Dear investors and well-wishers
The wholesale Frazis Fund declined -2.4% in June. We finished the 2026 financial year +46% with a rolling 3 year net return of 39% per annum.
Semiconductors
Only a few weeks ago, Anthropic was proudly boasting about the terrifying capabilities of their new model Fable, only for OpenAI, Kimi, Meta and Grok/SpaceX to release models matching or exceeding it in various dimensions.
This either caused or coincided with a semiconductor washout. Fortunately we had already taken profits across companies like Marvell, AMD, and AAOI, but we still had some exposure (mostly to mega cap like Nvidia, Broadcom).
It’s a good demonstration of why we have systematically harvest profits these days…
Part of this was just general market reversal after enormous moves and by at least one measure the most crowded positioning ever (from BofA’s July fund manager survey).
But there are quite a few new developments for markets to digest.
Firstly, LLM Competition and Price Wars
If there was a single large language model company serving intelligence, this company would indeed be worth trillions of dollars, and the global economy would reorganise around it.
Even, perhaps, with two. A cosy duopoly can be profitable - just ask the various Australian duopolies, which seems to be the default structure in this country (Qantas Virgin, Coles/Woolworths etc).
But there is a long list of disruptive technologies where a large number of competitors force down pricing and the industry profit margin collapses.
This is as true of innovative and revolutionary technologies as anything else.
These are the companies with SOTA models:
OpenAI
Anthropic
Google
Meta
Grok/SpaceX
And then you have the open models:
DeepSeek
Kimi
Minimax
Alibaba’s Qwen
And it’s not just Chinese companies. You have Nvidia’s Nemotron, Thinking Machines released Inkling last week, and there are a host of others like Arcee/Trinity.
Without taking away from the historic achievement in summoning forth the first large language models, it turns out they’re not so hard to build now - or at least, are within reach of many major tech companies.
There was no guarantee this would be the case, but that’s how it played out.
Which is a problem for the original duopoly, not to mention the scores of VC funds sitting on enormous unrealised gains, and requires careful thought from listed semiconductor investors, because much of the capex lined up over the next 2-3 years is directly earmarked for OpenAI and Anthropic.
A world where LLM pricing converges to electricity and datacenter costs does not justify the same level of buildout as a world with one or two companies with magical otherwise inaccessible technology generating massive monopoly profits.
And these investments depend directly on their profitability.
Secondly, the market is moving towards custom silicon
Custom chips are here. There are two threats to Nvidia. Firstly, that inference simply moves to purpose-built accelerators that are an order-of-magnitude or two faster, or more efficient, or both.
The world is already moving in that direction, which means the current generation that is under construction may be working with over-optimistic assumptions around pricing and residual value.
Google is reportedly working on a new ‘Frozen’ architecture, which involves printing part of the model directly onto the chip for a 6-10x efficiency, due in 2028. Which again, is an issue for everyone levering up to the hilt to buy this year’s general purpose GPUs.
The usual redirection when these concerns are raised is to point to the high prices commanded by old Nvidia GPUs today.
But that’s not a fair comparison. The advances over the last few years have indeed been substantial, but these improvements have been incremental, which is why older GPUs are still in demand.
If in the next 2-3 years there will be chips in the market that can run inference significantly faster and cheaper, workloads will move to them.
Every new announcement shows the threat to GPUs from custom ASICs is real.
But there’s another threat too: the optimal strategy for hyperscalers here is almost certainly to vertically integrate and run inference on their own chips - and where possible their own models.
They can optimize chips for their own stack, and better, not pay away a 75% gross margin to Nvidia or anyone else.
Google has TPUs, Amazon has Trainium and Inferentia, Microsoft has Maia, all in advanced generations now. Even OpenAi has Jalapeño.
Vertical integration is the logical step, and something we’ll see in the next few years. It’s mission critical for these companies to achieve this.
Thirdly, market positioning simply needed to reset
Semi exposure was way out of whack and had to be cured. Double and triple leveraged ETFs exploded (and have subsequently seen dramatic falls), and semiconductors became the ‘most crowded’ fund manager position ever.
The Koreans have provided a spectacle with their wholehearted adoption of leveraged ETFs. Some of the older versions have performed spectacularly over the last fifteen years or so, but many have been completely wiped.
The new latest craze, leveraged ETFs on single stocks, at typical stock variance, are all but guaranteed to fail.
You can see many examples of indices that (unlike US tech) didn’t grind up for 15+ years, and the returns have been catastrophic, as they bleed out from rebalancing (securities that trend do well, as the compounding effect can outweight the bleed, but that’s a topic for another day!
Fourthly, China is playing a clever game here
This is in line with the first point on competition, but is worth elaborating.
It seems the order for Chinese labs to open-weight their models came right from the top, judging from Chinese President Xi’s recent speech.
Releasing free open weight models massively weakens Anthropic and OpenAI, and through them the entire US ecosystem. That’s a win for the Chinese.
Without the pressure from these free open-weight models, US companies would be able to charge more and invest more.
The price war has already begun, and price wars are a Chinese specialty.
As we wrote at the time, the US Government restrictions on Nvidia have clearly been a massive own goal. They fostered China’s domestic industry, handing market share to new domestic entrants, and for all those lost revenues, profits and financial power ceded by the United States, Chinese models are currently state-of-the-art anyway. What was achieved?
Fifthly, there are clear signs of excess capacity and competition at the hyperscaler/neocloud layer
It’s old news now, but Elon Musk is renting out SpaceX/Grok capacity, and Meta is planning on becoming a hyperscaler. There are bullish takes on this, with some suggesting this is proof of the economics of neoclouds. But celebrating such well financed, highly technical new competitors is a stretch of reason.
The question around neoclouds comes back to competition. If AWS was the only cloud compute provider, they would certainly earn bumper profits. Similarly, they’ve done almost as well in a cosy oligopoly with Google and Microsoft.
But add in Meta, SpaceX, Nvidia…. And then maybe 100+ neoclouds, then you have a highly competitive, aggressive market, and industry profits could be wiped out, even as revenues dramatically exceed expectations. And this is a repeating market theme in technological innovations.
Similarly, if Nvidia was backing one neocloud, that would be interesting and a clear point of competitive differentiation.
But they are backing many, so there’s now a wide choice of which Nvidia-backed neocloud to use. They’re all offering the same thing, so they will have to compete on price.
And moving on from the neoclouds, lest we forget Econ 101, today’s higher prices are fostering new entrants across the supply chain.
As an example, the Memory triopoly is about to be challenged by CXMT, which is about to list in Shanghai.
Finally
Answer this question, how much has your life really changed in the last few years?
These models have been around for a while now, there’s been immense intelligence at your fingertips. How different is your life, really?
I can say from our side we’ve built some cool software, maybe hired one less person in ops, saved a little on software subscriptions. There are probably fewer spelling errors in our emails and investor notes. But mostly our lives are the same.
It will be interesting to see how our answer to this question changes over time. What’s yours?
Outlook
In summary,
Model competition is eroding scarcity and pricing power, putting a question mark over long term LLM profit margins,
Custom silicon will challenge GPU economics,
The rapidly growing number of compute providers will have to compete fiercely with each other for customers
China benefits from and is encouraging commoditisation
There are huge sums wagered in public and private markets on the end-state profit pool of these companies.
Adding in the waxing and waning of extreme investor enthusiasm (and not so long ago, despair), all supports the idea there will be multiple booms and busts over the next five years… which makes semis a great candidate for a trend-following strategy.
I certainly wouldn’t want to be a discretionary trader trying to catch every twist and turn, AI-assisted or not.
It’s very possible for a technology to change the world without every layer of the stack earning bumper profits.
For now, the market just needs to digest it all.
Mike