Necessity Breeds Innovation: Lessons from DeepSeek’s AI Breakthrough

Sangit Rawlley

Tags

Lets Discuss Your business goals

Key Points:

Innovation Through Constraints: DeepSeek leveraged US-imposed GPU restrictions as a catalyst for innovation, demonstrating how selective disadvantages can drive strategic breakthroughs in AI.
Blue Ocean Strategy in AI: By shifting away from traditional high-cost, closed-model approaches, DeepSeek introduced an open-source, cost-efficient Mixture of Experts (MoE) architecture, redefining AI development.
AI Disruption and Expansion: DeepSeek’s breakthrough is accelerating AI adoption, reducing costs, enabling new use cases, and intensifying competition, with major tech companies ramping up investments to stay ahead.

DeepSeek, a Chinese startup, has recently made waves in the tech world with its new AI model, rivaling the performance of US models like OpenAI. This achievement, especially given the limitations imposed by US GPU export restrictions to China, raises important questions about the drivers of strategic innovation. Speculation abounds about the actual cost of building this model, but rather than joining the debate, we take a different approach—a strategic lens.

What We’ll Cover:

The impetus that drove this innovation
The key differences in approach
The broader implications

The Catalyst for Innovation: Constraint

The US ban on high-powered GPU shipments to China presented a significant challenge for Chinese companies like DeepSeek. This constraint, however, became the catalyst for their innovation. As Michael Porter highlights in “The Competitive Advantage of Nations,” selective disadvantages can be powerful drivers of innovation. When resources are abundant, companies may become complacent. Conversely, constraints force them to find creative solutions.

History is replete with examples of innovation born from necessity. Japan’s limited land and natural resources spurred innovations like just-in-time production and product miniaturization. Israel’s challenging agricultural environment led to breakthroughs in drip irrigation and water management. Even the Apollo 13 mission demonstrated the power of creative problem-solving under extreme constraints by devising creative life-saving solutions, including an improvised fix for rising carbon dioxide levels.

DeepSeek’s situation mirrors these examples. Denied access to the same computational resources as their US counterparts, they were compelled to innovate and find alternative ways to train large language models.

“I think frugality drive innovation just like other constraints do. One of the best ways to get out of a tight box is to invent your way out.” – Jeff Bezos

From Red Ocean to Blue Ocean: DeepSeek’s Strategic Innovation

Having examined the drivers behind DeepSeek’s innovation, let’s analyze how this strategic innovation came about, using the Blue Ocean Strategy framework proposed by W. Chan Kim and Renée Mauborgne.

Breaking Free from the Value-Cost Trade-off

The conventional wisdom in business often dictates a trade-off between value and cost: offer high value at a high cost or reasonable value at a low cost. This “value-cost trade-off” is exemplified by the prevailing approach to Large Language Models (LLMs) in the US:

Invest billions to build high-quality LLMs
Restrict access to high-powered chipsets, limiting competitors to lower-quality models

Kim and Mauborgne term this a “red ocean” strategy, focused on competing within existing market boundaries. To transcend this trade-off and create a “blue ocean” of uncontested market space, Kim and Mauborgne suggest a framework of four key questions:

Eliminate: Which factors should be eliminated?
Reduce: Which factors should be reduced?
Raise: Which factors should be raised?
Create: Which new factors should be created?

Let’s apply this framework to DeepSeek’s model:

Eliminated:

Closed Model: DeepSeek eliminated the industry’s reliance on closed, proprietary models, opting for an open-source approach with disclosed architecture and parameter count.

Reduced:

Training Cost: Despite varying estimates, DeepSeek R1 demonstrably achieved a lower training cost compared to other large language models.
Parameters Activated: DeepSeek, despite having 671 billion parameters, activates only 37 billion per forward pass, suggesting a reduction in computational resources required for inference.
Response Refinement: DeepSeek’s responses, while powerful, are acknowledged to be less refined compared to dedicated chat models, indicating a potential trade-off in polish for other advantages.

Raised:

Transparency: DeepSeek’s open-source nature and disclosed architecture significantly raise the level of transparency in the LLM space.
Training Efficiency: DeepSeek employs an advanced optimization method, incorporating self-verification, surpassing the traditional supervised learning approach in potential efficiency.

Created:

Novel Architecture: DeepSeek introduced the Mixture of Experts (MoE) architecture. This fundamentally re-imagines how AI models process information, akin to a team of specialized experts called upon for specific tasks, rather than a generalist team tackling everything. This architecture is a key differentiator.

DeepSeek’s approach illustrates that constraints drive innovation. Instead of competing within traditional boundaries, the company reimagined AI development—lowering costs, enhancing efficiency, and pioneering a new model architecture that challenges industry norms.

The figure below presents a strategy canvas, highlighting how DeepSeek’s approach compares to traditional LLM providers.

The DeepSeek Effect: Unleashing the Next Wave of AI Innovation

In our article, “The Imminent AI Tsunami: How AI Is Following the Internet’s Disruptive Path”, we explored the parallels between the evolution of the internet and AI, highlighting AI’s potential to drive widespread disruption, similar to the internet’s impact over the past two decades. The introduction of the DeepSeek model serves as a compelling example of this accelerating disruption.

Contrary to initial fears that DeepSeek would reduce demand for AI infrastructure, leading to a market selloff that wiped out $800 billion in value from AI chipmakers like Nvidia and Broadcom in a single day, the reality is that DeepSeek will likely further fuel AI adoption.

Satya Nadella, in a recent LinkedIn post, astutely cited Jevons’ Paradox to explain this phenomenon. Named after the 19th-century economist William Stanley Jevons, this paradox highlights that increased efficiency can paradoxically lead to increased consumption. Jevons observed that improved steam engine efficiency, rather than leading to reduced coal usage, actually led to increased coal consumption as it made steam engines more economical to use.

This phenomenon is evident in other technology cycles. Moore’s Law, which describes the exponential growth of computing power, fueled the PC revolution. Similarly, Edholm’s Law, which describes the exponential growth of data transmission rates, played a crucial role in the dotcom boom.

“Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning into a commodity we just can’t get enough of.” – Satya Nadella

In a similar vein, the introduction of DeepSeek, and models like it, will accelerate AI adoption and disruption. Here are several ways DeepSeek will disrupt the AI industry:

Expanding AI Usage: The cost of using an LLM with equivalent performance is decreasing 10x annually. DeepSeek’s significantly lower costs will further democratize AI access, making it more affordable for individuals, startups, and smaller businesses. This will drive widespread adoption and fuel innovation.

New Use Case Explosion: Lower LLM costs will unlock previously unviable use cases. For instance, the cost of processing human speech with an LLM will become negligible, enabling new applications across various domains.
Agentic AI Acceleration: Lower token prices will drive innovation in Agentic AI, which relies on processing large numbers of tokens for complex workloads.
Increased Transparency and Trust: DeepSeek’s open-source nature fosters greater trust and transparency, addressing concerns about data privacy and model interpretability. This will accelerate community-driven improvements and innovation.
Mobile AI Revolution: DeepSeek’s efficient architecture will enable powerful AI models to run on mobile devices, ushering in an era of ubiquitous mobile AI, similar to the mobile internet.
AI Startup Boom: Reduced API costs and computational needs will empower AI startups to develop and deploy innovative products more rapidly and cost-effectively. This will foster a more competitive and dynamic AI ecosystem.
Vertical AI Specialization: DeepSeek’s open-source nature will benefit vertical AI startups by enabling them to customize models for specific industry needs, creating more tailored and effective solutions.

DeepSeek: Shifting AI into High Gear

DeepSeek’s emergence signifies a pivotal shift in the AI landscape. By embracing constraints and pursuing a novel “blue ocean” strategy – focusing on creating new value propositions rather than competing within existing market boundaries – DeepSeek has not only achieved impressive results but also demonstrated the power of strategic innovation. This approach offers valuable lessons for other players in the AI ecosystem.

Furthermore, the intense competition sparked by DeepSeek’s breakthrough, evidenced by major tech players like Meta, Amazon, Alphabet, and Microsoft increasing their AI investments to $320 billion in 2025 from $230 billion in 2024, underscores the urgency of embracing AI. The AI revolution is unfolding rapidly. Companies must proactively embrace AI or risk being left behind.

Sangit Rawlley is a Senior Partner and AI Practice Lead at JLA Advisors (www.jlaadvisors.io), a boutique consulting firm dedicated to empowering businesses with transformative strategies and customized solutions for success. JLA’s AI consulting practice helps clients leverage the power of AI to enhance existing products, unlock new opportunities, and deliver innovative solutions.