|
Green Ash Horizon Fund Monthly Factsheet - January 2025
|
|
The Horizon Fund’s USD IA shareclass rose +2.02% in January (GBP IA +2.16% and AUD IA +2.05%), versus +3.53% for the MSCI World (M1WO).
- We believe the DeepSeek moment has significantly cleared out crowding in AI related themes (including energy). This has coincided with clear signals from the AI research labs that we will see rapid acceleration in AI capabilities over the course of this year
- We don’t believe this is widely appreciated by the market – AI semi stocks have not had hyperscale capex raises incorporated into their forward earnings, nor are investors pricing for an inflection higher in cloud growth (topline growth numbers for Microsoft Azure and Google Cloud both missed last quarter due to capacity constraints)
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.
|
|
- DeepSeek was the hottest topic in AI in January, sparking a debate in AI circles about the extent of the US lead over China and a debate in financial circles about whether this spelled the end of the massive AI infrastructure capex cycle (we wrote about it here). Two weeks on, the upshot is:
- 1) The US labs still have a lead, both in terms of their latest model releases and internal models under development. Model iterations are going to accelerate across the board with the shift from pre-training to reinforcement learning and scaling test time compute
- 2) Hyperscaler capex plans for the year are unscathed. Since DeepSeek made the headlines Microsoft remain committed to their $80BN, Meta guided theirs to $62.5BN (+23% above forecasts), Google guided to $75BN (+25% above forecasts) and Amazon guided to $105BN (+24% above forecasts)
- On top of the hyperscaler capex, Softbank, Oracle and OpenAI announced the Stargate Project, with the ambitious goal to invest $500BN in AI infrastructure over the next four years. It isn't yet clear where this money will come from, though OpenAI is reportedly in talks to raise $40 billion at a $340 billion valuation
- We should emphasis that neither the recent hikes in hyperscaler capex, nor the Stargate Project are priced in to the forward estimates of AI semiconductor stocks - next year EPS estimates for NVIDIA/Broadcom are only up +9%/+2% on a three month basis and +18%/+3% on a six month basis
|
|
|
Hyperscaler capex plans for 2025 have risen +45% since last summer to $320BN, more than double the Pre-ChatGPT level of ~$150BN
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
|
Blackrock expect the AI infrastructure build out to last through the end of the decade, growing at a +20% CAGR over the next five years
|
|
|
|
Source: Blackrock; Green Ash Partners.
|
|
- So why are hyperscalers so committed to this unprecedented level of investment? As recently as last summer the market narrative had been one of AI progress slowing down, limited visibility on AI revenues/ROI and scaling laws yielding diminishing returns (we wrote about this here). On top of it all, DeepSeek showed frontier language models can be trained and served far more efficiently
- The answer is that AI progress is, in fact, accelerating, and pushing the scaling paradigm away from training and on to inference. Research from OpenAI's Noam Brown has shown that similar performance gains can be achieved by scaling 'thinking time' by 10x as can be achieved by scaling pre-training by 10x - a big deal given a model output at inference is 10 orders of magnitude cheaper in terms of compute than a pre-training run
- Hyperscalers have all reported being capacity constrained, with current infrastructure unable to serve AI demand
|
|
|
The Big 3 cloud providers all reported being capacity constrained, with AI demand expected to outstrip supply until the second half of the year
|
|
|
|
Source: Bloomberg; Green Ash Partners
|
|
- Reasoning models and multi-modal agents will increase inference demand by several orders of magnitude:
- An output from a standard language model might consume 4k tokens, while a reasoning model will use more like 30k (7.5x)
- Computer use agents like OpenAI's Operator or Google's Project Mariner need to sample screenshots (say once a second), use tokens for reasoning and planning, and then generate outputs to take action - this kind of use case could consume between 2.5k-25k tokens per second
- Physical agents will need an order of magnitude more again in order to match the 30-60 frames per second perceived by human vision in the real world
- Finally, scaling context windows is incredibly compute intensive - using standard transformer attention architectures, context length scales quadratically (so increasing context length by 8x would increase the model's training cost by 64x). Gemini's 2 million token context length can only handle two hours of video data at one frame per second
- Agents are still very early in their journey, but OpenAI's recent release of Deep Research for their $200/month tier has offered a glimpse of what will soon be possible. Sam Altman thinks Deep Research can perform "a single-digit percentage of all economically valuable tasks in the world"
|
|
OpenAI's Deep Research achieves double the performance of the leading reasoning models in Humanity's Last Exam, which is extremely hard - we would not be surprised if the average generalist human score was 0%
|
|
- Recent U.S. tariff moves have thrown a wrench into the aggressive digital ad strategies of major Chinese e‑commerce players. President Trump’s new 10% tariffs—and the suspension of the de minimis rule that previously exempted low‑value shipments—could force companies like Temu and Shein to either raise prices or reduce their billions‑dollar digital advertising budgets
|
|
|
The People’s Republic of China (PRC or China) has expanded its global e-commerce exports by more than tenfold over the past five years; PRC exports of low-value single packages expanded from $5.3 billion in 2018 to $66 billion in 2023
|
|
|
|
Source: via Stratechery
|
|
- It was interesting to see the level of correlation between AI semi stocks and electrification stocks in the DeepSeek share price reaction
|
|
|
US AI semiconductor and power infrastructure stocks are highly correlated
|
|
|
|
Source: Bloomberg; Green Ash Partners. Normalised.
|
|
|
|
|
|
|
|
|
