At a glance
Japan runs roughly 3.8GW of data centre IT load in 2026, more than 80% of it in Greater Tokyo and Greater Osaka, and researchers see it heading toward 7GW by 2031.
AWS (¥2.26 trillion), Oracle (US$8 billion) and Microsoft (US$2.9 billion) have pledged the headline platform capital; AirTrunk is the largest independent builder at about 530MW.
Power is the binding constraint, with grid connection running eight to ten years in Tokyo against three to five in Osaka, and data centre electricity demand forecast to more than triple by the mid-2030s.
The Japanese government is pushing capacity out to Hokkaido and Kyushu, with more than 100 regional authorities volunteering to host and new subsea cables landing beyond Tokyo.
Robin Khuda names Japan, Australia, the Singapore-Malaysia zone and India as the region’s four key AI hubs, putting Australia in the same bracket.
Khuda puts Japan in a four-market bracket that includes Australia
In a special interview published by the Kyodo-group wire NNA on 13 July 2026, AirTrunk founder and chief executive Robin Khuda set out where he thinks the region’s AI infrastructure will concentrate. In Asia-Pacific, he told NNA, four markets will carry the cloud and AI build: Japan, Australia, the Singapore-Malaysia zone, and India. Japan, in his account, is strategic beyond its own borders, the anchor for demand across North Asia.
That framing is the reason the interview matters to an Australian reader. Khuda places Australia in the same top tier as Japan, the market where AirTrunk has made its largest capital commitment outside its home base. It is the same investment logic we traced when Khuda used Narendra Modi’s Melbourne visit to press Australian super funds on the offshore opportunity, and it sits alongside our reads on India’s US$100 billion build and New Zealand’s NZ$35 billion pitch. Japan is the largest and longest-established of the four, and the market where the constraints AirTrunk manages are sharpest.
Japan’s installed base is large, concentrated, and about to double
Japan is among the largest data centre markets in Asia-Pacific. Market researchers put installed IT load at roughly 3.8GW in 2026, on a path toward about 7GW by 2031, with Arizton and Mordor Intelligence carrying figures in that band. More than 80% of it sits in two clusters, Greater Tokyo and Greater Osaka, with Tokyo alone holding about 41% of national capacity.
The platform capital behind that growth is largely foreign. Amazon Web Services has earmarked about ¥2.26 trillion, close to US$15 billion, through 2027 for Tokyo and Osaka. Oracle has committed US$8 billion over ten years, and Microsoft US$2.9 billion, its largest Japan investment in decades. Google has expanded its Inzai and Osaka footprint. The hyperscalers driving this spend are AirTrunk’s customers, not its rivals, a distinction Khuda drew in the interview: the market is growing faster than even the largest platforms can self-build, and in Japan the land, power and construction complexity makes local delivery partners essential. AirTrunk says it has run more than 100 staff in the country since entering in 2020.
Power, not demand, is the wall every operator hits
Demand in Japan is proven. The constraint is getting electricity to the rack. Grid connection can take eight to ten years in Tokyo against three to five in Osaka, on JLL’s figures, and analysts tracking the sector expect data centre electricity demand to more than triple over the next decade, from under 20 terawatt-hours a year to a figure some place above 60 by the mid-2030s. Utilities have started to respond, with more than ¥150 billion flagged from 2026 to upgrade Osaka-area substations and expand Greater Tokyo’s distribution network.
Khuda framed the power question as the defining one for the industry, and pushed back on the idea that data centres are simply large consumers. AirTrunk targets net zero by 2030 and, by its own account, already sources about 72% of its portfolio from renewables. Khuda argues the sector should be read as grid-supporting infrastructure, investing in the large battery storage that firms intermittent supply, not merely as a load bolted onto the network. That is the same battery-and-firming logic playing out in the Australian market, where firming assets increasingly gate new AI load, as our coverage of the NSW data centre consultation set out.
Tokyo and Osaka are full, so new capacity is dispersing to Hokkaido and Kyushu
Tokyo and Osaka are close to full on land and high-voltage power, so the Japanese government is steering new capacity to the regions. Through METI subsidies, it is backing builds in Hokkaido, centred on Sapporo and Ishikari, chosen for renewable resource, a cooler climate and available land, and in Kyushu, centred on Fukuoka, which offers low-cost and largely carbon-free power, subsea-cable connectivity and proximity to the rest of Asia. When METI opened its regional-dispersion program, more than 100 regional governments put up sites to host facilities. New subsea cables are landing beyond the traditional hubs, including a planned SoftBank station at Itoshima in Fukuoka for the East Asia to North America system, and METI has moved to treat subsea cables as economic-security infrastructure.
Khuda called the dispersion push “very interesting” and gave the technical reason it can work. Conventional cloud is latency-sensitive and clusters near the big cities, but AI training is not, so the viable radius from a demand centre stretches from about 20 kilometres to as much as 200. Regions with spare transmission capacity become live candidates the moment the workload no longer needs to sit next to the user, the same siting logic we set out in our explainer on what an AI data centre is. Dispersion also buys disaster resilience in an earthquake-prone country, a point Khuda made directly.
Why AirTrunk is committing US$8bn to Japan
AirTrunk has been building relationships in Japan since around 2015, and announced a 300MW campus at Inzai in Chiba in 2020, an outlier for scale at the time. It now runs four campuses, flagship TOK1 in east Tokyo scaling beyond 300MW, TOK2 in the Tokyo area at more than 110MW, and OSK1 and OSK2 in Osaka, for a platform heading to about 530MW on roughly US$8 billion of planned investment.
The reasons Khuda gives are structural. Japan is one of the world’s largest IT services markets with room to shift to cloud and AI, and its political and regulatory stability suits long-dated capital. Financing runs through Japanese institutions: by Khuda’s account, around 70 banks lend AirTrunk about A$30 billion globally, and its two largest lenders are SMBC and MUFG. In March 2026 AirTrunk closed a ¥191.6 billion, US$1.24 billion green loan for TOK1, the largest data centre financing completed in Japan, structured under AirTrunk’s green financing framework and led by SMBC and MUFG. On delivery, AirTrunk does not insist on buying land and building alone: its first Tokyo project paired with Daiwa House Industry on land and construction, and Khuda signalled openness to power companies, railways, trading houses and property groups that hold sites and grid access.
The capital behind that build is the part with the clearest Australian read-through. AirTrunk was founded in Sydney and sold in 2024 to a Blackstone-led consortium for about A$24 billion. In June 2026, Blackstone president Jonathan Gray told Nikkei the firm plans to invest US$30 billion in Japanese AI data centres over three to five years and build past 1GW, more than double its current roughly 500MW in the country. That current 500MW is essentially AirTrunk’s platform. AirTrunk’s US$8 billion Japan program is the on-the-ground expression of Blackstone’s US$30 billion Japan bet, and Khuda’s remark that long-dated ownership lets AirTrunk build speculatively ahead of customer contracts maps directly onto Gray’s public commitment.
Japan and Australia, side by side
Japan | Australia | |
Operational capacity | ~3.8GW IT load (2026 est.) | ~1,350MW live |
Outlook | ~7GW by 2031 (est.) | 3,100MW+ by 2030 |
Leading hubs | Greater Tokyo (~41%), Greater Osaka | Sydney, Melbourne |
Platform capital | AWS ¥2.26tn, Oracle US$8bn, Microsoft US$2.9bn | AWS, Microsoft, Google regions |
Largest independent builder | AirTrunk, ~530MW | AirTrunk, NEXTDC, CDC |
Core constraint | Grid connection, 8 to 10 years in Tokyo, 3 to 5 in Osaka | Grid connection and generation |
Government posture | METI subsidies steering builds to Hokkaido, Kyushu | Hosting Certification Framework; social-licence debate |
Siting pattern | Metro-clustered, dispersing by policy | Metro-clustered, customer-led |
Source: Arizton, Mordor Intelligence, METI, company disclosures and the NNA interview, July 2026. Capacity figures are researcher estimates.
The two markets are not competing for the same megawatt. Japan offers scale, deep institutional capital and a policy machine actively opening new regions, against a grid queue measured in years. Australia offers a cleaner grid on several measures, sovereign certification and regulatory predictability, and a market doubling toward 3,100MW. What Khuda’s four-hub framing says is that a serious operator does not choose between them. It builds in both, and moves capital and hardware along the corridor, the pattern behind the regional JLL Asia-Pacific supercycle worth US$772 billion.
What this means for Australia
For Australian operators, investors and policymakers, three points hold. First, the same firm now committing US$30 billion to Japan runs the largest data centre platform in Australia, so Japanese execution and Australian pipeline share a balance sheet and a management view. Second, Japan’s dispersion push is a live experiment in whether policy and AI’s latency tolerance can move gigawatt-scale load away from congested metros. Australia faces the same question from the other side of the ledger: AEMO’s disclosed 5.4GW data centre pipeline put about 60% of it in New South Wales, with Transgrid holding more than 6GW of connection applications around Sydney alone. Japan is steering load out to the regions by subsidy, while Australia’s clustering stays customer-led, so the open question here is whether AI’s latency tolerance moves the marginal campus to regional grids that still have headroom. Third, Khuda’s placement of Australia in the region’s top four is a market signal from the operator with the largest stake in the outcome, and it lands while the local debate is fixated on power and social licence.
What to watch: AirTrunk’s OSK2 build-out timeline and any move into a dispersed Japanese region, whether Blackstone’s US$30 billion translates into disclosed AirTrunk campuses, and whether the four-hub thesis shows up as fresh Australian commitments rather than words.