|
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 - July 2026
|
|
The GA Horizon Fund (USD IA) fell -10.02% in July (GBP IA -10.08% and AUD IA -10.13%), versus +0.51% for the MSCI World (M1WO)
- July saw a major pullback in momentum across a number of popular themes, most notably in semiconductors (SOX -20%), which had rallied +88% in the preceding three months, and still finished the month up +60% year to date. While a correction was due, the violence took many by surprise - it was the largest momentum drawdown since the GFC and the 2nd largest de-grossing in the last ten years (per GS Prime).
- We expect market sentiment to remain febrile until some kind of resolution is found between the US and Iran, but we remain constructive from a fundamental point of view
- Specifically, we remain very bullish on the AI infrastructure theme. The rush of enthusiasm over April and May was largely driven by Claude Code/Mythos releases in Q1 which imparted the sense of exponential improvement to the general public that had been prevalent in Silicon Valley for some time. It felt like this wore off a bit in June and July, but the exponential continues. We expect OpenAI to take the baton over the coming weeks with the release of GPT-6 and new multi-agent orchestration harnesses that not only unlock higher value use cases for AI agents, but possibly start accelerating their own development as well. In terms of the markets, we expect this will manifest in the form of rapid ARR growth in the frontier labs, insatiable demand for tokens in entreprise and consumer inference, and, consequently, massive ongoing investment in AI infrastructure capacity.
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:
|
|
- SpaceX held their first earnings call, in which they laid out plans to have 2GW of compute online by YE26, versus 1.4GW currently, and close to 10GW of cumulative compute by YE27. They also fully committed to NVIDIA's Vera Rubin platform - both for terrestrial datacentres and for their Starmind AI satellites which will function as an optimised Vera Rubin NVL72 rack
- Their terrestrial ambitions alone imply $350-420 billion in capex on NVIDIA chips over the next year and a half at ~$40-50 billion per GW, against current cash of $100 billion. This probably signals more large-scale vendor financing from NVIDIA coming down the pipe
- At the mid-point, it also amounts to about 57% of NVIDIA's forecasted datacentre revenues over the next six quarters. The servers for SpaceX/xAI's first 1.4GW were supplied at around a 50:50 split between Dell and SuperMicro - at last estimate, these two providers combined only amounted to 13% of NVIDIA's sales, suggesting SpaceX's planned emergence as a new hyperscaler is not priced in to estimates yet
|
|
|
SpaceX's potential 8.4GW of NVIDIA orders over the next 18 months are not in current bottom-up supply chain estimates
|
|
|
|
Source: Bloomberg estimates; Green Ash Partners
|
|
|
Since we last ran this chart, NVIDIA's long-term EPS growth forecasts have risen from +45% to +58% per year, yet its NTM P/E is at the lows of the last five years (PEG of 0.3)
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- Elon 'guesses' Vera Rubin's could monetise at $30-50 per watt, which marries up with recent benchmarking by CoreWeave which finds Vera Rubin systems around 10x more performant in terms of tokens processed per second per watt
|
|
|
CoreWeave released benchmark results showing Vera Rubin processing about 10x the tokens per second (TPS) per MW versus GB200s, which would tie in to Elon's ballpark estimate for $/watt (Silicon Data's B200 spot rental rates currently $5.58/hr)
|
|
|
|
Source: CoreWeave
|
|
- Microsoft, Amazon and Google all reported towards the end of July, and demonstrated another quarter of material estimate beats and growth acceleration in their cloud businesses. If we use a rule of thumb that hyperscaler cloud revenues from AI datacentre investment show up about 18 months after the capex, 1H26 growth would have been driven by the investments of 2H24 - some $115BN of capex deployed by Microsoft, Amazon, Alphabet and Oracle (+68% YoY)
- Street estimates currently forecast a material revenue growth deceleration in these cloud business in 2H27, however this period should reflect investments made in 1H26: $280 billion across this group, or +143% higher than the capex deployed in 2H24
- For context, these four cloud service providers were at a ~$400BN annualised revenue run rate in 2Q26 - even 'just' a sustained +56% annualised growth rate would double this every 15 months
|
|
|
Even after two consecutive quarterly beats in cloud revenues across the hyperscalers YTD and clear acceleration in revenue growth rates, street estimates a few quarters out remain conservative
|
|
|
Average of AMZN, GOOGL, MSFT, ORCL
Source: Bloomberg; Green Ash Partners
|
|
|
These four cloud service providers have a revenue backlog of $2.3 trillion
|
|
|
Average of AMZN, GOOGL, MSFT, ORCL
Source: Bloomberg; Green Ash Partners
|
|
- But what about open source? We wrote about this recently in The Orchestration Layer, and since then the open/closed source debate has continued to rise to the top of the conversation, not least due to the release of Kimi V3 - a Chinese open-source model with capabilities close to the publicly available frontier models of the closed source labs
- The leading open source models have largely achieved their capability improvements by chasing the same scaling laws as the US labs - Kimi is a massive 2.8T parameter model that requires an NVIDIA B300 to fit on a single rack (and also relies on Blackwell's native FP4 precision to run efficiently, Hoppers only natively support FP8)
- Increased usage of open source models is bullish for the AI infrastructure theme, not bearish. Running them on prem will require more AI compute, as owners won't benefit from the batching efficiency of public clouds; public cloud providers will still make good margins on the rest of the workloads (which will be the majority), and cheaper 'good enough' models will unlock more use cases, accelerating overall token demand growth
|
|
|
Open or closed source, AI models require infrastructure to run, and demand for tokens continues to grow exponentially
|
|
|
|
Source: Exponential View
|
|
|
The recent declines in token price indices have been driven by OpenAI cutting prices on their mid-size models (Terra, Luna) - open source token pricing is rising
|
|
|
|
Source: Silicon Data
|
|
|
AI infrastructure is running near full capacity, unlike the idle fibre-optic networks of the early 2000s
|
|
|
|
Source: Fiber optic cable utilization sourced from Harvard Business School study "Level (3) Communications in 2001: The 'Pivotal Year'" published on April 2001; Graphics processing units (GPU) utilization sourced from NBER paper "Flexible Data Centers and the Grid: Lower Costs, Higher Emissions?" published in July 2005; J.P. Morgan Asset Management assumptions of data center capacity from company reports. Data reflect most recently available as of October 17, 2025.
|
|
- OpenAI posted an interesting blog from their economics team, with data showing small businesses/teams are increasingly using AI to supply skills outside of their usual proficiency, be it financial analysis, legal work, marketing or IT support
|
|
|
Nearly half of non-generic ChatGPT use is outside a user's occupation
|
|
|
|
Source: OpenAI; Green Ash Partners
|
|
- We should probably highlight some recent cybersecurity developments here - the month of July saw several fairly worrying security lapses within the frontier labs' testing environment. The case that was broadly reported in the press involved an unreleased OpenAI model escaping its sandbox to connect to the internet and hacking Hugging Face to get the answers to an evaluation dataset
- Even more concerning, recent disclosures from the AI Security Institute (AISI) document examples of attempted social engineering of humans: "Our investigation found that... an AI agent took autonomous, unsanctioned action on the live internet, targeting real people and organisations. In total, we catalogued 19 such actions. Almost all of this behaviour (17 actions) came from a single model, Anthropic's Mythos 5, with 2 actions involving OpenAI's GPT-5.6-Sol... In the most serious case, an agent tried to insert malicious code into an open-source project. In an attempt to get the code approved, the agent engaged in social engineering — creating fake online identities and using them to pressure the project's maintainer to approve the code."
|
|
|
Disclosures of serious cybersecurity vulnerabilities in July were 5x higher than the previous monthly record pre-Mythos release, and +60% higher than June
|
|
|
CVEs from AWS, Apache, Apple, Cisco, Google, Linux, Microsoft, Mozilla, NVIDIA, Oracle, Red Hat, Adobe, IBM, Intel, AMD, Qualcomm, Samsung, SAP, VMware, GitHub, and OpenSSL. Reporting procedures, labeling, and cadence vary substantially between organizations.
Source: Epoch AI
|
|
|
AI Twitter has created FelonyBench, which tracks models escaping their internal evaluation sandboxes to hack websites online - all of these happened in the last few weeks
|
|
|
Chinese open source models are about seven months behind US closed source in cyber capability, suggesting we may see a Mythos-level open source model later this year
|
|
|
The U.S. trendline is composed of results from frontier U.S. models. A 400-point increase on the y-axis equates to a 10x increase in the odds of solving tasks. Error bars and shaded regions denote 95% CIs.
Source: AISI
|
|
- Clearly cybersecurity companies stand to benefit from customers paying up to defence against novel attack vectors, but there are broader implications relating to regulation, especially as open source model capabilities continue to improve and frontier labs get closer to cracking recursive self-improvement and automated AI research. What we have seen so far is just a fire drill for what is to come - as Demis Hassabis has said, we are in the foothills of the Singularity
|
|
- ERCOT's grid interconnection queue has grown to 474GW (90% of which is AI datacentre related). This amounts to 5.3x Texas' historical peak power demand of 91GW. Texas Governor Greg Abbot has ordered a comprehensive verification and audit of all datacentre projects advancing through the ERCOT interconnection process, effectively imposing an immediate moratorium on new grid approvals until the audit is fully completed
- This will encourage hyperscalers to BYOP (Bring Your Own Power) to new projects to avoid open-ended delays to construction timelines
|
|
|
The 427 GW of AI datacentre demand in the Texas interconnection queue is 18x larger than the 24 GW ERCOT expects to come online by 2031 in their long-term load forecast
|
|
|
|
Source: ERCOT; Green Ash Partners
|
|
- Circle Internet announced plans to launch the Arc blockchain on 16th September - a layer 1 blockchain network backed by some of the largest financial institutions and payment networks in the world, purpose-built for stablecoin finance, real-time global money movement, tokenised real-world assets, and institutional capital markets
- Adoption is somewhat in limbo, due to the ongoing back-and-forth over the CLARITY Act as it works through the Senate
|
|
|
|
