The decisions about where, how, and when to build AI datacentres being taken right now will shape AI development for decades. One entrepreneur who has spent years building at that physical layer explains what most of the industry is not yet asking.
There is a version of the artificial intelligence story that gets told often, and it is almost entirely about software. About models and benchmarks and the latest capability breakthrough from a laboratory in Silicon Valley, Bengaluru, or Beijing. It is a compelling story, and it is not wrong. But Saksham Kukreja thinks it misses something more fundamental: before any model runs, someone has to build the physical infrastructure that makes it possible. The decisions being made right now about where that infrastructure goes, how it is powered, and who it serves will shape the AI era long after today’s benchmark leaders are forgotten.
Kukreja is co-founder of Bitgrit, an AI infrastructure company whose work has taken him across three continents and into conversations with governments, sovereign funds, and industrial conglomerates. Most recently, he led his company to partner with the City of Akita in northern Japan to help architect what is planned as one of the country’s largest AI data centre facilities with a capacity of 500MW. His perspective is particularly relevant as India doubles down on building this physical foundation and makes decisions that will be difficult to reverse.
“Everyone is focused on which country has the best AI talent or the largest compute budget. But the more consequential question is which countries are making serious physical infrastructure decisions right now, before those decisions become urgent. By the time urgency arrives, the leverage is already gone.”
The variables that matter before anything else
When Kukreja evaluated Akita, he was not starting from a map and working outward. He was working from a checklist of physical requirements that most public conversation about AI infrastructure never reaches.
“Akita sits close to the landing point of an Arctic subsea cable connecting Japan to Europe by the shortest and most secure route available, bypassing the congestion points of the Red Sea. It has a renewable energy mix of offshore wind and hydroelectric generation. Its cold climate enables year-round ambient cooling that dramatically reduces operational costs. No single factor was sufficient. The combination is what made it worth building toward.”
The implication, he notes, is that genuinely viable locations for frontier AI infrastructure are considerably scarcer than the current volume of data centre announcements suggests. Countries that leave the question to market forces will find themselves dependent on whoever answered it first.
A new kind of bilateral conversation
AI datacentres have become a national priority for nearly every serious economy. But Kukreja has been watching a dynamic emerge, particularly among Gulf states, that points to where infrastructure strategy is heading.
“These are countries with enormous capital, genuine strategic ambition, and governments that move fast. What they cannot change is their climate. The structural cost of running datacentres in extreme heat is a geographic fact, not an engineering problem. And at the scale of serious AI compute, that fact shapes every operational decision.”
The response he has been exploring with partners in the region is a concept he calls a sovereign data embassy: a facility hosted in another country’s territory but operating under the legal jurisdiction and governance frameworks of the nation that owns it. Estonia used a version of this framework for its digital infrastructure years before AI made the idea strategically urgent. Whether the model translates cleanly to the scale and complexity of AI compute remains an open question.
A related question, one Kukreja has been pushing on in the Akita project, is whether public and private sectors build this infrastructure together or separately.
“Governments have the legitimacy and the long planning horizons. Private companies have the speed and the cross-jurisdictional relationships. The infrastructure that actually matters gets built where those two things meet.”
The Mayor of Akita City has welcomed the project, saying the data centre will bring companies together to use resources effectively and develop initiatives aligned with the city's urban plan. Whether Bitgrit can hold that public-private convergence together over the long build horizon remains to be seen. But if it can, the model may be worth watching.
Infrastructure for the many, or the few
There is a question Kukreja thinks most infrastructure conversations never reach: once the facility is built, who does it actually serve?
The default answer, in most capacity being developed today, is the organisations that were in the room when the decisions were made. Hyperscalers. Sovereign AI programs. Large enterprises with long-term contract capacity. The physical layer gets built, and the access it provides reproduces the concentration that already exists at every other layer of AI development.
Kukreja believes access to this AI infrastructure should somewhat be open to the masses. He is candid that his own position here is not neutral.
“I might be biased. Democratising AI has always been at the core of what we do. But I’ve watched thousands of developers on our platform build models that ended up deployed by NASA and SoftBank. They are people with no institutional affiliation and no conventional route to those clients. That happened because the platform was designed to let it happen. Infrastructure doesn’t do that by default. It has to be a choice.”
For Akita, that choice is embedded in the plan from the start. Alongside the data centre, the initiative includes an education and developer program designed to build local AI capability rather than simply import demand from elsewhere. The engineers it trains will work on the same infrastructure that their region helped build.
“Infrastructure that concentrates access produces concentrated outcomes. We are at the moment where that design choice still exists. It will not exist indefinitely. And once it is made by default, it is very hard to unmake.”
It’s becoming clear that before the models, we need the map. And the decisions being made on that map right now, about where to build, how to power it, who governs it, and who gets to use it, will outlast every benchmark being celebrated today. India has the talent, the policy intent, and the investment. What remains to be seen is whether those pieces come together before the window to shape that foundation on its own terms closes.
