|
Green Ash Horizon Fund Monthly Factsheet - February 2026
|
|
The Horizon Fund’s USD IA shareclass fell -0.86% in February (GBP IA -0.92% and AUD IA -0.93%), versus a +0.73% gain for the MSCI World (M1WO)
- While the MSCI World finished up in February, this masked considerable movement under the surface. Broadly speaking, this could be characterised as a rotation out of the world of bits into the world of atoms
- All eyes are currently on events in the ME, but we expect themes of AI investment and AI diffusion through the economy to re-assert themselves in time. Capability thresholds have been crossed in the last two months that are likely to accelerate the ROI side of the equation, alleviating fears of overbuilding, which is anyway being throttled by bottlenecks in energy, memory, and, ultimately, capital markets
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:
|
|
- Bellwether AI semi companies NVIDIA and Broadcom both posted strong earnings and issued next quarter revenue guidance around +7-8% ahead of street expectations. This was consistent with our view that forward estimates are too low in the context of large increases in hyperscale capex plans, but we continue to see significant scope for larger upside surprises to AI semiconductor revenues later in the year. $140BN in incremental capex dollars have been added by the hyperscalers per their full year guides, yet NVIDIA + Broadcom's revenue expectations have only increased by $40BN. We would roughly estimate around $60-80BN of the $100BN gap could flow to the two companies, the rest going to other areas of the AI semi supply chain and into the physical structure of the datacentres themselves
- A similar gap can be observed in FY27, which increasingly looks locked in as another strong year for AI infrastructure investment. We also highlight Broadcom's custom silicon program solidifying their #2 position behind NVIDIA for AI compute. Their AI revenues are set to grow 10x from 2024-2027e to $100BN
|
|
|
Estimate revisions for AI semiconductor revenues have significantly lagged estimate revisions for hyperscale capex
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
Broadcom is likely to increase their share of the AI semiconductor market by 2.5x over the next two years, relative to NVIDIA
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- There have been further bullish inference demand signals from neoclouds like Nebius and Coreweave. Both reported longer customer contracts (extending to multi-year deals), increasing examples of upfront payments outside of hyperscalers, and rising prices on older GPU families
- Logan Kilpatrick, Product Lead at Google DeepMind, commented, "the compute bottleneck is massively under appreciated. I would guess the gap between supply and demand is growing single digit % every day". CoreWeave co-founder Brannin McBee characterises 2026 as "probably broadly sold out in terms of billable compute capacity that's available into the market"
|
|
|
The GPU depreciation debate seems to have been largely settled, with previous generation H100 spot rental prices actually rising YTD, and stable demand for A100s - now 6 years old and fully deprecated
|
|
|
|
Source: Bloomberg
|
|
- With Anthropic dominating much of the discourse so far this year, OpenAI sought to reclaim the limelight by skipping GPT 5.3 and moving straight to 5.4. They maintained their focus knowledge work, making considerable progress on the model's ability to create and edit spreadsheets, presentations and documents. GPT 5.4 now scores 87.3% on OAI's internal junior investment banker benchmark; we notice significant improvements in financial modelling in our own testing, a key difference being GPT 5.4 actually outputs excel formulae, while Opus 4.6 typically uses code to make calculations and then pastes the values into cells
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- Anthropic published their latest paper on the labour market impacts of AI, highlighting a significant gap between the capabilities users are eliciting from AI and the model's theoretical potential
|
|
|
There is a notable gap between the theoretical capability and observed usage of AI models
|
|
|
We filtered the data down to knowledge work tasks, and found it implies AI users in these occupations are only capturing around a quarter of current AI models' potential to augment or automate their tasks
|
|
|
|
Source: Anthropic; Green Ash Partners
|
|
|
Agentic AI is even earlier in its adoption cycle, with software engineering once again leading the way
|
|
|
|
Data reflects tool calls made via public API. 95% CI < 0.5% for all categories, n = 998,481.Source: Anthropic
|
|
- There is increasing evidence that we are getting closer to recursive self-improvement in AI research. This helps explain OpenAI's recent move to release model updates on a monthly basis
|
|
|
|
Source: Andrej Karpathy
|
|
|
This chart shows autoresearch experimenting with 276 optimisations to improve a model's validation loss, with 29 successful hyperparameter changes kept (lower is better)
|
|
|
|
Source: Andrej Karpathy
|
|
- Software stocks have recovered somewhat from the SaaSpocalypse (IGV has retraced 38% of the YTD sell-off), but concerns linger and multiples remain under pressure
- Our current thinking is that large platforms, regulated verticals, cybersecurity, and proprietary data providers are largely insulated from AI disruption, and, for those that execute well, stand to benefit considerably from it
- Microsoft announced Copilot Cowork in their latest effort to capitalise on the AI opportunity. We haven't had a chance to test their latest offering, but it looks to be a step closer to integrating AI across the full Office 365 suite. Only 3% of Microsoft's 365's 450MM users are paying for a Copilot subscription, and the remaining 97% represent a $157BN annual revenue opportunity for Microsoft
|
|
|
Microsoft's new "Frontier Suite" E7 subscription is nearly 3x higher than their standard 365 offering pre-ChatGPT
|
|
|
|
Source: Green Ash Partners
|
|
- Last month, we wrote about the misconception that datacentre projects drive up local electricity prices for consumers. This topic came up on power contractor Quanta Services' earnings call, in which management noted their massive 3GW AWS datacentre project in Indiana was actually expected to reduce household bills by -$7 per month
- That said, grid capacity remains scarce and increasingly politicised, which is incentivising hyperscalers to build their own power generation (behind the meter). Cleanview analysed publicly disclosed permitting, equipment orders and press releases around these projects, and found three-quarters plan to use gas. Operators are getting around the 5-7 year backlog for turbines by turning to less efficient options such as mobile gas generators on trucks, aerodrive turbines (originally designed for aircraft and ships), reciprocating engines and refurbished turbines acquired from industrial operations
|
|
|
There are 46 datacentre projects in the US with a combined capacity of 56GW planning to build their own behind-the-meter power generation
|
|
|
|
Source: Cleanview
|
|
|
|
Source: Cleanview
|
|
- Circle Internet (USDC) posted very strong results, driving a +139% rally in the stock from the February low. We like Circle's breadth of relationships with financial institutions (e.g BNY, Blackrock, Visa), and the engagement of others like GS, State Street, Mastercard, DB and JPM on their Arc Blockchain Testnet. The testnet is a sandbox environment to experiment with use cases like instant settlement of tokenised securities, cross-border transactions and agentic commerce
- In December, Circle and Intuit announced a major, multi-year strategic partnership. By leveraging USDC, Intuit can bypass the delays and costs of legacy payment rails (like ACH or SWIFT), allowing them to offer their 100 million customers programmable, 24/7, near-instant money movement. This could drastically speed up events like tax refund disbursements, cross-border remittances, and small business settlements
- Most exciting is the opportunity for Circle to become the platform for agent-to-agent transactions, and could represent huge upside if there will one day be billions of agents transacting with each other online
|
|
|
Stablecoin transaction volumes grew +46% YoY to $3.7 trillion in the last quarter of the year
|
|
|
|
Source: Circle Internet; Green Ash Partners
|
|
|
|
