Stocks
of publicly traded companies including Coreweave, META, Microsoft, and
Oracle have lost 4% to 39% of their stock market value over the last
year. Coreweave's
core business is building data centers stocked with GPUs to rent AI
service providers for estimation and inference. Obviously, this makes
the standard data center look downright capital light. Racks must be
populated with servers and GPUs and capable of handling 120Kw power
loads. So power infrastructure and backup alone cost many multiples more
on a per square meter basis than the standard telecom hotel. Moreover,
these data centers must be bigger because a large language model might
have trillions of parameters to be estimated. Coreweave's massive debt
loads have led to a 37% decline in its market value over the last twelve
months.
Secondly,
customer switching costs are very low in the AI data center market. AI
service providers are software companies. They download their data into
the bare metal servers, estimate them, and then distribute them to
inference nodes to provide service to end users. No need for physical
colocation. Right now there is excess capacity in the market as owners
of installed GPUs have contacted me asking to offload their excess
capacity. The SpaceX deals vividly illustrate the problem. Its own
division, xAI, is floundering in terms of subscription growth and its
own data center lie fallow. This is why it is doing toothless deals (no
early termination penalties) with Google and Anthropic.
Although
META does not build AI data centers for rent to third parties, it does
so for itself. So it faces the same problem of huge investment spending
with uncertain returns. META forecasts its 2026 capex to be almost $100
billion. Yet its cumulative AI service revenue to date is zero. Indeed,
Zuckerberg just announced META's first services. They will be sold at
20% to 40% below the market. It is a desperate move that will move
market pricing down in an industry where no provider is booking a profit
on AI services. Facebook stock is down 4% year/year.
Microsoft
also has a big imbalance between revenues and capital spending. It
plans $190 billion in 2026 capex, but its annual AI service run rate is
only $37 billion. Its stock has fallen 23% over the last 12 months.
Many
of the fathers of modern AI have disowned their children. Yann LeCun,
winner of the 2016 Turing Prize, the highest honor in computer science,
has declared large language models to be 'silly'. He recently left his
position as head of META AI research to found a new company in Paris
pioneering a new approach to general artificial intelligence.
All-in-all,
the Big Picture looks a lot like fibre optic overbuilding during the
dotcom era that led to a crash in bandwidth prices and many
bankruptcies. The main difference is that many current players have
monopoly revenue sources in areas like operating systems, browsers, and
social media.
Comments
Post a Comment