Key points
- Chip and memory prices keep climbing
- 2000 had a parts shortage too
- Bond buyers are treating AI debt differently
- One warning sign hasn't shown up yet
AI hardware keeps getting more expensive. Memory prices have surged, and chipmakers and server suppliers are preparing further increases. I keep wondering how much of this looks like 2000, so I went back and checked. That boom had parts shortages, nervous lenders, and a tightening Fed too. The question is whether today's rush to buy leaves suppliers holding excess inventory later.
TSMC (TSM) plans to raise chipmaking prices by up to 10% starting in January 2027, Nikkei Asia reported in July. TSMC declined to discuss specific prices but said in a statement, "Our pricing strategy is strategic, not opportunistic." Samsung Electronics (005930), TSMC's main rival in making chips for other companies, has already raised prices on new orders placed since July, Reuters reported.
Server buyers are facing higher prices too. Some of the biggest customers of Nvidia (NVDA) were told in August that servers built on its AI chips will cost over 15% more from early 2027 because of memory costs, Bloomberg reported. Last week, a sales-channel report said Advanced Micro Devices (AMD) plans to raise prices on its GPUs, AI chips, and chipsets by about 10% in the fourth quarter, TrendForce noted. AMD hasn't confirmed it. Last week, we separated the reported price increases from company statements.
Memory is the wildest one. A 64GB DDR5 server module had a contract price of $1,500 on September 15, up 451% from $272 a year earlier, the Seoul Economic Daily reported. On the spot market, where buyers pay for immediate delivery, it sold for $3,100. My car lost value the day I drove it off the lot. Server memory went the other way.
Rising prices alone didn't end the dot-com boom. The damage came when demand weakened and suppliers got stuck with unsold inventory. Stocks didn't wait for the news. The Nasdaq peaked in March 2000, more than a year before Cisco (CSCO) reported its inventory write-down. Here's how it played out, and what I'm watching now.
2000 had a parts shortage too
People remember 2000 for cheap PCs and failed dot-com companies. The parts inside the boom told a different story. Tantalum ore, used to make capacitors for phones and PCs, sold for about $28 to $48 a pound in early January 2000. By early December, it was $140 to $230, according to U.S. Geological Survey price data. The USGS later pointed to over-ordering, including what it called "double and triple ordering" for capacitors.
That's the part I'm watching now. Customers worried about shortages can place extra orders to make sure they get supplied. Suppliers see a full order book, but some of it is the same demand counted more than once. When customers stop ordering, the shortage can flip into excess inventory.
Today's suppliers have some advantages. TSMC's advanced manufacturing and Nvidia's AI technology are harder to replace than a commodity such as tantalum. That may help them defend prices. It can't make customers keep buying equipment they no longer need.
The last bust showed how fast that can matter. Cisco took a $2.25 billion charge for excess inventory in the quarter ended April 28, 2001, according to its quarterly filing. CEO John Chambers said in the earnings release, "This may be the fastest deceleration any company of our size has ever experienced."
The damage spread across the chip industry. Global semiconductor sales fell 32% in 2001, from $204 billion to $139 billion, the industry's worst annual decline at the time, EE Times reported, citing the Semiconductor Industry Association. Average selling prices at Micron (MU) fell about 60% during its fiscal 2001. In the fourth quarter, they were about 85% below a year earlier, according to its annual report.
How the money cracked in 2000
The hungriest buyers in the telecom boom were upstart carriers laying fiber with borrowed money. When lenders pulled back, those carriers stopped buying equipment almost overnight. 360networks filed for bankruptcy in June 2001. Global Crossing followed in January 2002, and WorldCom filed in July 2002, the largest US bankruptcy at the time.
We're not there. But the bond market is already treating AI debt differently, and we covered that split in detail. AI-related bonds are trading at wider spreads than the broader investment-grade market. They've traded around 115 basis points over Treasuries, according to Goldman Sachs (GS) data cited by Reuters, against 78 basis points for the broader market, according to ICE BofA. Russell Brownback of BlackRock (BLK) told Reuters that some AI deals carried double-A ratings but were pricing more like triple-B credits.
Credit markets showed strain before the Nasdaq peaked, too. The Fed reported in early 2000 that default rates on junk bonds had risen in 1999 "to levels not seen since the recession of 1990-91," according to its monetary policy report.
The weaker borrowers are where I'd look for the first real break. About $18 billion of loans tied to an Oracle (ORCL) data center in New Mexico were quoted at 89 to 91 cents on the dollar this month as the banks struggled to sell them, the Financial Times reported. That's not a closed credit window. It's a crack in one that's still open.
The Fed is the other piece that lines up. In 1999 and 2000, it raised rates six times, from 4.75% to 6.5%, according to its rate history. The Nasdaq peaked on March 10, 2000, right in the middle of it. On September 16, the Fed raised its target range to 3.75% to 4%, its first increase since 2023. On Wednesday, traders put the odds of another increase at the October 27-28 meeting at 73%, according to CME FedWatch data cited by CNBC. That jump followed comments from Fed Governor Michael Barr and an S&P Global report showing inflation at its highest level in nearly four years. I think the traders are right, and a Fed raising rates into a spending boom is the most 2000 thing on this list.
What would have to happen next
So here's my scorecard. Chip and memory prices are rising at every layer. Bond buyers are demanding wider spreads on AI debt. The Fed is raising rates.
Our AI Bubble Index measures a different set of pressures, including valuations, momentum, and how much of the announced buildout is backed by firm commitments. It stood at 60 out of 100, in its "Heated" category, on Wednesday evening.
I haven't yet seen the companies discussed here report the kind of broad order cancellations and inventory buildup that followed the telecom boom. The word to listen for on earnings calls is "digestion." Alongside that, I'm watching for shorter lead times, memory spot prices rolling over, and an AI borrower that can't raise money at any price. Micron reports on September 30, so that's the first place I'll be listening.
Warning signs aren't a timer, though. The Nasdaq closed 1998 at 2,192.68 and peaked at 5,048.62 in March 2000, more than double. I made the same point about Monday's ugly breadth: a pattern from 1999 deserves respect, not panic. Being right about the ending doesn't pay much if you're more than a year early.
If orders start drying up, the comparison with the telecom bust gets more serious. But I won't assume stocks will wait for that evidence. For now, these are warning signs I'm monitoring, not an automatic reason to sell.
Prices, spreads, and rate odds are as reported on the dates noted. The 1999 to 2001 figures come from company filings, the Federal Reserve, the U.S. Geological Survey, and trade press, as linked.
I am not a financial advisor, and nothing here is investment advice.
