January 2024 Update

Dear investors and well-wishers,

The fund advanced 13% in December and closed the year +35%. We are having a positive January so far, though with a host of tech companies reporting next week the end of the month might be choppy.

It was a mixed year. On the one hand semiconductors performed well. On the other, biotech and small caps lagged. We landed somewhere in the middle. US small caps ended up 11%, cloud computing ended up 39% and biotech ended up ~3%. All these lagged big tech, with Nasdaq-100 up 54%.

As I write, after a strong few weeks we have closed our gap to the Nasdaq-100 from the beginning of last year to <2%. Still, a long way to go.

2023 will be remembered for three revolutions. This was the breakout year for large language models, the first time GLP-1 drugs became part of the popular lexicon and entered mass use, and at the end of the year we had the first CRISPR approval.

Where will these go in 2024?

Large Language Models

We can now speak to computers, and have them speak back. The impacts of this are unclear, but it’s a clear step-change in our ability as a species. We are all vastly more powerful, and not just in analytic domains: these models are particularly good at writing poetry, making artwork, and writing codes, areas which were presumed to be the last bastions of human techne.

The increase in computing has happened. Semiconductors were already a good bet, though are still clearly cyclical.

We wrote this was a step-change in demand which would land in leading chipmakers, which have been suffering - and are still suffering - from an extended consumer slowdown in PCs and Mobile.

This put NVIDIA, which for most of the year was the only game in town, into the leading seat.

It really is fascinating how Graphics Processing Units developed for gaming turned out to be the perfect chips for a certain type of crypto mining, and shortly after, the ideal chip for training large language models.

The company presciently invested heavily in software to allow developers to access the multiple processes on each chip, and perhaps more importantly, built a community of developers around it. This is not something that can be easily replicated.

AMD is trying. They are still a clear number #2 but have been helped by two factors:

Firstly, the willingness of major buyers to foster an NVIDIA competitor, (notably Microsoft, you’ll have seen AMD computers if you’ve been to a Microsoft store).

And secondly - and this is more of a prediction - a smart decision to design their latest top end chip to outcompete NVIDIA at inference rather than training. AMD’s MI300 is 1.4x better at inference than NVIDIA’s leading H100, according to the company.

This has sparked a decent rally.

In semis, all roads lead to Taiwan Semiconductor. Companies like Apple who designed their own chips still manufacture them in the same place, by Taiwan Semiconductor.

This is cheap for a blue chip company, though NVIDIA’s 2024 PE was <25x only weeks ago.

There are a host of other companies in the supply chain. ASML just reported solid results, flagging an uptick in orders.

Sam Altman is apparently gaining Middle Eastern backing to build his own semiconductor company (all roads lead to…) and this would involve immense Saudi capital expenditure on ASML’s machines and the other duopolies and monopolies that make the ultra-high tech components required to build advanced chips.

AI in 2024

The inference demand shock has barely started.

There are easy straight-line predictions. Now, you can build basic apps with natural language. Soon, coding copilots will enable anyone to build and deploy complex apps at scale.

Voice-to-imaging will extend to voice-to-video. Google released this only yesterday, and managed to give it an appropriate name too ‘Lumiere’.

Much better than ‘Bard’ which sounds like a goat.

Finally there is still low hanging fruit in research. I’m struck by how similar our own lines of thought are to language models. The next word we think or say just kind of appears in front of us, in a consistent, single-threaded train of thought.

If you want to know what it would feel like to be a language model, you might just have to look within.

The ability to talk to computers, and have them talk back, has opened up robotics entirely. Before large language models, robots needed to be precisely and agonizingly programmed. Now, with computers understanding natural language and accurately interpreting live video feeds, we may soon see the futuristic robots science fiction promised us. This is all going to require a lot of compute.

There are some dark clouds. The US has restricted Chinese access to Western technology. This is an agonizingly short-sighted move, as it guarantees within several years there will be state-backed Chinese competitors for all the market-leading Western companies.

Usually it’s the other way around! When China wanted to foster domestic internet champions the first thing they did was block Western companies that would have quickly controlled the market, as they did in the EU and the rest of the world.

Now the US is doing it for them, for basically no gain, other than a minor edge in the short term. Even that is arguable: drones and aircraft carrier-destroying missiles are going to be sufficiently capable regardless. And ASML, NVIDIA and the like are missing valuable revenues that could have been spent maintaining their lead.

The impact is years away, but I’m sure by 2030 it will be clear that the West would have done better to maximize capital formation in their own companies, rather than handing those revenues to State-backed Chinese competitors in such a strategic area.

There are also going to be considerable gains in efficiency. Every day it seems someone has made a large language model work on a smaller device, with less and less compute.

My personal bet is that both things occur: a massive demand shock, and vastly improved efficiencies. Either way, it’s clear where capital is going to flow and which companies are going to lead this current regime.

Last night ASML reported 9 billion euros of new orders, more than triple last quarter.

GLP-1s

In 2024, I expect these to morph from weight loss drugs, to health improvement drugs.

So far, stomach symptoms aside, which is the whole point, larger and larger studies have failed to show negative health consequences. In fact, the opposite: better cardiovascular and kidney outcomes.

As an analyst from Morgan Stanley commented on Novo’s 2 year trial of semaglutide:

This year will see major trial read-outs across a range of conditions, notably fatty liver disease and sleep apnoea.

The next generation will also come of age, including oral formulations (oral Ozempic already exists, but is not widely available given higher doses are required to pass the digestive system, and there’s a shortage), and we will see whether Eli Lilly’s retatrutide achieves the 24% weight loss seen in earlier trials (vs Ozempic at 15% and Mounjaro at 15-21%).

We’ll see long-term safety data from Mounjaro as well, and also see the results of early trials testing GLP-1s in neurodegenerative diseases. We will know a lot more this time next year.

It’s also worth remembering these franchises have an end date. They will become generic in 2031. That’s several lifetimes in the stock market, but it’s an important factor. So far Eli Lilly seems to have a significant lead in the next generation, though Ozempic, in Australia anyway, is the household name.

Gene therapy

CRISPR and Vertex’s gene therapies for sickle cell disease and beta-thalassemia, approved late last year, will be rolled out in 2024. I expect payers to pay, but the financials to look ugly for some time. Huge investments will be required to get these treatments to patients. After an initial pop, the market is giving CRISPR little love.

These are still gnarly treatments, requiring chemotherapy and months of hospital visits, perhaps marginally better than bone marrow transplants, only with fewer immune issues (no foreign transplant) and no need to find a matched donor.

Apparently this will be worth it to avoid horrible ‘pain crises’.

One prediction for 2024 is that focus moves to next-generation treatments that don’t require chemotherapy, and where the gene editing is done inside the body (in vivo, as opposed to ex vivo).

This sector has not seen a recovery I expected.

One company which has seen precious little love is Intellia, which is sitting on what looks like two one-time cures, one for ATTR-amyloidosis, and the other for hereditary angioedema (HAE). I understand the heavy discount for ATTR-amyloidosis. This is a competitive market, with RNAi silencers like Onpattro from Alnylam, and effective TTR stabilizers. But the orphan treatment for HAE should be very well received, and would be the first in vivo treatment of any kind. Trials of both treatments should read out this year.

Probably because they had such an eruptive peak, these ‘platform’ companies have been withering for three years now. Meanwhile, their technologies are three years ahead, and we are seeing approvals and late stage data.

Key to a near-term recovery will be the progress Vertex and Crispr make in commercializing their new products. Strong take-up will justify significantly higher values, at the moment it looks like these are going to trade down to cash.

In short

The revolutions of 2023 will continue to play out in 2024, though the demand shock may weigh more to inference than training in AI, GLP-1s will be seen more and more as health-improvement drugs rather than simply weight loss, and CRISPR/gene therapy sector will face a long-awaited test: will these long-awaited cures make money?

Best wishes

Michael