Don't bet on the ASI timeline — capability can be bought, the foundation can't

"Superintelligence (ASI)" is a phrase you will have seen far more often in the last six months. Until 2024 the industry's watchword was AGI — human-level general AI. Now the major labs speak, in unison, of a goal beyond it: superintelligence. It makes for dramatic news. But to an executive, the question that matters is not when it will arrive. It is whether you are already making the moves that pay off whether it arrives or not. That is what this piece is about.

We set out the definitions of ASI and AGI — and the four routes AGI might take to get there — in the previous piece, AGI is not the goal — DeepMind's path to superintelligence (ASI). This is the sequel: why the term is suddenly everywhere, what it means for your business, and what to start building today.

Why "ASI" became part of the vocabulary

Three currents have converged.

• The label moved up. OpenAI said its sights are now set on a superintelligence beyond AGI; since then Meta ("Superintelligence Labs") and Microsoft ("Humanist Superintelligence") have written superintelligence into the names and missions of their organizations — all on the record.
• The scale of money and talent. Safe Superintelligence, with nothing shipped, was reportedly valued near $32 billion. Anthropic went from $183 billion to $380 billion in early 2026, and has reportedly risen further since. OpenAI was valued at $852 billion in a funding round that same year (reported; some figures officially announced). Superintelligence is at once a technical goal and a banner for gathering capital and talent.
• The opposition grew louder too. A statement calling for a halt to superintelligence development "until there is broad scientific consensus that it can be done safely" was published in October 2025 and has gathered more than 70,000 signatures, including leading AI researchers. From both sides — for and against — superintelligence has become a matter of public debate.

"It's near" and "it's a bubble" come from the same people

Lab leaders say that within a few years AI will surpass humans across many domains. The same figures also warn that some investors will get badly burned — an acknowledgment that a bubble is possible. Both remarks are made in the middle of enormous fundraising and IPO preparation. The safest reading is not to treat these as verdicts, but simply to note that this is what they said.

On the engineering side there is caution too. Some argue that today's large language models will plateau even if you keep scaling them. Apple published a paper arguing that reasoning models hit clear limits; others countered that the real problem was how the models were evaluated. The exchange itself is a good illustration of how hard it is to measure what AI can really do.

History is instructive. The forecast that "general AI is ten to twenty years away" has been made, and missed, repeatedly for decades. Don't stake limited resources on the will-it-won't-it prediction game.

Even the giant labs struggle with organizational design

Money and talent alone do not seem to close the gap to superintelligence. Even Meta, for all its resources, has reportedly restructured its new superintelligence group more than once within months, with staff cuts. We treat these accounts as reporting, not established fact — but the pattern is telling: money and talent alone don't solve organizational design.

The same picture appears inside companies. Adoption of generative AI has spread, but examples that connect through to the bottom line are limited. One study reported that roughly 95% of corporate generative-AI pilots produced no measurable results. The methodology has drawn criticism, but the trend is broadly consistent with other studies. What decides success is less the technology itself than the design of where, and how, you embed it in the work. That leads to the next point.

What executives should hold is the foundation, not the prophecy

From a business standpoint, one thing carries high confidence. Per-unit, foundation AI models keep getting cheaper and steadily more capable — though total spend is another matter once usage grows. The staggering capital investment by each lab is what underwrites this. In other words, the raw ability of the model itself is becoming something anyone can buy on equal terms.

When that happens, the difference lies not in the model but in what surrounds it: the business data only you hold, the mechanism that grounds AI in your real operations, and the operating loop that grows smarter the more it is used. That is why the only things you cannot buy are your own data and the way AI is grounded in your operations. A company with the foundation can swap in new capability cheaply; a company without it rebuilds from zero each time. The more capability rises, the wider this gap opens.

Regulation is part of the foundation too. The EU's AI Act carries real penalties and can reach companies operating in Japan — foreign-owned ones included — if you place AI systems on the EU market or your AI's output is used there. Japan's own regime is non-binding guidance for now, but in practice the pressure lands as requirements from your customers and their supply chains. It isn't something to defer.

Nihonbashi AI Lab supports building this foundation along four axes.

• Business design. Rather than wait for more capable AI and then start thinking, map the roles and handoff points between people and AI in your workflow now. Once the grounding is decided, new capability is a matter of swapping in.
• Data structure and access design. Put your organization's know-how and outside data into a form AI can read and people can use. This is where compounding returns are largest — and the part most often deferred. In an era of multiple AIs at work, where you draw the lines on who can reference what becomes the boundary of your business risk.
• AI use design. Don't bet on a single strongest model; bundle several AIs by role and keep models swappable. Avoid over-dependence on any one vendor, and keep the surrounding design in your own hands.
• Operations. Build the loop that grows smarter with use. Store AI's judgments, failures, and corrections as the organization's memory, and put them to work next time. The organizations that turn this loop fastest gain the most as capability rises.

A small step you can take this week

It takes no large investment. Each of the four axes has an entry point for this week.

• Business design: pick one task you'd like to hand to AI, and write down the handoff points with people.
• Data structure and access design: take stock of where the information that task needs currently lives — in whose head, in which files.
• AI use design: check whether you're locked to one specific model, or built so you can swap.
• Operations: choose one place to record what worked and what didn't.

Rather than predict whether superintelligence arrives, put your company in the position to gain the most if and when it does. As a management decision, that's the one that reliably pays off. We work in both English and Japanese, and if you're unsure where to start on the foundation, we can begin by taking stock of your operations together. Reach us through Contact.