It's heating up out there

Z.ai released GLM-5.2, and already it looks more consequential than the latest upgrades from Anthropic and OpenAI.

Cheap/free models are commoditising intelligence, which a) threatens major lab economics and capex, but b) supercharges the memory/datacenter cycle, as cheaper inference means more usage, and more memory demand. This kind of cyclical setup, with clear long-term growth trends, but equally clear risks, is exactly what we built our strategy for.

I can’t speak to whether the latest free models can hack the CIA, but I can assure you that for our use cases, which seem typical for most businesses (quant research, fund operations, general software development for clients at Ace), the free models have caught up.

And this is now the base level of intelligence available to all. And already groups have taken the weights and post-trained them into better models.

There’s a race to see who can serve this model best on US infrastructure. Baseten.co seems to be winning with the fastest inference and a decent blog on what it takes to run a model like this properly.

It’s an important read for those investing in the hundreds of neoclouds popping up around the world. This is what it’s going to take to differentiate.

Brian Armstrong, the CEO of Coinbase, shared this chart showing how Coinbase cut their overall spend, even while increasing usage.

Optimizing spend is the early-mover/AI-forward play right now, as most companies are still figuring out how to manage OpenAI/Anthropic.

But by the end of this year, I’m sure every company with a large AI bill will be undergoing cost optimization.

Which is reminiscent of how the most threatened SAAS companies are those with large customers spending tens of millions of dollars annually. Pre-AI, this kind of large enterprise customer was a sign of strength, and proof of the value of the software. Post-AI, these large bills are just an obvious target and cost opportunity for management.

H100 prices are falling. SemiAnalysis - the surprisingly large research house - posted this chart, with the commentary that falls are not reflective of long-term contract prices, where pricing is still tight. But we shall see… Australians are no strangers to cyclical commodities.

Memory

The biggest story in markets is the exceptional result from Micron and the massive increase in memory prices.

Micron’s datacenter revenue was $25 billion, annualizing to $100 billion, which is a serious chunk of global capex (they are one of three memory providers, plus a state funded Chinese player trying to break in).

A good way to understand the winners and losers in the space is to divide them into givers and receivers. The givers (hyperscalers) are diverting immense free cash flow and raising/borrowing more to fund datacenter builds. Memory providers are on the receiving end of this gush of cash, as are GPUs and all kinds of industrials and energy providers which have an unprecedented boost to demand.

But prices may finally be hitting something of a limit, at least reaching the point of marginal demand destruction.

Apple announced a series of price increases across their product range to protect margin from memory price hikes. But there’s little sympathy from Micron et al, who only a few years ago had pricing ground to the dust in aggressive Apple negotiations, even while Apple charged exorbitant amounts for necessary memory upgrades to mobiles and PCs.

Only a few weeks ago I was about to buy a Mac with 512GB RAM, which can run most SOTA open-weight models locally. But several weeks later, the most RAM I could get on a fully upgraded MacBook Pro was 128GB. The 512GB options were unavailable at any price. Checking in on the apple store now, the most memory available on a Mac Studio, again at any price, is 96GB. Demand destruction in plain sight.

This is now a political issue.

Apple is appealing to the US Government to lift the ban on Chinese memory, which will no doubt soon be flooding the market regardless.

There is a mad rush to build memory manufacturing in Korea, China, Japan, and the United States. One to watch is Elon Musk’s Terafab, jointly owned by his various entities (in partnership with Intel), including freshly and lavishly funded SpaceX.

So this will be resolved… but until then the demand is great, and the price at which that demand is cleared and prices settle is high.

This shows how fast power can shift in a cyclical industry. The underinvestment in factories is due to low prices negotiated hard by the likes of Apple. Now, higher prices will change the politics, bring well-funded new entrants into the market (SpaceX) and cause existing suppliers to ramp up capacity. It will all hit the market at the same time.

But until then, in a hyper charged demand environment prices are going to have to stay high to destroy demand and balance the market.

This is neutral to the debate around LLM economics… as the free models actually increase demand.

Micron has been a bit of a conundrum for fund managers.

On the one hand, it’s the cheapest stock around for its size, scale and certainly growth rate, trading at a single digit PE (it’s about 7x forward now). But on the other, value investors pride themselves on buying things that are down, and mostly seem to have missed it.

Which is similar to how we saw the industry react to Nvidia when it began its run from the lows of 2023.

Coincidentally Nvidia is trading at an FY28 PE of under 16 again… which is closer than it seems (time is moving way too fast).

These stocks are perfect candidates for our quant strategy, equally dramatic opportunities and cyclical risks.

But first we’ll watch the market price in these new models, the new capacity coming online (including from new entrants like SpaceX and China), and also see whether the growing market share of free LLMs takes the wind out of Anthropic and OpenAI’s capex plans.

Government

While writing this OpenAI launched (to a limited set of customers) their next suite of models, amusingly and presumably coincidentally named Terra, Luna and Sol (Terra/Luna was a spectacular crypto collapse, and Sol has also crashed from its highs. Cool words though to be fair).

This is crystallising a long-feared threat to OpenAI/Anthropic, as the US Government is controlling access to these models, for now at least.

Giving more time to the good guys before State-of-the-Art intelligence is widely available isn’t necessarily a bad thing… but it certainly is for Anthropic. But what does it mean for capex and spending plans if they can’t sell their latest models? It’s clearly important to maximize revenue for the brief periods a particular model is on top.

OpenAI’s new mid level model is apparently as good as their previous flagship 5.5 (which is excellent) and priced at half the cost. So pricing is rapidly coming down, likely due to a mix of open-weight and competitive tension with Anthropic. Regardless of which APIs end up in production, they all need datacenters stacked with memory and GPUSs.

Portfolio

Our strategy is somewhat unusual, designed to outperform in long trends in both directions. Finding securities that display these characteristics is a different challenge to more typical styles of investing. But it does seem well-placed for the current environment.

We had to close a number of our software holdings as the SAAS-resurrection reversed, but it’s interesting to see that cybersecurity and networking/data plays like Rubrik and Snowflake held up well, as well as quite a few companies that aren’t really software but tend to trade with the group, like Reddit, Hims&Hers (new peptide opportunities), and Robinhood. Even some Australian names bucked the trend, while household names like Xero and realestate.com.au languished.

We also worked on a number of new names in healthcare, adding Agios, Omada, BillionToOne, BridgeBio, Caris, and even some on the ASX. It’s about time these companies had a run.

Mike

Please note that all information in this note refers to the wholesale Frazis Fund. Our listed ETF, ASX:ROAR operates a different strategy. Relevant information, including the PDS and TMD is available separately here.