In partnership with

Tech stocks peaked on March 10th, 2000, and didn't reach that level again until April 2015. The two categories of tech investors are the ones for whom that drawdown was a formative experience, and the ones for whom it's as abstractly historical as the Panic of 1873. But it was a memorable experience. In retrospect, we tend to mentally draw fairly straight lines between big market inflections, but 2000 actually featured several shifting narratives. The year kicked off with the continuation of the vertiginous late-90s rally, when basically every tech stock was ripping, especially the smaller and more speculative ones and a few sporting what were then considered extreme price/sales ratios.¹

The really interesting phase was after the big March selloff, when prices started to creep back, but unevenly. The new, sophisticated claim was: who knows if Pets.com will be the one in twenty online pet stores that actually makes money? It's a sucker bet to figure it out. But, whoever does eventually win is still going to need networking equipment, and bandwidth to provide it. So, stop betting on the companies panning for gold, and bet on the ones selling picks and shovels like a responsible adult!

The only big problems with this were:

1. The tech financial system at that time was basically a machine for turning venture dollars and dot-com IPO proceeds into revenue for companies that sold servers, routers, databases, etc. There wasn't a good bridge between when this funding disappeared and when the first dot-com winners were generating enough operating cash flow to replace that demand. Amazon, for example, slammed on the brakes hard enough that by Q3 2001, their annual growth was +0.2% Y/Y, slower than GDP growth, and they were still far from generating positive cash flow.² It turned out that starting a dozen, well-funded companies in parallel to go after each vertical generated a lot more spending than having a handful of scrappy companies doing the same thing.

2. One of the big capex categories was fiber. If there's a shortage of bandwidth, it's great to own scarce wires. But the marginal cost of transmitting additional bits on a line that's below capacity is very low indeed, and it turned out that the market-clearing price for fiber was pretty close to free. By the time some fiber projects were complete, there wasn't any reason to turn them on. (That fiber did end up getting bought: when capex overshoots, there's net wealth destruction, but some people make money from selling early while others make money from buying very, very late.³ )

It's tempting to map this straight to AI. Consumer-facing AI applications have a tendency for their peak ARR to exceed their lifetime revenue: they grow very fast, but people churn fast, either because their product turns out to make more sense as a feature in an established company's app, or because the labs subsume it, or just because it was a fun toy that didn't find a business model in time. The investors backing datacenters are much more serious and much more careful.

But the supply-glut piece can't easily be replayed, because inference isn't a zero-marginal cost activity. There can be a glut that pushes prices down to the point that companies are reporting negative GAAP income by pricing tokens at a slim margin above electricity prices, but what those companies are really doing is inefficiently liquidating their assets in order to pay back bondholders. Meanwhile, since the power part of the supply chain lags more than the chip production part, inference can get cheaper again as power gets cheaper.⁴ And, unlike in the dot-com boom, we do have a pretty good idea of where to direct money in order to bet on current trends continuing. Even though AI companies will have weirder economics than most businesses, their business models are pretty legible to investors—if investors in 1999 had thought about cohort economics the way modern investors do, and had convinced companies to report them accordingly, Amazon would have had a much smaller drawdown.

But the parallel to worry about is thesis drift during drawdowns. If investors start getting cautious about AI, and capital shifts from applications and models to infrastructure, there's always the risk that they're betting on revenue that's from other companies' venture funding, even as that funding disappears. But this, too, is something that people who survived the dot-com bubble remember well (or anyone in a cyclical industry): a drop in demand somewhere will ripple through the relevant supply chain and be very bad news to whichever link in that supply chain most overinvested. If there is a big bear market in AI soon, there will be brief periods when some wishful thinking and special pleading makes it seem that it won't apply to some places. And it'll be good to remember that these have a shorter half-life than the narratives they replace.

The summer 2000 rally has many lessons for investors, and echoes other events in financial history where a trend looked like it was reversing but was actually just getting started. We've written some related pieces in The Diff:

• Here's more detail on what specific bet neocloud investors are making.

• Thoughts on what AI economics would look like after a bust.

• One of the things that keeps these rallies going is opportunistic M&A, which often doesn't work out. But the acquisitions that matter are the ones that would be blocked on antitrust grounds if regulators had as much foresight as buyers ($).

• You can hide but not remove most volatility ($).

• One feature of the economic cycle is that some industries get more capital-intensive as they grow ($).

GTM Atlas, by Attio

Your GTM motion is creative. The thinking behind it should be too.

GTM Atlas is the ultimate resource on AI GTM for early-stage builders, providing foundational knowledge for teams navigating growth from scratch. Curated by Attio, the AI CRM, Atlas gives you:

• Systems thinking for every stage of the customer journey

• Frameworks and templates that scale with you

• Conversations with GTM operators at Clay, Lovable, and Vercel.

Mapped by operators. Curated by Attio.

Explore now

Join the discussion!

1  At the time, much more tech revenue came from one-off purchases rather than subscriptions, so for a given growth rate, a typical tech company today does deserve a higher multiple. Some of that multiple will just come from the fact that, even if nothing changes about their underlying economics, they're recognizing revenue over the life of a customer relationship instead of recognizing that revenue upfront. But in practice, subscription revenue is more durable (and more likely to expand) than that implies.

2  In fact, one of the things investors liked about Amazon was that their high inventory turnover and ability to source the long tail of books from wholesalers meant that they didn't need much working capital. But for any business that's collecting cash faster than the typical term of its payables, that means that slower growth also needs more capital.

3  Of course, the presence of this alpha in a trade that produced negative beta over that timeframe means that gross losses were even bigger.

4  Incidentally, this means that the position that AI is both a bogus technology and making power bills rise is incoherent. A new power plant built to power a datacenter is going to have a much longer useful life than the chips in that datacenter, so the correctly-calibrated way to say this is "AI is a boondoggle, but fortunately VCs are throwing so much money at it that our power bills will go down as soon as they realize their mistake." Unfortunately, the earliest cohort of AI skeptics were so brilliant that they had almost no common ground with the newer variety.