The Great Gen AI Race Continues

Where The Machines Start To Reason

Curated perspectives and research for business leaders who look outside the box. The Insight Loop is your essential guide to the latest market, business, and technology forces shaping the future of AI-led enterprises. We hope you enjoy this month's edition!

If you’ve noticed your favorite large language model (LLM) taking a moment to “think” before answering, you’ve already met the latest evolution in AI: reasoning models. These advanced models, trained through reinforcement learning, break down your questions into logical steps and test different approaches before arriving at a more thoughtful response. Simply put, they’re designed to think critically and methodically before providing answers.

Last September, OpenAI unveiled o1, the world’s first public reasoning model. This model “thinks” before it answers—producing a deep, internal chain of reasoning. So far, the results have been remarkable: o1 has soared to the 89th percentile on Codeforces programming questions, ranked among the top 500 math students in the U.S., and outperformed human PhD-level accuracy on a physics-biology-chemistry benchmark, according to OpenAI.

While most traditional AI models are prone to “hallucinations,” Reasoning models distinguish themselves by effectively fact-checking their own outputs. As a result, they tend to deliver more reliable answers—especially in fields like physics, science, and mathematics.

As you might expect, it was only a matter of time before the tech hyper-giants jumped headfirst into this race. Shortly after Open AI's release, Google responded with Gemini Flash Thinking, and within just three months, China joined the fray. Alibaba introduced QwQ (Qwen with Questions), while Hangzhou-based DeepSeek released R1, with 671 billion parameters—built for research tasks requiring logical thinking, self-verification, and reflection.

So, what makes this model so special—and why is it garnering so much attention? DeepSeek achieved pure reinforcement learning by autonomously verifying its own work, without relying on supervised datasets.

It also introduced a completely new approach to reward modeling. While traditional models often exploit loopholes and game the system to earn rewards, DeepSeek implemented two distinct reward systems—one based on the accuracy of the final answer, and another focused on the model’s reasoning structure. This technical breakthrough is what allows the model to generate long 'chains of thought' and apply complex step-by-step reasoning to tasks.

As the Global AI arms race heats up, who’s emerging on top?

In the U.S. and across Europe, private-sector heavyweights are primarily driving major efforts in LLM development. Despite well-known challenges like high costs, power consumption, and frequent “hallucinations”, AI research is spurred forward only by commercial interests, often outpacing academic and government-led efforts.

Meanwhile, China is pursuing a more diverse, state-driven AI strategy, to weave “values” into AI from the start, ensuring safety and alignment with societal priorities. Their strategy reflects a more wholistic approach, such as their ‘Interim Measures for the Management of Generative AI Services’ initiative to address emerging risks while fostering technological advancement, or the ‘AI Safety Governance Framework’ which fosters experimentation with AI applications while ensuring alignment with ethical principles.

It’s hard to say who’s truly winning at this point. But, we will have to give the edge to China on this one simply for their bold long-term vision to improve governance and drive economic growth, together.

More formidable challengers are entering the arena.

Back to DeepSeek. They also released a general-purpose LLM called Deep Seek-V3, a massive 671-billion-parameter model—bigger than any free downloadable model to date, bigger than even Meta’s Llama 3.1 which has 405bn parameters.

AI models leverage attention mechanisms, a technique inspired by how the human brain selectively processes relevant information. The V3 model takes this a step further with its Multi-Head Latent attention mechanism, a breakthrough that significantly reduces the memory bottleneck for LLMs. By focusing on multiple parts of the input simultaneously, the model can analyze large, complex data with greater depth and accuracy.

What’s even more impressive is that the V3 model took only a tenth of the computing power and expense that went into Llama 3.1, requiring just 2,000 chips, whereas the former used 16,000 chips! 

The V3, has a context window (the amount of text that an LLM can process at once) of up to 128,000 tokens (the basic units of LLM data–words or individual characters–used to understand the nuances of natural language).

DeepSeek’s V3 is a formidable rival to Open AI’s GPT-4o model, excelling at processing complex tasks with high accuracy and coherence. V3 is also trained on 14.8 trillion high-quality tokens, making it exceptionally good at understanding long-form content, making it ideal for projects with large data processing requirements. Keep in mind, all this from a company that has just 200 employees compared to OpenAI's 3,500, as of January 2025.

The encouraging news is that both the cost and speed of training these models are steadily improving, thanks in large part to the use of synthetic data, which is significantly driving down overheads. In just two years, the cost of processing 2 million tokens—both input and output—has dropped 240x, from $180 to just $0.75, making advanced AI more affordable than ever.

UC Berkeley’s Sky Computing Lab recently unveiled a reasoning model that rivals OpenAI's o1 reasoning model on several benchmarks. Their open-source model was trained for under $450 proving that high-level reasoning models can now be built affordably.

Do businesses see real ROI from Gen AI advancements?

Gartner estimates more than 10% of businesses in China are already using Generative AI, up from 8% about six months ago. India is way ahead in terms of adoption—with 93% of students and 83% of employees actively using Gen AI for everything from editing content to educational research, according to a recent Deloitte report. India ranks first in Asia for Gen AI adoption and second globally, behind the U.S.

For businesses, Deloitte’s research shows the most advanced Gen AI deployments focus on IT (28%), operations (11%), marketing (10%), and customer service (8%). Industries also lean into AI initiatives differently: manufacturers target operations (23%), while life sciences prioritize R&D (21%). A promising 74% of enterprises say their Gen AI projects are meeting or exceeding ROI, with cybersecurity efforts especially likely to outperform.

This is only the beginning. The next phase of Gen AI will test how quickly private-sector and government efforts can adapt, innovate, and collaborate. In next month’s edition of The Insight Loop, we’ll explore emerging use cases that are delivering impressive results across manufacturing, healthcare, financial services, insurance, and consumer enterprises. Be sure to look out for our upcoming report to stay ahead of these evolving trends.