|
FOR PROFESSIONAL INVESTORS ONLY
This is a marketing communication. It has not been prepared in accordance with legal requirements designed to promote the independence of investment research and is not subject to any prohibition on dealing ahead of its dissemination. It expresses Green Ash's views on a market theme and does not constitute investment research, advice, or a personal recommendation. Green Ash and funds or accounts it manages hold, or may hold and deal in, positions in companies referred to below - see holdings disclosure.
|
|
|
Green Ash Horizon Fund Monthly Factsheet - August 2026
|
|
The Horizon Fund’s USD IA shareclass rose +3.67% in August (GBP IA +3.67% and AUD IA +3.70%), versus +2.58% for the MSCI World (M1WO).
- The latest semiconductor results give us more confidence than ever that the AI infrastructure cycle is still accelerating. The main constraint is not demand, but the supply bottlenecks, which are spreading from leading-edge silicon into memory, networking, optics, power and the financing of the datacentres themselves
- Looking ahead, September could be choppy, especially given the added uncertainty over the midterms. September drawdowns in midterm years tend to be deeper, though they usually bottom by early October and the ensuing seasonal rally into year end tends to be stronger than typical years. Of course there is no such thing as a typical year, and this time around we are in the midst of a generational capex cycle, as AI datacentre investment approaches 1.5% of US GDP (on a last quarter annualised basis)
- The quiet periods in between earnings seasons are normally dominated by macro narratives, but these may well get sidelined by the imminent Anthropic IPO launch, which will provide investors with their first peek into the financials of a frontier AI lab. We also expect another step change in model capability with the release of Astra by OpenAI. It may take time for the full significance to be appreciated by general users, but the model’s rumoured architectural changes could unlock generally capable, long-running agents, and cross the thresholds of the looser definitions of AGI
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 of factsheet). 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 or future returns and you may consequently get back less than you invested. Fund performance is presented net of fees. Benchmark used is M1WO Index
|
|
|
Here are some tidbits on the themes:
|
|
- Last month we had hyperscaler earnings, which showed an acceleration in top line growth as datacentre capex of 18-24 months prior started to generate revenues. These were followed up in August by bellwether AI semi companies, all of which signalled capacity limitations in areas like HBM and optics on the manufacturing side and power availability on the installation side both limiting the ability to fully serve demand
- Despite these bottlenecks, the main players in AI compute and networking expect +74% growth next year - amounting to an extra +$331 billion in revenues versus 2026. It is the first time we've had longer-term guidance from NVIDIA and Broadcom, both of whom are uniquely positioned to see demand signals and supply chain capacity about two years out
|
|
|
AI semiconductor revenue is growing at a +69% CAGR from 2024 through 2028e
|
|
|
Datacentre segment revenues for NVIDIA, AMD and MRVL, AI revenue sub-segment for Broadcom, total revenues for Cerebras
Source: Bloomberg estimates; Green Ash Partners
|
|
- There were numerous new AI model releases in August, many of which are comment-worthy in their own right, whether it's Meta's Spark 1.3 vying with Google's Gemini Flash 3.8 for fast, cheap tokens at a given level of intelligence, or Anthropic's Fable 5.1 release which attempts to claim the entire intelligence per token curve at different effort levels. But all of these releases have been side-lined by OpenAI's release of GPT-6 Astra
- Astra is rumoured to use "recurrent depth" in its reasoning. The concept of recurrent depth has been popular in ML research for four decades, coming back into fashion every 6-8 years. Most recently, architectures such as recurrent-depth transformers, iterative latent chain-of-thought models, and looped transformers have been in vogue due to their potential to decouple model parameter size from inference compute
- This is beneficial to cloud providers hosting models for inference - serving LLMs is highly memory-bandwidth bound during the token-generation (decoding) phase. For every single token generated, the GPU must fetch billions of parameters from HBM into local SRAM. By looping through identical weights, you keep those weights resident in fast SRAM cache or reduce the total gigabytes of data that must be pulled from HBM
- Beyond hardware economics, the major non-memory benefit is implicit reasoning. Instead of forcing an LLM to waste tokens writing out long, verbose chain-of-thought paragraphs in human language, recurrent depth allows the model to "ponder" mathematically across multiple internal loops before expending tokens to emit a single, concise answer
- The downside to this is observability. Given the cyber capabilities of frontier models (and the recent Hugging Face incident), allowing models to reason in latent space, outside of their readable chain of thought, makes them difficult to monitor
|
|
|
GPT-6 Astra appears to be a massive jump in opaque reasoning ability - it can solve hard competition math problems entirely in its head (i.e. without verbalised reasoning) while earlier LLMs could barely solve basic word problems
|
|
|
|
Source: AISI; Green Ash Partners
|
|
|
Even within readable chain of thought, in their drive for token efficiency frontier models are trending towards a highly compressed form of language, parsable only with the help of other LLMs
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- Astra tops nearly every evaluation benchmark, in some cases by a significant margin, but the main advance for everyday users is a massive improvement in computer use and long-horizon tasks. This effectively enables the model to use any piece of software on a computer for days, just as a human would
|
|
|
GPT-6 Astra has set a new Epoch Capabilities Index record, with a score of 169. This is a substantial jump from the prior best (163), and an acceleration above the trend line
|
|
|
|
Source: Epoch AI; Green Ash Partners
|
|
- Perplexity's WANDR benchmark is one way to showcase Astra's step change in long horizon research capability. WANDR (Wide ANd Deep Research) evaluates AI research agents across 500 complex data-collection tasks totalling 170,495 required records, where the median task demands gathering 50 distinct members across a multi-tier hierarchy with 245 verifiable records overall. It is exceptionally difficult for LLMs, because agents must cast a wide net to find every relevant example that fits a rule, rather than just naming the few prominent ones in the dataset, while satisfying strict, multi-branch evidence verification
|
|
|
Astra scored significantly higher than Fable 5.1 on the WANDR benchmark, while costing -6.1% less per task
|
|
|
|
Source: perplexity.com
|
|
- Frontier model capability is now broadly considered to have crossed the threshold necessary to meaningfully accelerate scientific discovery. This recent paper from Google DeepMind offers a glimpse of this, and the results were achieved with a significantly less capable model than Astra or Fable. DeepMind's "Co-scientist" is essentially a specialised harness orchestrating AI agents in parallel, but could eventually be given control of lab equipment directly, automating experimental research across all scientific fields. Some real-world examples from the paper:
- Advanced Materials Discovery: Designed a non-hazardous chemical vapour deposition precursor pathway using hexachloroethane, successfully synthesising 2D MXene lattice structures while removing toxic reagents from the laboratory - reducing chemical safety hazards in the lab and finding novel, cleaner chemical pathways
- Semiconductor & Nanomaterial Fabrication: Generated lab-tailored synthesis recipes in real time, enabling successful single-attempt, "one-take" growth of high-quality monolayer TMD flakes in roughly one hour - this potentially saves months of trial-and-error recipe calibration in materials fabrication
- Synthetic Biology & Microbiology: Accurately predicted complex emergent swarming phenotypes and colony geometries of engineered E. coli across chemical gradients using sparse imaging data, drastically reducing required wet-lab screening cycles - this accelerates phenotypic screening and reduces the volume of wet-lab assays needed to map genetic/chemical variations
- Clinical AI & Healthcare Systems: Autonomously engineered a novel inference-time scaling architecture (Agent H) that outperformed frontier models on medical benchmarks while demonstrating statistically significant reductions in potential clinical harm during blinded physician trials
- Research Integrity & Autonomous Safety: Integrated deterministic log-verification modules that cross-reference claims against physical instrument telemetry, dropping severe scientific hallucinations from 100% to 24% and automatically intercepting 98.7% of dual-use or unsafe research trajectories - brings us a step closer to operating autonomous wet labs to run scientific experiments at scale
- We see AI drug discovery, personalised health and automated wet labs as one of the highest-value and most beneficial applications of AI for society. We would love to see this nascent "ChatGPT moment" in science gain momentum, giving us the signal to make it a much larger exposure in the fund
|
|
|
Automating scientific research is the ultimate ambition of all the frontier AI labs
|
|
|
|
Source: Google DeepMind; Green Ash Partners
|
|
- At the G20, Elon Musk quoted independent analysis from Funda that there will be a 15 GW shortfall in power in 2027 in the US, given the amount of datacentre capacity slated to come online. SpaceX are seeking to alleviate the gas turbine bottleneck by making their own vanes - something only a handful of companies in the world can do
- There are a number of other short-term solutions to the looming energy bottleneck, including repurposing grid-connected crypto mining sites, behind-the-meter solutions (reciprocating engines, fuel cells), grid & generation uprating, nuclear restarts and load flexibility (curtailment)
|
|
|
There are a number of short-term solutions to the looming energy bottleneck
|
|
|
|
Source: Green Ash Partners. Ranges derived from estimates from Morgan Stanley, Jefferies, Funda, Semianalysis, S&P Global
|
|
|
Morgan Stanley estimates 85 GW of new datacentre capacity set to come online in 2026-2028e, from these four companies alone
|
|
|
|
Source: ERCOT; Green Ash Partners
|
|
- Take-Two released 26 minutes of Grand Theft Auto VI gameplay footage last week ahead of the game's launch in November, which has already been watched 20 million times on YouTube alone (also streams on Netflix)
- GTA VI reportedly cost between $1-2 billion to develop (versus $265 million for GTA V), and its 125 sqkm open world map is roughly the size of Boston
- 13 years after its release, GTA V is still generating $500-700 million annually for Take-Two, with estimated lifetime revenues of $9 billion
- Analysts estimate $3-5 billion in sales in the first week of launch
|
|
|
|
Source: Wikipedia list of best-selling games, 2026. Tetris (~520M, brand-level across dozens of releases) excluded as not a single title.; Green Ash Partners
|
|
|
|
Source: Take-Two quarterly earnings (sold-in / shipped units). Curve interpolated between reported milestones.; Green Ash Partners
|
|
|
|
