Millions Watched Zuckerberg Sip A Guinness. Nobody Heard What He Said Next.
Authored by Patrick Feeley via Substack,
The Sip
On the October 12 episode of John Collison's Cheeky Pint, Mark Zuckerberg picked up a full pint of Guinness, took what looked like a mouthful of foam and held it there long enough for the clip to travel around X and pick up millions of views. Zuckerberg eventually responded himself with "that's just my face," which is fair, and I will admit I have made a similar face reading through a capex budget.
Most people who saw the clip probably did not watch the rest of the episode, which is a shame, because later in the conversation Zuckerberg said something that matters a great deal more for anyone trying to understand the AI buildout than his reaction to an Irish stout.
"At some point the GPUs will be fully depreciated, but we're probably still going to be using them for a while, because the value of even an H100 at this point is higher than it was last year... compute continues being constrained."
In other words, the CEO of a company running one of the largest fleets of Nvidia chips in the world said on camera that a used H100 has gone up in value over the past year, and he said it almost as an aside while making a separate point about why Meta owns its own data centers.
Land
The common assumption among investors is that AI chips behave like a perishable good, where a company buys the hardware, the hardware loses most of its value within three or four years, and the whole bet depends on revenue showing up before the chips are worthless. Much of the bubble commentary I read about the hyperscalers rests on some version of that assumption, whether or not the writer says so directly.
Zuckerberg described the hardware differently. He explained that owning data centers is "a financial advantage for us" compared with "a bunch of the other companies that are paying clouds," and he put the cost of capacity at roughly "15 billion a gigawatt." When he described the downside, he said that if AI works overall, "plenty of other people want all our compute," which sounds a lot more like a landlord describing a building than a company worried about a depreciating machine.
I think history supports him, at least partly. Canals, railroads, the electric grid and long-haul fiber all looked like manias while they were being built, and many of the investors who financed them lost money, though the assets themselves lasted for decades and kept producing for whoever owned them after the shakeout. The mistake investors kept making was treating the fate of the first owner and the life of the asset as the same question.
The Math
Zuckerberg's own figure is a useful place to start. If a company writes off $15 billion per gigawatt over 5.5 years, the annual depreciation is roughly $2.7 billion, while the same investment written off over four years comes to about $3.75 billion, so a single assumption about how long the equipment lasts moves the yearly cost by about a billion dollars per gigawatt. Since his figure also includes buildings and power equipment that last much longer than chips, the sensitivity on the servers themselves is even larger than that.
Meta has already made this choice in its own accounting. In January 2025 the company extended the estimated useful life of most of its servers and network equipment to 5.5 years, and according to its 10-K that change alone cut 2025 depreciation by $2.92 billion and added $2.59 billion to net income, or about $1.00 per diluted share.
A company that rents from a cloud provider is in a very different position, since it pays for the depreciation, the provider's margin and the power every year it rents, and that bill does not shrink once the underlying chip is paid off. The owner, once the chip is written off, is mostly paying to keep it running, and if compute stays scarce the owner can also choose to keep using the chip, move it over to inference work or rent the capacity to someone else, options the renter simply does not have.
The Dissent
Amazon took the opposite view, and I do not think that should be brushed aside. Effective January 1, 2025, Amazon shortened the useful life of a portion of its servers and networking equipment from six years to five, citing the pace of development in artificial intelligence and machine learning, which means two of the most experienced operators of computing hardware in the world looked at similar equipment in the same month and came to opposite conclusions. I find that disagreement more informative than most analyst models on the subject.
Amazon's position has real merit, because Zuckerberg's argument depends on compute staying scarce, and scarcity rarely lasts. Each new generation of Nvidia chips does more work per watt, and in a world where power is the real constraint, a data center operator may pull an older chip that still works perfectly well simply to make room for a more efficient one. An H100 is valuable in a shortage and far less so in a glut, and it is worth remembering that the railroads lasted for over a century while plenty of railroad investors were still wiped out.
My view is that the industry has not yet agreed on what kind of asset these chips are, and I think the market has largely adopted the cautious view as though the question were already settled. The answer will come from utilization rates, prices for used chips and the useful-life disclosures in company filings over the next few years, and of those three I would pay the most attention to used chip prices, since that is the one number management cannot set.
Downstream
The part of this that matters most to me sits further down the chain. If older chips have a long second life, they will eventually move out of the largest companies' fleets and into cheaper inference work and into businesses that would never buy a new frontier GPU, which means a regional distributor or a specialty manufacturer could put AI to work for a fraction of today's cost. That is the adoption gap I spend most of my time studying, and a durable H100 would help close it sooner.
Zuckerberg also said Meta built inference capacity "ahead of knowing that the product was going to be ready," which is the kind of bet a company only makes if it believes the hardware will outlast the original plan. He may well be right, or supply may catch up faster than he expects, but either way he told the audience exactly how he thinks about the asset, and he did it between sips of Guinness.