The market has been highly efficient at sniffing out bottlenecks in the AI trade – GPUs, data, infra, and electricity. But could the next bottleneck be…cash?

▶️Big Tech’s Big Capex

The combined CapEx of the Big4 (msft, meta, amzn, aapl) tech companies in FY24E is $210B, a 40% increase from FY23 (or +$60B).

Over the next five years, cumulative CapEx is going to be $1.2T.

Consensus 2024 CapEx estimates have been revised up by as much as 40% since Jan-24, and the numbers are still too low.

~70% of CapEx is est. to be spent on technical infrastructure (i.e., AI-related).

As a % of revenue, CapEx is at 15% in FY24E, higher than the recent highs of 12%. These recent highs already include big projects such as $AMZN doubling its fulfillment centers and $META transitioning from CPUs to GPUs for its ranking system.

h/t AllianceBernstein for the summary:

▶️The LLMs’ Perspective Indicates a Staggering ~5x Increase in Training Costs Annually

GPT4-equivalent: ~$300M in training costs (~25000 A100 equivalent)
GPT5-equivalent: ~$1B in training costs (~30000 H100 equivalent)
Model trained in 2025/26: ~$5-10B in training costs. (X.ai building a 100k H100 cluster, then a 300k B200 cluster)
Post-2025/2026 models: $50-100B…? Jensen already talking about 1M GPUs connected together OpenAI & MSFT plotting $100B in Project Stargate: shorturl.at/eD978

These numbers won’t translate linearly into Big Tech’s CapEx, but the direction is clear – that $210B still needs to grow significantly as models keep getting bigger!

▶️ROI? FCF!!

The market has focused on finding the ROI for AI investments. I happen to be extremely optimistic about ROI in the long run, but I am more concerned that Big Tech’s FCF might deplete and become the imminent bottleneck.

The Big 4’s combined FCF is ~$250B today. In the near term (until AI investments get reflected in revenue), I expect the FCF number to decline.

The market is not correctly modeling D&A and AI’s impact on margins. D&A in the outer years could be up to 50% too low (AI servers depreciate over a shorter useful life with higher D&A).

▶️The Outcome of This Arms Race?

Right now, Big Tech is stuck in a high-stakes game of AI poker. As long as one player keeps raising the stakes, the others have no choice but to call, since none of them can afford to fold. They’ve already put too much money into the pot.

If the current trend continues—each new AI model gets better but not enough to rake in big bucks—Big Tech could find themselves “cash poor” in 3-4 years. The “weak hands” are the first ones to fold.

$AAPL the smart outsider?
$AAPL’s CapEx has been remarkably stable (1H FY24 CapEx of $4.4B is down YoY from $6.7B in 2023 and a 2015-23 avg. of $4.4B), in part because the company has been increasingly using third-party data centers to de-risk their AI investments.

P.S. Reflecting on my tweet from 11 months ago reminds me of two things: 1) how quickly things can change, and 2) the simplicity of the concept that one guy’s revenue is another’s expense.
x.com/FredaDuan/stat…