Ben Horowitz went on All-In last week and relayed a conversation he’d had with Andy Jassy the day before. Horowitz said everyone was out of compute until 2028. Jassy corrected him.

“I said, yeah, like, we’re out of everything, you know, till 2028. He’s like, oh, it’s beyond 2028. Like, I don’t think we get stuff back, like, end of 2028, maybe.”

That’s secondhand, off a podcast, from a guy with a very large AI book to talk up. Discount it if you want. It still lines up with what Jassy says on the record. On Amazon’s Q2 call in July, he said AWS won’t have enough capacity to meet 2026 demand even after this year’s spend, expects the same in 2027, and called the demand already in hand for 2028 “striking.” Amazon is spending roughly $200 billion on capex this year. The big four hyperscalers together are near $700 billion.

I’ve been saying this for months, to anyone who’ll listen: there is no AI bubble, and there’s nothing to burst. Demand is still through the roof. You can’t even buy the compute, and when you can’t buy the thing, you are not in the kind of bubble we had in 2000. Jassy just handed me the best evidence yet.

The timing is worth noting. AI infrastructure names have spent October getting hammered, and “bubble” is back on every financial show. Before we all start reliving 2001, it’s worth being precise about what actually broke last time.

What actually happened to the fiber

Most people remember the dot-com crash as Pets.com and the sock puppet. The bigger financial wreck was underneath it, in telecom.

The 1996 Telecom Act blew the market open. Cities were required to let basically anyone dig, and they did. The Seattle Times reported in 2002 that Global Crossing started tearing up Aurora Avenue weeks after the state had repaved it, and that three different carriers had trenched the same boulevard in Bellevue. Telecom equipment spending grew about 17% a year from 1995 to 2000, then fell 24% in 2001.

By early 2002 the paper put the share of that new cable sitting completely unused at roughly 95%. The FDIC’s own read that June was that only about 2 to 5 percent of installed fiber was lit. More than 60 telecom companies failed between 2001 and 2003. Global Crossing went bankrupt holding $25.5 billion in assets. Corning, which made the glass, went from over $100 a share to about $2 in two years.

Here’s the part people skip. The fiber was fine. It still works. Much of it carries your Netflix right now. Nobody overestimated the internet. They overestimated how fast the customers would show up, and they financed that guess with debt that couldn’t wait.

Horowitz put it simply: people had to get broadband, buy PCs, pay for a browser. “There weren’t enough consumers on the internet to absorb all the bandwidth that had gotten built out.” Supply showed up years ahead of demand, and the interest clock ran faster than the adoption clock.

Price is the tell

If you want to know whether you’re in a glut, don’t read the op-eds. Look at the price of the thing.

In a glut, price collapses. Horowitz lived this. At LoudCloud he locked in bandwidth at $1,500 a megabit on a three-year contract in 1999, right before bandwidth prices fell off a cliff. That’s what oversupply looks like from the buyer’s side: you signed at the top and the market walked away from you.

Now look at GPUs. SemiAnalysis tracks one-year H100 rental contracts, and the rate climbed almost 40%, from a low of $1.70 per GPU-hour in October 2025 to $2.35 by March. The H100 is a 2022 chip. It was supposed to get cheaper the minute Blackwell shipped. It didn’t. Nebius raised its on-demand H100 rate from $3.85 to $4.50 an hour on October 1.

Old hardware getting more expensive is not what an oversupplied market does. Gavin Baker had the cleanest version of this on a16z’s feed: at the 2000 peak, 97% of America’s fiber was dark. Today there are no dark GPUs.

The obvious pushback: bandwidth was scarce in 1999 too. That’s exactly why Horowitz was paying $1,500 a megabit. Scarcity at the peak is normal. It’s what triggers the overbuild in the first place.

So the real test is what happens when the new supply lands. In 2000 it landed and prices fell off a cliff. Today Blackwell is shipping and three-year-old H100s are still getting more expensive. The new capacity is being absorbed as fast as it comes online. That’s the difference.

Sold out isn’t the same as paid for

The best bear case isn’t that demand is fake. It’s that demand is circular. Labs raise from investors, spend it on compute from hyperscalers, and the hyperscalers invest back in the labs. A queue for 2028 capacity isn’t contracted revenue, and a queue can evaporate.

Fair. So follow the money to the end customer.

Anthropic told prospective IPO investors last week that Q3 revenue was $16.8 billion, up from $4.7 billion in Q1. It said it was cash-flow positive and brought in $1.70 of revenue for every dollar it spent on compute. Those are adjusted, unaudited, pre-IPO numbers, so haircut them. Even with a haircut, that’s a buyer of compute paying for it out of what its own customers pay.

The market’s reaction tells you where we actually are. Chip stocks sold off on that report because 45% quarter-over-quarter growth was a deceleration from the quarter before. A company that grew revenue more than 3.5x in six months disappointed people. That’s an expectations problem. It isn’t a demand problem.

And Jassy keeps making the same point about the enterprise side: it’s still early for inference running in real production apps. In 2000 the customers didn’t exist yet. Today they exist, they’re paying, and most of them haven’t really started.

The debt is real. The bubble isn’t.

None of this means every balance sheet in AI is pretty. Horowitz said it himself: a pathological debt situation can wreck anybody, and a16z stayed out of the neoclouds for exactly that reason. But bad debt and a bubble are different things, and the bears keep blurring them.

CoreWeave carries roughly $50 billion in debt and lease liabilities. Last quarter revenue more than doubled to about $2.6 billion and the net loss still widened to $626 million, mostly on interest. Its five-year credit default swaps hit a record 1,194.5 basis points on October 15, which the market reads as better-than-even odds of default within five years. Its unsecured bonds traded below 80 cents, with $4.4 billion due before year-end. Microsoft, for comparison, runs a debt-to-equity ratio of 0.29. CoreWeave’s is 8.94.

That’s the telecom pattern in miniature: a real asset, real demand, and the wrong capital structure. If CoreWeave breaks, it won’t be because nobody wanted GPUs. It’ll be because the debt came due before the cash did, in a market where last week’s hot inflation print put another Fed hike back on the table. That would be a credit event at one company, felt by the private credit funds and insurers who wrote the paper. It would not be an AI bubble bursting, any more than a leveraged landlord going under in a tight rental market proves nobody needs apartments.

The other line you’ll hear is that this is subprime all over again: speculative data center debt, sliced up and sold to pensions, endowments, and insurers. The plumbing rhymes. The collateral doesn’t.

Most data center securitizations are backed by lease payments, typically on 12-to-20-year leases with Microsoft, AWS, Meta, or Google. In 2007 the weak link was a borrower who couldn’t make the payment. Here the tenant is one of the strongest credits on the planet. The market is also a fraction of the size: about $61 billion of data center ABS and CMBS outstanding as of mid-2026, against roughly $1.3 trillion in subprime mortgages in 2007. And nobody is pricing it as riskless the way they priced subprime AAA tranches. AA-rated data center ABS have paid spreads of 155 to 175 basis points since early 2025.

Who holds it matters too. 2008 turned into a crisis because leveraged banks funded that paper with short-term money that could run overnight. Today insurers mostly take the senior loans on stabilized, already-leased buildings, after private credit funds take the construction risk. A life insurer holding a 15-year Microsoft lease doesn’t face a run.

The part that does look like subprime is GPU-backed lending to neoclouds, where the collateral loses value every year. That’s north of $20 billion, and Nvidia has reportedly been talking to insurers about covering loans made against its own chips, which should make anyone a little uneasy. It’s real. It’s also a small, identifiable slice, and it’s the same layer I already said could break. The lease paper is the safest debt in the stack. Calling all of it subprime is like calling every mortgage in 2006 a liar loan.

The hyperscalers are a different animal. Yes, their numbers are rough right now. Om Malik ran them in the spring: Alphabet’s free cash flow was down 38% year over year, and Meta had $12 billion of quarterly free cash flow against $145 billion of planned capex. The debt loads are heavy and getting heavier.

They’ll be fine. These are trillion-dollar companies with some of the most profitable businesses ever built sitting underneath the spend, and every GPU they install is already spoken for. Depressed free cash flow during the biggest build-out in corporate history is what a build-out looks like.

The real ceiling on this cycle isn’t demand. It’s electrons. Transformers, turbines, interconnection queues, permits. That’s a government problem, and frankly it’s one my own party should be loudest about. If you want the data centers, the construction jobs, and the tax base in your state, you have to let people build the power. Right now the binding constraint on the most important industrial build-out of our lifetime is a permitting calendar.

What would change my mind

I could be wrong, so here’s what I’m watching.

  1. H100 contract pricing. If one-year rates break back below the $1.70 low while Blackwell capacity ramps, a glut is forming. That’s the dark fiber signal, and it’ll show up in rental prices before it shows up in anyone’s earnings call.
  2. Capacity language. Every hyperscaler is “capacity constrained” right now. The day Amazon or Microsoft starts saying supply and demand are “in balance,” pay attention.
  3. Who’s lending. If the neocloud credit stress spreads from the junk layer into investment-grade hyperscaler paper, this stops being one company’s balance sheet and becomes a system problem.

Until one of those moves, I’m sticking with what I’ve been saying. Fiber companies in the late nineties built for customers who hadn’t shown up yet. AWS is turning customers away. That’s a shortage, and shortages don’t pop. The only thing that can break here is a balance sheet that borrowed too much, and if you know which part of the stack holds the debt, you know exactly where to stand.

And one more possibility almost nobody on CNBC will say out loud: maybe these companies are just worth this much. A bubble is a price that has come loose from anything the business will ever earn. When the product is sold out, the customers are paying, and most of the buyers haven’t really started yet, a huge number on a stock screen isn’t proof of anything. Sometimes a big number is just the right number.