Nvidia lost a leader. So did the entire U.S. stock market. Investors looking for guidance on the AI trade can turn to ARPDC. (Don't worry, it's defined below.)
Jensen Huang is out at the company he founded in the early 1990s, the one that makes the chips enabling the technology driving the U.S. economy.
The scandal that abruptly ended his tenure couldn't have come at a worse time for the AI trade, with investors worried about capital markets' ability to finance the trillion-dollar buildout and Democrats' willingness to let the buildout happen.
Bullish investors have taken solace in tracking capital spending at the companies building AI data centers: Amazon.com, Alphabet, Meta Platforms, Microsoft, SpaceX, Oracle, CoreWeave, and others.
That group of eight is expected to spend about $1 trillion in 2026 and is expected to spend $1.5 trillion in 2027. There is a range of estimates. Baird fears spending could be closer to $1.3 trillion. JPMorgan sees spending closer to $1.8 trillion.
Whatever happens, spending is expected to grow. Does anyone really think the Nvidia crisis means Alphabet will stop trying to grow its AI business to offset any weakness in its flagship Google Search business?
But with Nvidia stock floundering -- down roughly one-third from its 52-week high through midday trading on Friday -- hyperscaler CapEx isn't enough to power the AI bulls.
Try annualized revenue per AI data center. ARPDC for short. Which Barron's would like to pronounce as arp-a-da-k. It is, or should be, the new metric investors watch to make sure demand for Nvidia chips, and all the ancillary products and services AI data centers drive, remains robust with firm pricing.
ARPDC is easier to understand than AI token costs, which fall as AI computers get cheaper and new open-source AI models gain traction. It also answers the question: Are people using all this AI computing being built? (Eventually, the question will morph into: Is the computing power being used generating economic value?)
Getting one ARPDC metric isn't easy. There is no true spot price for a gigawatt of AI computing power the way investors can look up the price of copper. Investors have to infer rates from various companies and agreements.
Take SpaceX. In May, it agreed to provide Anthropic access to roughly 220,000 Nvidia GPUs for $1.25 billion per month, following a ramp-up period. In June, it agreed to provide Alphabet access to about 110,000 Nvidia GPUs for $920 million per month, with the full monthly fee beginning in October.
The power used by 330,000 Nvidia GPUs in a data center is roughly 500 megawatts, or half a gigawatt. Anthropic is paying about $3.8 billion per gigawatt per month, for an ARPDC of about $45 billion. Alphabet is paying closer to $5.5 billion per gigawatt per month, for an ARPDC of $65 billion.
ARPDC isn't rising 45% month over month. Each agreement is a little different. But the idea that one gigawatt of AI compute can generate $40 billion or more per year is a very useful thing to know, especially since one gigawatt costs in the range of $50 billion to build.
SpaceX also reports AI revenue and AI computing capacity at the end of each quarter. Its SI business (SpaceX now calls AI SI) generated third-quarter revenue of $7.6 billion against estimated average computing capacity of 1.55 gigawatts. That's an ARPDC of about $20 billion. The segment includes advertising and software revenue, however, limiting comparisons with compute-rental contracts. Wall Street projects fourth-quarter SI revenue of $16.3 billion. Average compute should be about 2.2 gigawatts, implying an ARPDC of about $30 billion.
The exact ARPDC isn't as important as the trend.
SpaceX isn't the only source of ARPDC. Embattled neocloud AI services provider CoreWeave said pricing was moving higher on its third-quarter earnings conference call. Its quarterly revenue per gigawatt of average "active power" was about $2.3 billion in the third quarter, up from $2.1 billion in the second quarter. (That's like ARPDC, but isn't annualized, and "active power" isn't exactly the same as data center capacity.)
Anthropic is about to go public. Its revenue and year-over-year growth will be another important source of information. Third-quarter revenue of roughly $17 billion grew more than 1,100% year over year.
Interestingly, Anthropic's margin after compute and infrastructure expenses was about 42% in the third quarter, down from 52% in the second quarter.
That suggests AI computing prices are rising, but with 1,100% revenue growth, Anthropic is likely to keep paying.
Comments