Mexico, Canada work hard to avoid Trump’s 25% tariffs - Chinese DeepSeek reality check for the AI stocks - ECB 25bps cut
WEEKLY TRENDS
WEEKLY TRENDS
- The third week of Q4 earnings releases were heavily disturbed by a Chinese AI bombshell (DeepSeek) that rocked the AI stocks early last week. Most of those stocks managed to recover by Friday close but the question remains whether cheaper AI solutions will impact existing AI revenues models in the long run (Nvidia ended the week at -3%, Microsoft at -2%)
- Trump implemented his new tariffs onto Mexico & Canada (25%) and China (+10%) starting Feb 1st
- The FED stayed put as expected (Powell said no need to be in a hurry to lower rates) and the ECB cut by 25bps its deposit benchmark rate, as expected, now standing at 2.75%. The December US PCE inflation was released in line with expectations at +2.6% vs 2.4% in Nov (Core at 2.8%) still well above the 2% FED target
- Earnings varied a lot (IBM ended the week at +15%, Meta at +10% while General Motors and LVMH ended the week at -9 and -4%)
MARKETS
Equities
Q4 earnings released (weekly stock performance):
Meta (+10%), IBM (+15%), ASML (+11%), Apple (+5%), Novartis (+5%), Tesla (+2%), Roche (+3%), Shell (+2%), Sanofi (+3%), SAP (+3%)
UPS (-14%), Comcast (-10%), STMicro (-10%), GM (-9%), Microsoft (-2%), LVMH (-4%), ABB (-4%), Exxon (-3%), Chevron (- 4%)
Analysts:
ThyssenKrupp (GS ’sell’ target €8.7); LVMH (HSBC ‘buy’ target €850 and MS ‘o/w’ at €820); Richemont (HSBC ‘buy’ target CHF 200); Hermès (HSBC ‘buy’ target €3000); SG (Barclays ‘o/w’ target €243)
Rates
US curve (2-10 years) steepening remained at 35bps. Bond yields slightly lowered across the board, HY corporate spread were somewhat stable
Commodities
Oil price down, on expected commercial tensions due to Trump’s tariffs, note Rotterdam TTF gas is at its highest level since Oct 2023 (+6% WoW)
Gold price up, on expected commercial tensions due to Trump’s tariffs, lower US bond yields and despite a stronger USD
China
Jan PMI Manufacturing broke the 50 mark at 49.1 (weakest since Aug) vs 50.1 in Dec, while non-manufacturing came out as 50.2 vs 52.2 in Dec
Cryptos
New incoming flows for the BlackRock IBIT (BTC Spot ETF), market awaits Trump’s action towards US reserves while Lagarde said No.
Under the watch
EU high yield corporate bonds (30% of outstanding notionals are due to be refinanced in 2025 and in 2026)
Nota Bene
Financials and Healthcare sectors ended the month well ahead of the Tech and Consumer Staples sectors; EU50 finished at +8% SP500 at 2.7%
New US Treasury Secretary Bessent has to tackle $10trn maturing debts in 2025 ($2.5trn last year and $3.5trn in 2026)
CALENDAR
Monetary Policy releases:
UK: MPC/BOE (6 Feb) likely to cut by 25bps from 4.75% at the moment
Macro data releases :
China Caixin Jan PMI (5 Feb); US Non-Farm Payroll job report for January (7 Feb)
Q4 Corporate earnings:
US: Palantir (3 Feb), Alphabet, Merck, PepsiCo, AMD, Pfizer (4), Amazon, Eli Lilly (6)
EU: Novo Nordisk, TotalEnergies, ARM Holdings (5), L’Oreal, Linde, Astra Zeneca (6), Vinci (7 Feb)
WHAT ANALYSTS SAY
UBS, 31 Jan 2025 - Information technology
Authors: Ulrike Hoffmann-Burchardi, Head CIO Global Equities
· Headlines around DeepSeek’s new R1 model have rattled financial markets from stocks to indices. In an attempt to separate facts from narratives, we share our interpretation of what happened and how to think about investing in equities going forward.
· We expect capex growth across hyperscalers to continue and semiconductors to be beneficiaries.
· We recommend keeping diversified exposure to the different layers of the AI value chain.
· We reiterate our Attractive view on China internet companies.
1. The Large Language Model (LLM) training cost curve is diminishing at an expected rate
Observations
· The history of technology is one of lower costs and higher volumes (one of the best examples is Moore’s law from 1965 that has so far accurately predicted the number of transistors on a microchip will approximately double every two years).
· For LLMs, the empirical curve for training cost approximately decreases 75% per year right now. DeepSeek version 3 (v3) is in line with the industry cost curve where we would expect a model equivalent to GPT-4o to currently be about 75% cheaper. Adjusting for a 2x performance differential to GPT-4o, Dario Amodei suggests that DeepSeek v3 should be close to 90% cheaper. Another way is to compare pricing of open source GPT-4o class models, like Llama 70B and 90B, to DeepSeek v3, which looks very similar.
· The decrease in cost is driven by new algorithmic techniques (including distillation where input/output tokens from an existing model are used to train the next model) and more powerful hardware.
· With R1, a well performing low cost open-source reasoning model has emerged that will drive more artificial intelligence (AI) adoption.
UBS CIO investment view
· Algorithmic improvements do not make hardware scaling redundant. We expect capex growth across hyperscalers to continue and semiconductors to be beneficiaries.
· Because of the technical complexity of AI, narratives can drive prices in the short term, lead to increased stock and equity index volatility, and present trading opportunities.
· We recommend keeping diversified exposure to the different layers of the AI value chain.
2. Hardware scaling laws remain intact Observations
· Hardware and algorithmic scaling are not mutually exclusive; on the path to general artificial intelligence (AGI), the two will likely move in lockstep.
· With re-enforcement learning (RL) as the key innovation in DeepSeek’s R1 model, a new hardware scaling law has emerged.
· At the same time, there is a legitimate question of the terminal value of frontier models when lagging edge models can be replicated quickly. It is too early to draw conclusions about a defensible moat in AGI.
UBS CIO Investment View
The AI enabling layer remains an attractive investment opportunity to capture the AI opportunity set.
3. China has played and will continue to play a strong role in AI innovation Observations
· Yi-34B-Chat by 01.AI, launched in November 2023, was ranked just behind GPT-4 and outperformed Llama on various benchmarks largely attributable to innovation in data curation and processing.
· Alibaba’s Qwen 2.5-Max model, released on 28 January 2025, is outperforming GPT-4o in preference based and general AI tasks.
· DeepSeek’s model v2, released in January 2022, contained innovative features :
• Mixture of Experts (MoE) variation of experts with more specific and broader knowledge,
• Multi-head Latent Attention (MLA), which reduces memory usage by compressing the key value store.
· R1 uses reinforcement learning on top of v3 for advanced reasoning, which we consider the most significant innovation of DeepSeek
UBS CIO Investment View
We re-iterate our attractive view on China internet companies, which have developed an impressive cadence of open-source models.
These allow for better customization and lower cost offerings.
4. Start-ups with a small number of employees and a large number of AI agents will become commercially relevant and disruptive to software incumbents
Observations
· The DeepSeek app became the top free app on the Apple App store within days of its R1 model release, surpassing ChatGPT.
· DeepSeek was established in 2023 and employs reportedly only about 200 people.
UBS CIO Investment View
Focus software investments into companies with proprietary data or customer moats.
5. Software will be created faster and released more rapidly.
Observations
· DeepSeek released R1 on 20 January 2025, less than a month after its v3 model.
· OpenAI released 12 shipments of code in December 2024 (‘Shipmas’), including o1 as the most performant reasoning model.
UBS CIO Investment View
Focus on physical investments over legacy software investments, as physical resources (compute + power) are needed to run AI and the physical layer cannot be disrupted at the same rate as software.
6. AI innovation will become increasingly unpredictable. AI is the first technology in human history that is self-improving.
Observations
· Llion Jones’ paper on LLM2 (LLM squared) leverages LLMs to propose, refine, and implement new algorithms that enhance their own training and performance.
· This self-referential approach should accelerate the discovery of new effective algorithms.
UBS CIO Investment View
Go long volatility through structured products and avoid investments in non-physical assets with no moat.
Equities
Q4 earnings released (weekly stock performance):
Meta (+10%), IBM (+15%), ASML (+11%), Apple (+5%), Novartis (+5%), Tesla (+2%), Roche (+3%), Shell (+2%), Sanofi (+3%), SAP (+3%)
UPS (-14%), Comcast (-10%), STMicro (-10%), GM (-9%), Microsoft (-2%), LVMH (-4%), ABB (-4%), Exxon (-3%), Chevron (- 4%)
Analysts:
ThyssenKrupp (GS ’sell’ target €8.7); LVMH (HSBC ‘buy’ target €850 and MS ‘o/w’ at €820); Richemont (HSBC ‘buy’ target CHF 200); Hermès (HSBC ‘buy’ target €3000); SG (Barclays ‘o/w’ target €243)
Rates
US curve (2-10 years) steepening remained at 35bps. Bond yields slightly lowered across the board, HY corporate spread were somewhat stable
Commodities
Oil price down, on expected commercial tensions due to Trump’s tariffs, note Rotterdam TTF gas is at its highest level since Oct 2023 (+6% WoW)
Gold price up, on expected commercial tensions due to Trump’s tariffs, lower US bond yields and despite a stronger USD
China
Jan PMI Manufacturing broke the 50 mark at 49.1 (weakest since Aug) vs 50.1 in Dec, while non-manufacturing came out as 50.2 vs 52.2 in Dec
Cryptos
New incoming flows for the BlackRock IBIT (BTC Spot ETF), market awaits Trump’s action towards US reserves while Lagarde said No.
Under the watch
EU high yield corporate bonds (30% of outstanding notionals are due to be refinanced in 2025 and in 2026)
Nota Bene
Financials and Healthcare sectors ended the month well ahead of the Tech and Consumer Staples sectors; EU50 finished at +8% SP500 at 2.7%
New US Treasury Secretary Bessent has to tackle $10trn maturing debts in 2025 ($2.5trn last year and $3.5trn in 2026)
CALENDAR
Monetary Policy releases:
UK: MPC/BOE (6 Feb) likely to cut by 25bps from 4.75% at the moment
Macro data releases :
China Caixin Jan PMI (5 Feb); US Non-Farm Payroll job report for January (7 Feb)
Q4 Corporate earnings:
US: Palantir (3 Feb), Alphabet, Merck, PepsiCo, AMD, Pfizer (4), Amazon, Eli Lilly (6)
EU: Novo Nordisk, TotalEnergies, ARM Holdings (5), L’Oreal, Linde, Astra Zeneca (6), Vinci (7 Feb)
WHAT ANALYSTS SAY
- UBS - Investment takeaways post DeepSeek’s R1 release
UBS, 31 Jan 2025 - Information technology
Authors: Ulrike Hoffmann-Burchardi, Head CIO Global Equities
· Headlines around DeepSeek’s new R1 model have rattled financial markets from stocks to indices. In an attempt to separate facts from narratives, we share our interpretation of what happened and how to think about investing in equities going forward.
· We expect capex growth across hyperscalers to continue and semiconductors to be beneficiaries.
· We recommend keeping diversified exposure to the different layers of the AI value chain.
· We reiterate our Attractive view on China internet companies.
1. The Large Language Model (LLM) training cost curve is diminishing at an expected rate
Observations
· The history of technology is one of lower costs and higher volumes (one of the best examples is Moore’s law from 1965 that has so far accurately predicted the number of transistors on a microchip will approximately double every two years).
· For LLMs, the empirical curve for training cost approximately decreases 75% per year right now. DeepSeek version 3 (v3) is in line with the industry cost curve where we would expect a model equivalent to GPT-4o to currently be about 75% cheaper. Adjusting for a 2x performance differential to GPT-4o, Dario Amodei suggests that DeepSeek v3 should be close to 90% cheaper. Another way is to compare pricing of open source GPT-4o class models, like Llama 70B and 90B, to DeepSeek v3, which looks very similar.
· The decrease in cost is driven by new algorithmic techniques (including distillation where input/output tokens from an existing model are used to train the next model) and more powerful hardware.
· With R1, a well performing low cost open-source reasoning model has emerged that will drive more artificial intelligence (AI) adoption.
UBS CIO investment view
· Algorithmic improvements do not make hardware scaling redundant. We expect capex growth across hyperscalers to continue and semiconductors to be beneficiaries.
· Because of the technical complexity of AI, narratives can drive prices in the short term, lead to increased stock and equity index volatility, and present trading opportunities.
· We recommend keeping diversified exposure to the different layers of the AI value chain.
2. Hardware scaling laws remain intact Observations
· Hardware and algorithmic scaling are not mutually exclusive; on the path to general artificial intelligence (AGI), the two will likely move in lockstep.
· With re-enforcement learning (RL) as the key innovation in DeepSeek’s R1 model, a new hardware scaling law has emerged.
· At the same time, there is a legitimate question of the terminal value of frontier models when lagging edge models can be replicated quickly. It is too early to draw conclusions about a defensible moat in AGI.
UBS CIO Investment View
The AI enabling layer remains an attractive investment opportunity to capture the AI opportunity set.
3. China has played and will continue to play a strong role in AI innovation Observations
· Yi-34B-Chat by 01.AI, launched in November 2023, was ranked just behind GPT-4 and outperformed Llama on various benchmarks largely attributable to innovation in data curation and processing.
· Alibaba’s Qwen 2.5-Max model, released on 28 January 2025, is outperforming GPT-4o in preference based and general AI tasks.
· DeepSeek’s model v2, released in January 2022, contained innovative features :
• Mixture of Experts (MoE) variation of experts with more specific and broader knowledge,
• Multi-head Latent Attention (MLA), which reduces memory usage by compressing the key value store.
· R1 uses reinforcement learning on top of v3 for advanced reasoning, which we consider the most significant innovation of DeepSeek
UBS CIO Investment View
We re-iterate our attractive view on China internet companies, which have developed an impressive cadence of open-source models.
These allow for better customization and lower cost offerings.
4. Start-ups with a small number of employees and a large number of AI agents will become commercially relevant and disruptive to software incumbents
Observations
· The DeepSeek app became the top free app on the Apple App store within days of its R1 model release, surpassing ChatGPT.
· DeepSeek was established in 2023 and employs reportedly only about 200 people.
UBS CIO Investment View
Focus software investments into companies with proprietary data or customer moats.
5. Software will be created faster and released more rapidly.
Observations
· DeepSeek released R1 on 20 January 2025, less than a month after its v3 model.
· OpenAI released 12 shipments of code in December 2024 (‘Shipmas’), including o1 as the most performant reasoning model.
UBS CIO Investment View
Focus on physical investments over legacy software investments, as physical resources (compute + power) are needed to run AI and the physical layer cannot be disrupted at the same rate as software.
6. AI innovation will become increasingly unpredictable. AI is the first technology in human history that is self-improving.
Observations
· Llion Jones’ paper on LLM2 (LLM squared) leverages LLMs to propose, refine, and implement new algorithms that enhance their own training and performance.
· This self-referential approach should accelerate the discovery of new effective algorithms.
UBS CIO Investment View
Go long volatility through structured products and avoid investments in non-physical assets with no moat.
Contacts
Main office
1-5, № 53, 12 Charents Str., Yerevan, 0025
+374 43 00-43-82
Broker
broker@unibankinvest.am
research@unibankinvest.am
Disclaimer
The information presented in the document contains a general overview of the products and services offered by Unibank OJSC (registered trademark – Unibank Invest, hereinafter referred to as the Bank).
The information is intended solely for the attention of the persons to whom it is addressed. Further dissemination of this information is allowed only with the prior consent of the Bank.
The information is only indicative, is not exhaustive and is provided solely for discussion purposes. The information should not be regarded as a public offer, request or invitation to purchase or sell any securities, financial instruments or services. The Bank reserves the right to make a final decision on the provision of these products and/or services to a specific customer, including refusing to provide products and/or services if such activities would be contrary to applicable law.
No guarantees in direct or indirect form, including those stipulated by law, are provided in connection with the specified information and materials. The information presented above cannot be considered as a recommendation for investing funds, as well as guarantees or promises of future profitability of investments.
