|
Green Ash Horizon Fund Monthly Factsheet - March 2026
|
|
The Horizon Fund’s USD IA shareclass fell -3.18% in March (GBP IA -3.27% and AUD IA -3.36%), versus -6.37% for the MSCI World (M1WO).
- There is a marked difference in how tech has performed in this latest geopolitical crisis compared to 2022. We went into this year with much lower valuations - the Mag 6 (ex. Tesla) had NTM P/Es of 33.4x on a market cap-weighted basis on 31/12/21, versus 26.7x on 31/12/25 and 22.2x at the end of March
- Most importantly, this time tech has the secular tailwind from AI which is driving the largest co-ordinated infrastructure build out in history, encompassing large swathes of the economy from construction, heavy industry and advanced manufacturing to knowledge work and the digital world
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:
|
|
- NVIDIA held their annual GTC event - as always, the highlight was Jensen Huang's 2hr+ keynote. Noteworthy moments were:
- Comments on more than $1 trillion in purchase orders for Blackwell and Rubin systems through CY27 (NVIDIA disclosed $500BN through CY26 last October). This implies another $500BN in datacentre revenues in CY27, vs. consensus of $438BN. Additional orders likely to build on this backlog over the next 6-9 months - street estimates are much too low
- Groq LPUs for fast inference set to be incorporated into Vera Rubin multi-rack pod systems from 3Q26
- More detail was given on NVIDIA's roadmap through to the 2028 Feynman system. The Blackwell->Rubin->Feynman generations showcase NVIDIA's strengths in extreme co-design, with parallel innovations across compute, memory, storage and networking, all the way up to the design of the rack itself
- Using Hopper (2022) as a baseline, these innovations are expected to increase tokens/$ by 500-1000x and tokens/watt by 250-500x over a six-year period
- Each rack has ~1.3 million components made by >200 suppliers, so every design choice or architectural change sends ripples through NVIDIA's supply chain
|
|
|
NVIDIA's 'extreme co-design' started in earnest with the Blackwell generation
|
|
|
|
Source: NVIDIA; Green Ash Partners
|
|
- One such architectural change is the looming need to replace copper with optics in the scale up fabric (which provides the all-to-all connectivity between GPUs in a single rack). It looks like copper will still play a role in Rubin and Feynman racks over short distances, but photonics will win out over the medium term. The fund has been adding exposure to this theme over the last few weeks
|
|
|
Copper hits physical limits over rack distances (~2m) at bandwidths over 1.6T
|
|
|
|
Source: Lumentum; Green Ash Partners
|
|
- Higher memory demand is also locked in by NVIDIA's roadmap - the Feynman generation (2028) will likely have 5-10x the HBM per GPU than Hopper (2022). This trend will apply to AMD's systems and custom XPUs also. HBM is about 3-4x more intensive to make in terms of bits/wafer versus standard DRAM, and so datacentre demand will crowd out traditional consumer electronics markets for the next several years
|
|
|
Memory share of total hyperscaler capex is forecast to nearly triple between 2025-27, implying about $250BN in spend - about a third of Micron and SK Hynix's combined market cap
|
|
|
|
Source: Semianalysis; Green Ash Partners
|
|
- Elon Musk announced his plans to build a Terafab, under the joint banners of Tesla, xAI and SpaceX. His argument goes that the need for compute will far outstrip the conservative multi-year roadmaps for capacity expansion set by TSMC and ASML
- In the limit, leading edge wafer capacity is an insoluble bottleneck over the 2026-30 timeframe. ASML can only make about 500 EUV machines over this period, held back by Carl Zeiss, whose mirrors are polished to atomic-scale tolerances in factories that take up to 7 years to build. Even in the most aggressive scenario, where AI datacentre chips crowd out smartphones and PCs to take a 60% share at the leading edge, and there are no bottlenecks on the energy or the physical construction side, the world can only deploy a maximum of 120-140GW in AI datacentre capacity through 2030 (this still would represent $6-7 trillion of cumulative investment)
- True to form, Elon Musk would like to think bigger than this, planning a semiconductor fab with an initial output of 100k wafers per month, scaling to 1 million wafers per month over time (about 70% of TSMC's total global capacity)
- TSMC consumed 25TWh of electricity in 2024, about 8% of Taiwan's electricity consumption (this would equate to about 0.6% of US consumption)
- Oh, and 80% of Terafab's output will be for datacentres in Space
|
|
|
Musk has suggested Terafab could grow to 100 million sqft, 10x larger than Giga Texas (Tesla's HQ) which is one of the largest buildings in the world
|
|
|
|
Source: Green Ash Partners (made with Claude)
|
|
- Token demand continues to significantly outstrip supply, causing upward pressure on spot GPU rental rates. This is being driven by increased adoption of agents, especially in coding, where employers are starting to monitor token usage as a proxy for productivity (Jensen also recently commented that he would be ‘deeply alarmed’ if one of his engineers earning $500,000 a year did not consume at least $250,000 in tokens)
- Agentic workflows require larger context windows, which, in turn, explodes memory requirements. With this in mind, Google's TurboQuant algorithms couldn't have come at a better time - far from being bearish for memory demand, the ability to compress the KV cache and reduce per token memory usage by 6x makes >1 million token context windows feasible for everyday use cases. These kinds of algorithmic efficiencies arrive regularly - DeepSeek's Multi-head Latent Attention (MLA) reduced the KV cache size by 57x, and quickly replaced traditional MHA after publication in May 2024
- If you asked frontier lab insiders what everyone else isn't discussing enough, they might say automated research. Current models are playing an increasingly large role in designing their next generation, which is noticeably speeding up release cadences
|
|
|
Ex-OpenAI researcher Leopold Ashenbrenner flagged automated AI research as a catalyst for a 'fast take-off' scenario in his Situational Awareness essay (June 2024)
|
|
- Anthropic announced reaching the $30BN ARR milestone (up from $9BN at the start of the year), while also unveiling their new Mythos model, via a 245 page technical report and Project Glasswing
- Mythos Preview has "already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser". Project Glasswing is a limited release of Mythos to 40 partners to give software infrastructure partners time to harden their security against systems with much higher cybersecurity capability
- It is a much larger model - estimated at ~10 trillion parameters versus ~5T for Opus and ~1T for Sonnet. OpenAI, Google and xAI are all gearing up to release their own models in this weight class in the coming weeks
|
|
|
Mythos is far more expensive to serve on a $/token basis, but is much more token efficient (about 5x better than Opus 4.6 on agentic search tasks)
|
|
|
|
Source: Anthropic
|
|
- Mythos has set off another round of weakness in software stocks, on the realisation that the next batch of frontier models will not just be very good at creating new software, but finding cybersecurity exploits in currently deployed software products as well (Bessent and Powell convened a meeting of bank CEOs this week to discuss the potential risks Mythos poses to the banking system)
|
|
|
A tale of two tech sectors - software has significantly underperformed semiconductors YTD
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- Planet Labs successfully used an NVIDIA Jetson Orin module to run an AI image processing model in space, aboard one of their Pelican satellites. These satellites capture multiple spectral images down to 30cm resolutions, and are capable of generating hundreds of terabytes of raw data every day. This is 4-5x downlink capacity, made worse by orbital mechanics resulting in the satellites only being in a position to broadcast to ground stations about 10% of the time
- By processing images onboard the satellite, Planet can reduce the amount of data that needs to be transmitted to Earth by several orders of magnitude, and reduce time to answer from 2-4 hours to minutes
|
|
|
Imagery captured by Pelican-4 on March 25, 2026 over Alice Springs, Australia, demonstrating the first successful deployment and execution of AI-driven object detection directly onboard Planet spacecraft.
|
|
|
|
Source: Planet Labs
|
|
- NVIDIA is also starting to get involved in the power side to AI datacentres, developing software optimisations designed to manage grid loads in times of high demand. 'Six nines' of uptime have become a mantra in cloud contracts, but this may not be necessary for AI workloads. For example, they see ways to to dynamically reduce power draw by ~10% in times of stress, while perhaps only giving up -2% in system output
|
|
|
Nvidia is working on algorithmic solutions to improve the power usage efficiency (PUE) of large clusters
|
|
|
|
Source: NVIDIA; Green Ash Partners
|
|
- Last summer, the Trump administration announced a trade deal with Japan that included a vague and somewhat puzzling commitment from Japan to "invest $550 billion directed by the United States to rebuild and expand core American industries"
- This seems to materialising, to some extent, in the form of a massive 10GW datacentre project in Ohio, led by Softbank and AEP Ohio. It will be powered by 9.2GW of new natural gas generation capacity, backed by $33.3BN in Japanese investment. Phase 1 is targeting 800MW by 2028, but the full 10GW could ultimately require $350-500BN in capital
|
|
|
Softbank's 10GW datacentre is will be built on the site of the Portsmouth Gaseous Diffusion Plant in Piketon, Ohio
|
|
|
|
Source: US Department of Energy
|
|
- Permitting is one of the main bottlenecks holding back datacentre construction and new power generation. Texas' relatively lighter touch in this regard is attracting a lot of both
|
|
|
Texas is growing solar and storage at 2x the rate of the other 49 states and its share of total US capacity is set to reach 26% and 42% respectively by 2027e
|
|
|
|
Source: EIA-860; Green Ash Partners
|
|
- The CLARITY Act in the US continues to lack clarity, due to intensive lobbying from banks to prevent interest-like features in stablecoins which could compete with traditional deposits. The White House came out on the side of stablecoins this week, with an estimate from the CEA contending that allowing interest on stablecoins would only increase banking lending by $2.1 billion, or 0.02% of total loans
|
|
|
Crypto has outperformed gold as a geopolitical risk hedge this time around
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
|
