|
Green Ash Horizon Fund Monthly Factsheet - May 2026
|
|
The Horizon Fund’s USD IA shareclass rose +18.13% in May (GBP IA +18.16% and AUD IA +18.10%), versus +4.55% for the MSCI World (M1WO).
- May was a solid month for equities generally, supported by positive noises about a US/Iran deal in the latter half. Geopolitics aside, the AI theme remains the dominant force in the markets, and is still being led by semiconductors: the SOX Index gained another +22.14% in May, while the Mag7 only rose +6.64%
- In addition to our 8% in cash, we have added costless collars to a further 20% of the portfolio. We still see considerable scope for upside in the equity book across all five themes, but a pullback would be healthy as we navigate the thinner liquidity of the summer months, and we are well-positioned to add back risk should one occur
Please click below for monthly factsheet and commentary:
|
|
|
|
Source: Bloomberg; Green Ash Partners. The Green Ash Horizon Strategy track record runs from 30/11/17 to 08/07/21. Fund performance is reported from 09/07/21 launch onwards (USD IA: LU2344660977; performance of other share classes on page 3). Strategy Track record based on managed account held at Interactive Brokers Group Inc. Performance calculated using Broadridge Paladyne Risk Management software. Performance has not been independently audited and is for illustrative purposes only. Past performance is no guarantee of current of future returns and you may consequently get back less than you invested. Benchmark used is M1WO Index
|
|
|
Here are some tidbits on the themes:
|
|
- World Semiconductor Trade Statistics raised their FY26 forecasts for the semiconductor market by +55% to $1.5 trillion, implying growth of +89% YoY, and are forecasting a further +27% growth in FY27 to $1.9 trillion. This is largely driven by memory, which is forecast to grow at a +114% CAGR versus FY25, and logic which is expected to grow at +32%. Everything else is forecast to grow at a 2Yr CAGR of just +10% through FY27.
|
|
|
WSTS raised their FY26 forecast for the semiconductor market by +55% to $1.5 trillion. This was largely driven by memory
|
|
|
|
Source: World Semiconductor Trade Statistics (WSTS); Green Ash Partners
|
|
- When it comes to bottlenecks slowing down the AI infrastructure build out, much of the attention has been on the physical construction side - permitting, grid interconnections, availability of labour - but there are also bottlenecks re-emerging on the silicon side, both in terms of leading edge wafers which are crowding out smartphones and PCs, and memory
|
|
|
AI demand is increasingly crowding out smartphones and PCs from leading edge wafer capacity at TSMC
|
|
|
|
Source: Semianalysis; Green Ash Partners
|
|
- Memory remains the most acute bottleneck on the silicon side of things. By 2027, AI memory demand will represent nearly 70% of the total DRAM market, versus just 12% in 2023. GS recently joined the higher-for-longer club, expecting tightness through 2028, with 2027 being tighter than 2026. Sell-side reports are starting to acknowledge the increasing proportion of future supply being contracted via long-term agreements (LTAs), which should support continued multiple expansion in the peer group
- Micron trades at NTM P/E of 11.4x, compared to AI compute/networking semis like NVIDIA and Broadcom at 21.2x and 26.8x respectively.
|
|
|
AI demand is crowding out other DRAM end markets too, with HBM requiring 3x the wafer capacity of traditional DRAM
|
|
|
|
Source: Semianalysis; Green Ash Partners
|
|
|
Micron's stock has risen ~10x in the last year, but so have forward earnings expectations, so its NTM P/E remains low at 11.4x
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- While we remain bullish on the outlook for memory stocks, we also recognise the moves in recent months have been unusually large. We started the year with 8% exposure to memory, and by mid-May this had grown to 20%, driving over a third of the YTD return of the fund. Heading into the thinner liquidity of the summer months, we thought it prudent to reduce our weighting by about a third, and also placed costless collars around two of the positions (Sandisk and Micron) - still allowing for decent upside (the short call strikes were +50% higher than the spot stock prices), but providing some protection to the downside in the near term
|
|
- Increasing clarity is emerging on the economics of very large AI datacentres. As a baseline, a 1GW datacentre can generate about $14BN in revenues and $5 billion in operating income per year (on 5Yr server and 15Yr datacentre depreciation schedules), and costs around $38 billion/GW in capex (per Epoch AI). Server costs could be lower, if custom accelerators like TPUs or Trainium are used, or higher, maybe up to $50 billion/GW for full NVIDIA clusters. IBM's CEO has been on the podcast circuit recently quoting costs $80 billion/GW. It is unclear how he arrived at this number, and IBM haven't built any large AI datacentres
|
|
|
Base case assumptions for a 1GW datacentre show potential to generate AWS-like operating margins
|
|
|
|
Source: Epoch AI, Semianalysis, ARK Investment Management; Green Ash Partners
|
|
|
Datacentre supply remains so constrained that older generation chips are renting well above their expected pricing at this stage in their depreciation curve. Nebius announced a +31% hike to their H100 rental pricing to $3.85 in May
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- Perhaps GPU depreciation rates will re-steepen once more capacity comes online, though there is an interesting new hypothesis emerging - most recently articulated by Gavin Baker - that the disaggregation of prefill and inference could extend the useful lives of HBM-rich servers like H100s for many years (10-15 years, or, "until it melts"). AI inference consists of two parts. "Prefill" is the stage where the model takes in and understands the prompt/context, which is fundamentally a memory capacity-bound problem. "Decode" is the stage where the model generates new tokens, which is a memory bandwidth-bound problem. Because of these different constraints, you do not need to rely on a single chip to do both efficiently. Newer, specialised chips like Cerebras systems or Groq LPUs can be put in front of older Nvidia GPUs (like Hoppers or Amperes) to handle the decode phase, while older NVIDIA GPUs with lots of HBM may remain perfectly well suited to handle the memory-heavy prefill stage, even if their compute speed is outclassed by newer chips for decoding
- This would drastically change the unit economics of the massive AI infrastructure buildout. It lowers financing costs and, to quote Gavin, could "single-handedly save private credit" markets that heavily underwrote the initial 3-to-4-year loans for these GPUs
|
|
- Anthropic's limited release Mythos model has uncovered thousands of software vulnerabilities over the last three months. On Palo Alto Networks recent earnings call, the CEO commented that Mythos "has "increased the terminal value of the entire cybersecurity industry" - the proliferation of AI agents will keep demand for structural protection permanently high
|
|
|
Mythos has increased the pace of uncovering "severe" and "high" software vulnerabilities by 2-3x
|
|
|
|
Source: Anthropic, Peter Wildeford; Green Ash Partners
|
|
- Before Mythos, previous generation models were already having a noticeable effect on "bug bounty" prices, at least in the low to medium impact categories. Mythos and its successors will move up the value chain to more serious vulnerabilities. There is material economic value at stake - resolving a vulnerability via a bug bounty program is roughly 30x less expensive than dealing with a patch after a breach occurs
|
|
|
|
Source: HackerOne, Immunfi, Sherlock XYZ. Green Ash Partners
|
|
- We are paying close attention to the upcoming shift to 800V DC architecture that will begin in 2028. We have already been positioning for solid-state transformers outside of the IT cluster, and in May supplemented this with some new positions in power semis, as silicon carbide and gallium nitride semiconductors will play increasingly important roles as power density per rack continues to rise
- These companies have been suffering from a slowdown in renewables/EVs, which have been the primary sources of demand for DC components and equipment. This makes it a bit different to the bottleneck theses elsewhere - they have capacity to ramp, and should be able to demonstrate strong operating leverage when this transition gathers steam
|
|
|
Gallium Nitride and Silicon Carbide are key enablers of 800V DC datacentre architectures, driving improvements in PUE and allowing much higher energy density per rack
|
|
|
|
Source: Green Ash Partners
|
|
|
NVIDIA's rack power density is set to increase by 25x over five generations of chip and systems design
|
|
|
|
Source: Green Ash Partners
|
|
- Cryptocurrencies have been out of favour lately, though one silver lining is so-called TreasuryCos - companies whose main activity is to accumulate coins like Ethereum or Bitcoin - trade very close to their NAVs now, and are reasonable proxies for the underlying coins, which are still inaccessible to many investors, even via ETFs (especially in Europe)
|
|
|
Bticoin and Ethereum TreasuryCos have become much more closely correlated to the underlying coins
|
|
|
|
Source: Green Ash Partners
|
|
|
|
