The case is that AI looks like a bubble only if prices are detached from demand; the evidence here says the opposite, with token growth, compute utilization, backlog, and rental prices all still tight. The parallel to dotcom is framed as imperfect: in the late 1990s bandwidth was collapsing in price as fiber sat idle, while now GPUs, memory, and power are still constrained and getting more expensive even on older hardware. Michael Bur’s shorts on AI names are set against a market where NVDA, AMD, ARM, MU, and SNDK have all ripped, but the argument is that the real warning sign would be utilization rolling over and prices falling. The last point is that the market can debate valuations, but the underlying AI buildout is still being prepaid years ahead: $1.7T of contracted revenue, 98% chip utilization, and rising power-auction prices are presented as the checkable facts.
ARM and AMD are up over 150% YTD; MU is up over 200%; SNDK is up well over 500%.
Michael Bur, who bet against housing before 2008, is now short AI stocks; in May he wrote that AI stocks feel like the last months of the 1999-2000 bubble.
Bur’s last public filing was in November, with about 80% of the portfolio in shorts on NVDA and PLTR; a week later he deregistered the fund and now discloses shorts through a paid Substack.
Bur’s current short list includes ORCL, PLTR, NBIS, CRWV, and MU.
From Netscape’s August 1995 IPO, the NASDAQ rose 400% over the next 4.5 years, turning $10,000 into $50,000; after VA Linux’s 698% first day in December 1999, the NASDAQ peaked 3 months later and then fell 78% over 2.5 years.
The NASDAQ stayed below its 2000 peak for 15 years after that decline.
Greenspan’s December 1996 'irrational exuberance' warning was followed by a nearly 4x NASDAQ gain over the next 3 years.
An April 1998 US government report said internet traffic was doubling every 100 days, but that estimate traced to one internet company and was wrong by an order of magnitude; real traffic was only doubling about once a year.
By summer 2001, a Wall Street estimate said data carriers were using less than 3% of the fiber they weighed, and bandwidth prices had been falling about 70% a year for 3 straight years.
A New York-to-Los Angeles data line fell from $1.8M-$8M a year in early 2000 to under $150,000 in 2002, even though internet traffic kept doubling every year through the crash.
Google’s token count rose 7x between last May and this May; OpenAI went from 6 billion tokens per minute last October to 15 billion about 6 months later; Open Router says traffic has grown at least 10x every year.
MSFT, GOOGL, AMZN, and META spent $147B on buildings and equipment in 2023, then $511B over the last 12 months, up 75% year over year.
GPT-4 launched in March 2023 at $30 per million tokens; 16 months later GPT-4o Mini was described as good or better at most tasks at $0.15 per million tokens, about 200x cheaper.
DeepSeek’s V4.1 Flash launched September 10 at as little as $0.60 per million output tokens; 8 days later it was the second-most-used model on Open Router, and demand for DeepSeek’s older models also kept growing.
NVDA compute is described as fully utilized across every cloud NVDA serves, GOOGL said last fall even its 7- and 8-year-old TPUs were at 100% utilization, and MSFT and AMZN said demand still exceeds supply and expect that to continue in 2027.
OpenAI stopped taking new signups for its $200/month pro plan on September 10, one week after launching GPT-6 Astra, because of capacity limits; MSFT’s CEO also said last year it had chips it could not plug in because there were not enough buildings and electricity.
H100 rental rates fell from $6.60 an hour to $2.80 by mid-last year over two years, but one-year H100 contracts jumped about 40% from last October to this March.
CRWV raised prices across its lineup by about 25% in July, NBIS is raising prices 20% starting October 1 for the second time in 3 months, and ORCL said GPU contracts were renewing or reselling 20% above prior levels even though most of the chips are 4 years old or older.
DRAM contract prices nearly doubled in Q1 and were still climbing; NBIS bills RAM separately and is raising memory prices by over 40%, while MU stock more than tripled and SNDK is up over 500%.
PJM’s contracted capacity payment jumped from about $29 per megawatt-day in 2024 to $270 in 2025 and $329 this year, an 11x increase in 2 years; regulators capped auction prices and the latest auction cleared 6,800 MW short, largely because of data-center demand.
The dotcom-era tell was falling prices on unused infrastructure; the AI tell here is the opposite: 98% chip utilization, rising rental and memory prices, and $1.7T of contracted revenue across MSFT ($678B, up 84% YoY), GOOGL Cloud ($514B), and AMZN ($496B), excluding neoclouds, NVDA, AMD, ARM, and AVGO.
CSCO passed MSFT as the most valuable company in March 2000 at about 200x earnings and 130x forward earnings; today NVDA is the most valuable company at 28x earnings and about 18x forward earnings, while SK Hynix, MU, and SNDK trade under 9x next year’s earnings.
The checkable AI-bubble test is framed as this: if GPU or memory utilization starts to fall and prices start to drop, supply is finally outrunning demand; until then, token growth, utilization, and pricing point to real demand rather than a bubble.