AI Bubble Rumors: Why This Time Is So Different from 2000

There is a trillion-dollar question in the tech world: Are we in an AI bubble?

Looking at the headlines, it's hard not to be alarmed. The $3 to $4 trillion planned to be spent over the next five years to expand data center capacity. OpenAI's more than one trillion dollars in committed deals. The fact that in just the last three years, more dollars have been invested in data center capacity than the entire U.S. interstate highway system, which took 40 years.

These figures recall one of the darkest moments in tech history: the 2000 dot-com (or, more accurately, Telecom) bust. That era was also defined by new economy promises, mind-boggling spending and a euphoria that ended in disaster.

So, is history repeating itself? Or are the fundamental dynamics truly different this time?

At a recent a16z event, Gavin Baker , Managing Partner and CIO of Atreides Management and David George of a16z tackled this taboo question. Baker, as someone who personally lived through the 2000 bubble as a tech investor, gives a clear answer: "No, we are not in an AI bubble today."

In this article, we will delve into Baker and George's analysis, highlighting the key difference that separates today's capex frenzy from that of 2000 and discuss the disruptive effects of AI not only on hardware but also on the model layer, SaaS applications and future business models.

Understanding this topic is critical not just for your next investment decision, but also for grasping the fundamental dynamics that will shape the economic and technological landscape for the next decade.

The 2000 Bubble vs. 2025 Reality: The "Dark Fiber" Paradigm

The crux of Gavin Baker's analysis rests on a powerful analogy that summarizes the fundamental difference between 2000 and today: Dark Fiber

If you are a veteran of the year 2000, you know this term well. Back then, companies like WorldCom and Global Crossing laid tens of thousands of kilometers of fiber optic cable to meet the anticipated massive future demand of the internet. They constantly told investors: "We laid 200,000 miles of dark fiber this quarter, this is so amazing!"

The critical word here was "dark." Fiber is a useless piece of glass without the optical hardware, switches and routers on either end. It was laid, but it was not "lit up."

So, at the peak of the bubble, how much of this cable was dark? Baker recalls a shocking statistic:

“At the peak of the bubble, 97% of the fiber that had been laid in America was dark.”

In other words it was not being used. This was the definition of a supply glut based purely on speculation, not demand.

Now, let's contrast this picture with today. Let's look at GPUs, which are today's "fiber" Baker's observation is sharp:

“Contrast that with today. There are no dark GPUs.”

There is no such thing as a "dark GPU" today. On the contrary, when you read any technical paper, one of the biggest problems during a training run is that the GPUs are "melting" Demand is pushing supply so hard that every available unit is being used to its fullest extent.

This isn't just an anecdotal observation; it's reflected in the numbers:

1. Valuations

In 2000, at its peak, Cisco was trading at a P/E ratio of 150 to 180 times trailing earnings. Today, NVIDIA, the company at the center of the AI revolution, has a P/E ratio of around 50x. Valuations are at a much more reasonable level.

2. Return on Invested Capital (ROIC)

More importantly, one must look at the balance sheets of those doing the spending (the major tech companies). Baker points to the heart of these expenditures: the Return on Invested Capital (ROIC).

The ROIC rates of the massive public companies buying these GPUs have seen an approximately 10 point increase since they ramped up capex spending. This shows we are in a fundamentally different world from the 97% idle capacity of 2000. That was speculation; today, we have massive utilization and a positive return generated by that use.

3. The Spenders: History's Most Powerful Companies

David George highlights the other side of the equation: those writing these trillion-dollar checks. In 2000, the spenders were often unprofitable startups, inflated by venture capital. Today, these buyers are the most profitable companies in history: Google, Meta, Microsoft, Amazon.

In George’s words: “They happen to be the best companies in the history of the world” Collectively, these companies generate around $300 billion in free cash flow annually and hold $500 billion in cash on their balance sheets.

They have the power to spend this money. And why are they spending it? Because this isn't a luxury; it is an existential necessity. A quote attributed to Larry Page within Google summarizes the mentality:

"I'm happy to go bankrupt rather than lose this race."

"Round Tripping" Concerns and NVIDIA's Real Competitor

But what about the claims that some of this spending is round tripping? (i.e., Microsoft investing billions in OpenAI and OpenAI using that money to buy compute from Microsoft Azure. Or NVIDIA investing in various AI labs, which then use that money to buy NVIDIA GPUs).

This is a worrisome concept because it recalls the sham revenue tactics of 2000.

Gavin Baker acknowledges that this is objectively happening, but with two important nuances:

1. Scale: It is currently happening at a very small scale.
2. Motivation: The primary driver is not financial engineering, but competitive dynamics.

The Model Layer: The "Netscape Moment" and the Risk of Sustaining Innovation

Moving from hardware to software that is, to the models (OpenAI, Anthropic, Google Gemini, etc.) Baker's first warning is humility.

He shares an analogy:

“...if we're going to make an analogy and say that ChatGPT is to AI, has Netscape Navigator was to the internet. At this point in the internet boom, Google had not been founded. Mark Zuckerberg was in middle school. Travis Kalanick was in kindergarten. So it's just very early.”

If the release of ChatGPT was the internet's Netscape moment, the moment the public first encountered the technology en masse, then it is impossible to predict who the winners will be right now.

So, will this new wave be disruptive like the internet (creating new winners and destroying incumbents) or will it be a sustaining innovation (allowing the current giants to become even stronger)?

Baker notes there are reasonable arguments that AI could be a sustaining innovation. This is because the core ingredients needed to win this game (data, capital to buy compute and distribution) are all possessed in spades by today's biggest tech companies (the Mag7).

Data, distribution, compute, dollars and talent... They have it all. “They have every right to win.” Of course, this is valid only on one condition: if they execute well. If they don't, they could share the fate of IBM. At this point, Baker offers another great analogy:

“To me, ChatGPT was Pearl Harbor for Google.”

However, the business model of these model companies (OpenAI, Anthropic, etc.) will not resemble the SaaS (Software as a Service) models we know. Due to scaling laws and their compute-intensive nature, their gross margins will be structurally lower. It is unrealistic to expect the 80-90% SaaS margins of the 2021-2022 period in this space.

The Future of SaaS and Gross Margins as a "Badge of Honor"

This low margin observation brings us directly to the application layer: the SaaS world. You have likely seen the frequent "SaaS is dead" debate on X recently. Baker holds a more nuanced view on this. Although he initially thought all application SaaS (excluding infrastructure SaaS) might go to zero, he now believes there could be some very big winners, especially those serving the SMB market.

However, he thinks incumbent SaaS companies are making a critical mistake: Trying to preserve their existing high gross margins.

"It is definitionally impossible, given what we just discussed, to succeed in AI without gross margin pressure"

He states that success in the world of AI is, by definition, impossible without gross margin pressure. If your margins aren't falling, you are probably not using AI as you should be. He also finds it unnecessary for SaaS companies to fear this situation. Because we already have an existence proof: Microsoft and Adobe. When these two giants transitioned from on-premise licensing to the cloud, investors worried that margins would fall. Yes, cloud margins were lower, but Microsoft, which successfully completed this transition, became one of the best-performing stocks of the last 10 years.

SaaS companies need to look at declining gross margins not as "something to be feared" or a "badge of shame" but as a "mark of success." David George makes a fantastic point here. He says that when an AI company presents to them now, they look at the gross margins and low margins have become almost a "badge of honor."

You have to ask this simple question: Would you prefer $10 in revenue with 90% gross margins, or $50 in revenue with 60% gross margins? The answer is simple.

The Consumer, Chip Wars and Future Business Models

The final part of the conversation focuses on where the future is being shaped: consumer applications, the next step in the chip wars and the evolution of business models.

1. The Consumer Layer and the "Reasoning" Revolution

Google was the portal to the internet; its purpose was to understand your intent and direct you to another website. AI, however, doesn't direct you; it does the job for you.

At this point, Baker issues a warning to the AI companies launching their own browsers (likely referring to Perplexity, OpenAI etc.): Google, which owns Chrome with its 5 billion users, might let these companies test the market for 3-6 months and then release something far better, saying, "We had to do this too."

But a fundamental development has changed the game in AI models: "Reasoning." According to Baker, before reasoning, a frontier model without access to unique data and internet-scale distribution was “the fastest depreciating asset in history.”

However, reasoning capabilities have unlocked the user feedback loop (flywheel). Now, user responses to the model (RL - Reinforcement Learning) make the model better. This enables the "more users -> better algorithm -> better product -> more users" cycle, which was at the heart of every great consumer internet company (Google, Meta), for AI as well. This fundamentally changes the economics for independent labs like OpenAI, Anthropic and xAI.

2. The Final Act in the Chip Wars: NVIDIA vs. Google vs. (Broadcom + AMD)

Returning to the chip layer, Baker emphasizes that the fight here is more complex than it appears. NVIDIA is no longer just a semiconductor company; it's a software company with CUDA, a systems company with rack-level solutions and now a data center company with its network architecture.

The war is between NVIDIA and Google (TPU). However, there is a "third way" emerging against this duo: Broadcom and AMD. Broadcom is going to companies like Meta and saying, "Against NVIDIA's proprietary NVLink or InfiniBand, I will build you a 'fabric' (network fabric) on open standards over Ethernet. I will also make you your own version of a TPU (an ASIC). And you know, it took Google three generations to get the TPU working... If your ASIC doesn't work well, no problem, you can just plug AMD right in."

This means Broadcom and AMD are effectively going to market together. However, Baker's personal prediction is that most of these custom ASIC programs will be canceled, especially if Google starts selling TPUs externally.

3. The Evolution of Business Models: From Subscription to "Outcome"

The biggest opportunities in technology emerge when a platform shift is combined with a business model shift. We are experiencing exactly that right now. We are moving from SaaS's "per-seat" monthly model to an "outcome-based" or "consumption-based" model.

• Example 1: Customer Service. It is now possible to pay "per resolved issue" instead of "per agent"
• Example 2: Consumer. Baker foresees that his personal AI (which he says will be a version of Grok, as he is an xAI shareholder) will plan his next vacation, know the hotels he likes, get the best price and room from them and take an affiliate fee for the transaction.

This may be less efficient than Google's advertising model (because advertisers "systematically overpay"), but it is an inevitable change.

Conclusion: Not a Bubble, but the Birth of a New Ecosystem

The conclusion from this in-depth conversation with Gavin Baker and David George is clear: No, we are not in an AI bubble.

The 2000 crash was defined by speculation on unused infrastructure (97% dark fiber). Today is defined by investment in utilized infrastructure that is providing a positive ROI (no dark GPUs). In 2000, those burning money were unprofitable startups. Today, those spending the money are the richest, most profitable companies in history and they view this as an existential necessity.

However, this does not mean everything will stay the same. This new revolution has a different economic structure. Gross margins will not be like they were in the SaaS era. Business models will evolve from "subscription" to "outcome."

Instead of "SaaS is dead" it is more accurate to say, "SaaS is undergoing a massive transformation, just as it did when it moved from on-premise to the cloud." And in this transformation, the winners will not be those trying to protect their high margins, but those who accept lower margins as "proof of utilization" and a "badge of honor" and focus on volume.

We are not witnessing a bubble, but the painful, yet thoroughly rational, birth of a massive new ecosystem where the rules are being rewritten.

Reference:

"Is there an AI bubble?” Gavin Baker and David George