|
Green Ash Horizon Fund Monthly Factsheet - August 2025
|
|
The Horizon Fund’s USD IA shareclass rose +0.75% in August (GBP IA +0.69% and AUD IA +0.61%), versus +2.61% for the MSCI World (M1WO).
Please click below for monthly factsheet and commentary:
|
|
|
In August, we completed a significant overhaul of the fund's themes, to better reflect the way we think about the exposure of the underlying stocks to major trends like AI semis, infrastructure, energy, and the beneficiaries of the huge investment cycle currently underway. For more information:
|
|
|
|
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 (new) themes:
|
|
- NVIDIA earnings announcements are now more important risk events for the broader markets than inflation or payrolls data (as measured by options hedging volumes). This quarter saw another beat, despite street estimates creeping higher in the run-up
- Some viewed the Q3 guide as a little light (despite being in-line with street estimates) - this can be explained by no China sales included in the guide
- NVIDIA was preparing a low cost/low performance chip for the Chinese market that complied with the strict US rules on compute and memory bandwidth (to be called the B30/B40), but recent messaging suggests the Trump admin might be open to issuing licenses for a chip half as as performant as the B300 (expected to be called the B30A). This would still be significantly better than the best domestic chips from Huawei
|
|
|
Even at 50% of the performance of NVIDIA's B300s, the rumoured B30A chip would be significantly better than any domestic Chinese alternative
|
|
|
|
Source: Semianalysis, public reporting; Green Ash Partners
|
|
- GPT-5 finally arrived. We sent out some thoughts the day after release, and since then there have been ebbs and flows of reactions, ranging from disappointment, to wonder, to the perennial 'deep learning has hit a wall/we are building too many datacentres'. A few weeks on, our updated thoughts are:
- GPT-5 is a strong frontier model, with a much more 'agentic' feel to it. It seems far better at weaving tool calls into its outputs, for example, by blending information from web search in with uploaded documents when we use it to augment our research work
- For most of ChatGPT's 700 million users, it will be their first interaction with a reasoning LLM - Sam Altman tweeted that only 7% of paid users and <1% of free users had experimented with the previous reasoning models like o1 and o3 prior to the launch of GPT-5
- Even when set to 'thinking' mode, GPT-5 is a pared back and heavily optimised model - a necessity for OpenAI's current infrastructure capacity constraints - and so tells us little about the remaining mileage to be had from scaling (whether that be pre-training, post training/synthetic data/RL, or test time compute/search/parallelisation)
|
|
|
The real leap was from base models to reasoning models - GPT-5 is incomparably better than the original GPT-4 (and a bigger improvement than GPT-3 to GPT-4), but more of an iterative improvement versus other reasoning models released this year
|
|
|
|
Source: Artificial Analysis; Green Ash Partners. Artificial Analysis Intelligence Index v2.2 incorporates 8 evaluations: MMLU-Pro, GPQA Diamond, Humanity's Last Exam, LiveCodeBench, SciCode, AIME, IFBench, AA-LCR
|
|
- Google DeepMind released several state-of-the-art models in other modalities. The two most significant were:
- Genie 3 - a general purpose world model for generating dynamic worlds that can be navigated in real-time, retaining consistency for several minutes (versus 10-20 seconds for Genie 2, 8 months ago). World models have lots of potential for things like videogames and virtual reality, but GDM see them as a critical stepping stone on the path to AGI, making it possible to train AI agents on an infinite source of simulation data, in an environment that mimics the physics of the real world
- Nano-banana - re-named post release to the much more prosaic Gemini 2.5 Flash Image - dramatically advanced the frontier of generative image editing. The model pairs image generation with Gemini’s world knowledge for more semantically correct results, supports multi-turn editing, and can maintain character/object consistency across numerous iterations
|
|
|
An example of Gemini 2.5 Flash Image arranging 13 different elements into a single composition
|
|
|
|
Source: @MrDavids1
|
|
|
And another converting some of the Bayeux Tapestry into the style of modern war photography
|
|
|
|
Source: @emollick
|
|
- MIT released a paper that the markets pounced on to justify some seasonal August jitters. The paper was initially paywalled, so all that circulated was the headline "95% of enterprise generative AI pilots are failing". The paper is now publicly available and a full reading leads to completely different conclusions:
- General LLMs are performing far better than narrow custom AI solutions within entreprises
- There is a much higher failure rate when entreprises try to build their own tools (contradicting the consensus 'AI will destroy software' narrative
- There is already a great deal of 'shadow AI' use in entreprises, with 90% of employees using LLMs regularly, but only 40% of companies having purchases an LLM subscription
- Of the 7 sectors evaluated, Media & Telecom and Professional Services have showed rapid adoption, while other sectors are lagging behind - this highlights a GenAI divide between different types of entreprises. A similar divide exists within entreprise business functions, with current AI deployments most impactful in Sales & Marketing, Customer Service, and Operations (this all makes intuitive sense)
|
|
|
Rather than showing a 95% failure rate in entreprise AI pilots, the paper showed an 80% success rate for general-purpose LLMs, though only a 25% success rate for narrower systems
|
|
|
|
Source: The State of AI in Business 2025 - Aditya Challapally et al
|
|
- Meanwhile, a paper from Stanford, which receive far less attention, showed the clearest signs yet of AI beginning to impact the labour market, specifically in entry level roles (22-25 year olds). This pattern is apparent across roles, but most pronounced in software development and customer service
- The results of the study were confirmed anecdotally by Fed Governor Chris Waller, who said recently: "Supplementing this hard evidence of falling labour demand is the consistent story from my business contacts that they are not hiring. The mix of reasons is not only uncertainty over tariff policy and slowing demand for their goods and services, but also, increasingly, uncertainty over how to use artificial intelligence, which is especially freezing hiring for some entry-level jobs."
|
|
|
Software Developer Headcount Over Time by Age Group (normalised)
|
|
|
|
Source: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence - Erik Brynjolfsson, Bharat Chandar, Ruyu Chen
|
|
|
Customer Service Headcount Over Time by Age Group (normalised)
|
|
|
|
Source: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence - Erik Brynjolfsson, Bharat Chandar, Ruyu Chen
|
|
- AI model training runs have a very unusual load profile, unexpectedly rising and falling from full load to nearly idle in fractions of a second. This could become a serious issue as cluster sizes approach 1GW (equivalent to the output of a nuclear power plant). Fluctuations at this scale could potentially lead to cascading blackouts across the ageing US electricity grid
- Battery Energy Storage Systems could help mitigate this - Fluence Energy estimate a potential TAM of $8.5BN from AI datacentre demand. This market could ultiamtely be much larger - a 1GW datacentre would need about $1BN of BESS, just for 4hr back-up coverage. If a gigawatt-scale datacentre was powered by solar (which would require 10,000 acres of solar panels), it would require about $9BN worth of BESS to cover a 'two bad nights' scenario
|
|
|
Battery Energy Storage Systems (BESS) can smooth power fluctuations related to AI training runs
|
|
|
|
Source: Company Estimates; McKinsey
|
|
|
|
