deep Dive into Global Financial Markets 2025 and the Great Game of 2026: Huge Transitions and New Equilibria
1. technological deflation and geopolitical inflation collide
in 2025, global financial markets were shaken by a complex set of dynamics that are unprecedented in history. on the one hand, artificial intelligence (AI) technology has sparked an "efficiency revolution" that is dramatically lowering the cost of production, exerting technical deflationary pressures, while on the other, protectionism and supply chain fragmentation have stimulated structural inflation, driving up costs. The collision of these two massive forces has led to extreme volatility in asset prices and presented investors with an unprecedentedly challenging market environment.
the "cost-effectiveness shock" unleashed by Chinese AI startup DeepSeek at the beginning of the year fundamentally questioned Silicon Valley's capital-intensive growth model and led to a reassessment of tech stock valuations. more than just one company's success, it signaled an inflection point for the AI industry, moving from the era of "training" to the era of "inference" and "application".
at the same time, the 'Reciprocal Tariffs' policy of Donald Trump's second term administration, originating from Washington, D.C., has shaken up the global trade order. Since the day of the tariff declaration, dubbed "Liberation Day," global supply chains have been reorganized around security and resilience rather than efficiency, with differential impacts on companies' margin structures and economic growth rates across countries.
in the midst of this turmoil, the South Korean stock market has been buoyed by a semiconductor supercycle and a robust government valuation program, ushering in the KOSPI 4,000 era for the first time ever. nvidia reached an all-time high of $5 trillion in market capitalization, marking the pinnacle of AI infrastructure investment.
this in-depth report analyzes the key themes that shaped the market in 2025 through micro data and macro insights, and looks ahead to the investment landscape in 2026, including the structural changes in the AI industry brought about by the Deep Shock, the ripple effects of Trump tariffs on the real economy, the path of the Federal Reserve's monetary policy, and the megatrends that will drive 2026, including HBM4 and humanoid robots as the next battleground for semiconductor technology. we are at the entrance of a new era, where the success equations of the past no longer apply.
2. the DeepSeek Shock: A Structural Shift in the AI Paradigm
2.1. Cost-effectiveness revolution: David overcomes Goliath's moat
in January 2025, the DeepSeek-V3 and R1 models unveiled by Chinese AI startup DeepSeek shocked the global AI market with more than just technological advances: the key was not performance, but overwhelming "cost-effectiveness". while demonstrating inference and coding capabilities on par with OpenAI's latest model, O1, DeepSeek succeeded in dramatically reducing training and operational costs.
according to the data, DeepSeek-R1's input token cost is $0.55 per million tokens, which is only about 3.7% of the $15.00 cost of OpenAI's o1. the output token cost is also $2.19, compared to O1's $60.00, which is about 96% cheaper. this was not just price disruption, but a fundamental challenge to the traditional "scaling law" of "bigger and more expensive is better.
[Table 1] DeepSeek-R1 vs. OpenAI o1 cost-effectiveness comparison
categoryDeepSeek-R1OpenAI o1cost Savings input cost (per 1M tokens) 0.55 15.00 approximately 96.3% reduction output cost (per 1M tokens) 2.19 $60.00 approx. 96.4% savings operating Cost Ratio Approximately 5% of OpenAI 100% (baseline) -
source: Datacamp and compilation of key technical reports
deepSeek says that the final training cost of its DeepSeek-V3 model was just $5.6 million, which is the result of training 14.8 trillion tokens in two months using 2,048 Nvidia H800 GPUs. at a time when Google, Meta, and OpenAI are racing to build clusters of tens of thousands of GPUs with hundreds of billions to trillions of dollars in funding, DeepSeek's announcement opens up the possibility that even undercapitalized startups and research labs can develop cutting-edge AI models.
this dramatic cost reduction suggested the possibility of reducing reliance on expensive Nvidia GPUs, which immediately hurt Nvidia's stock price. the market dubbed it the "dipshock," and Nvidia suffered its largest single-day loss in history, wiping out nearly $600 billion in market capitalization in a short period of time.
2.2. Technical Background: MoE Architecture and Secondary Lossless Load Balancing
the key to DeepSeek's success was not a hardware onslaught, but rather a sophisticated innovation in algorithms. DeepSeek-V3 has a total of 671 billion parameters, but only 37 billion (about 5.5%) are activated at token generation. this is a result of our advanced Mixture-of-Experts (MoE) architecture.
2.2.1. DeepSeek MoE Uniqueness: Auxiliary-Loss-Free Strategy
during the routing process of selecting an "expert" neural network to process the input data, traditional MoE models added an "auxiliary loss" to the objective function to avoid overloading a particular expert. While this aids in load balancing, it had the side effect of sacrificing some performance by interfering with the main objective of the model (accurate next token prediction).
DeepSeek-V3 adopts a groundbreaking strategy of removing this secondary loss and instead introducing a Bias term to control load balancing. this resulted in maximizing the model's performance while maintaining the stability of the learning process. We also set a Multi-token Prediction goal to train the model to predict multiple future tokens in a single computation, further increasing learning efficiency.
2.2.2. FP8 Precision Training and the Introduction of MLA
deepSeek is the first open-source LLM to apply FP8 (8-bit floating-point) precision from the pre-training stage. while lower precision typically results in faster computations at the expense of accuracy, DeepSeek minimizes the loss of accuracy through fine-grained quantization techniques. We also introduced a Multi-head Latent Attention (MLA) architecture to reduce the amount of Key-Value Cache (KV Cache) memory used during inference, resulting in higher performance and lower cost for long context processing. this technical approach is how we were able to reliably train 14.8 trillion tokens using only 2,048 NVIDIA H800 GPUs.
2.3. Market reaction and the transition to the "Inference" era
the rise of DeepSeek was a catalyst for the AI market to shift its center of gravity from training to inference. up until 2024, the market was focused on who could train the biggest models with the most data, but after the DeepSeek shock in 2025, the focus shifted to "who can run their models more efficiently to provide real services."
investors began to worry that the moat that tech giants had built through astronomical Capex might not be as deep as they thought. with the open-source camp outperforming closed-source models by more than 90%, the prospect of commoditization of AI models themselves is gaining momentum. this suggests that by 2026, the competitive landscape of the AI market will shift from building hardware infrastructure to optimizing software and providing agent services.
in particular, the expansion of the inference market will accelerate the growth of edge AI and on-device AI. highly efficient models such as DeepSeek will enable high-performance AI to run on limited hardware environments such as smartphones and robots, which will be a key driver for the commercialization of humanoid robots and autonomous driving technologies in 2026.
3. the Trump 2.0 era and the rise of neo-traditionalism
3.1. Reciprocal Tariffs and "Liberation Day"
on April 2, 2025, President Donald Trump officially declared "reciprocal tariffs" and dubbed it "Liberation Day. this is a policy that states that the United States will impose tariffs on U.S. products that are mathematically equal to the rate of tariffs imposed by other countries on U.S. products in order to address the chronic U.S. trade deficit and to counter unfair trade practices by other countries.
3.1.1. Details and legal basis for tariff policy
the Office of the United States Trade Representative (USTR) has imposed a broad range of tariffs based on Section 301 of the Trade Act and the International Emergency Economic Powers Act (IEEPA). the Trump administration invoked IEEPA's powerful authority by declaring a national emergency, stating that "the large trade deficits of the United States pose an unusual and extraordinary threat to national security."
major country tariffs: Tariffs ranging from 35% to 50% were imposed on key trading partners, including Brazil and Canada, and threats of tariffs as high as 30% were made against Mexico and the EU.
tariffs onpopular countries: China was subjected to additional retaliatory tariffs on top of its existing high tariffs, resulting in a stack of IEEPA tariffs and Section 301 tariffs, resulting in a number of items with effective tariff rates exceeding 50-60%.
3.1.2. Economic impact: Lower GDP and higher prices
according to an analysis by the World Bank and the Tax Foundation, the weighted average tariff rate in the United States is estimated to soar from 1.5% in 2022 to 15.8% by the end of 2025. these tariff barriers have driven up import prices, squeezing profit margins for businesses and eroding the purchasing power of consumers.
analysts estimate that these tariff policies will reduce U.S. GDP by about 0.7% in the long run. if companies pass on the cost of tariffs to consumers, demand will contract, and if companies absorb the costs, profits will decline, leading to a vicious cycle of falling stock prices. in fact, in April and June 2025, the S&P 500 experienced its largest daily and weekly volatility of the year, coinciding with tariff policy announcements.
3.2. Reshaping global supply chains and the "Silicon Heartland"
the Trump administration's tariff policies had the effect of forcing companies to relocate production to the U.S. mainland, which accelerated the rise of the "Silicon Heartland" centered in the U.S. Midwest. companies have increased the number of automation factories in the U.S. instead of Mexico or Asia to hedge against tariff risk, which has led to a surge in demand for robotics and industrial automation equipment.
countries like Vietnam have seen their exports to the U.S. soar as a result of being used as a transit point for Chinese goods, prompting the U.S. to escalate the trade war by imposing high tariffs of 40% on goods diverted through Vietnam. with Vietnam's exports to the US accounting for more than 20% of its GDP, these tariffs could severely impact the emerging economy.
3.3. Fears of reigniting inflation and the Fed's dilemma
higher tariffs are inevitably accompanied by higher import prices. in the second half of 2025, the US consumer price index (CPI) and producer price index (PPI) came in higher than expected, raising concerns that inflation could reignite. while technological innovations like DeepSeek are exerting deflationary pressure, artificial cost increases from tariffs are offsetting or overwhelming them. This is a major factor in slowing the Fed's pace of rate cuts and is one of the biggest risks to the 2026 economic outlook.
4. federal Reserve (Fed) Monetary Policy: The Rate Cut Dilemma and Path to 2026
4.1. The 2025 Rate Cut Cycle and the "Insurance Cut"
in 2025, the Fed began its rate-cutting cycle in September in response to slowing inflation and cooling signals from the labor market. At the September Federal Open Market Committee (FOMC) meeting, the Fed cut its benchmark interest rate by 25 basis points to a range of 4.0-4.25%. the move was in line with market expectations, and was largely based on weaker-than-expected August employment data.
chair Jerome Powell characterized it as a "risk management cut," suggesting that it was more of an insurance policy to prevent a recession than a move to full-scale monetary easing. the October meeting followed with another 25 basis points, bringing rates into the 3.75-4.00% range.
but as the end of the year approached, the Fed's tone became more cautious. at the December meeting, Fed members scaled back their forecast for the number of rate cuts in 2026, signaling a "hawkish cut" stance. this was in response to renewed inflationary pressures from Trump's tariffs and the fact that US GDP growth remained robust at 3.1% in the third quarter. Opinions within the Fed were divided, with members like Stephen Miran arguing for a bold 50 basis point cut, while others like Jeffrey Schmid and Austan Goolsbee argued for a pause or cautiousness.
4.2. 2026 rate outlook: slow march toward neutral rates
major investment banks are predicting a very slow pace of rate cuts by the Fed in 2026. j.P. Morgan sees only one additional rate cut by the Fed in 2026. the Fed's Dot Plot also suggests only one cut in 2026, so the rapid monetary easing that markets have been expecting is unlikely.
[Table 2] Key economic indicators and Fed forecasts for 2026 (median)
indicator2025 Outlook (Revised)2026 Forecast (Revised)notes GDP Growth Rate 1.6% → 1.7 1.8% → 2.3
reflects the effects of Trump tax cuts and deregulation
PCE Inflation 3.0% → 2.9 2.6% → 2.4 modest decline expected despite tariff impact unemployment Rate 4.5 4.4 full employment level maintained interest rates (median) 3.75 to 4.00 3.50 to 3.75 one additional cut (25 basis points) in 2026 is implied
source: Federal Reserve Summary of Economic Projections (SEP)
this will be a 'double-edged sword' for equity markets in 2026: while the lack of a sharp rate hike is a positive, the 'Higher for Longer' stance will continue to weigh on companies' financing costs. Especially for tech companies that use a lot of debt to invest in AI infrastructure, rising interest costs could be a profitability killer. morgan Stanley warns that this could lead to wider spreads on investment-grade corporate bonds.
5. war of the Tech Giants: Toward the $5 Trillion Club
5.1. Nvidia: the first $5 trillion milestone and volatility
in October 2025, Nvidia became the first company in the history of the world to temporarily surpass $5 trillion in market capitalization. a keynote speech by CEO Jensen Huang, along with the announcement of a lineup of next-generation AI products and big partnerships with Oracle and Finland's Nokia, were the triggers for the share price spike. In particular, the announcement of a $1 billion investment and AI infrastructure collaboration with Nokia confirmed that Nvidia was more than just a chipmaker, but a dominant player in the AI ecosystem.
however, the Dipshit shock, Trump tariffs, and AI bubble theories have caused Nvidia's stock price to correct back to the $188 level, retracing its $5 trillion market capitalization. nevertheless, Nvidia is cementing its position as a utility for AI infrastructure. in 2026, the company is expected to attempt to re-enter the $5 trillion mark again, driven by its "Rubin" platform with HBM4 memory and full-scale adoption of its Blackwell architecture. analysts value Nvidia at a 52-58x forward P/E ratio as of December 2025, which is still an attractive range given its explosive growth (PEG ratio of 0.98).
5.2. The Chasers: microsoft and Alphabet
microsoft and Google's parent company Alphabet are also close to joining the $5 trillion club. microsoft currently has a market capitalization of about $3.6 trillion and is projected to reach $376 billion in revenue by 2026. based on its dominance in the cloud and AI agent market, analysts suggest that if revenue growth exceeds 20%, it could reach $5 trillion in 2026.33
alphabet also maintains its "Magnificent Seven" status with a steady rise from the $3.8 trillion level. despite challenges from open-source models like DeepSeek, the solid cash-generating power of the search ad market and the sophistication of its own AI model, Gemini, are driving the stock higher.
5.3. 2026 Capex Forecast: A $500 Billion Gamble?
goldman Sachs predicts that AI hyperscalers will spend more than $500 billion in capex in 2026. aI investment by big tech companies like Microsoft, Alphabet, Meta, Amazon, and others is on par with that of telecom companies during the dot-com bubble of the 1990s.
however, some in the market have expressed doubts that these massive investments will translate into tangible return on investment (ROI). BCA Research warns that depreciation costs for AI assets can reach 20% per year, leaving hyperscalers on the hook for $400 billion in depreciation costs in 2025 alone. 2026 will be the "ROI Phase," when AI companies will have to justify their investments with "results" rather than just "hopes," and the jockeying for position will be intense.
6. korea's Rise: KOSPI 4000 and the Semiconductor Super Gap
6.1. Historical Milestone: The KOSPI 4000 Era Begins
at the end of October 2025, South Korea's KOSPI index surpassed the 4,000-point mark for the first time ever, closing at 4,214.17 on December 30, the year's closing day, marking a historic year. this is a phenomenal V-shaped rebound of 75.6% in annualized gains, considering that the index plunged as low as 2,284 points at the start of the year on Trump tariff concerns.
there are three key factors behind this rise.
AI semiconductor supercycle: hBM and AI memory semiconductor performance has exploded, led by Samsung Electronics and SK Hynix.
government's value-up program: Strong government policies and tax incentives to encourage shareholder return have led to companies burning shares and increasing dividends.
foreign capital inflows: Foreign investors were net buyers of nearly KRW 10 trillion in Samsung Electronics alone, driving the index higher.
6.2. Samsung Electronics vs. SK Hynix: Round 2 of the HBM War
6.2.1. Samsung Electronics' resurgence and turn-key strategy
samsung Electronics successfully reached the "100,000 electronics" mark in 2025, with its stock price rising 123.8% to close at 119,000 won. After losing the lead to SK Hynix in the early days of the HBM market, Samsung emerged as a strong competitor with its 'turn-key' solution that combines design, memory, foundry, and packaging from HBM4 onwards, passing Nvidia's quality tests. in particular, Samsung plans to start mass-producing HBM4 at its Pyeongtaek campus in February 2026, which will be used in Nvidia's next-generation AI accelerator, Rubin. in addition, the company has diversified its customer base by supplying HBM4 to Google's 7th generation TPU.
6.2.2. SK hynix's technology leadership and TSMC alliance
SK Hynix solidified its position as the second-largest market capitalization with a stock price increase of 273.8% in 2025. SK hynix allied with TSMC, the world's No. 1 foundry, to produce HBM4, and applied TSMC's 12nm logic process to the base die, which doubled the bandwidth and improved power efficiency by more than 40%, maintaining its technological edge.
6.3. 2026 Semiconductor Technology Roadmap: HBM4 and 16-stack stacking
the year 2026 will be a technological inflection point for the semiconductor industry.
HBM4 mass production: both Samsung Electronics and SK Hynix are targeting mass production of HBM4 in early 2026. With twice the bandwidth and 40% less power than the existing HBM3E, HBM4 will be a key component in lowering the cost of inference for increasingly large AI models.
16-Hi race: Nvidia has called for the development of16-Hi HBM4 for Q4 2026 delivery, a challenging technology that requires reducing wafer thickness to 30 micrometers, requiring the adoption of next-generation packaging technologies such as hybrid bonding. Process innovation to meet the JEDEC standard of 775 micrometer thickness limit will be a key theme for the semiconductor equipment market in 2026.
7. investment themes and outlook for 2026
7.1. The arrival of physical AI: humanoid robots
one of the key investment themes for 2026 is "humanoid robots. morgan Stanley sees 2026 as the year when expectations for humanoid robots will peak. expect the unveiling of Tesla's third-generation Optimus and the entry of major big tech companies into the robotics market. however, they are likely to remain in the data collection phase via tele-operation rather than full autonomy, and will be overshadowed by the difficulty of hardware manufacturing.
7.2. Energy Infrastructure: SMRs and data centers
The surge in power consumption in AI data centers is forcing investment in energy infrastructure. according to Deloitte, data center power demand is expected to increase fivefold by 2035. nuclear power, specifically small modular reactors (SMRs), is gaining traction as a solution. the U.S. Department of Energy (DOE) is providing significant funding for early deployment of SMRs, and big tech companies will expand their partnerships with nuclear power companies to secure reliable power. 2026 could be the year that SMR-related stocks move beyond just a theme to actual orders and project starts, with projects to convert retired coal plant sites to SMRs gaining momentum.
7.3. Risk Factors: AI Bubble Bursting Theory and Inflation
it's not all rosy: BCA Research warns that the AI boom could burst like the dot-com bubble in 2026. the emergence of efficient models like DeepSeek could paradoxically reduce demand for expensive AI hardware, and a massive selloff could emerge if AI fails to deliver productivity gains for companies. In addition, a resurgence of inflation due to Trump tariffs is a potential brake on the Fed's rate cuts and could trigger a recession.
8. conclusion: Investment strategies for the "post-hype" era
if 2025 was a year of "don't ask, don't tell" and excitement about AI, 2026 will be a year of "performance and efficiency" dominating the market. the DeepSeek shock has reoriented AI development away from "blind enormity" and toward "economic efficiency," which will be a challenge for hardware providers like NVIDIA, but an opportunity for application layer companies that leverage AI to create real-world services.
investors should pay attention to three strategic points
semiconductor sub-division (materials, components, and equipment) jade stone: Focus on companies with next-generation technologies such as HBM4, 16-stack stacking, and hybrid bonding. companies in the value chain of Samsung Electronics and SK Hynix are promising.
expanding infrastructure: Companies with physical infrastructure to power AI, such as power grids, SMRs, and cooling systems, will benefit from continued demand growth regardless of the efficiency of AI models.
defensive portfolios: maintaining an appropriate allocation to defensive sectors such as healthcare, consumer staples, and alternative assets such as gold willbe necessaryto hedge against the expected economic slowdown and inflationary risks in the second half of 2026. 30
The Korean market in the KOSPI 4000 era has a clear growth engine in semiconductors, but it is also the market most sensitive to the ramifications of the global trade war. It is time to enjoy the volatility, but be selective and data-driven.
[Disclaimer] This report is for informational purposes only and does not constitute a recommendation to buy or sell any specific asset. Investments are made at your own risk.