|
Green Ash Horizon Fund Monthly Factsheet - June 2026
|
|
A bit of housekeeping to start off - the Horizon Fund's GBP IA share class is now available on Interactive Brokers UK. USD RA and AUD IA share classes to follow shortly.
The GA Horizon Fund (USD IA) fell -2.12% in June (GBP IA -2.10% and AUD IA -2.05%), versus -0.72% for the MSCI World (M1WO).
- We expressed some caution in our commentary last month, which proved well founded, as equity markets were much choppier in June. The VIX spiked up above 20 on two occasions during the month, each time settling back down into the teens a few days later
- July is typically a seasonally strong month for equities, as fresh capital gets deployed at the half-year mark. In a few weeks, attention will turn to Q2 earnings season. We note that 1H26 earnings growth in the US is expected to come in at +20% YoY, twice the S&P’s +10% rise 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 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:
|
|
- Bottom-up analysis from Morgan Stanley suggests NVIDIA's latest Vera Rubin systems are ~1.5x more expensive than Google's most recent TPUs and 2.7x more than Amazon's Trainium 3 per GW. However, measured in compute power efficiency (TFLOPs/watt) they are up to 8x ahead of custom ASICs (in FP4)
|
|
|
NVIDIA systems are considerably more expensive in $/GW than custom ASICs
|
|
|
|
Source: Morgan Stanley
|
|
|
However, measured in compute power efficiency (TFLOPs/watt) they are up to 8x ahead of current ASICs
|
|
|
|
Source: Morgan Stanley
|
|
- In addition, NVIDIA's hardware improves significantly over time, due to software optimisations - in just two months, NVIDIA's software stack has reduced token costs for DeepSeek V4 by up to 5x on Blackwells. Software improvements are often backwards compatible with prior generations of NVIDIA systems, extending their useful life
|
|
|
NVIDIA’s full-stack inference software continuously improves hardware performance
|
|
|
|
Source: NVIDIA
|
|
- Systems-level co-design through GPUs, CPUs, networking and storage is increasingly important for AI agents, which operate much more complex workflows than traditional datacentre loads. With complexity comes opportunity for differentiation and optimisation within the various hardware ecosystems
|
|
|
Agentic workflows are much more complex than traditional datacentre loads
|
|
|
|
Source: NVIDIA
|
|
- Rising complexity makes domain knowledge crucially important when navigating the semiconductor investment landscape. Architectural decisions at system level have material investment implications for compute, memory and networking with increasing heterogeneity across hardware ecosystems, datacentre operators and AI labs. NVIDIA remains the force behind many of these trends due to its rapid iteration cycles and active management of AI semi supply chains
|
|
|
There is a prevailing market narrative that NVIDIA may retain their near-monopoly on training, but cede share in the inference market. There are no signs of this so far
|
|
|
|
Source: The Information; Green Ash Partners
|
|
- A recent Bloomberg story about Meta potentially following in SpaceX/xAI's footsteps and renting out some of their AI compute clusters caused a sell off in datacentre stocks and raised the by now familiar spectre of AI datacentre oversupply. We note that:
- Meta has leased 5GW of neocloud and co-located compute year to date (per Semianalysis). Nebius has a $27 billion deal with Meta, and just in May raised their FY26 capex guide to urgently bring capacity online in 2027. About half of this is to be dedicated to Meta, and Meta has agreed to act as a guaranteed buyer for the other half, which Nebius is permitted to sell to third-party enterprise clients first (which typically command much higher margins) but has Meta as a guaranteed safety net if the capacity remains unsold. Meta's agreements with neoclouds like Nebius and CoreWeave strictly prohibit the subletting of third-party infrastructure - Meta cannot legally act as a discount reseller of Nebius's hardware.
- Just last week there were reports that Google was having to pare back Gemini access, as they didn't have enough capacity to meet Meta's demand (Meta have been using Gemini to moderate/classify social media posts)
- So then, why is Meta thinking about selling compute? This largely comes back to heterogeneity - very large, coherent clusters of GPUs are very scarce - this has been demonstrated by SpaceX renting out 100s of MW clusters way above market rates. LLM training requires running massive GPU clusters at 100% capacity for months at a time but once a frontier model finishes training, utilisation drops sharply as the compute shifts to inference or fine-tuning, which requires vastly less raw horsepower. The Bloomberg report and subsequent analysis note that external companies are actively approaching Meta and bidding at a premium for this idle capacity
- If they follow SpaceX's strategy, there is very little downside - no other hyperscalers or clouds would rent out huge clusters with 90-day cancellation clauses (by either side), and just 200MW could command $10 billion in annual revenues - transformative for SpaceX which only had c.$19BN in revenues last year, but enough even to move the dial for Meta (revenues of $200BN in FY25)
|
|
|
SpaceX has demonstrated the scarcity value of coherent clusters in the 100s of MW with immediate availability
|
|
|
|
Source: Semianalysis; Green Ash Partners
|
|
- Underestimating the demand for AI compute has been a consistent theme of the last three years - street estimates for FY25-FY28e hyperscaler capex have been revised higher by a cumulative +$1.6 trillion (+70%) versus expectations 12 months prior
|
|
|
Street forecasts for datacentre capex have consistently and materially undershot the reality over the last few years
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- This is partly due to extreme uncertainty and scepticism over the revenue trajectory of the leading frontier AI labs over the next few years. OpenAI and Anthropic have reached a combined ARR of $100 billion - up from ~$2 billion just 18 months ago, and are both eyeing $300 billion by YE30 which would make them amongst the largest software companies in the world by revenue. This may sound outlandish, but attaining this goal only requires a +57% CAGR from here - not far off some of the more bullish datacentre capacity growth forecasts - and could even prove conservative if AI capabilities continue to follow an exponential rate of improvement
|
|
|
OpenAI + Anthropic's shared $300 billion ARR by YE30 goal only requires a +57% CAGR from here
|
|
|
|
Source: Past ARR numbers and annual targets through YE30 combine company disclosures with industry rumours, reporting from The Information; Green Ash Partners
|
|
- OpenAI posted an interesting blog about internal use of Codex (as a proxy for agentic AI use) which may serve as a leading indicator for entreprise AI adoption. So far, AI labs have understandably followed the steepest adoption curve, followed by tech companies and then finance. Yet even in a frontier lab, departments like finance, HR and legal lag software engineers by about six months
|
|
|
Share of work at OpenAI on Codex by department since August 2025
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- Agentic AI use via Codex is now the primary AI tool used at OpenAI, but is still in the very early innings amongst OpenAI's customers
|
|
|
Share of active users using Codex
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- For the average OpenAI worker, Codex usage now accounts for more than 85% of output tokens. Since Codex users tend to use more tokens than non-users, its share of overall tokens is even higher: Codex accounts for 99.8% of weekly output tokens generated within OpenAI
|
|
|
Change in combined output tokens by OpenAI department since November 2025
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- The same is true for organisations and individuals who use Codex. By May 2026, 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, 70.2% made one estimated to exceed one hour, and 25.6% made at least one Codex request estimated to exceed eight hours
|
|
|
Share of output tokens from Codex
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- Of course, from a customer perspective, all of these tokens come at a price. The brief sugar-rush of token-maxxing has been replaced by fiscal discipline - Uber have reportedly capped software engineers at a $1,500 per month token budget, and Tesla have been even more frugal with a $200 per week cap. The market debate has shifted from how rapidly entreprises are adopting agentic AI to how much of a benefit they are actually extracting from it, and at what cost
- It should be noted though that these levels of expenditure are somewhat outlier examples, and would put the companies well above the 90th percentile of AI expenditure, per Ramp, which itself over-indexes to tech companies, so the true enterprise AI expenditure distribution is likely to be substantially lower still
|
|
|
The top 1% of Ramp's (tech-heavy) userbase are spending $7,500 per month on AI tokens
|
|
|
|
Source: Coinbase
|
|
- As AI use transitions from $20/month subscriptions to serious incremental opex, organisations are starting to get much more serious about value. Coinbase recently shared their strategy to control costs not by restricting token consumption, but by optimising the model orchestration layer - defaulting to cheaper models for basic queries, custom harnesses to route and combine models (e.g. frontier model for planning, cheaper models for tool calls etc), improved caching, context optimisation (clearing context for new tasks) and closer scrutiny of high usage (to evaluate the ROI). These measures have reduced their internal AI expenditure by ~-40%, even as overall token consumption has grown
|
|
|
Model routing and usage optimisations at Coinbase supported continued overall token usage growth at ~-40% lower cost
|
|
|
|
Source: Coinbase
|
|
- Just as huge intellectual resources are poured into proprietary asset allocation models in the fund management industry in the search for the Pareto frontier of risk-adjusted returns, so the orchestration layer between models and entreprise workflows will become an area of intense focus, as companies strive to maximise the productivity benefits of AI while minimising cost. There is no single answer to this - AI is the ultimate horizontal technology, with limitless applications and equally limitless potential for combination within harnesses, scaffolds, parallelisations and loops
|
|
- Datacentre developers are increasingly turning to on-site gas generation to avoid long grid connection queues. For a gas plant serving a 500MW data center in the US, gas engines are currently the lowest-cost option, with a levelised cost of electricity (LCOE) of $103/MWh, according to BloombergNEF analysis. Their lower capital requirements outweigh higher fuel costs. Fuel cells are the most expensive option today at $140/MWh, while open- and combined-cycle gas turbines (CCGT) sit between the two
- BNEF estimate 124GW of on-site gas power could be supplied to datacentres, based on announced capacity - the wrinkle is air permits. Nebius' $2.6 billion/328MW deal with Bloom Energy for solid oxide fuel cells is reportedly driven by difficulty getting air permits for their datacentre in New Jersey, which is critical for them to bring online to meet their FY26 ARR guide. SpaceX's Colossus datacentre in Memphis is operating without air permits, creating a row with the local population which they are trying to defuse with discounted Starlink subscriptions
|
|
|
Gas engines are the lowest cost option for on-site power generation
|
|
|
|
Source: Green Ash Partners
|
|
- Real world asset tokenisation has reached an inflection point, supported by the technological maturity of blockchain networks and stablecoin payment rails, regulatory momentum, and institutional demand. US Treasuries and private credit account for two thirds of the market (47% and 19% respectively), and while the overall size of $32 billion is a drop in the ocean versus the trillions represented by these asset classes, it is growing rapidly (10x in the last two years)
|
|
|
Real world asset tokenisation has reached an inflection point
|
|
|
|
Source: RWA.xyz, Apollo Chief Economist
|
|
|
|
