|
Green Ash Horizon Fund Monthly Factsheet - April 2026
|
|
The Horizon Fund’s USD IA shareclass rose +23.50% in April (GBP IA +24.21% and AUD IA +24.53%), versus +9.59% for the MSCI World (M1WO).
- It was the largest monthly gain in the 102 months of the strategy track record to date, the second highest being last September
- The markets staged a recovery in April, following a ceasefire in the Middle East. That isn’t to say there’s been any resolution – the Strait remains closed to shipping, and we are weeks away from acute shortages in critical refined products, especially in Europe. But the left tail of War is being more than offset by the right tail of AI
- We expect our heavy weighting to tech to prove resistant to geopolitics, as it did in March, however we raised our cash allocation to 8% in early May from smaller, high beta positions without any positive catalysts expected in the near term. We maintain full exposure to the main themes that have driven the bulk of fund performance YTD
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:
|
|
- CPUs are having their day in the sun. Intel is up ~4x since the US government took a 10% stake last August, resulting in an unrealised gain of nearly $40 billion
- This is due to the explosion in agentic AI workloads. These increase CPU use in many ways:
- Orchestration: the CPU manages the agent loop - planning steps, routing tasks, tracking state, deciding which tool to call next and handling retries
- Tool calls: when an agent uses search, code execution, databases, APIs, calendars, email or enterprise software, those calls are usually CPU-driven, not GPU-driven
- Waiting and coordination: the GPU often sits idle while the CPU waits for external systems, parses results, handles errors and decides the next action
- Retrieval and memory: agents constantly fetch context from vector databases, document stores, caches and long-term memory systems. That means more CPU work for indexing, lookup, ranking and data movement
- Serialisation and parsing: agents pass around lots of structured data - JSON, function-call outputs, tool responses, logs and intermediate states - which creates CPU overhead
- Multi-agent fan-out: one user request can spawn several sub-agents or parallel tasks. The more fan-out, the more scheduling, threading and coordination work shifts to CPUs
- Security and permissions: entreprise agents need authentication, policy checks, access controls, audit logs and sandboxing, all of which add CPU-side work
- Data movement: CPUs help move data between storage, memory, network, databases and accelerators
- It's important to note that this has been widely anticipated. NVIDIA's Hopper HGX systems (2022) had a GPU:CPU ratio of 4:1. This moved to 2:1 in GB200 NVL72 racks (2024/25) and the Vera Rubin systems coming out towards the end of this year have whole standalone CPU racks as part of the overall system design (256 Vera CPUs per rack)
- Also important is that, given the multi-year visibility the industry has had on this trend, companies like NVIDIA, Alphabet/Broadcom, Amazon and AMD have largely sewn up key supply chain capacity several years out. This may be leading node and advanced packaging capacity at TSMC, or more esoteric layers like testing, substrates or 3D bonding lines. In theory, Intel, being a foundry themselves, can sidestep the leading node and packaging bottlenecks, but they have been plagued by execution issues, so we'll have to see
- We'd also like to give a bit of framing on the investment potential for CPUs going forwards:
- Yes, server CPUs are likely to go from no growth to good growth over the next five years. AMD doubled their 2030 forecasts in their earnings call this week to $121BN, implying a growth CAGR of +35% (this may even prove conservative)
- However, CPUs would still only represent 12% of the AI accelerator + server CPU TAM by 2030. The revenue opportunity is dwarfed by GPUs/XPUs, and, of course, memory - in an Nvidia rack, the CPU is 3-6% of system ASP
- We should also note that very high volumes of CPUs can be made from a 12-inch silicon wafer, relative to AI accelerators (a Blackwell GPU uses 30x more wafer area than a smartphone SoC). Global CPU volumes outstrip XPU/GPU production by 100:1
- It isn't entirely a fair comparison to compare Blackwell GPUs to consumer CPUs - chip-to-chip, AMDs flagship Turin CPUs are more like a quarter of the cost of a Blackwell GPU, and only 4x more of them can be produced per wafer - but we would expect a significant chunk of the CPU work in agentic tasks to be offloaded to the user's PC or laptops (Claude Code already does this)
- Finally, a decent chunk of existing AI leaders already make and deploy CPUs for AI workloads: NVIDIA's Grace/Vera, Google's Axion, and Amazon's Graviton. There was a rumour recently that an AI lab tried to buy out Amazon's entire Graviton capacity last quarter
|
|
|
AMD have doubled their server CPU forecasts for 2030, implying CPU share of AI datacentre TAM will rise to 12%, up from 6%
|
|
|
|
Source: AMD; Green Ash Partners
|
|
|
There isn't much 'catching up' to do in the main CPU stocks from a valuation perspective
|
|
|
|
Source: Bloomberg; Green Ash Partners. The PEG ratio is NTM P/E divided by long-term EPS growth forecasts (lower is better)
|
|
- Elon Musk's ambitious TeraFab project is in motion with initial filings from SpaceX indicating a $55BN initial investment, which may scale to $119BN for the full build-out. The campus is to be located on a 6,000 acre site in Gibbons Creek Reservoir, which was previously a coal plant, so has access to grid connections and water. Tesla will be leading the R&D, and is also the primary customer, buying chips for FSD and Optimus. xAI will also buy chips for AI datacentre clusters, and ultimately TeraFab will supply chips for orbital datacentres. Pilot production is targeted for late 2026, with full-scale operations planned for 2027
|
|
- Demand for inference continues to outstrip supply, and this was evident in hyperscale cloud earnings (AWS and GCP top line growth accelerated +480bps to +28% YoY and +1,500bps to +63% YoY respectively). Along with Azure, the big three hyperscale cloud revenue run-rate has reached $340 billion and if current growth rates were maintained revenues would be approaching $1.5 trillion by 2030
- No one is feeling the compute shortage more acutely than Anthropic, who signed a 5 year deal with Google last week to spend $40 billion a year on TPU chips and cloud capacity, and this week signed a deal with xAI/SpaceX for exclusive use of the full 300MW Colossus 1 datacentre. Dario Amodei has said Anthropic's compute demand has risen by 80x this year, versus 10x in their original planning
|
|
|
Steet estimates still expect flat to declining hyperscale cloud revenue growth rates through CY27e
|
|
|
|
Source: Goldman Sachs Investment Research
|
|
|
This is hard to square with a doubling of datacentre capacity over the next two years, even after adjusting for delays and cancellations
|
|
|
|
Source: Goldman Sachs Investment Research
|
|
- DeepSeek V4 was finally released, and while the team once again demonstrated their prowess in wringing efficiencies (At a 1M-token context, DeepSeek-V4 Pro requires only 27% of single-token inference FLOPs and 10% of the KV cache compared to DeepSeek-V3)
- Some important innovations to highlight are their Engram Conditional Memory module, which separates static knowledge from dynamic reasoning, and two new attention mechanisms which compress context which makes KV caching to SSDs feasible. Together, these techniques could increase NAND content in AI datacentres by 10-25x if widely adopted
- As with DeepSeek-V3, these new algorithmic techniques have been openly published will benefit the whole field. What DeepSeek failed to demonstrate this time, however, was further closing of the gap with frontier US labs. In April 2026, the Center for AI Standards and Innovation (CAISI) evaluated DeepSeek-V4 Pro and found the capability lag between open Chinese labs and the US frontier widening from 4 to 8 months since DeepSeek R1's release in February of last year
|
|
|
CAISI evaluations indicate that DeepSeek-V4’s capabilities lag behind the frontier by about 8 months
|
|
|
Comparison of aggregate capabilities over time of the most capable publicly released U.S. and PRC models according to a suite of benchmarks covering five domains.
Every 200-point increase on the y-axis equates to a 3x increase in the odds of solving a given task. Model capability was fitted using an approach inspired by Item Response Theory (IRT), as detailed in the Appendix. 16 benchmarks across 35 models were used to produce this figure. Trend lines were fit with least squares regression on frontier models. Error bars denote 95% CIs.
Source: CAISI
|
|
- OpenAI released GPT 5.5, which is on par with Mythos in one of the AISI's cybersecurity tests. 5.5 is a brand new pre-train, based on two years' of research, and is far more token efficient than Mythos (around 20% of the cost to run)
|
|
|
Mythos and GPT 5.5 (green box) show clear acceleration away from the pack (blue box) when provided with very large token budgets
|
|
|
|
Source: AI Security Institute (UK), h/t @scaling01; Green Ash Partners
|
|
- The ability to alleviate bottlenecks like the memory wall with algorithmic efficiencies combined with frontier models working on long horizon tasks with 10-100 million token budgets show Jevon's paradox in full effect: While frontier models of a few months ago showed linear performance improvements with logarithmic test-time compute scaling, the latest models show exponential gains in capability at these very large token budgets
- Industry participants have spoken of the supply/demand shortfall in AI compute widening even under the current paradigm. Agentic models working at ever longer time-horizons will further exacerbate this imbalance, and with physical constraints on the pace that more compute can be brought online, there are strong incentives for more progress on the algorithmic front
|
|
- GS performed a detailed analysis of AI's impact on biopharma so far, and estimate a +370bps improvement in FDA approval rates (equating to 28 additional new drugs per year) and up to 3 years shorter development times across the various development stages. At an 8% discount rate, they model a PV of $31 billion over one year's pipeline and $412 billion in value uplift from 10 years of pipelines
- Healthcare is a key lever employed by frontier AI labs to make the case for AI improving the human condition rather than just replacing jobs. There has been lots of activity YTD to try to instantiate this reality, from NVIDIA's partnership with Lilly to OpenAI's acquisition of healthcare technology start-up Torch, the release of their life sciences model, Rosalind, or Anthropic's acquisition of Coefficent Bio last month
|
|
|
The use of AI can speed up drug discovery by 2.3x, and the full discovery through to FDA testing and development process by 1.3x (still a long time: 9.8 years total)
|
|
|
|
Source: FDA, BCG, NIH, Tufts, Goldman Sachs Global Investment Research; Green Ash Partners
|
|
- Anthropic held a financial services event, unveiling 10 new AI agents designed to take on tasks like building pitchbooks, reviewing earnings, drafting credit memos and auditing statements
- Financial firms now represent 40% of Anthropic top 50 customers, and they announced a $1.5 billion JV with Blackstone, Goldman Sachs and Hellman & Friedman to make an AI-native financial services firm
- The lesson since the launch of reasoning models is that when frontier labs target a certain domain with curated training datasets and reinforcement learning, model capability in that domain improves extremely quickly - we would expect to see rapid iteration and improvement in both healthcare and financial services in the coming months, as we have seen with coding in the last year or two
|
|
- OpenAI paused their flagship Stargate datacentre project in the UK over high costs of energy and regulation. This is despite efforts by the UK government to attract investment through its AI growth zones policy, which offers subsidies of ~$32/MWh
|
|
|
France and Scandinavia are competitive with the US in terms of large load energy costs
|
|
|
|
Source: BNEF, EIA; Green Ash Partners
|
|
- The Senate released new (bipartisan) text on the CLARITY Act, upholding a ban on interest being being paid on stablecoin deposits, but allowing payments in the form of "bona fide" rewards tied to real transactions, payments, platform usage, loyalty programs, etc.
- This is something of a goldilocks outcome for Circle Internet, as they can keep the yield on treasuries held as collateral, while also allowing them to design financial products and services that reward users adopting their stablecoin platform
|
|
|
|
