AI：錯誤的泡沫類型

Hacker News

大約 1 個月前

AI 生成摘要

本文認為當前的人工智慧泡沫並非如歷史上能奠定關鍵基礎設施的「生產性」泡沫。作者批評了將所有泡沫都視為有益的過度簡化觀點，並準備透過 Carlota Perez 的技術浪潮週期模型來分析 AI 泡沫。

AI: The Wrong Kind of Bubble. AI might change the world, but not in… | by Sameer Singh | Jan, 2026 | Breadcrumb.vc

Breadcrumb.vc

Follow publication

Network effects are among the most powerful, but also the most misunderstood, forces that shape technology startups. Breadcrumb.vc is my attempt to lay a “trail of breadcrumbs” to help founders and investors.

Follow publication

AI: The Wrong Kind of Bubble

AI might change the world, but not in the way you think

Listen

This post is bit of a history lesson, but an important one. Lately, there has been a lot of discussion about “productive” bubbles in technology — manias that focus capital and talent around a vision of the future. The argument is that bubbles create critical infrastructure that entrepreneurs leverage after the crash — the dotcom bubble is a key example. This is an oversimplified view of bubbles. All technology bubbles are not necessarily “productive”. They don’t always create infrastructure that is used after the bubble. Let’s take a look at whether this applies to the current AI bubble — and it is one, despite real utility.

The Two Types of Technology Bubbles

For those unfamiliar with Carlota Perez’s Technology Surge Cycle, it is based on a 200 year history of technology revolutions — canals, railways, electricity, automobiles and the current computing era. According to the model, new technology revolutions go through the following phases:

The entire cycle, from irruption to maturity, lasts about 50–60 years. Here is how this model applies to the current computing era:

The computing era began in 1971 with the invention of the microprocessor — the “cheap input”. It was originally designed for calculators before ending up in mass produced automobiles and home appliances — products of the previous technology surge reaching maturity. However, the true revolution began when hobbyists used the microprocessor to invent the personal computer — the Apple II in 1977 was the first realization of this vision. Mainstream adoption of the PC created an initial set of winners (Apple, Microsoft, Dell, etc.) and laid the groundwork for the internet — leading to the dotcom bubble and crash. After the crash, the true potential of the internet and then mobile was realized, driving a long period of economic growth post-2008. Today, the internet and mobile are ubiquitous and the winners of the era are looking for the next technological push to keep the good times going. This has led to unprecedented levels of investment in AI and data centers.

Why do I think AI is part of the maturity phase and not a whole new cycle? Simple, it carries all the hallmarks of a maturity innovation (h/t to Jerry Neumann) — the biggest incumbents in the technology world immediately jumped on it and adoption was lightning fast as the internet is already fully diffused. Whole new technology cycles begin on the fringes with hobbyists — like unknown nerds creating the Apple I. That is absolutely not what we’re seeing with AI.

Mid-Cycle vs. Late-Cycle Bubbles

A key takeaway from this model is that technology diffusion causes not one, but two types of financial excess — a mid-cycle bubble during the Frenzy phase and a late-cycle bubble during the Maturity phase. The two types of bubbles are strikingly clear in the chart below. The chart shows the Shiller PE Ratio, also known as the Cyclically Adjusted PE Ratio (CAPE Ratio), for the S&P 500 over the past 150 years. This ratio swaps out one year earnings (the E in the PE Ratio) with 10 year average inflation adjusted earnings to smooth out temporary fluctuations and account for business cycles (earnings can also be inflated during bubbles).

Let’s take a deeper look at the causes and after-effects of the two types of bubbles:

Mid-Cycle (Frenzy) Bubbles

A typical mid-cycle (Frenzy) bubble follows the first visible winners of a new technology surge, like Apple, Microsoft and Dell in the late 1980s. Investors begin to believe that every technology innovation will deliver exponential returns. This leads to frenzied speculation on new, unproven companies with questionable earnings. Talk of a New Era is everywhere, and investors come up with creative ways to value these companies. This is exactly what happened in the dotcom era, with scores of internet companies going public with minimal revenue and valuations based on “eyeballs". This environment fuels intense speculation and violent upswings in stock price multiples before a dramatic crash.

On the positive side, this period of speculation also results in significant infrastructure investments. While the crash results in steep losses for investors, it leaves behind a physical foundation that enterprenuers can build on. During the dotcom bubble, telecom companies laid millions of miles of fiber optic cable to profit from the sudden surge of interest in the internet. After the collapse, this “dark fiber” became the backbone of cheap internet access, fueling the success of internet companies in the post-dotcom era. Similarly, the 1929 bubble led to massive investments in road paving and radio networks, creating critical infrastructure for the automobile and mass marketing boom in the post-WWII era. This pattern also occurred during in 1840s, with nearly 10,000 miles of track laid during peak “Railway Mania” before the Panic of 1847.

Late-Cycle (Maturity) Bubbles

In contrast, a late-cycle (Maturity) bubble looks very different. At this stage, the technology has won and the dominant companies of the era are so entrenched that they are thought to be infallible. It is this belief that leads to another wave of speculative excess and overinvestment.

Any new technology breakthroughs are adopted at lightning speed. The incumbents begin to invest aggressively to harness its potential and strengthen their positions. The strong cash position of established companies leads to a heavy amount of financial engineering. Exuberance also leads to risky debt issuances. This is exactly what is happening with AI — rapid adoption, aggressive investments from incumbents like Google, financial engineering with Nvidia investing in companies that buy its products, and companies like Oracle and Coreweave taking on unprecedented levels of high-risk debt to finance data center construction. The presence of companies with real earnings makes this bubble less violent on the upswing, but no less speculative.

Crucially, maturity bubbles do not leave behind infrastructure for future use. Instead, they trigger overconsumption of the input that created the surge in the first place — leading to severe shortages, skyrocketing prices and geopolitical friction. This supply shock inevitably leads to abandoned infrastructure projects, sparking a collapse (or a series of collapses). The root of the collapse is a breakdown in the core assumption driving the technology surge — unlimited cheap inputs.

The Memory Crisis

This is exactly what we are seeing with the current memory crisis. Random Access Memory (RAM) is a foundational component of all computing hardware. Before this crisis, it typically accounted for roughly 40–60% of a GPU’s bill of materials (BOM). Aggressive expansion in data center construction has triggered a severe supply shock, with prices going up 3–4x in a few months with no end in sight. This crisis has been created by the need for high bandwidth memory (HBM) for data center GPUs, which consumes 3x as many wafers as conventional DRAM.

Get Sameer Singh’s stories in your inbox

Join Medium for free to get updates from this writer.

The most recent parallel was the oil crisis of 1973 — triggered by the OPEC embargo, which ended the Nifty Fifty bubble during an era of gas guzzling muscle cars. Previous examples include copper shortages from the early 1900s (caused by industrial demand and later WW1) and the coal famine in 1872 (caused by a surge in industrial and railway demand).

Let’s take a closer look at the similarities between the current memory crisis and the oil crisis in the 1970s. To state the obvious, the price appreciation looks remarkably similar.

What’s driving this supply shock? Three companies — Micron, SK Hynix and Samsung — control 95% of global RAM capacity. Increasing capacity is exceptionally challenging as new fabs take 3–5 years from planning to first production, at the cost of $10B-$20B. Case in point, Micron just broke ground for a new facility in New York (in Jan 2026) that was announced in Oct 2022 — a month before ChatGPT launched!

Previous ramps in capex have resulted in years of losses for manufacturers — the risk of a demand shift during this prolonged capex cycle is simply too great. So they are understandably careful of taking on commitments to rapidly scale capacity. Instead, their incentive is to gradually increase capacity, while maintaining complete price (and margin) control — much like the OPEC countries in the 1970s. Even if they (or smaller Chinese manufacturers) wanted to massively scale up capex, it wouldn’t have an impact on availability and prices until 2028/2029 at the earliest.

New approaches like Deepseek’s Engram attempt to offload some memory retrieval from HBM to NAND flash memory. This wouldn’t solve the crisis either. NAND is experiencing the same supply shock because of data center demand and has similarly long capex cycles. Aside from a collapse in demand, there is no near-term solution to this supply crisis — which also means there is no theoretical upper limit to memory prices.

In previous late-cycle bubbles, unsustainable input consumption and price shocks eventually resulted in a painful shift towards efficiency — as we saw with fuel efficient Japanese cars after the oil shock of the 1970s. An equivalent today would be a gradual and painful shift to on-device (or edge) inference. Edge inference has zero marginal costs as the user already owns the hardware. Of course, the problem is that it does not generate any revenue for AI labs and the hyperscalers. It is also more challenging to market (or hype) as on-device models are smaller and less capable.

Where Are We Today?

By now, it should be obvious that we’re in the midst of a classic late-cycle maturity bubble. The spike in RAM prices is poised to make already challenging AI economics even worse. Unlike software that came before it, AI inference has meaningful marginal costs. This is the reason why AI applications have negative gross margins, despite a series of investor subsidies flowing through the value chain.

These high marginal costs are a direct consequence of unprecedented levels of data center spending. In 2025, companies spent nearly $450B+ on data center spending to generate just $65B in AI software/services revenue. As Harris Kupperman explained, the industry would need to generate >$480B in revenue just to generate a return on 2025 infrastructure spend. If that looks steep, the planned spend for the next few years will raise the bar even higher. With >$1T of planned data center spending over the next few years, the industry will need trillions of dollars in revenue to generate a return. Most importantly, this math was worked out before the explosion in memory prices — we are already seeing H100 (previous generation) GPU rental prices go up because of shortages. The gap between hope and reality is staggering, and there are no good answers to how this will be closed.

If that wasn’t a big enough problem, repurcussions are now being felt beyond the AI world. Memory chips are a critical component for any end product with a microprocessor. This supply shock will have a direct impact on the price of smartphones, laptops, consoles, appliances, TVs and even cars. Manufacturer margins are already slim, so they have no choice to pass on the costs to end consumers. On top of that, new (non-GPU/AI) cloud servers will be more expensive, which will have an impact on subscription prices. I’ll let you assess the likelihood of a backlash from consumers and regulators, especially when the cause is so easy to identify.

What Happens Next?

This is not to say that AI is a dead end as a technology. The utility is clear, but the only way to make the economics work is to moving inference to the device. Of course, that will be harder for models with trillions of parameters running autonomously for hours. Instead, we will need to work with smaller, more efficient, open source models for specific use cases. Again, think of the transition from the American gas guzzling muscle cars of the ’70s — which at the time were the best selling cars in history — to the fuel efficient Japanese cars of the ’80s.

However, this may still be a limiting view of technology evolution. Late-cycle bubbles do not leave behind lasting infrastructure, but they can create the spark for the next technology surge. Every new technology surge is created by the combination of a new cheap input and a novel invention to harness it. Steam power needed the locomotive, electricity needed the light bulb, petroleum needed the internal combustion engine and the automobile, and of course, the microprocessor needed the PC.

Intelligence, or rather inference, has a case for being the new cheap input. Once inference moves to the edge, it actually can be cheap — zero marginal cost. And so, the most important question of our era is this: What is the invention that will truly unlock the value of this input? So far, the industry has tried to churn out AI laptops, “pins”, pendants and soon earbuds. These are attempts to squeeze a new input into the logic of the existing technology, e.g. microprocessors in calculators, washing machines and automobiles in the 1970s or the “horseless carriage" from the 1890s.

If inference moves to the device — as it will need to for AI to realize its potential — it will also need an entirely new product category to unlock its potential. Fortunately, we already have a vast ecosystem of open source hardware that can be combined with open source models. The Raspberry Pi, its growing ecosystem of robotics kits and AI HATs, combined with the vast library of open source models on Hugging Face, are one set of early building blocks for the next "hobbyist’s toy”. Some signals to look for are early experimentation and adoption, combined with dismissiveness from today’s industry leaders. Mainstream skepticism is a key indicator because the invention needs to break so many assumptions of the current computing era. Imagine how executives at Ford and ExxonMobil would react to the Altair 8800 and the Homebrew Computer Club in 1970s.

That said, we will need a reset in both capital markets and ideas to get there. On-device inference and open source models are diametrically opposed to the business models of today’s major players — from the AI Labs to Nvidia. This transition will not be easy or painless, but it is inevitable because the economic assumptions behind the AI data center rollout are unsustainable. In this environment, many data center projects will likely end up stranded or abandoned — the 21st century equivalent of rusting iron railway tracks left behind after the Panic of 1873.

In that sense, AI is the wrong kind of bubble, but still a necessary one to liberate capital chasing the wrong ideas.

Note: All opinions are my own.

What you should do next…

1. Reach out: No warm intros required

I invest in pre-seed/seed-stage startups as a Partner at Speedinvest. If your startup is built on network effects and based in Europe, you can reach me here or here.

If you’re not ready to pitch yet, sign up for my weekly office hours to talk through any questions about building network effects or consumer products.

2. Take my Reforge course

Want to learn more about network effects? Take my Reforge course on identifying, bootstrapping, measuring and enhancing network effects. Start the course for free via this link!