Neoclouds and an article

Hello all,

The Australian Financial Review wrote about us earlier this week:

Meta made an interesting announcement two days ago, announcing their entry as a supplier of compute, joining the ranks of AWS, Google Cloud, and Microsoft’s Azure.

This caused a sharp sell-off in neoclouds, many of whom were hoping to count Meta as a customer rather than competitor.

We now have Meta and SpaceX, two extremely well-funded companies, entering the space. Even Masayoshi Son is building a neocloud.

Only a week ago I was debating this with an investor, saying if you were Zuckerberg, are you going to call the 100^th marginal supplier of compute in a random country, or are you going to call Elon Musk? Who would you prefer supplying your compute?

But even that framing may already be dated, as soon Zuck won’t be calling anyone, other than to sell his own excess capacity.

The steady stream of Nvidia ‘partnership’ announcements with neoclouds deserves scepticism as they’re more like supply and financing arrangements. They’re not partnerships, they are buying Nvidia chips! This is like saying I’m in partnership with my local coffee shop. The fact Nvidia is investing a small amount in equity is beside the point. Like getting your 20^th coffee for free.

Meanwhile there are almost daily announcements of new efficiencies in memory, processing, and inference. Agentic AI is still new, we have just crossed the six month mark since the release of Claude Code. There is immense potential and demand for cost opimizations from here, and the entire stack is being attacked: models, chips, memory, networking, software, and deployment.

I’ve written about to the threat to the current generation of GPUs from Cerebras and Groq, who are getting >1,000 tokens a second. And they too are threatened by companies etching entire models onto chips which claim 500,000 tokens/second on Llama 70B (this is a recent supplier claim from Etched). These may form a small part of the market now, but the direction is clear.

These could make the current generation obsolete, which matters because this is the generation being locked into current financing rounds of many neoclouds. It also explains the Nvidia incentives and ‘partnerships’, and the willingness of hyperscalers to rent the current generation and hand back the chips when the time is right.

This is entirely consistent with the current market pricing of both listed neocloud equity valuations and GPU rental rates. That’s why cyclical industries are cyclical.

If you need memory today you’re going to have to pay up (if you can even find it), regardless of the demand balance in two years time, when massive Korean, Japanese, Chinese, and certainly US supply comes online.

The same logic applies to compute, as Meta and other hyperscalers might happily rent capacity in the near term as they continue their own buildout and advance their own ultra efficient chips, where they won’t have to pay a substantial markup to an AMD or Nvidia.

The hyperscalers, who are among the largest customers of neoclouds, are vertically integrating, and this is likely the future of the industry. Over time, they will capture as much of the GPU margin as possible, and optimize their stack for their own needs. Not in the next 1-2 years, but quite possibly by the time 5 year GPU rental contracts come up.

Google’s Ironwood TPU is explicitly designed for the inference era. AWS Trainium3 is a 3nm in-house AI chip for agentic, reasoning and video workloads, with AWS claiming up to 4.4x compute, 4x energy efficiency and nearly 4x memory bandwidth versus Trainium2. Meta says it is developing and deploying four new generations of MTIA chips within two years. Microsoft has introduced Maia 200 as an inference accelerator with FP8/FP4 tensor cores, 216GB HBM3E and 7TB/s memory bandwidth.

Which is why we personally would be very cautious undewriting long-term demand from hyperscalers for generic third-party GPU rental.

This is also why, when analyzing neoclouds, it’s so important to scrutinize the actual deal structures, lending terms, utilisation assumptions and exit clauses in these contracts. These are extraordinarily capital intensive businesses, and the optimal play for hyperscalers is to rent in the near term, option up capacity just in case, but optimize their own economics as soon as possible.

Microsoft, Amazon, Meta, Google, SpaceX, etc all have a lower cost of capital and better access to the leading talent in the industry than neocloud entrants., and the strongest reasons to own more of the stack themselves.

Of course, at times it makes sense to shift the financing burden off-balance sheet to others, but that’s consistent with this view, and another reason to be sceptical - there’s certainly a scenario where neoclouds forced to operate at an all-in loss to secure customers, which is one reason hyperscalers would continue to use them… if they’re selling $1 for 90c.

We are as bullish as anyone on AI, and it has been a major driver of our recent returns. But there’s a real risk in this part of the market that customers vertically integrate and capture more of the margin themselves, while leaving the neoclouds with the massive interest, operational, and capital costs in the meantime, as well as residual risk.

There are many other great ways to play the AI theme. These chips will all require manufacturing, packaging, memory, networking and power infrastructure regardless of whether they are off-the-shelf or custom ASICs. There’s a parallel in the various crypto booms and busts, when all compute moved to custom chips and it no longer became economical to (for example) mine bitcoin on an unoptimized computer. Ironically, many of the neocloud operators are the same people.

There are ways to invest in the AI infrastructure buildout without taking a bet on the residual value of today’s GPUs in 5 years time.

Good luck out there

Mike