Until December 2024, conventional wisdom held two certainties about AI: U.S. tech giants would maintain their lead through massive capital deployment, and closed-source models would dominate through superior resources. On January 27th, 2025, those assumptions collapsed. Nvidia’s market capitalisation plunged $600 billion—the largest single-day value destruction in stock market history—after a Chinese company demonstrated both premises wrong.

The Foundation Shift

AI is the next major paradigm shift in computing. It’s a loose collection of many approaches, but the current shift is currently happening with Large Language Models (LLMs) – the models that can both read and write human language (including computer code), and can be used to power everything from software development to customer service. Historically, training these models—teaching them patterns from vast datasets—required massive computing power. Running them (inference) demanded expensive hardware. The industry accepted this as immutable: better performance required more computing power, which meant more expensive chips.

This doctrine shaped the entire industry. OpenAI and Google poured billions into training increasingly powerful models. Nvidia, selling the chips that made this possible, enjoyed 90%+ profit margins. The U.S. government, seeking to maintain its lead, even restricted chip exports to China. The assumption? Without access to cutting-edge hardware, Chinese AI development would stall.

The Catalyst Timeline

The dominoes began falling on December 26th, 2024, when DeepSeek released its V3 model to minimal fanfare. Most observers missed its significance amid holiday distractions, but early technical reviews on platforms like Twitter and Reddit hinted at a seismic shift. Then came January 20th, 2025: As Washington focused on Trump’s inauguration, DeepSeek quietly released R1, their reasoning-focused model. The following day, Trump announced Project Stargate—a $500 billion AI infrastructure initiative primarily benefiting OpenAI.

By January 27th, as investors fully digested the implications, Nvidia’s stock dropped 16%. The market finally grasped what AI researchers had been buzzing about: DeepSeek had achieved what many thought impossible, training a world-class AI model for $6 million—roughly 1/20th the industry standard cost of $100+ million.

The Pre-Disruption Power Structure

Before January 2025, the AI landscape resembled a very expensive oligopoly. OpenAI boasted a $5 billion annual run-rate by December 2024. Nvidia commanded massive profit margins in their AI segments, while their CUDA software (the software platform used to develop the models) ecosystem locked in the majority of the world’s AI researchers.

The industry operated on what insiders called “The Scaling Doctrine”—a set of seemingly immutable laws governing AI progress. Compute requirements doubled every six months. Training data expanded from 300 billion tokens for GPT-3 to 15 trillion tokens for 2024’s leading models. Costs increased 150% year-over-year, while training times stretched from weeks to months. For an industry obsessed with exponential growth, efficiency seemed almost an afterthought.

DeepSeek’s Triple Disruption

DeepSeek’s big achievement is that the company has achieved comparable model performance at a fraction of the cost. Their innovations exemplify a fundamentally different approach to AI development. Here are three examples from their broader set of breakthroughs:

• Efficient Training Architecture: By simplifying calculations with fewer decimals, Deepseek achieved comparable results while using 75% less memory. This is like discovering you can build the same skyscraper with a quarter of the materials. It sounds incredibly trivial, but when paired with the company’s other innovations, it delivered just the required performance.
• Multi-Token Prediction: While traditional models predict one word at a time, Deepseek’s system predicts multiple words (tokens) simultaneously with 85-90% accuracy, effectively doubling inference speed. The efficiency gain is remarkable—imagine if your car suddenly required half the fuel while maintaining 90% of its horsepower.
• Mixture-of-Experts (MoE) Architecture: While GPT-4 activates all of its estimated 1.8 trillion parameters for every task, DeepSeek’s approach is more selective. Their 671 billion parameter model activates only 37 billion parameters at any time—just the experts needed for each specific task. It’s like having a company where, instead of calling all 1,800 employees into every meeting, you bring in only the 37 specialists most relevant to the problem. Simple. Effective. Brilliant.

On their own, these advances might have led to an inferior model. However, combined, they produced the same quality output with drastically lower computational costs—running on two consumer-grade NVIDIA 4090 GPUs costing under $2,000 total versus GPT-4’s requirement for multiple $40,000 H100 GPUs.

The Deepseek team also did many other clever things to change the game, including a major focus on reinforcement learning rather than human fine-tuning (meaning that they taught the computer how to become smarter rather than relying on laborious fine-tuning by experts). This difference in approach is much more than just a technical tweak —t’s a philosophically different view of how to develop the best AI.

Best bit: DeepSeek open-sourced their entire system. It’s all there, for everyone to use, to tweak and to learn from. It’s hard to overstate the disruptive impact of this.

Commercial Earthquake

DeepSeek’s breakthrough reshapes the AI economics fundamentally:

Training costs for large language models have collapsed from $100+ million to $6 million – a reduction that transforms the competitive landscape. Inference costs follow the same trajectory, with DeepSeek’s architecture delivering a 95% reduction in operational expenses compared to current market leaders.

This cost revolution creates three immediate market impacts:

1. Incumbent Pressure: OpenAI and Anthropic face an existential choice – match DeepSeek’s 95% lower pricing and accept razor-thin margins, or maintain premium pricing and watch their market share evaporate.
2. Cloud Provider Ripple: When a $200/month ChatGPT Pro subscription faces competition from DeepSeek’s open source (=free) R1 model, the implications cascade to cloud infrastructure. This could impact Microsoft, Amazon, and Google’s AI near-term revenue projections.
3. Hardware Vulnerability: NVIDIA’s dual revenue streams – from both model training and inference – face immediate pressure. Even if the overall volume of AI work continues to grow rapidly, there will be near-term implications for the amount of hardware required.

As such, the promised “Year of Agentic AI” in 2025 arrives with an unexpected plot twist: the established players now face fierce competition from new entrants who can match or exceed their performance at a fraction of the cost.

Geopolitical Chess Game

The timing couldn’t be more pointed. A day before Trump’s $500 billion Project Stargate announcement—meant to cement U.S. AI supremacy—a Chinese company demonstrated that capital alone doesn’t determine AI leadership. Despite U.S. export controls blocking access to Nvidia’s most advanced chips, DeepSeek leapfrogged the entire industry.

Initial reactions in Silicon Valley included speculation about potential deception—suggestions that DeepSeek must have accessed more computing power than disclosed. However, their open-source release and subsequent validation by independent researchers seem to confirm the breakthrough’s legitimacy. The U.S. giants weren’t outspent; they were out-innovated.

While Nvidia’s 16% drop likely represents an overcorrection—partly triggered by broader market dynamics seeking a reason to correct—the underlying technological shift remains significant.

European Implications

For Europe, DeepSeek’s breakthrough is particularly significant. Over the past year, as Mistral and other European AI companies struggled to match their American counterparts, the prevailing narrative blamed resource constraints. DeepSeek’s success fundamentally challenges this excuse. The limitation wasn’t capital—it was innovation.

This serves as both a wake-up call and an opportunity. Europe must confront the uncomfortable truth that blaming insufficient resources masked deeper innovation gaps. However, DeepSeek’s success also demonstrates that the AI race remains wide open. Despite remarkable U.S. progress, technological leadership can shift rapidly through creative approaches to fundamental problems.

Looking Forward

Expect a frantic response over the coming weeks. The U.S. tech giants will scramble to match DeepSeek’s efficiency gains. Watch for major announcements from OpenAI, Meta, and Google.

The short-term implications will be volatility and potentially a full-blown market correction, but the longer-term implications will be more significant. Unlike search engines or mobile operating systems, which developed into entrenched oligopolies, AI appears headed toward a more dynamic competitive landscape. The ease of switching between models, combined with the demonstrated potential for technological leapfrogging, suggests a future where innovation matters more than market position. This inherent flexibility—a core strength of software—may prevent any single player from establishing lasting dominance.

The near-term market turbulence masks a larger truth: AI just became 95% cheaper while growing more capable. When a technology’s cost drops this dramatically while improving, adoption doesn’t just grow—it explodes. Contrary to what Trump’s Stargate project will have us believe, sometimes the most profound disruption comes not from doing things bigger, but from doing them smarter. And that, it seems to me, is at the heart of what the startup and venture ecosystem is all about.