Neoclouds

I’m going to leave the election predictions to others - there’s been far too many surprises the last few years for anyone to have confidence!

But following up on my last note, it’s clear that neoclouds are going to be a defining feature of this cycle in both directions.

As background on the most prominent, Coreweave is the largest and was founded in 2017 as a crypto miner and pivoted to supplying Nvidia GPU capacity for AI-related workloads. They have thirteen datacenters in the US and two in the UK, financed with US$12.7 billion of debt and equity over the last eighteen months alone. Notably US$7.5 billion in debt from traditional investors like Blackstone, and an equity cheque from their main supplier, Nvidia. The equity was recently priced at US$19 billion.

GPU rental companies are now positioned as suppliers of excess demand, reminiscent of equipment and labour rental companies in the mining boom from 2009 to 2012/2013.

After a wave of Chinese and global stimulus after the 2009 financial crisis, major mining companies were doing well. But the juniors fared even better, as undeveloped marginal assets flipped into profitability and a new set of multimillionaires were minted in Perth.

Capital flowed freely and well-funded NewCos drove intense demand for people and equipment. ASX-listed companies like Emeco and Boart Longyear made record profits on the back of high leasing rates, and executives at BHP and RIO found they could earn more as independent consultants than employees and left to set up firms.

Investment banks produced charts like those below, arguing (for example) that the cost curve implied iron ore prices were unlikely to fall below set levels, as this would cut out supply from all but the lowest cost assets.

Of course, the market turned hard, and commodity prices crashed through these levels like a knife through butter.

And as it did, the cost curve itself pushed down as the price for equipment, labour and materials collapsed.

Consulting contracts were cancelled and employees who stayed in critical roles faired better than those that left.

And companies that supplied excess capacity in the demand squeeze found themselves in a market where there was no need at all for their equipment or services.

This was an archetypal value trap which I remember this well as I was on the losing side of it, though luckily Apple and Tesla at the time were having some of their best years.

Neoclouds are playing a similar role today, offering excess capacity to companies as large as Microsoft.

Why would Microsoft use Coreweave? They can build datacenters as fast as anyone.

The answer is it makes sense to use these companies short term as they have the capacity today, and b) they will be left with all the long term risk.

When the contracts roll off, Microsoft will be in a position to buy the latest Nvidia GPUs - now upgraded annually - and the neoclouds will be left holding soon-to-be obsolete GPUs.

In other words, in return for satisfying excess demand, neoclouds are taking all the residual risk on the GPUs, and they are depreciating exceptionally fast.

Earlier this year SuperMicro was the top market performer among companies of its size in the United States.

The auditors resigned and gave an amusing response to SMCI’s announcement.

SMCI has accused of all kinds of misbehaviour from channel stuffing to using related parties to fake revenues and goose the share price. Ernst&Young’s resignation suggests truth to the accusations.

SMCI was at one point the third largest buyer of Nvidia’s GPUs. And Coreweave was SMCI’s largest customer. Coreweave is also the beneficiary of both equity investment by Nvidia, and apparently vendor financing. So Nvidia is contributing the equity and the debt to buy their own chips.

Which is why this corner of the market is so interesting.

Semiconductors have been in a funk since their highs five months ago, down 18%.

Under the surface, most have sold off more, with only a handful (notably Nvidia, Broadcom and Taiwan Semiconductor) keeping the sector from more significant falls.

Semiconductors are now only up 15% year-to-date, which is quite an achievement given the capex in the sector. Since the start of the year, FY25 capex estimates has actually increased by $94 billion, from $194b to $288b:

Big tech is certainly not slowing down.

In the commodity bust over ten years ago the providers of excess supply all went to zero. There was some residual value, but for most of these companies the enterprise value cut in below the value of the debt.

This all emphasizes how dominant Nvidia’s position is today. Sure, the current generation of chips may soon be obsolete and Coreweave’s asset base rapidly depreciated - but only because everyone is buying Nvidia’s latest supply. This is a uniquely strong position. And is vastly superior to the market position of commodified neoclouds.

There are only a handful of people making decisions on the bulk of this capex spend - the leaders of less than ten firms in the United States.

And when Zuckerberg or Sergey Brin or another at that level talks (everyone seems to be doing a lot of podcasts these days) it’s they are determined to use every resource at their disposal to drive progress forward.

And what’s driving these entrepreneurs and CEOs?

Ultimately it’s the research teams at their own companies, which are still showing that incremental compute is improving the performance of models.

As long as this continues, the investment cycle will go on.

So while the outlines and key players of the next downturn are becoming clearer, the wave of capex is likely to continue, which suggests the weakness in semis is an opportunity.

Mike