Hello!

Between 2025 and 2026, AI-related investment in Japan has been moving fast.

Domestic LLM development, a data center construction boom, expanding government support programs. The sense of urgency—"Japan cannot afford to fall behind in AI"—shows up clearly in the numbers.

This momentum is welcome, and it is far better than doing nothing. That said, as we spend our days

on the front lines of AI adoption, something occasionally feels slightly off.

The budgets are moving. The will is there.

But is this really the right direction?

Of course, no one knows the future.

Still, when you line up the publicly available data, you find several pieces of evidence suggesting it might be worth pausing to think for a moment.

In this two-part series, we organize that evidence.

Part 1 covers three themes: domestic LLMs, data center investment, and data sovereignty.

Part 2 looks at the structural shift behind "SaaS is Dead" and considers what positioning strategies are possible in this environment.

Chapter 1: Where Domestic LLMs Stand Today — Let's Talk About Scale

Domestic LLMs developed by major Japanese telecom carriers and IT vendors tend to be relatively small to mid-sized in parameter count, with designs that clearly target on-premises and closed-network environments.

"Lightweight," "low cost," "runs on a single GPU," and "on-premises ready" are the selling points, and the direction itself—building practical products with limited resources—is a rational one.

Meanwhile, there are also more ambitious efforts. Several startups and research institutions, supported by the GENIAC (Generative AI Accelerator Challenge) project, have developed 100B (100 billion) parameter-scale models entirely from scratch.

GENIAC is a framework led by the Ministry of Economy, Trade and Industry (METI) and NEDO to strengthen Japan's generative AI development capabilities, centered on securing compute resources, supporting foundation model development, and accelerating real-world deployment (METI / NEDO).

What is distinctive about this "domestic LLM movement" is that its institutional design also provides policy-level support for compute resources, data, and knowledge sharing.

Some developers supported by GENIAC have reported high scores on certain Japanese-language benchmarks (for example, Jaster's 0-shot and 4-shot evaluations).

There are also cases showing strong results on evaluations specialized for Japanese business domains. Achieving this with limited resources is genuinely impressive. That said, an advantage on benchmarks does not immediately imply parity in general-purpose reasoning capability. Results vary greatly depending on evaluation conditions (target tasks, prompt settings, dataset composition), and there is considerable distance between scoring high on specific Japanese tasks and matching frontier models in general-purpose performance across languages and domains. Overseas frontier models are estimated to be in the hundreds of billions to one trillion parameter range. Parameter count is not everything, but the gap in capital and compute investment is large, and competing head-on in general-purpose performance is, realistically, a steep challenge.

The Investment Gap — Getting a Sense of Scale

Why did this gap emerge? The blunt answer is that the invested capital differs by orders of magnitude.

US hyperscalers have all issued guidance for substantially increased AI-related capital expenditure (CAPEX) in 2026.According to Reuters, guidance from multiple companies suggests total 2026 investment could reach the $600 billion range.

At the same time, the market reacted negatively to Amazon's 2026 investment plan (guidance of roughly $200 billion), and its stock price fell. In other words, even the US players are not certain this level of investment is sustainable, and investor concerns remain strong. One caveat is needed here: "US CAPEX" and "Japanese policy support" are numbers of a different nature.

The US figures are private-sector capital expenditure (data center construction, GPU procurement, power infrastructure, and so on), while the Japanese figures are policy support—subsidies, funds, institutional frameworks, and procurement—led by programs like GENIAC. A direct comparison can be misleading, but it is still useful for getting a sense of the scale of funding flowing into AI infrastructure.

With that caveat, Japan's AI-related government support—including GENIAC, the government AI pilot project "Gennai," the government guidelines for generative AI use, and several other programs—totals in the hundreds of billions of yen. Compared with US hyperscalers' annual CAPEX (a projected $600 billion), the figures differ in nature, but the amount of capital that can be mobilized for AI infrastructure differs by more than two orders of magnitude.

Looking at these numbers, the "lightweight, low-cost" strategy that domestic LLM vendors have taken looks more like a level-headed judgment than anything else. That a head-on collision is unlikely to be rational is something the vendors themselves probably understand best.

The question is how to accept this reality and where to place the next bet.

The Revenue Behind the "Domestic LLM Business" — A Peek Inside the Numbers

One major domestic telecom carrier has announced quarterly orders of 67 billion yen and a cumulative total of more than 1,800 orders for its AI business built around its own LLM.

It projects 150 billion yen for the full year and over 500 billion yen within a few years. Impressive numbers indeed.

However, these order figures include not only LLM license revenue but also consulting, systems integration (SI), infrastructure build-out, and operations and maintenance. The LLM is the "hook" that wins the deal, while most of the revenue comes from conventional full-stack SI. Another major vendor has set a target of "over 50 billion yen in sales over three years" for its own LLM, but how much of that is standalone LLM license revenue is unclear from public disclosures.

Earning revenue through SI is a perfectly respectable business. Still, it is worth calmly noting that many of the numbers cited as the "commercial success of domestic LLMs" are, in substance, the "success of SI with an LLM as the hook." If you have been in the IT industry for a while, this pattern may look familiar. Every time a new technology appears, SI projects bearing its name emerge, and the numbers get counted as the "market size" of that category.

It happened with cloud, and it happened with big data. The name of the technology changes but the structure does not—one of the curiosities of Japan's IT industry.

That said, the lightweight, low-cost strategy of domestic LLMs is not meaningless either. Demand for closed-network environments certainly exists. There is data showing that a majority of inquiries about one company's own LLM ask for closed environments.

It is a fact that the pitch "a domestic model is necessary for security" resonates in a certain market.

The question is where the ceiling of that market lies—and how long "industry-specialized lightweight models" can remain a point of differentiation as general-purpose LLMs keep improving. We will return to this point a little later.

The Delicate Relationship Between the SIer Structure and AI Adoption

Let us touch on one more topic that rarely gets discussed head-on.

So far we have looked at upbeat numbers such as "order value" and "deal count" for domestic LLMs. Here, let us deliberately shift the perspective.

When the commercial success of domestic LLMs is discussed, the figures are often tallied not just as "model licenses" but as orders for a full adoption package—consulting, requirements definition, system construction, infrastructure build-out, and operations and maintenance.

In other words, the success or failure of domestic LLMs depends less on the quality of the model alone and more on "the ability to embed it in corporate front lines and tie it to business outcomes."

And in Japan, the players responsible for that on-the-ground adoption are, in most cases, SIers (systems integrators). That is why the structure of the SI industry is not a side note to the domestic LLM discussion—it sits at its very center.

Japan's IT industry has a deeply rooted multi-tier subcontracting structure of primary contractors, subcontractors, and sub-subcontractors. Often called the "IT general contractor" model, this structure—where major SIers win contracts while actual development and construction are handled by multiple layers of partner companies—has long been debated as a uniquely Japanese business practice.

The structure has arguably worked for "running large projects stably," but when it comes to AI adoption, it introduces a different kind of difficulty.

One survey points out that most of Japan's IT engineers are concentrated on the vendor side, with a relatively low ratio of in-house engineering at user companies. Recent reports also suggest that business conditions for small subcontractors are getting tougher.

How this structure will play out in the AI era is, frankly, a thorny question.

What really matters in AI adoption is not simply "deploying" a model.

It is understanding the business deeply, then deciding which model to apply to which task, what data to feed it, how to evaluate it, how to handle exceptions, and how to keep improving it in operation—in short, assembling "business context × design × validation × operations" as a single whole.

But in a multi-tier subcontracting structure, that integrity breaks down easily. The primary contractor closest to the customer understands the business but has difficulty getting into the details of models and implementation; the technically capable subcontractor is strong at implementation but rarely gets to touch the depths of the customer's business.

Business context and technology end up split across separate organizations, making it easy for projects to move forward "with design and validation left thin."

The observation that upstream-side staff, buried in management and paperwork, struggle to accumulate hands-on implementation and evaluation-design skills is not unrelated to this split structure.

What does this split cause?

The most likely outcome is that AI adoption becomes a project that merely "traces the conventional SI process carefully."

The process itself—requirements, design, build, test, operate—is not the problem. The problem is the risk that what AI adoption genuinely requires—model performance validation, evaluation metric design, iterative improvement from field feedback—gets pushed outside the process and absorbed into a purely formal rollout. As a result, the client company never develops internal AI capability, dependence on outsourcing becomes entrenched—and the project gets marked "complete" without AI ever taking root as an organizational capability.

Structurally, this pattern is all too easy to fall into.

Incidentally, while we are not a contract development firm, we tackle AI-driven development of our own software head-on, and we believe that gives us a head start on this problem—we also offer consulting services in this area.

Now, back to the main thread.

What makes things even trickier is that SIers' revenue model and AI's productivity gains are structurally prone to collide. Most SIers have long run businesses based on person-month billing—put bluntly, a model where "more hours means more revenue."

AI's value, on the other hand, tends to show up as "fewer hours, more output."

The same "productivity gain" that a user company welcomes can be a conflict of interest for the SIer's short-term revenue logic. Of course, human roles—quality assurance, requirements clarification, process management, operations design—remain significant on the ground, and AI cannot automate everything. But the more AI streamlines implementation and operations, the harder it becomes for "stacking up person-months" to remain the center of value.

There is a structural tension here.

It is natural to read the major SIers' push into consulting as a response to precisely this tension.

Indeed, it has always been the case that involvement through SI attracts consultations in adjacent areas, generating consulting-like value.

But there is a pitfall here too—one that invites overreach.

The role user companies have sought from SIers has mostly been "the company that builds our systems," and their brand as "the company that proposes strategy" is not necessarily strong. So it is dangerous to assume that the further upstream SIers move, the more they will succeed. The transition takes time, and depending on how customers receive it, the value may not be recognized as much as hoped. This should honestly be left standing as an uncertainty. Of course, there must be many engagements that work well, and many companies for which SI-style support is indispensable. There is a real gap between "deploying" AI and "implementing it in society," and one can argue that filling that gap is where SIers' reason for being lies.

The point here is not that "SIers are bad." Rather, the issue is that the bottleneck determining the outcome of domestic LLMs and AI investment lies not only in model performance or GPU counts, but also in "fragmented implementation" and "misaligned incentives." The more domestic LLMs sell as full adoption packages, the more that implementation structure determines outcomes. That is why this risk deserves a place among the central issues, not in the wings.

On the Culture of Technical Verification

In 2025, one of Japan's most-watched AI startups became the subject of debate over a technical claim. After it published a paper claiming dramatic speedups from automated GPU code optimization, overseas researchers pointed out that the claimed improvements could not be reproduced in external verification, and the company found itself having to explain. We will not judge the specifics here. What matters is that reproducibility and verification mechanisms determine the credibility of the entire ecosystem.

In Japan's AI sector, flashy announcements and large funding rounds make the news easily, while the culture of third-party technical verification and replication experiments is still maturing.

A cycle in which the community verifies published papers and openly debates any problems found would be just as important to the health of the ecosystem as raising capital.

The Problem of Differentiation Melting Away

Finally, one more structural challenge facing domestic LLMs. Domestic LLM vendors differentiate on "lightweight and cheap," "strong in Japanese," and "industry-specialized."

But general-purpose LLMs are evolving frighteningly fast.

They are rapidly improving on industry-specific tasks as well, which risks eroding the differentiation of "industry-specialized lightweight models."

On top of that, open-source models—Meta's Llama chief among them—are improving remarkably. Once freely available open-source models match or exceed "lightweight, cheap domestic models," the cost pitch loses its bite as well.

As for the pitch that "a domestic model is necessary for data sovereignty," given that closed-network, dedicated, and in-country-region deployment options exist, the logic that "data cannot be protected without a domestic model" is not necessarily accurate.

We cover this in detail in Chapter 3. Of course, there are domains—defense and intelligence, for example—where "domestic at every stage of the supply chain" is mandatory, and development aimed there is meaningful. But for most ordinary companies, how often "it must be a domestic model" actually applies deserves a cool-headed look.

Chapter 2: What Is Inside Data Center Investment — A Look at the Substance of Growth

The Numbers Look Great. But Look Inside...

Japan's data center market is certainly booming. Multiple research firms project the market roughly doubling through the late 2020s. As for power consumption, Wood Mackenzie's analysis projects more than a threefold increase from 19TWh in 2024 to 57–66TWh by 2034, and the IEA likewise notes that data centers will drive more than half of Japan's electricity demand growth over the same period. Judging by the numbers alone, it looks like "Japan's AI industry is growing rapidly."

But break down what is behind this demand, and a somewhat different picture emerges.

Foreign Cloud Region Expansion Is Driving Construction Demand

Among operational data center racks in Japan, the share of hyperscale racks (serving major overseas cloud providers) keeps expanding and is projected to exceed half by 2028.

Meanwhile, the number of retail-type racks (conventional data centers used by domestic companies) has stayed roughly flat. Major overseas cloud providers have announced large-scale investments in the Japanese market.

AWS has announced a plan to invest 2.26 trillion yen in cloud infrastructure in Japan through 2027, and Oracle has committed more than $8 billion in Japan over ten years.

In other words, one of the main drivers of Japan's data center construction boom is foreign providers expanding their regions into Japan, with Japanese companies using cloud services on top of them. That creates construction jobs and power demand, but whether it amounts to autonomous growth of Japan's AI industry is a somewhat different question.

Recall the US figures above. Compare a world where hyperscaler guidance implies roughly $600 billion of investment in 2026 alone with Japan's entire data center market, and the difference in scale is considerable.

Power and Grid Bottlenecks

Whenever data centers come up, the power problem comes up with them.

Gartner forecasts that global data center power consumption will double from 448TWh in 2025 to 980TWh in 2030.

Of that, consumption by AI-optimized servers is estimated to grow from 93TWh in 2025 to 432TWh in 2030, accounting for 44% of total data center consumption.

Japan has its own particular problems. Data center demand is concentrated in the Tokyo and Kansai metropolitan areas, where data centers are expected to account for roughly 7% of electrical load by 2030. Hyperscalers typically aim to deploy data centers within five years, but building the gas turbine power plants that supply them can take up to ten. This time lag has already become a delay factor for some large projects.

Grid build-out is another challenge. As the Agency for Natural Resources and Energy's strategic energy plan and OCCTO's long-term grid plan indicate, reinforcing the transmission infrastructure that carries renewable power from rural areas to cities takes many years. The mismatch between data center construction speed and power infrastructure build-out is a challenge shared worldwide, not just in Japan—but in Japan's case, delayed renewable adoption makes the situation even more complicated.

What a Swing in GPU Supply and Demand Reveals

Apart from the long-term challenges of data center investment, there has been a recent case in which the GPU supply-demand balance swung sharply.

One domestic cloud provider, with subsidies from METI, deployed a large fleet of the latest GPUs and rode the wave of AI demand to rapid growth in both revenue and profit.

Demand was so strong that capacity "sold out instantly" upon deployment, making the company a symbol of the AI boom. The following fiscal period, however, a large contract it had expected to continue came to an end, and the company sharply cut its GPU service revenue forecast.

Its operating profit forecast was cut by more than 90%, and the stock fell about 20%, hitting its daily limit.

The company's CEO reflected that "leaning too heavily on physical GPU servers was the structural problem."

This is one company's case, and generalizing it into "GPU demand across Japan is weak" would be inappropriate. But at minimum, it suggests that domestic AI GPU demand can be concentrated in specific large projects, and that swings in utilization can shake earnings hard.

We operate our own GPU server cluster for training and inference workloads, and frankly we do not have the capacity to keep hundreds of H100s running around the clock. We often find ourselves wondering who could possibly use—and afford to use—that many GPUs.In a separate post, we noted that training GPT-class LLMs is on a different order of magnitude entirely—far beyond what a venture company can realistically take on.

Recouping the investment will depend on whether a stable, ongoing base of demand takes shape.

Cloud-Based Usage Dominates — With Caveats

Most Japanese companies tend to use AI "through the cloud" rather than "procuring and running GPUs themselves."

If all you do is call the ChatGPT or Claude API, you do not need your own GPU server cluster.

To repeat: procuring large numbers of GPUs in-house to train models is for a limited set of players—foundation model developers and large research institutions.

That said, declaring that "all the compute flows overseas" is not accurate either.

As overseas hyperscalers expand their Japanese regions, part of the inference workload will run in domestic data centers. And if demand for closed-network usage grows, a certain amount of domestic data center demand could emerge. To put it

more precisely, how the investment gets recouped depends not only on the formation of steady domestic corporate demand but also on the hyperscalers' Japan region strategies.

Whether Japan's data center investment is "for Japan's AI industry" or "for overseas hyperscalers' Japan expansion" is hard to judge until we see how the demand structure evolves.

Chapter 3: Thinking About "Data Sovereignty"

The Need for Closed Environments Is Real

"We don't want to send confidential data to overseas clouds." "We want to avoid the risk of internal information leaking through AI." Finance, healthcare, government, manufacturing design data—the demand for closed environments in these domains is real and should not be dismissed.

There is data showing that a majority of inquiries about one domestic vendor's own LLM ask for closed-environment usage.

The Digital Agency is also running a pilot for government use of domestically developed LLMs (the government AI platform "Gennai"), and models meeting certain security requirements are called for when handling highly confidential government information. Up to this point, the problem framing is entirely correct.

It Is the Leap That Follows That Concerns Us

What concerns us is how often the argument leaps, in a single bound, from this problem framing to the conclusion "therefore we should develop domestic foundation models."

Think about it calmly: what companies want to protect is "their data," not "the nationality of the model." The two look similar but are entirely different.

What determines safety is not "the model's nationality" but the mechanisms that control "where the data goes, how it flows, and who can touch it."

Getting slightly technical: the means of achieving data control are not limited to "using a domestic model."

For example, the following options exist. First, deploying a high-performance open-source model in your own environment (on-premises or a private cloud) keeps the data under your control.

Second, choosing a major cloud provider's closed-network, dedicated, or in-country-region deployment options can pin down where the data resides and, in some cases, let you use LLMs in environments physically isolated from other customers' workloads.

Third, even when using APIs, combining monitoring, masking, and control technologies for input and output data lets you build a mechanism that blocks confidential information before it is ever sent to the LLM.

In short, between the need to "use it in a closed environment" and the conclusion "it must be a domestic model," multiple technical options exist.

If anything, the risk that being locked into a lower-performing model undermines the very effectiveness of AI adoption cannot be ignored. When model quality drops a few notches, it can fall below the line of being usable at all—anyone working in the field will recognize this feeling. "Safe but unusable" defeats the purpose.

Data Control Mechanisms Are the Essence

The point is this: the essence of protecting data sovereignty is not which LLM you use, but having mechanisms to know and control where your data goes and how it flows. Detect whether input data contains confidential information. Monitor output for unintended leaks. Block when problems arise. Having this "data control layer" may be a more effective approach than fixating on the model's nationality.

Our AI security solution LLM-Audit targets exactly this problem.

"It's a domestic model, so it's safe" is a strange piece of logic on closer inspection: even a domestic model sends data outside if you put it in the cloud, and even on-premises, internal misuse cannot be prevented without proper auditing. Conversely, even with an overseas model, the risk of confidential information leaking is manageable with proper data controls in place.

None of this means "domestic models are pointless." The proposal is that the industrial-policy goal of "nurturing domestic AI" and the security requirement of "protecting data" should, by rights, be treated as separate issues.

Conflate the two, and the discussion stops connecting—and it becomes harder to arrive at the best solution.

Part 1 Wrap-Up

Let us organize what we have seen so far.

Domestic LLMs are mostly small to mid-sized models for on-premises and closed-network use, with several players taking on 100B-class models under GENIAC support.

High scores have been reported on some Japanese benchmarks, but the gap in capital and compute makes it hard to match overseas frontier models in general-purpose performance.

Much of the revenue in the "LLM business" is effectively SI business, and how the SIer-specific multi-tier subcontracting structure and person-month billing model affect AI adoption warrants close attention.

Data center investment: the market is projected to grow, but one main driver is overseas hyperscalers expanding their Japan regions. There has been a case where GPU supply and demand swung on the back of specific large deals, and whether a broad base of steady demand forms remains to be seen.

Power and grid bottlenecks also remain as long-term challenges.

Data sovereignty: the underlying problem is real. But the means of protecting data are not limited to "developing domestic models"—options include deployment configurations and building a data control layer. Separating the industrial-policy debate from the security debate matters.

None of this is meant to say "Japan is hopeless" (in fact, if you pretend the US and China do not exist, Japan's AI environment looks quite good). We have laid out these facts as material for thinking, with an accurate view of reality, about "where, then, is it smart to place our bets?"

In Part 2, we will look at the "SaaS is Dead" structural shift underway in the software industry and consider what positioning strategies are possible in this environment.

→ Part 2: "Becoming a Nation That Wins with AI" is next.

Appendix

Definitions

CAPEX (capital expenditure): Private-sector investment in data center construction, GPU procurement, power infrastructure, and the like. Accounting definitions and scope may vary by company.

Policy support: A collective term for subsidies, funds, tax incentives, government procurement, guideline development, and so on. Different in nature from CAPEX.

Data sovereignty: A concept that includes not only where data resides but also access control, auditing, and management across the entire supply chain.

Sources and References

GENIAC (overview, institutional design, selection)

• METI, "GENIAC Phase 3 (Compute Resource Support) Selection Press Release (2025-07-15)"
• GENIAC Official Portal (METI)
• NEDO GENIAC-related calls for proposals

AI investment scale of US Big Tech (2026 outlook and market concerns)

• Reuters (2026-02-06) "Big Tech's $600 billion spending plans exacerbate investors' AI headache"
• Financial Times (2026-02-06) "Big Tech's 'breathtaking' $660bn spending spree reignites AI bubble fears"

Japan's data center investment, demand structure, and power constraints

• Reuters (2025-12-19) "Japan's biggest data centre hub planned in Toyama prefecture"
• Wood Mackenzie (2025-08) "Japan's data centre gold rush: the battle to power a data-driven future"

Investment in Japan by major overseas cloud providers

• AWS, "Plan to invest 2.26 trillion yen in cloud infrastructure in Japan (through 2027)" (2024-01-19)
• Reuters (2024-01-19) "Amazon's AWS to invest $15 billion to expand cloud computing in Japan"
• Oracle, "Over $8 billion investment in cloud and AI in Japan" (2024-04-17)
• Reuters (2024-04-17) "Oracle to invest over $8 billion in Japan in cloud computing, AI"

Government use of generative AI (institutional and operational overview)

• Digital Agency, "Government AI / Generative AI Environment 'Gennai' Explainer"
• Digital Agency, "Guidelines on Procurement and Use of Generative AI for the Evolution and Innovation of Government" (2025-05-27)
• Digital Agency, "Generative AI Usage by Digital Agency Staff"
• Digital Agency, "FY2024 Technical Verification and Environment Development for Business Use of Generative AI"