Goku - Another DeepSeek on the Horizon
Fresh & Hot curated AI happenings in one snack. Never miss a byte 🍔
This snack byte will take approx 4 minutes to consume.
Remember when everyone thought China's AI ambitions would be crushed by chip restrictions?
Well, someone forgot to tell Shanghai Goku Technologies.
While the rest of us were still debating whether DeepSeek was a fluke, this quantitative trading fund quietly submitted a research paper to NeurIPS—the so-called "AI Olympics"—claiming their training method beats both DeepSeek and OpenAI at their own game.
And yes, they named their company after an anime character who gets stronger by fighting increasingly powerful opponents. The metaphor writes itself.
Shanghai Goku Technologies, operating under the delightfully uncompromising slogan "logic and truth are the only principles we obey," has thrown down the gauntlet with their SASR (Step-wise Adaptive hybrid training framework for Sequential Reasoning) method.
Their paper, co-authored with Shanghai Jiao Tong University researchers, doesn't just critique existing approaches—it claims to fundamentally reimagine how we train AI models.
The timing is exquisite.
On the same day they published their breakthrough research, Goku registered a new AI subsidiary called AllMind.
Their approach addresses the limitations of supervised fine-tuning (SFT) and reinforcement learning (RL)—the bread and butter techniques that OpenAI and DeepSeek rely on. Instead of these static methods, SASR mimics human cognitive development, adapting its training approach step-by-step.
Goku's path mirrors DeepSeek's remarkably closely. Both originated from quantitative trading funds (DeepSeek from High-Flyer, Goku from its eponymous trading operation).
Both established independent AI subsidiaries. Both focus on algorithmic efficiency over brute-force computational power.
And both have managed to achieve breakthrough results despite—or perhaps because of—hardware restrictions.
DeepSeek claims to have trained its model for just $6 million using 2,000 Nvidia H800 graphics processing units (GPUs) vs. the $80 million to $100 million cost of GPT-4. This isn't just cost optimization—it's a complete paradigm shift.
While American companies throw money and compute at the AI problem, Chinese researchers are pioneering elegant solutions that do more with less.
The secret sauce?
DeepSeek's "mixture of experts" architecture, which essentially means that it comprises several specialized models, rather than a single monolith. This allows it to give answers while activating far less of its "brainpower" per query. It's computational efficiency at its finest—why use your entire brain when you only need the relevant parts?
While Western observers debate whether China can compete without access to the latest chips, Chinese companies are quietly writing massive checks. Chinese smartphone maker Honor will spend $10 billion over the next five years on developing AI for its devices, and that's just one company's commitment.
The scale becomes clearer when you zoom out.
AI is rapidly changing real life in China and creating a market that could reach $1.4 trillion in the next five years as the country becomes one of the global leaders in the new technology. We're not talking about modest investments here—this is nation-scale commitment to AI supremacy.
China recently announced a state venture capital guidance fund expected to attract nearly 1 trillion yuan ($138 billion) in capital over 20 years from local governments and the private sector. When a country commits $138 billion to emerging technology development, you pay attention.
Even established tech giants are doubling down. China Unicom reported that spending on computing power, including data centres, rose 19 per cent in 2024, underscoring the company's commitment to AI. This isn't venture capital moonshots—this is infrastructure-level investment from companies that already dominate their markets.
Jensen Huang, Nvidia's CEO, recently made a statement that should make policymakers uncomfortable: "China has 50 per cent of the world's AI developers". If you control half the world's AI talent, hardware restrictions become less about blocking progress and more about forcing innovation.
The irony is palpable. Export controls designed to hamper China's AI development may have inadvertently accelerated it. When you can't rely on the latest hardware, you're forced to develop smarter algorithms. When you can't brute-force solutions, you have to innovate. DeepSeek's breakthrough wasn't despite hardware limitations—it was because of them.
Goku's SASR method represents the next evolution of this efficiency-first philosophy. Instead of simply optimizing existing training methods, they're questioning the fundamental assumptions about how AI models learn. Their step-wise adaptive approach suggests that perhaps the problem isn't computational power—it's computational intelligence.
There's something poetic about quantitative trading funds leading China's AI revolution. These firms have spent decades turning market inefficiencies into algorithmic advantages. They understand better than anyone that the most elegant solution isn't always the most obvious one.
Goku manages over 15 billion yuan (US$2.1 billion) in domestic and global assets, using AI-driven strategies. They've already proven they can deploy AI profitably at scale. Now they're sharing their methodology with the research community, potentially accelerating AI development across the entire ecosystem.
While Western AI labs chase ever-larger models and higher compute budgets, Chinese researchers are playing a different game entirely. They're proving that intelligence isn't just about scale—it's about efficiency, architecture, and algorithmic innovation.
The submission to NeurIPS isn't just academic posturing. It's a statement: Chinese AI research has moved beyond catching up to setting the pace. When trading funds start publishing breakthrough research that outperforms Big Tech, the balance of AI innovation is shifting.
We're witnessing the emergence of what might be called "constraint-driven innovation"—breakthrough AI development born from resource limitations rather than resource abundance. If Goku's claims hold up to peer review, we might look back at 2025 as the year China didn't just achieve AI parity, but pioneered a fundamentally more efficient approach to artificial intelligence.
The question isn't whether China can compete in AI despite hardware restrictions. The question is whether anyone else can compete with China's efficiency-first approach to AI development.
DeepSeek was just the beginning. Goku might be the next chapter in a story that's reshaping the global AI landscape.
And somewhere in Shanghai, a company named after an anime character is quietly revolutionizing how we think about machine intelligence. Sometimes the future arrives with less fanfare than you'd expect.
About the author: Rupesh Bhambwani is a technology enthusiast specializing in the broad technology industry dynamics and international technology policy.
When not obsessing over nanometer-scale transistors, energy requirements of AI models, real-world impacts of the AI revolution and staring at the stars, he can be found trying to explain to his relatives why their smartphones are actually miracles of modern engineering, usually to limited success.