The claims seem to be true.
> A typical one-gigawatt AI data center requires $38 billion in up-front capital expenditure (CapEx) and $0.9 billion in annual operating expenses (OpEx). If CapEx is annualized over each asset’s lifespan, the total cost of ownership equates to $8.5 billion per year. Servers dominate this cost at $5 billion per year, or 60% of the total. Operating costs are small by comparison: even energy, the largest OpEx category, costs only $0.6 billion per year.
If we steel-they the point in the essay, it still doesn't matter because a bubble is setup in the economics of the system, not whether the underlying tech works and/or is profitable. It can both work, be profitable, but have been overcapitalized before the knock on effects are in-place in the rest of the economy. This is a crazy analogy, but what if 1700s earth figured out, theoretically how to build a fusion reactor that could supply all the energy needs of the whole planet. But it cost in their equivalent dollars, 1 quadrillion dollars? You can't even integrate under their money curve to scrape together that many dollars. The only economy you can make about that is in the research side to push the cost down. Then when its dialed in, you can launch that rocket AND the rest of the system has had time at many different levels to integrate AI at various levels. OpenAI dry barebacked its way into the economy and literally screwed everything up. If they really were an altruistic AI safety lab, they should have kept working on ChatGPT, but demoed it to the world including government, so we could as a planet decide where we are taking this.
It is a huge bubble, AI is over capitalized, we have no idea what the knock on effects are and really the only benefit right now is in programming, which is huge and civilization changing. But it also has massive downsides in that it has given the most powerful even longer levers to move the earth.
they seem to think that debt markets are infinite. "There's Still Plenty Of Capital To Raise"
> The ones that have issued public debt, like Alphabet and Meta, have excellent bond ratings. If they need more capital, they won't struggle to find people to give them a loan.
Yes, and that bond rating is based on the idea that they really don't have that much debt, and have a net profit of something like 30%. If they hoover up a load of debt, the rating will plunge.
> I don't think the rhetoric about the AI bubble is really about finances
Its exactly about finances. The market participants that are not bollocks deep in fake Private Equity valuations are worried about circular debt obligations that are off books. The SEC has been rather lenient about reporting of non-traditional obligations (hence why spaceX was allow to do what ever the fuck it did)
For example anthropic has about 450 billion in non-cancellable obligations. the S-1 puts it at 895 billion. all of that is someone's debt.
If openAI are doing the same thing, thats pretty close to 1.6trillion in debt, both for the same customers, at the time when customers are going "hmm this isn't giving us what we hoped for the price"
It could be that a slowing of data centre build (and hence a slowing in the finance needed) takes the heat out of the situation avoiding a dramatic burst. Even then there will be casualties. If someone knows who and when ... they could do very well. More likely, Microsoft and Google will buy up anything distressed that's worth having.
If you have 20 years to pay off 5 trillion dollars, it doesn't come out to a neat 250 billion per year as that suggests there's 0% interest and no servicing cost. Servicing costs would probably be at least 3 times the stated figure, corresponding to 10% interest paid. Some of the debt may obviously carry less interest, but the financing is opaque (as the author admits to), we have seen interests rates creep across the economy and big tech has had hits to their credit ratings.
I hope the Liberal Party can find a better candidate than the author of this piece ...
The DC push may be the thing that finally gets renewables over the finish line in the US. For example, TX is around 35% renewables right now. If that continues to rise, we may see power fall in price. It’s all about incentives.
Another issue is that while e.g. Anthropic and OpenAI seems to be interested in building their own datacenters, they are mostly just renting capacity. If they can shop around, then they can get the price down, but there's a base price, due to power, hardware and staff.
The author also suggest that there's still capital to raise... no there's not. Anthropic is going for an IPO, that means institutional investors, not venture capital or borrowing and Anthropic, CoreWeave and Oracle are rated as junk. That doesn't mean that they can't raise capital exactly, but it will be very expensive and the interest payments needs to be factored into the price of the products.
The current-gen hardware is expected to have only 2-3 years of useful life. New tech, and especially new architecture is going to make inference much faster and more efficient in the near future, and the big vendors are going to be more limited by datacenter floor space and power/cooling, so that the hardware will be replaced very quickly by better models.
Meanwhile, Meta is selling their spare AI data centre capacity.
Nobody is getting fired for thinking they need to buy scarce H100s in case they aren't able to get them in the future. Nobody is even getting fired for keeping their companies out of Jensen Jail — the fear that if they don't buy the whole server stack from Nvidia they might not get future GPU allocations.
There really are analysts who think Nvidia is channel-stuffing.
If the market deflates there will be a glut of very cheap compute, both in terms of data centres (though many of them will be able to offer managed model training services instead, I suspect) and in terms of unused hardware being made available for sale.
Nvidia won't lose out; everyone else is going to.
Supply and demand bring what it is means that (alleged) cheap compute won’t stay cheap or go unused for very long.
The problem is that "scarcity of GPUs and RAM" is being used as a proxy for "demand for AI".
It's a broken proxy; the scarcity of H100s is artificial.
Are the contracts on the income statements of the sellers? If Anthropic promised to buy $200B of GPUs from NVIDIA, is it recorded on NVIDIA's income statement (other than a footnote "we have $200B orders lined up")?
'Cause if it is, that'd be $200B GDP increase _literally_ conjured out of thin air.
If they haven't been manufactured, things are potentially quite sketchy.
Expected by whom and why?
Google hasn't retired any of its now ten-year-old TPUs, or any of the subsequent generations. Maybe TPUs are special or Google doesn't understand compute economics but either argument requires data.
If your argument is depreciation, you're confusing a tax concept (depreciation) with useful life. Yes, they're supposedly related, but in practice, the relationship is rather loose. (Residential housing depreciates in 28 years yet typically lasts 2-3x as long and 4x is not rare.) Google's V1 TPUs
also, power and cooling agreements are usually 10y too.