Key points
- DeepSeek’s R1 AI model triggered market volatility but hyperscalers’1 continued investments highlight long-term AI growth potential
- The Trust shifted modestly from AI infrastructure leaders to AI beneficiaries who may enjoy tailwinds from lower inference costs
- AI’s rapid cost reductions and innovation are expected to boost adoption across multiple industries, reshaping markets beyond technology through 2025
Market review
Global equity markets delivered strong returns in January, with the MSCI All Country World Net Total Return Index rising +3.9% and the S&P 500 Index up +3.3% while the DJ Euro Stoxx 600 rose 7.7% (all returns in sterling terms).
Headline numbers masked significant volatility. AI stocks saw considerable intra-month volatility following the release of the DeepSeek R1 AI model. Meanwhile, market concerns around inflation caused US 10yr Treasury yields to jump up to 4.8% before subsequently retracing to end the month broadly unchanged. In addition, the new Trump administration announced tariffs on Canada, Mexico and China causing additional uncertainty.
The US economy remains a bright spot, with real Gross Domestic Product (GDP) rising +2.3% in Q4 and inflation slowly trending downwards. This was confirmed by macroeconomic data including strong US jobs numbers, while ISM data surged to its highest in nearly two years. Consequently, US yields initially rose but then fell into the end of the month as US core CPI (Consumer Price Index) inflation slowed to +0.2% in December.
The Q4 GDP report also flagged consumer/household spending growing +4.2% (from +3.7% in Q3), the strongest rise since Q2 2023 and the first back-to-back quarterly increase above 3% since late 2021. US consumer strength was also apparent in results from Visa and Mastercard with affluent consumers appearing to benefit from the ‘wealth effect’ with a robust jobs market and continued wage growth supporting the majority.
The US Federal Reserve (Fed) unsurprisingly left its target rate for the Fed funds (overnight interest) rate unchanged at 4.25-4.5% at the end of the month. The accompanying press release made clear the Fed would not be in a hurry to lower interest rates further (having already cut by 100basis points2 (bps), but still plans to bring them down gradually when conditions allow. Supporting this decision, the Fed’s preferred measure of inflation showed core Personal Consumption Expenditure (PCE) price index at +2.8% and 0.2% month on month (m/m), in line with expectations but uncomfortably above the target 2% level.
President Trump’s ‘America First’ policy agenda has so far been broadly supportive of US equities (stronger economy/dollar). The agenda skews pro-business/growth (lower taxes; deregulation) and focused on incentivising domestic manufacturing. Another element is ensuring the US sustains AI/technology supremacy and January saw the announcement of the Stargate Project, a new company planning to invest $500bn over the next four years to build AI compute infrastructure for OpenAI in the US. The project is backed by Japan’s Softbank, OpenAI, Oracle and MGX with the aim of securing US leadership in artificial intelligence.
Tariffs also clearly remain a key pillar of Trump’s strategy and on the last day of the month he confirmed across-the-board tariffs would be imposed on Mexico (25%), Canada (25%) and China (10%) within days. The dollar strengthened and markets sold off intra-day on concerns that higher import prices might challenge both the strength of the global economy, particularly ex-US, and the trajectory of falling inflation (limiting the Fed’s scope to cut rates). However, Trump subsequently agreed to a one-month extension for Canada and Mexico allowing negotiations to continue, following apparent concessions from both countries to try to better stem the flow of illegal immigrants and drugs into the US.
DeepSeek and AI developments
The big news in technology during the month was Chinese AI developer DeepSeek’s latest R1 model release which sent shockwaves through the AI industry. The severity of the correction in AI infrastructure-exposed stocks reflected the gains DeepSeek has made, as well as a lack of ‘pushback’ by executives at OpenAI, NVIDIA, Microsoft and Meta Platforms (Meta) to the algorithmic advancements DeepSeek disclosed. In addition, the fact that a model near parity with GPT-4o came from an unfamiliar Chinese entity and was released ‘open source’ brought additional challenges to the prevailing narrative that Western AI labs and the hyperscalers were leading the AI race. The magnitude of profit-taking was also likely exacerbated by strong year-to-date returns in AI infrastructure stocks, the earlier Stargate announcement and extended investor positioning.
Fortunately, stocks rebounded as deeper evaluation suggested the impact may not be as stark as first appreciated. There is still considerable scepticism related to training cost claims and details such as the hardware used for training and distillation model sources (likely OpenAI’s GPT-4/Meta’s Llama) remain undisclosed. Research provider SemiAnalysis has suggested that the $6m DeepSeek claims is a “very narrow portion of the total cost” and should be seen in the context of >$500m in overall GPU (graphics processing unit) investments, $1.3bn in total server capital expenditure (capex) and >$700m in operational costs.
DeepSeek appears to have brought significant innovations to both pre-training and inference, especially in terms of the efficiency of NVIDIA GPU and memory usage apparently allowing them to achieve significant performance gains.
Despite these uncertainties, DeepSeek appears to have brought significant innovations to both pre-training and inference3, especially in terms of the efficiency of NVIDIA GPU and memory usage apparently allowing them to achieve significant performance gains. DeepSeek also used a reinforcement learning-only approach for post-training (i.e. no human feedback or supervised finetuning).
Where the advancements and performance claims seem strongest are on the inference side and DeepSeek R1 (hosted in the US/EU) is now available on Microsoft, Amazon and Perplexity among others. Meta CEO Mark Zuckerberg cited “a number of novel things they did” including “advances that we hope to implement in our own systems”. Satya Nadella, CEO at Microsoft, added: “DeepSeek had some real innovations” and “as AI becomes more efficient and accessible…we see exponentially more demand”. Even Sam Altman, CEO of Open AI, acknowledged: “It’s a very good model” before adding: “We will produce better models, but we maintain less of a lead than we did in previous years”.
Elsewhere, at least for now, training of the largest frontier models will continue at pace with hyperscaler capex apparently unaffected as these companies race towards AGI (artificial general intelligence). Zuckerberg pre-announced Meta’s capex plans for 2025 ($60-65bn), a forecast well ahead of expectations, fully aware of the R1 model. The same week, OpenAI announced the $500bn Stargate Project, with $100bn to be deployed “immediately”, which implicitly assumes AI progress will remain dependent on more capital and compute capability. Alphabet also raised its full-year capex expectations to $75bn, +43% year on year (y/y) and well above the $58bn expected. Microsoft confirmed plans to spend c$80bn, largely on cloud and AI-enabled data centres and China’s own five-year AI plan calls for $137bn in capex. In aggregate, Bank of America estimates that aggregate capex estimates for Microsoft, Amazon, Alphabet and Meta were revised up by $50bn (+17%) for this year following the DeepSeek news.
Fund performance and activity
The technology sector lagged the broader market in January, the Dow Jones Global Technology Net Total Return Index (W1TECN) returning +1.5%. The Russell 2000 Technology Index (small cap) returned +0.4%, while the Russell 1000 Technology Index (large cap) was flat (all figures in sterling terms).
At a Trust level, we made some adjustments to positioning, but we consider them to be evolutionary and more concerned with risk management, tied to areas where uncertainties remain.
While it is far too early to know if DeepSeek’s innovations pose a meaningful threat to NVIDIA’s dominance, the outlook is more complex and the potential for tighter export controls has increased. To that end, the Trust has modestly rotated away from AI enablers/infrastructure and towards AI beneficiaries as lower inference costs offer both a new path for AI scaling (test time compute) and a lower cost input when embedding AI in applications. We will be watching closely to see the implications of dramatically lower inference costs as we (and others) expect AI volumes to expand significantly as the use of AI becomes more efficient, as per Jevons paradox4. If so, any short-term disruption to demand tied to export controls or greater capital efficiency could be short-lived and drive even greater demand for AI infrastructure given earlier HBM (high bandwidth memory) supply constraints.
Technology review
Technology stocks were impacted by a flurry of announcements from the Trump administration including the Stargate Project and policy changes such as revoking Biden-era policies that the new administration says “hampered the private sector’s ability to innovate in AI by imposing government control over AI development and deployment" and called for the development of an AI Action Plan to sustain and enhance America’s AI dominance.
The Philadelphia Semiconductor Index (SOX) gained +1.3%, weighed by NVIDIA concerns and weak non-AI-related semiconductor demand (autos; industrials). The NASDAQ Internet and iShares Software indices gained +7.1% and +3.3% respectively, as investors rotated from the semiconductor sector and AI infrastructure enablers into internet and software, considered more agnostic to tariffs and potential beneficiaries of DeepSeek developments.
The market was also impacted by the start of Q4 earnings season. In the internet sector, Meta reported strong results, with advertising revenue +21% y/y in constant currency5 (cc) terms. Four million advertisers are now using at least one generative AI (GenAI) creative tool, up from one million just six months ago. Next-quarter guidance was a touch light of expectations, and no 2025 guidance was issued. However, the company announced that Meta AI now has 700 million monthly active users (MAUs; from 500-600 million in December and 400 million in September). Zuckerberg also sounded excited about the upcoming Llama 4 models, arguing Llama 3 was to prove that open source can compete, but for Llama 4 the ambition is to lead it, with natively multimodal, agentic6 capabilities.
Netflix delivered another robust quarter with a record 19 million net subscriber additions, fuelled by the company’s strong content slate. Advertising-based video on demand (AVOD) accounted for 55% of the signups (10 million), with AVOD membership growing +30% quarter on quarter (q/q). Next-quarter guidance was underwhelming due to advertising seasonality and the timing of price increases, but FY25 (fiscal year 2025) revenue and operating margin guidance were raised.
Alphabet (underweight) continued to benefit from a succession of product releases, including Willow chips (for quantum computing) and the new Gemini 2.0 as well as video and image-generation models Veo 2 and Imagen 3.
Within software, Microsoft reported solid headline results with AI revenue at a >$13bn run rate (+175% y/y), but Azure disappointed due to non-AI services (Azure AI was better than expected despite capacity constraints). Next quarter, Azure is expected to grow +31-32% y/y cc, with an increasing contribution from AI, but this was modestly below expectations. Microsoft 365 Copilot appears to be gaining momentum, with the number of people using Copilot daily doubling q/q and usage increasing +60% q/q.
ServiceNow reported a mixed 4Q24 with upside to estimates lower than recent quarters, while FY25 guidance for subscription revenue growth of +19.5-20% (cc) was slightly below consensus. Management factored in extra prudence given the potential impact of the new US presidency on federal business in the first half of the year, as well as the shift towards a hybrid subscription/consumption pricing model. Elsewhere in software, SAP reported a solid quarter and raised 2025 guidance based on strong cloud backlog growth. Cloudflare continued to benefit from positive channel checks and investor excitement around edge AI.
In the semiconductor subsector, TSMC reported solid results, with strength in high performance computing (AI) more than offsetting weakness in automotives, the Internet of Things and consumer electronics. Guidance for mid-20% revenue growth in 2025 was better than expected and management gave a buoyant long-term growth outlook for a five-year compound annual growth rate (CAGR) of 20%, with AI growth of mid-40%, a positive read across to other AI semiconductor companies including NVIDIA and Broadcom. 2025 capex is expected to be $38-42bn, 6% above market forecasts, supporting semiconductor production equipment stocks. Celestica, an electronics manufacturing services (EMS) company, reported strong results and raised 2025 guidance. It is a leading provider of networking switches for AI applications and is seeing growing demand for end-to-end full-rack solutions which include custom ASICs (application-specific integrated circuit), networking and liquid cooling. The company disclosed that it has won two new large programs which should drive growth in 2026 and 2027.
Elsewhere,Tesla had a softer 4Q24, with deliveries only +2% y/y, while revenue and gross margins were impacted by lower average selling prices due to pricing, attractive financing options and mix. The company expects vehicle deliveries to grow in 2025, benefiting from a refresh of its Model Y and the launch of a lower cost model in the first half of the year. However, Elon Musk framed 2025 as a year to continue to lay groundwork for an “epic 2026, and a ridiculous 2027 and 2028” driven by improvements in the company’s real-world AI capabilities and the rapid scale-up of unsupervised full self-driving (starting with a ride-hailing service in Texas in June) and Optimus humanoid robots.
Apple was weak during the month due to soft iPhone channel checks, but reported an in-line quarter with Services, Macs and iPads offsetting worse-than-expected iPhone sales, particularly in China where there was an inventory drawdown. Next-quarter guidance for low to mid-single-digit growth was better than feared.The stock rallied at the end of the month, as the company is perceived as a potential beneficiary of edge AI which received a potential boost from DeepSeek.
Outlook
While DeepSeek captured the AI headlines during the month, it may have limited impact on the road to AGI given the current lack of native multimodal capabilities and a short context window. In addition, DeepSeek’s disruptive pricing compared to existing models may be better understood as ‘just’ the acceleration of an existing path of rapidly declining inference costs. For instance, the cost of ChatGPT-4-level intelligence has fallen by 1,000x in the past 18 months.
What is clear is the pace of innovation is accelerating and dramatic progress in early 2025 should make this increasingly apparent.
Beyond DeepSeek, the pace of AI innovation remains furious as evidenced by a series of other model announcements including OpenAI’s o1 and o3 reasoning models, Google Gemini 1.5 Deep Research and this week’s announcements of Gemini Advanced 2.0 models, as well as Open AI’s Operator (a computer-using agent (CUA)) and its own Deep Research offering. New AI models trained on much larger 100,000 GPU clusters are likely to be forthcoming from Meta (Llama) and xAI (Grok) in coming months. Open AI and other US innovators are also likely to be accelerating their next model rollouts in response to DeepSeek.
There is considerable hardware innovation ahead too: NVIDIA’s Blackwell architecture will be rolled out more widely and the company is expected to announce details of the second-generation B300 at its GPU Technology Conference in March before the new Rubin platform expected in 2026. We are confident that 2025 will prove another extraordinary year for AI innovation, even as it becomes more complex.
Meanwhile, the US economy is in good health as consumer spending remains robust, SMB (small and medium-sized business) confidence is strong and the deregulatory/pro-business agenda should help support economic activity. Inflation is also (slowly) trending lower – we are still optimistic that tariffs and political developments will not derail an otherwise supportive backdrop and our base case remains a supportive backdrop for US equities and for continued AI investment.
Overall, we believe the collapsing cost of intelligence is a crucial part of the AI adoption story which can democratise AI and drive a corresponding explosion in volumes as the more efficient use of AI allows it to be put to much greater use. Indeed, Microsoft Azure’s AI revenues (the vast majority of which is inference-based) have grown 11x since March 2023 even as inference costs (in dollars per million tokens) have declined 20x pre-DeepSeek.
However, volatility is a persistent feature of new technology cycles, when the innovation curve is at its steepest and both the pace of progress and scale of the opportunity are hard to define. For example, there were seven >15% corrections between 1995-1998 (i.e. before the 1999 ‘melt up’), while the NASDAQ gained 354% over the period (in dollar terms). The DeepSeek episode is an important reminder of this as well as the benefits of diversification (AI cannot only be about NVIDIA and related suppliers) and highlights the increasing complexity of the AI story.
Thecase for market broadening appears stronger too, as more companies appear able to access leading-edge AI at ever lower inference costs. It is notable that January saw the equal-weighted S&P 500 enjoy its best month versus the market-cap-weighted S&P 500 since July last year (when the cooler CPI started the rotation to small-cap stocks). The equal-weighted NASDAQ 100 also significantly outperformed. This should provide a tailwind to active management while indices remain dominated by the largest constituents.
The Mag7 group dominated returns in 2024 (+67.3%), benefitting from positive earnings revisions and excitement about AI, accounting for almost 60% of the S&P 500’s return. The Mag7’s fortunes have diverged in 2025 and there have been early signs of market broadening with the S&P 500 ex-Mag7 up +5.3% in January, well ahead of the S&P 500’s +3.3% return. Mag7 was up just +3% in January, with only three of the seven (Meta; Amazon; Alphabet) showing positive returns year to date to 6 February, and lagging the equal-weighted NASDAQ’s +5.4%. It is too early to say whether this trend will continue, but if history is a guide leadership changes when new cycles take hold and the Trust is well positioned to take advantage of any AI broadening.
What is clear is the pace of innovation is accelerating and dramatic progress in early 2025 should make this increasingly apparent. We expect this year will prove to be when the impact of AI becomes unavoidable and will likely shape investor returns well beyond the technology sector.
1. The largest cloud service providers (AWS; Microsoft Azure; Google Cloud; Meta Platforms; Apple; TikTok)
2. A basis point is a common unit of measure for interest rates and other percentages in finance. One basis point equals 0.01%
3. An idea or conclusion drawn from evidence and reasoning
4. Economist William Jevons found that when technological improvements increase the efficiency of resource use, instead of reducing overall consumption of that resource, it leads to increased consumption
5. Constant currency reporting is an accounting technique used by companies to present financials year-over-year for comparative purposes without the effects of currency movements
6. Agentic AI refers to an advanced AI system that autonomously takes actions, adapts in real-time, and, solves multi-step problems based on context and objectives.
