The Worst Stocks To Own When The AI Bubble Bursts

If you own index funds, you’re all-in on AI — whether you intend to be or not.

Today, 32% of the S&P 500’s entire market value sits in just seven stocks: Apple, Amazon, Google, Meta, Microsoft, Nvidia, and Tesla. That’s the highest concentration in the index’s history.

And these megatech companies are in the midst of making profoundly large bets on AI datacenters.

So if AI is in a bubble, the fallout won’t just hit tech bros, it could affect nearly everyone.

Mind Boggling Scale

Since ChatGPT launched in fall 2022, Amazon, Google, Meta, and Microsoft have increased annual capital spending by $200 billion, from $150B to $350B per year. And analysts think that number could approach half a trillion in 2026. 🤯

These companies are engaged in a historic AI arms race, building out multiple multigigawatt scale datacenters that cost tens of billions each.

The key problem is that no one really knows if there will ever be enough revenue generated to justify these massive investments.

Yet the big tech companies are committing increasingly larger portions of their cash flow to AI datacenters:

While the big tech companies have recently grown earnings at a faster rate than the rest of the S&P 500 index, higher depreciation expense from datacenters could slow down earnings growth in the years ahead.

Something similar happened back in 2022 when several of the big tech companies let expenses and investments grow too fast, which depressed earnings and cash flow.

The resulting stock drawdowns were quite painful for many investors:

This time around the investments are even larger in scale since the “Magnificent 7” CEOs view AI as the most fundamental technology shift of our lifetimes.

Newcomers Pile In

It’s not just megatech that is committed to building AI infrastructure either.

New AI focused cloud providers (aka “neoclouds) like CoreWeave and Nebius have emerged to try to meet the red hot demand for AI compute.

CoreWeave sports a $50 Billion+ revenue backlog, and Nebius recently signed a $17 Billion contract with Microsoft.

Oracle has also gotten into the AI cloud game in a big way. The enterprise software company recently signed a $300 Billion agreement with OpenAI that skyrocketed their revenue backlog:

Of course there’s a lot of skepticism in the market about that future potential revenue since OpenAI remains unprofitable and dependent on ongoing capital raises.

There are other players building out datacenter capacity as well, including xAI ($20 Billion recently raised), Chinese companies like Alibaba ($50 Billion+ committed), and the Kingdom of Saudi Arabia which intends to spend many billions more.

Global AI datacenter capacity is set to expand rapidly for the next several years:

DotCom 2.0?

This naturally has a lot of people worried about an AI datacenter / CapEx bubble.

With so many different companies building out datacenter supply at the same time, it feels like we’ll inevitably end up with too much capacity at some point.

Each player is looking at their own supply & demand and making the rational decision to expand capacity but few are considering the overall supply outlook a few years out.

This can inevitably lead to a vicious “capital cycle”, just as we saw with the early internet boom and bust in the late 1990’s & early 2000’s.

The result of that episode was a 40%+ drawdown in the S&P 500 index and a nearly 80% drawdown in the Nasdaq-100 index.

Many individual stocks went down over 90%, and even previous leaders like Cisco took more than a decade to recover to new highs.

One interesting insight about the Cisco example is that the company’s revenue and profits had recovered and started growing again by 2004 but the stock struggled for a decade due to valuation multiple compression.

Looking at datacenter spend today, it has an eerily similar magnitude to the fiber build out that peaked in early 2000:

No surprise then that we’ve seen every major news publication call this an “AI Bubble”.

We’ve also had investing firms like Apollo and Sequoia calling out the bubble potential.

Even OpenAI’s Sam Altman has admitted that we might be in the middle of an AI spend bubble:

That’s a wild admission from the CEO of the company at the center of the AI boom. OpenAI itself has committed to more than $1 TRILLION of spending over the next decade:

Nvidia invested in OpenAI recently which some see as proof that OpenAI has a clear path to make good on its contracts while the skeptics see it as a circular deal that allows Nvidia to basically pay for their own GPUs.

Let’s just say OpenAI’s funding and commitments are a key risk to the AI boom.

Datacenter Economics

One critical factor in determining whether we’re in a bubble or not is how much revenue is being generated by AI datacenters versus the cost of said AI factories.

A single one gigawatt datacenter can cost somewhere between $25 and $50 Billion dollars to build out and equip.

Industry-wide AI CapEx is tracking to $500 Billion for 2025.

To earn even a modest 9% return on this year’s build outs, we estimate that datacenter owners need to generate more than $200 Billion of new recurring revenue.

And next year’s global datacenter commitments (10-20 additional Gigawatts) add another $300 to $400 Billion to the required revenue base.

This is a massive revenue hurdle that grows by the month!

AI Revenue Projections

So how much revenue is currently being generated by AI? And can it grow enough in the future to avoid a nasty bubble pop?

OpenAI will do $12 Billion+ this year alone. Despite the company’s insane spend commitments, we can’t deny its success to date with products like chatGPT.

ChatGPT reached 100 million active users faster than any independent app in history, and was also fastest to 500 million active users. Today it has over 800 million weekly users, up 4X in just a year.

It’s also now in the top 10 most visited websites in the world and was by the far the fastest company to break in.

OpenAI is projecting more than $100 Billion of revenue by 2029:

How could it get there?

One path is to start taking a cut of transactions that were originated from within chatGPT, like when a user is recommended a product or a travel itinerary for example.

ChatGPT is already partnering with Shopify and Booking for this, and the chat application is already a major source of traffic for online retailers like Walmart.

Another path for chatGPT is integration into hardware like phones, cars, and headsets, which they are already working towards with famous designer Jony Ive.

Yet another path is software integration like what was recently announced with Intuit.

It’s certainly possible that OpenAI could go from being deeply unprofitable to one of the most profitable companies in the world.

However the path to get there is still very uncertain given the trillion plus of spending commitments and competition

Speaking of the competition, there are other major LLMs including Google’s Gemini, Anthropic’s Claude, xAI’s Grok, Deepseek, Alibaba’s Qwen, Meta’s Llama, and Mistral. As we expand our view to the overall industry, the revenue pie does expand quite a bit.

Anthropic projects it could be at a $9 Billion annual revenue run rate by the end of this year, and hopes to triple that by next year.

Google’s Cloud business will add over $10 Billion of revenue this year at a more than 30% growth rate, driven heavily by AI enhancements. Azure, AWS, and Oracle are also growing their cloud businesses significantly on the back of AI usage.

There are a lot of AI startups building their companies on these platforms including ones you’ve probably heard about like Cursor, Lovable, n8n, ElevenLabs, Runway, HeyGen, Perplexity, and many others.

Most existing software companies are also building in AI features, from large enterprises like Adobe and Salesforce to younger upstarts like Canva and Notion.

Between the cloud providers (Amazon, Google, Microsoft, Oracle), neoclouds, and startups we estimate that industry wide revenue could be in the neighborhood of $100 Billion in 2025:

If industry revenue doubles or triples next year it might be enough to deliver an acceptable return on 2025 industry CapEx.

However, there will likely be between 10 to 20 gigawatts of additional global AI datacenter capacity built in 2026. That raises the required revenue bar by another $300 to $400 Billion, which is the equivalent of adding the combined current annual revenue of AWS, Azure, and Google Cloud.

Revenue needs to grow very RAPIDLY and continue growing rapidly or we’re going to end up with datacenter overcapacity & low CapEx returns.

Investing Implications

There are hundreds of publicly traded stocks that have benefitted from the AI boom in the past three years. Here you can see the blistering performance of just a subset of these stocks:

Winning stocks include semiconductor hardware (Nvidia, AMD, Broadcom, Micron, TSMC), neoclouds (CoreWeave, Nebius, Oracle), electrical equipment & utilities (Eaton, Quanta, GE Vernova, Vistra, Constellation Energy), crypto miners turned datacenter operators (Iren, Core Scientific), and big tech (Google, Meta, Microsoft, Amazon).

If the world finds itself with too much AI datacenter capacity later this decade, some of these companies will fare much worse than others.

Here’s our educated guess at a “Tier List”, with the best reward-to-risk in the “S” and “A” tiers and the worst reward-to-risk in the “C” and “D” tiers:

If you want detailed commentary on why we categorized them this way, check out this YouTube video (starting roughly at the 27 minute mark):

Many of the companies mentioned in the tier list are benefitting from the first wave of the AI boom, the infrastructure build out. And many of their stocks have already priced in a bright future.

But ultimately we think much of the long term value from AI will be generated downstream by the companies providing AI applications to consumers and businesses.

In the dotcom boom, the early winners might have been infrastructure plays like Intel, Cisco, and Lucent but the real long-term winners were companies like Amazon, Booking, Facebook, Salesforce, and Uber.

We see opportunity in shifting focus to the companies that could start generating huge revenue growth by implementing the best AI features.

Sadly many of the most exciting companies like Cursor, ElevenLabs, HeyGen, n8n, Descript, Suno, and others are not yet public.

But there are some existing public companies out there which could have under-appreciated potential for future growth. We’ve been taking a close look at companies like Atlassian, Hubspot, Five9, Uipath, Adobe, Duolingo, and others.

We even wrote up a trade idea on one of these recently:

We expect to have a lot more thoughts to share on this theme in the near future!

The TL;DR (Takeaways)

1. An AI spending bubble is likely inevitable, BUT compute demand is still rising faster than supply right now. We might still be years away from the reckoning. Don’t get wrecked trying to call the top too early!

2. Most of the magnificent 7 stocks will be able to survive the inevitable fall out given their immense cash flows and ability to readjust Capex. The eventual hangover for these stocks could look a lot more like the 2022 correction than the 2000 bust.

3. Unproven concept stocks, neoclouds, and players that are solely dependent on AI infrastructure build outs to drive growth carry the most long term risk.

4. There’s still tremendous value to be generated in AI applications. If too much computing capacity is ultimately built, it's going to drive costs down and benefit the AI application developers the most.

There will be plenty of long term value generated in the aftermath of an AI infrastructure boom and bust. There’s always a bull market somewhere!