The Japanese AI Boom Needs A Little More Ambition

SUMMARY

Asianometry's Jon Y explores Japan's overlooked AI boom through recent Tokyo conversations, highlighting deployment challenges, development pitfalls, and the need for greater ambition in a landscape dominated by foreign tools and conservative corporate practices.

STATEMENTS

Japan has launched a leading-edge AI lab, semiconductor foundry, and integrator startups, yet receives less attention than the US and China in the global AI narrative.
Generative AI deployment among Japanese consumers shows buzzword marketing proliferation, with tools like ChatGPT used for image generation, translation, and coding, though adoption levels require further study.
Corporate AI use in Japan remains mixed, with underutilization in basic IT processes like tax form automation, and generative AI often dismissed in favor of human oversight.
Japanese firms risk repeating past software mistakes by focusing on niche, custom AI models through system integrators, potentially leading to Galapagos syndrome and vulnerability to global LLMs.
Revenue perspectives differ: Japanese companies view AI primarily as a cost-cutter for productivity, unlike American firms that see it as a revenue generator.
Promising areas for Japan include AI in materials and drug discovery, with startups like Matlantis, MI6, and Eiktol leveraging existing data and systems.
Compute constraints limit Japan's AI development, but reactivating idle nuclear plants could address power shortages, while hardware strengths enable custom chips like Preferred Networks' MN-Core series.
Japanese talent possesses strong raw skills in math, science, and problem-solving, inspiring global tools like PyTorch via Chainer, though top researchers often migrate to the US for better pay.
Government support through programs like GenIAC aids LLM training but emphasizes sovereign AI over global competitiveness, potentially hindering broader innovation.
Japan's digital deficit, reaching about $24 billion in early 2025, underscores reliance on foreign software, calling for more ambitious domestic products to close the gap.

IDEAS

Despite prolific AI marketing in Japan, popular generative tools remain entirely foreign-made, raising questions about the absence of domestic equivalents in consumer spaces.
Automating a basic tax form online wizard feels outdated, highlighting Japan's lag in even fundamental IT adoption within HR and outsourcing sectors.
System integrators train small AI models to mimic human tasks like button presses in power plants, saving labor but echoing costly, niche software customization pitfalls from the past.
Japanese executives treat information technologies as productivity enhancers rather than revenue drivers, contrasting sharply with the growth-oriented mindset in American tech firms.
Startups like Matlantis and Eiktol position Japan advantageously in AI-driven materials discovery, capitalizing on untapped corporate data troves for innovation.
Half of Japan's nuclear plants remain idle post-Fukushima, presenting a straightforward power solution for scaling AI data centers amid global compute races.
Preferred Networks' custom MN-Core chips avoid the Nvidia premium, evolving into 3D-stacked processors for LLM inference by 2027, potentially enabling a diverse model-running ecosystem.
PyTorch's design drew heavy inspiration from Japan's Chainer framework, underscoring overlooked contributions to foundational AI tools despite talent migration trends.
Sovereign AI initiatives prioritize Japanese-controlled data and infrastructure for resilience, yet may foster insularity, as seen in the 1980s semiconductor decline from sticking to domestic tools.
Japan's digital deficit ballooned to $24 billion in early 2025, fueling pessimism about competing with American software giants and stifling entrepreneurial drive.
Kaggle competitions reveal strong Japanese performance in data science, indicating untapped potential beyond the elite, Twitter-active AI research circles.
Government programs like GenIAC share LLM training costs democratically, dubbed the "AI Squid Game" for their competitive, fair allocation of public funds.
Rapidus' rapid progress in 2nm fabs, including adopting superior ASML tools over Nikon, exemplifies hardware execution that could inspire bolder AI strategies.
The allure of US salaries draws top Japanese AI talent abroad, yet cultural preference for staying home could retain a core of committed innovators.

INSIGHTS

Japan's AI ecosystem thrives in hardware and niche applications but risks obsolescence by prioritizing customization over scalable, general-purpose models that drive global disruption.
Cultural emphasis on human oversight and cost-cutting in tech adoption perpetuates a digital lag, contrasting with ambition-fueled revenue innovation that propels ecosystems like Silicon Valley.
Sovereign AI builds resilience through localization but limits competitiveness by restricting access to diverse, high-quality global data essential for world-class performance.
Untapped nuclear capacity and hardware expertise position Japan to overcome compute barriers, enabling custom solutions that bypass foreign dependencies without sacrificing quality.
Strong foundational talent, evidenced by influences on tools like PyTorch, suggests Japan could reclaim AI leadership by fostering domestic ambition over emigration to higher-paying markets.
Government-industry collaboration accelerates R&D but falls short on commercialization, highlighting the need for private-sector risk-taking to transform prototypes into market-dominating products.

QUOTES

"Everything is AI now, often with ample sidehelpings of ICT."
"In their eyes, the human should always be in charge."
"My biggest worry is that the big LLMs or the agent layers wielding them eventually enable competitors to vastly outperform these little AI models."
"Japan really does know how to do hardware."
"The thing that I think Japan most lacks compared to Silicon Valley is ambition."

HABITS

Japanese programmers rapidly integrate coding AI tools like Claude into daily workflows, mirroring US peers to boost efficiency.
Companies train small classifiers for routine tasks, such as categorizing thousands of client emails monthly, to streamline internal operations.
Forward-thinking firms mandate top-down AI adoption among employees, pushing generative tools into everyday corporate processes.
System integrators collaborate closely with clients to customize AI models, replicating specific human actions like power plant operations.
Executives view AI primarily through a productivity lens, habitually deploying it to cut costs rather than exploring revenue-generating applications.

FACTS

Japan's digital deficit hit about $24 billion in the first half of 2025, nearly two-thirds of the full-year 2023 total of $37 billion.
Half of Japan's nuclear power plants remain idle due to post-Fukushima regulations, limiting potential for large-scale AI data centers.
A planned Toyama data center cluster offers 3.1 gigawatts of future capacity, dwarfed by projects like Stargate's 10 gigawatts.
Preferred Networks developed the MN-Core chip series starting in 2017, pivoting post-ChatGPT to 3D-stacked L1000 processors ready by 2027.
Japan's high school students rank highly in international math, science, and problem-solving assessments, indicating strong raw AI talent potential.

REFERENCES

Richard Katz's Substack "Japan Economy Watch," which analyzes Japanese firms' cost-focused approach to information technologies.
PyTorch AI framework, inspired by Preferred Networks' earlier Chainer framework.
Startups including Matlantis and MI6 for AI materials discovery, Eiktol from Nagoya University, and Sakana AI's AI scientist tool.
GenIAC project by METI for cost-sharing in generative AI training, enabling models like Rakuten's.
Preferred Networks' Playo Translate model, adopted by the Japanese government for administrative documents.

HOW TO APPLY

Assess current AI tools in your organization by surveying usage for tasks like image generation or translation, then integrate foreign models like ChatGPT to benchmark domestic gaps and accelerate adoption.
Identify underutilized processes, such as manual form filling or email sorting, and pilot small classifiers or online wizards to automate them, measuring time savings to build internal buy-in.
Partner with system integrators to customize niche AI models for specific operations, like replicating operator actions, while evaluating scalability against global LLMs to avoid over-customization traps.
Leverage existing data assets in sectors like materials or pharmaceuticals by collaborating with startups such as Matlantis, training domain-specific models to uncover new discoveries and revenue streams.
Advocate for compute expansion by exploring nuclear reactivation or custom hardware like MN-Core chips, starting with small data center pilots to run diverse open-source models for inference services.

ONE-SENTENCE TAKEAWAY

Japan's AI boom shows promise in hardware and talent but urgently needs bolder ambition to rival global leaders beyond conservative, niche deployments.

RECOMMENDATIONS

Shift corporate mindsets from viewing AI as mere cost-savers to revenue innovators, exploring applications in materials discovery to grow markets.
Invest in custom hardware like Preferred Networks' chips to sidestep Nvidia dependencies, building inference platforms for both proprietary and open-source models.
Encourage top talent retention through competitive incentives and quality-of-life perks, fostering a domestic AI research ecosystem over US migration.
Prioritize global data integration in training over strict sovereign AI limits, adopting best-in-class tools like ASML equivalents for competitive edge.
Expand government R&D funding into commercialization support, running competitive programs that reward ambitious, high-impact AI productization.

MEMO

Japan's Quiet AI Awakening, Tempered by Caution

In the shadow of America's flashy AI spectacles and China's relentless scaling, Japan has been crafting its own path in artificial intelligence—one marked by hardware prowess and incremental innovation, yet hampered by a reluctance to dream big. During a recent visit to Tokyo, tech analyst Jon Y of Asianometry sat down with industry leaders, startup founders, and policymakers to gauge the pulse of this underappreciated boom. What emerged was a landscape buzzing with potential but constrained by cultural conservatism and a digital deficit siphoning billions abroad.

Consumer adoption tells a familiar story: generative AI hype permeates Japanese life, from taxi ads featuring enigmatic rabbit puppets to social media flooded with ChatGPT-fied Studio Ghibli art. Programmers here, much like their Silicon Valley counterparts, lean on tools like Claude for coding efficiency, and marketers generate images to slash costs. Yet, the tools are overwhelmingly imports—no homegrown rival to ChatGPT dominates the scene. Corporate deployment lags further, with even large HR firms boasting of automating tax forms via simple wizards, feats that feel decades overdue. Generative AI? Often dismissed as secondary to human judgment, a mindset that could delay widespread integration for years.

Development efforts reveal deeper structural echoes of Japan's software past: an army of system integrators peddling bespoke AI for tasks like mimicking power plant operators. These niches pay well but risk "Galapagos syndrome"—isolated, maintenance-heavy solutions outpaced by versatile large language models. Revenue views compound the issue; Japanese executives eye AI for productivity gains, not bold expansion, unlike their American peers chasing disruption. Promising bright spots include startups like Matlantis in materials discovery and Preferred Networks' custom MN-Core chips, poised for LLM inference by 2027, dodging Nvidia's premiums.

Government intervention adds layers of support and subtlety. Programs like METI's GenIAC share costs for training models, enabling efforts from Rakuten, while "sovereign AI" pushes localized data and infrastructure for resilience amid geopolitical tensions. Yet this inward focus may stifle global competitiveness—Japan's 1980s chip decline stemmed from similar loyalty to fading domestic tools. Talented youth, ranking high in math and science, have already influenced global frameworks like PyTorch via homegrown Chainer, but many top minds flee to U.S. salaries.

Ultimately, Japan's AI trajectory hinges on injecting ambition into its meticulous execution. With a digital deficit swelling to $24 billion in early 2025 and idle nuclear plants holding untapped power, the nation could leap forward by embracing the best technologies worldwide, much like Rapidus did with ASML lithography for its 2nm fabs. Without that spark—beyond outliers like SoftBank or Sakana AI—the boom risks fizzling into efficiency tweaks rather than transformative leaps.