业则面临配额限制或更高的成本。 最终,这种物理限制可能会暂时逆转人工智能成本下降的趋势,并可能减缓该技术向小型企业的普及。随着软件规模与芯片、散热和电力等有形资源紧密挂钩,人工智能竞赛的赢家将不仅由算法决定,还将取决于他们控制支撑这些技术所需稀缺物理基础设施的能力。

Hacker News 最新 | 过往 | 评论 | 提问 | 展示 | 招聘 | 提交 登录 日益增长的算力短缺 (apollo.com) 12 点,由 dmitriy_ko 发布于 1 小时前 | 隐藏 | 过往 | 收藏 | 1 条评论 帮助 oezi 4 分钟前 [–] 至少能有一张展示全球数据中心算力总量随时间变化及未来预期上线情况的图表会有所帮助。 回复 考虑申请 YC 2026 年秋季批次!申请截止日期为 7 月 27 日。 准则 | 常见问题 | 列表 | API | 安全 | 法律 | 申请 YC | 联系 搜索:
相关文章

原文

At the semiconductor level, TSMC’s advanced-node capacity—particularly N3, which underpins much of the AI accelerator ecosystem—is approaching full utilization through at least 2027.4 Advanced fabs take years to construct, and the EUV lithography machines required to equip them have similarly extended production timelines.

Memory has emerged as another major bottleneck. High-bandwidth memory (HBM), a specialized type of dynamic random-access memory (DRAM) essential for AI accelerators, remains structurally undersupplied as demand from hyperscalers exceeds existing production capacity. But AI systems also rely on large amounts of ordinary DRAM, and as manufacturers shift capacity toward HBM, the broader tightening of the memory market has driven spot DRAM prices up approximately 8x since early 2025.5 Importantly, this dynamic not only impacts AI infrastructure providers. Industries competing for similar memory supply, including smartphones and consumer electronics, are already seeing upward pressure on component costs.

Even if silicon constraints ease, power infrastructure remains a critical limiting factor. Forecasts for global datacenter electricity demand now imply 200-300 gigawatts of data center power demand globally by 2030.6 Hyperscalers are pursuing gigawatt-scale campuses while simultaneously racing to secure turbines, transformers, switchgear, and grid connectivity. GE Vernova and Siemens Energy have both said they are nearly sold out of gas turbines through 2029.7 Transformer lead times have stretched to multiple years, and utilities and regional grids are increasingly struggling to respond to the rising demand from AI datacenters.

Scarcity Is Repricing Assets

As a result, the market is beginning to reprice anything perceived as a bottleneck to the AI buildout.

One of the clearest examples is bitcoin miners. For years, many crypto mining assets were viewed as structurally impaired following the collapse in crypto valuations and increasing competition within the sector. But in a compute-constrained world, those same assets suddenly look strategically valuable. Many miners already possess what hyperscalers need most: Access to power, land, and grid connectivity. Some companies like Crusoe and Core Scientific, that started as crypto-focused enterprises, have already made that transition, but more are likely to follow the same path now.

In other words, scarcity is creating “Lazarus assets”—assets once viewed as obsolete that are being revived by their relevance to AI infrastructure.

This rerating dynamic has spread across the broader economy. Power equipment manufacturers, cooling companies, electrical component suppliers, networking providers, and even industrial and brownfield infrastructure assets are increasingly valued through the lens of their ability to alleviate compute constraints.

Compute Is Becoming a Competitive Moat

In a compute-constrained world, access itself becomes a competitive moat. Some of the largest AI platforms and model builders have committed enormous amounts of capital to secure long-term compute supply. What looked like an overreach only months ago increasingly looks like strategic foresight. Companies with secured capacity can serve demand today; those without it could face rationing, slower deployment, or significantly higher costs.

This dynamic may also alter the economics of AI itself. For years, the dominant assumption surrounding AI was that models would become rapidly cheaper over time as hardware improved and efficiency gains accumulated. That may still prove true eventually. But in the near term, scarcity could temporarily reverse that trend. If demand continues to outpace supply, the immediate solutions are either materially higher pricing or some form of usage rationing for the leading-edge models. Early signs of both are already emerging across parts of the industry.

Higher inference costs would also likely slow the diffusion of AI across the economy, concentrating compute resources toward the highest-value applications first. Companies with the ability to pay for compute may gain an advantage, while smaller firms and experimental use cases could face increasing barriers to access.

Several developments could ultimately ease the compute shortage narrative: A sharper-than-expected improvement in model efficiency, a cyclical slowdown in AI demand, a faster ramp in semiconductor and power infrastructure capacity, or breakthroughs that reduce dependence on today’s constrained hardware stack. But for now, the opposite dynamic appears to be dominating. Demand is accelerating faster than the physical world can respond.

The Physical World Still Matters

For the better part of two decades, the dominant assumption in technology was that software scaled infinitely while the physical world slowly faded into the background. AI is reversing that logic. The next phase of the technology cycle will not simply be determined by better models or smarter algorithms, but by access to scarce physical resources: Chips, memory, electricity, land, cooling, and industrial infrastructure. In other words, the AI race is no longer just a software race. It is an industrial-scale, infrastructure race.

And increasingly, the winners may not just be the companies with the best ideas, but the ones that secured the compute to turn those ideas into reality.

In Dune, spice scarcity reshaped the balance of power across the galaxy. In today’s economy, compute may be beginning to do the same.

联系我们 contact @ memedata.com