|
Green Ash Horizon Fund Monthly Factsheet - January 2026
|
|
The Horizon Fund’s USD IA shareclass rose +5.77% in January (GBP IA +5.70% and AUD IA +5.68%), versus +2.24% for the MSCI World (M1WO).
- The first month of 2026 has packed in enough events to fill an entire year. Despite this, it was a pretty good start to the year for risk assets, with the major equity indices in Europe and the US all finishing in the green, and most reaching record highs intra-month. Europe outperformed the US, led by European defence stocks and banks.
- Looking ahead, we are gaining increasing conviction that 2026 will be the year of accelerated AI diffusion into entreprise. With most AI semi and cloud hyperscaler stocks well off their highs, we see considerable scope for upside in the coming months, as earnings season progresses
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:
|
|
- Hyperscaler capex estimates rose into the end of last year, but the FY26 guides from the big three + Meta in the last couple of weeks have blown through even these elevated expectations. We wrote a piece called Is there a $600BN hole in GenAI? in July 2024 which included street capex forecasts for FY25e and FY26e; FY25 realised spend was +70% higher and the new guides for FY26e are nearly 3x higher than what was expected back then. And far from flattening off as most expected, the 2nd derivative - the rate of change - has actually accelerated every year for the last three years
- Many talk about how hard it is for humans to think in exponentials, but even as we nod in agreement, we generally fall back into linear thinking when forecasting trends. Someone who definitely does think in exponentials is Elon Musk. This week, he announced the merger of xAI with SpaceX, with plans to launch a million AI datacentre satellites into space, eventually adding "hundreds of GW" per year (the installed base of AI datacentres on Earth is currently 50-100GW in total). There are numerous bottlenecks that need to be addressed to achieve this (discussed in detail here):
- Chips - it takes five years to make a chip fab, so Musk plans to build 'Terafabs' to produce 100GW of chips per year (his volume target of millions of wafers is roughly equal to the entire world's current output of advanced logic)
- Energy - solar panels in space generate 5x the power of those on land. Musk's target of 100GW/year roughly a 1/6th of the world current installed base
- Mass to orbit - Musk's target of 10-30,000 Starship launches per year compares to 263 orbital launches in 2024 (134 of which were SpaceX). 10,000 launches/year is roughly one per hour
|
|
|
Hyperscaler capex is expected to exceed $650BN this year, using the mid-points of company guidance, and is accelerating
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- All of this is extremely bullish for AI semiconductors - 2026 demand is well anchored by the hyperscalers (and we haven't even mentioned the Neoclouds, Sovereign AI demand etc), and if the longer term vision is even partially realised, these companies will be in a secular boom for several years to come
- And yet, valuations do not reflect this. NVIDIA's NTM P/E is 3 turns lower than the market lows of October 2022, and about the same as the S&P 500, despite long-term EPS growth forecasts of +48% - about 5x higher than the S&P 500. EPS estimates for NVIDIA have risen +16% in the last three months, but have not reacted at all to the recent capex guides, which beat street forecasts by +28% in aggregate and added ~$142BN in incremental capex dollars (this compares to currently forecasted NVIDIA datacentre revenues of $307BN for this calendar year)
- We just use NVIDIA as an example for simplicity - the new hyperscaler capex guides will materially rebase forward estimates across all of the bellwether AI semiconductor companies as well as flowing down to dozens more across the datacentre infra ecosystem
|
|
|
NVIDIA's NTM P/E is 3 turns lower than the October 2022 market lows
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
The Street raised their calendar FY26e forecasts by +16% last quarter, but there have been no revisions to reflect the huge upside capex surprises in the last two weeks
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- All of this hyperscaler capex needs to translate to ROI. One of our highest convictions calls for this year is that hyperscaler revenue growth estimates are too low, and do not reflect the ~20GW of AI datacentre capacity that is coming online this year (the global datacentre capacity of AWS was just 2GW in 2019, and Amazon added 1GW in Q4 alone). While all of the big 3 clouds beat estimates, only Google decisively demonstrated this inflection in the most recent quarter
- There are some idiosyncrasies amongst the clouds - Microsoft commented that they allocated more compute internally last quarter to support the growth of their Copilot product - this implies further distancing from their OpenAI relationship, but this should balance out via more capacity for Oracle
- But probably the bigger factor is simply the lumpiness of capacity coming online. The first 1GW datacentre clusters will begin operations this year, with each one representing single-digit billions in revenues
- Every AI lab and cloud operator report capacity constraints as the main bottleneck to revenue growth
|
|
|
Of the Big 3, only Google showed the growth inflection we have been calling for in hyperscale cloud revenues this year
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
OpenAI's CFO published a chart showing the smooth relationship between datacentre capacity and revenues - this should apply across the board, so long as demand exceeds capacity
|
|
|
|
Source: OpenAI
|
|
|
Anthropic raised their three year revenue projections, which now anticipate them overtaking OpenAI some time in 2027, and are about +50% higher by YE27 than their 'optimistic' scenario' forecasted last summer
|
|
|
|
Source: The Information
|
|
- Continued AI progress on the training side and adoption on the inference sides are the lynchpins of this pent up demand. For us at least, this trend is not under debate. We tried convey a sense of this in our Applied AI piece last month, and since then there has been further acceleration, with the release of new frontier models from OpenAI (GPT 5.3) and Anthropic (Opus 4.6) just two months after the previous versions advanced the state-of the-art and enabled breakthrough agentic AI products like Claude Code and OpenClaw
|
|
|
The METR time horizon benchmark has become the chart of the singularity (it took METR two months to test and publish results for GPT-5.2 and Opus 4.5, and just hours later, Opus 4.6 and GPT 5.3 were released)
|
|
|
|
Source: METR
|
|
|
Opus 4.6 soars above the SotA models of two months ago in Vending-Bench (the testers noted Opus 4.6 deployed tactics that "range from impressive to concerning: Colluding on prices, exploiting desperation, and lying to suppliers and customers"
|
|
|
|
Source: Andon Labs
|
|
- Claude Code has started to hit the mainstream and cause a stir in the broader non-coding areas of knowledge work. This has been aided by the release of Claude CoWork - ultimately the same thing, but with a nicer visual interface at the front end (built in ten days by Claude Code)
- Anthropic also released a set of plugins aiming at various areas of knowledge work like legal, customer support, marketing etc. There is limited evidence that anyone is actually using these (they have 5k stars on Github, versus 170k for OpenClaw which came out around the same time), but they have had a cataclysmic effect on investor sentiment on software stocks
|
|
|
It took the press 5 days to notice Anthropic's knowledge work plugins, but when they did, they enthusiastically ran with it - this is similar to the 'DeepSeek Moment' last year, when it took five weeks for the n DeepSeek v3 research paper to become market-moving news, eventually becoming not market-moving after all
|
|
|
A representative view point on Claude Code/Clawdbot from finance professional outside of the AI/tech info bubble
|
|
- Most of the exposure in Horizon is hardware and infrastructure focussed, so our main interest is AI adoption beyond coding and what this might mean for inference demand. We think 2026 is the year this will truly take off, for reasons articulated well by Eric Jang in his As Rocks May Think essay (which is a bit technical, but worth a read)
|
|
|
This essay from a former DeepMind researcher well articulates our view that inference demand is going to rise by orders of magnitude as swathes of entreprise workflows are reshaped into massively parallelised agentic loops (with human oversight)
|
|
- The virality of OpenClaw - an open-source, autonomous AI assistant developed by Peter Steinberger and released a couple of weeks ago - highlights another major source of inference demand. All the labs are making a major push into productising LLMs as personal assistants. Some inference workloads will run at the edge, on PCs and smartphones, but for complex tasks these agents will need frontier models at the >1 trillion parameter scale. We have spoken to Openclaw users of Claude models consuming well over 500 million tokens of Opus per week (c. $450/day in API costs)
|
|
|
This post from an Economist journalist highlights the voracious token consumption of constantly running personal agents
|
|
|
|
Source: X
|
|
- 2026 may also be the year AI makes a major impact on biotech. This has been long overdue (we have had positions in small AI- drug discovery companies for at least five years)
- OpenAI partnered with Gingko Bioworks to connect GPT‑5 to a cloud laboratory - an automated wet lab run remotely through software, where robots execute experiments and return data—and used that lab-in-the-loop setup to optimise cell-free protein synthesis (CFPS). After three rounds of experimentation, GPT-5 established a new state-of-the-art, lowering protein production costs by -40% and reducing reagent costs by -57%
- Separately, there are rumours that AI labs may take stakes in AI-first drug discovery companies, granting them access to frontier models to accelerate research - potentially a major catalyst for a long-beleaguered sector. A cynical interpretation is that they feel pressure to demonstrate grand societal benefits at the 'cure cancer' scale to offset the imminent labour force disruption they anticipate
- Goldman Sachs told CNBC that they have been working with Anthropic over the last six months to automate some back office and finance roles - we think legacy finance has a shot at being one of the earliest and most material beneficiaries of AI productivity gains, if their leadership is willing to take the leap
- JPMorgan released a financial model for Waymo, forecasting a 1% share of total US taxi rides in 2026, and 6% by 2030. This would imply about 30k vehicles in operating, undertaking 5.3 million rides per week and generating about $5.8 billion a year in booking revenue (each vehicle earning just over $500 per day)
|
|
|
Waymo is scaling rapidly, albeit these are still small numbers given the very large TAM of the US ride-sharing market
|
|
|
|
Source: JPMorgan Research; Green Ash Partners
|
|
- Rising electricity prices are being associated with the datacentre build out and is becoming an issue with voters. President Trump has sought to get ahead of this tweeting "I never want Americans to pay higher Electricity bills because of Data Centers, therefore, my Administration is working with major American Technology Companies to secure their commitment to the American People, and we will have much to announce in the coming weeks... Microsoft will make major changes beginning this week to ensure that Americans don’t 'pick up the tab' for their POWER consumption, in the form of paying higher Utility bills"
- There have been several studies, such as one from Lawrence Berkeley National Laboratory at the end of last year, and one this year from Charles River Associates that push back on the narrative that rising electricity prices are being driven by datacentres. The main arguments are that:
- The US national price index is heavily influenced by regions like California which suffer from wildfires and other regions with high volatility due to natural gas prices - energy prices have been fairly stable across most of the US
- Sharing generation capacity with large customers has economies of scale that help reduce the consumers' share of transmission and distribution costs, which are an increasingly proportion of utility capex
|
|
|
Hyperscalers are mounting PR campaigns to avoid becoming a political target in the regions they are building out datacentre infrastructure
|
|
|
|
Source: Microsoft
|
|
- Polymarket adopted Circle Internet's USDC stablecoin for settlement (joining Kalshi). Polymarket and Kalshi now settle billions of dollars of bets every month, and continue to grow rapidly. These betting platforms have become an important alternative data source in financial markets, and are now integrated with Bloomberg
|
|
|
Market participants now often look to prediction markets when gauging political and geopiltical risks
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
|
