Glenn

Because of this, the unit economics of such investment demand a much faster return on capital. The weighted average cost of capital will be much higher — you can’t use a ton of LT debt to finance “compute”. It will have to be mostly equity.

So the “compute” business model will have to find and scale massive use cases (hundreds of billions of $ a year) to justify the massive datacenter capex spend. And we have already seen how open-source models like Deepseek can wipe away the moats of existing models in a virtual instant.

China will figure out EUV lithography and advanced chip fabrication at some point. Let’s ponder what happens if it is commercialized in the next three years. We know what often happens when Chinese firms enter a new tech/industrial space. Huge economies of scale and cost deflation.

Now you start seeing competitive AI chips from companies like Huawei at a third the cost. This only further depreciates the remaining value of the hundreds of billions invested in compute. These are further asset writedowns against what’s already a rapid depreciation rate.

IOW faster-than-expected obsolescence might mean the original 3-5 year useful life assumption was too optimistic. This just means you have to find these huge monetizable use cases even faster.

And maybe that’s the bet of companies like Microsoft and Oracle. Of course reducing the cost of compute would be great for the cloud software companies whose business model is about finding those massive use cases. So it makes sense that they are so interested in being a part of this.

Finding hundreds of billions in annual revenue in such a short period of time seems like a daunting task. Microsoft’s entire productivity and business processes segment (e.g. Office) is ~$77 billion per year.

People think China’s HSR investment was a massive money pit. But the difference in the nature of the asset (short vs. long-lived) can help put this monetization problem in perspective.

China Railways carries total PP&E on its balance sheet of ~$1.1T. Most (but not all of it) is HSR. The annual depreciation charge against it is about $28B per year. Depreciation is ~16% of China Railway's revenue.

I don't know how the $500B would split between compute and physical datacenter costs, but let's just assume 50/50. $250B of "compute" depreciating over a 5-year schedule equates to $50B a year of of depreciation that needs to be recovered.

Moreover, rail assets are a de facto monopoly. There is minimal risk of technology or competitive disruption here. Compute is a pure commodity. Massive risk of obsolescence and technology-based disruption. The risk profile here is fundamentally different.

Tech has always been about capital efficiency and building moats through sticky use cases. It's no surprise that Elon is expressing some major skepticism here! https://x.com/elonmusk/status/...

My skepticism is actually less about whether they could raise the money. Softbank's involvement points to a variant of Softbank Vision Fund, which has raised an aggregate $134B across two funds, a significant amount from Middle East investors.

The reference to "four years" also suggests alignment with Trump's second-term administration and expectation of significant government (i.e. taxpayer) support.

I don't really care much about these companies raising hundreds of billions from accredited or foreign investors to potentially light on fire. I do care about taxpayer funds though.

What is clear (and I think what Elon is referring to) is that the companies in question are not coming up with the $500B. They will provide a tiny sliver of equity (1-2%?) and the rest will be "other people's money" (OPM).

If there is any government support involved at all, what we really need to think deeply about as a society are the alignment of interests. How is risk and reward allocated between the "general partners" of this Fund, the limited partners and any taxpayer-funded support?

$500B is not quite $7 trillion but it is still a tremendous amount of capital, which will have significant impacts on this country's allocation of economic resources. Do we have the institutional and state capacity to ensure/monitor that decisions on allocating massive quantities of societal resources are aligned with the public interest?

This is ultimately what that comes down to. When you are handing OPM (including direct or indirect taxpayer support) to a small group of society to manage, incentives, alignment of interests and risk allocation are absolutely crucial.

"Heads I win massively, tails I don't lose much" is one of the least-understood aspects of these structures. It's one thing for such a risk allocation strategy to government small sums of private capital. Its wholly different when we are talking about hundreds of billions.

This is no longer a private capital question. At these levels of scale, it's a societal question that falls under the purview of the state and general voting public.

And what happens when the voting public loses faith that the government: (i) not only lacks the institutional capacity to properly manage such a widescale allocation of resources, (ii) but where the interests of individuals within government have misaligned personal incentives?

Anyway, I am like you just trying to read the tea leaves here. We don't really know what the details are. But hopefully this highlights some parameters that we should carefully consider as more details are revealed in the coming days/weeks/months.

Maybe we should consider a better process for settling major hundred-billion dollar societal resource allocation questions than fighting it out on social media. This does not lend confidence to "state capacity". Surely there is a better way. https://x.com/RushDoshi/status...

After years of hearing about the lack of capital efficiency and massive capital misallocation in China, the juxtaposition of these headlines is shocking.

All right this is helpful in clarifying some details on what is part of the $500B. Re-allocation of existing datacenter capex budget, per Satya.