Japan Says 6G Has Started. What Ran Was Three Test Towers

Japan says 6G has started. What actually ran was three experimental base stations in one block of Ginza. The radio band they used has no 6G allocation yet. Alongside it, a flying base station crossed the Pacific and hovered over Kochi Prefecture carrying a 4G LTE radio. The commercial date for any of this is 2030.

That gap between the headline and the hardware is the interesting part. It hides what 6G is actually being built for. At the CEATEC 2026 conference on October 14, five of Japan's largest operators sat on one panel with Nvidia. They agreed on something that has nothing to do with speed. They are changing the unit of measurement for mobile networks. The metric is no longer megabits per second. It is tokens per second, and time to first token.

What Actually Ran in Tokyo

SoftBank and Nokia announced in July 2025 that they had started outdoor trials for 6G. The details are specific. SoftBank obtained an experimental test station license covering 7,180 to 7,280 MHz. It installed three experimental base stations in the Ginza area of Chuo-ku, Tokyo. Massive MIMO antennas were used for the test. The question was whether a 7 GHz band could build a coverage area comparable to 5G in a dense city. It was the first outdoor 6G trial by a Japanese carrier, as of June 13, 2025.

The catch is the frequency. 7 GHz is only a candidate for 6G. It is on the agenda for discussion in 2027, at the ITU Radiocommunication Sector World Radiocommunication Conference. It sits in the FR3 range, which covers 7 to 24 GHz. Until that conference decides, the band SoftBank is testing in Ginza has no legal status as a 6G band. It is a bet on an outcome that has not been settled.

Why 7 GHz, and not the 28 GHz millimeter band used in 5G. The 7 GHz band sits close to sub-6, so it is faster and easier to reach. The 28 GHz band is far faster, but its range is much shorter and it is easily blocked. That makes 7 GHz a compromise. Speed above sub-6, reach below millimeter. It also sits much closer to the band Wi-Fi already uses, so that frequency fight is not finished.

The Flying 4G Tower That Crossed an Ocean

In August 2026, SoftBank ran its own commercial test of HAPS, a High Altitude Platform Station. A lighter-than-air airship built by Sceye in New Mexico left on August 9 at 22:00 Japan time. It spent 13 days riding the winds across the Pacific. It parked over Kochi Prefecture at roughly 17 km altitude, about 10 km offshore, and ran the test from there. It returned the way it came, by a more northerly route. The round trip was near 30,000 km across 30 days.

The payload is the part worth sitting with. The airship carried a base station equivalent to 4G LTE. Measured throughput was 36 Mbps down and 4.8 Mbps up. SoftBank tested voice calls, SMS, web browsing, LINE, Zoom video calls, and YouTube playback. It also ran a reliability check, 100 calls in a row, all of which had to connect. Emergency earthquake and tsunami warnings were broadcast to handsets. A call to 118, the Japanese emergency number, was tested with the Coast Guard.

Two of the results were claimed as world firsts. The first was edge computing on board the airship. Data from a handset was processed by a server on the HAPS and returned. That cut latency by more than 40% against processing it in the cloud. SoftBank dates that claim to public information as of August 23, 2026. The second was laser ranging. The airship was tracked in flight from the ground, and the distance measured through a corner reflector on board. That was SoftBank with Hitotsubashi University and the National Institute of Polar Research, as of September 8, 2026.

Commercialization is targeted for 2027 and later, starting with disaster response. SoftBank took ten years to get here, and it began the program in 2017.

The Metric Just Changed

This is the part that matters for anyone building software, and it is also the least covered. At the CEATEC 2026 session on AI-Native 6G, SoftBank's Hidehito Funayoshi put up a slide. It reframed what a network is for. The session was hosted by Japan's Ministry of Internal Affairs and Communications, with XGMF and the radio industry association ARIB.

The current network metrics are peak speed in bits per second, plus the antenna pictogram. That is the little 5G or 6G indicator on a phone. Funayoshi said those are about to be replaced by two token metrics. The first is TTFT, time to first token, how long until the first token comes back. The second is TPS, tokens per second. A network that scores well on throughput and badly on TTFT is a bad network for AI. Whatever the speed test says.

He also described running inference inside the carrier's closed network. Put a GPU in the closed network and a cloud response of hundreds of milliseconds drops to tens of milliseconds. The radio link stops being a bottleneck and becomes the delivery mechanism for an edge AI service.

This reframing is not a Japan-only idea, and it is not new. But Japan's operators converging on it at the same time, in public, at a trade show, is a marker. When four carriers and a chip vendor describe the network in tokens, the network is being redefined as compute infrastructure.

Where the AI Actually Goes

The same panel produced concrete numbers, and they are worth having on record because none of them are about radio.

Two details from the session deserve emphasis. KDDI's own researcher Toshiaki Yamamoto said the current state is not good enough. AI as of 2026 is only an assistant reaching in from outside the network. AI-Native, in his framing, means AI becomes part of the network and works autonomously. He wondered aloud whether the idea of generations would survive that. And Nvidia's Makoto Noda pointed out that the axis data center networking is measured on has already shifted. The question used to be how many bits per second could be moved. Now it is how many GPUs can be kept fed without stalling. That changes reported GPU utilization by a factor of two to eight.

The Timeline Nobody Is Hiding but Everybody Misses

Commercial 6G lands in 2030. Docomo has said it targets commercial service that year. Ericsson expects first commercial services around 2030, with pre-commercial trials from 2028. No country has a public commercial 6G network on the market, and none will before the end of the decade.

The standard is what sets the date, not any single country. China is moving fast on the same track. Its Ministry of Industry and Information Technology has said it is accelerating 6G demonstration experiments. The Fifteenth Five-Year Plan for 2026 to 2030 names 6G a key future industry. It sits alongside embodied AI, quantum technology, and fusion energy.

Japan's own preparation runs on a longer clock than most people assume. The National Institute of Information and Communications Technology demonstrated many-connection performance for the 6G era in January 2026. The demo used a hybrid of quantum and classical computing. In June 2026, Docomo, NEC, 1Finity, NTT, and Fujitsu announced a four-year program. It covers fiscal years 2026 to 2029, funded under the Ministry's program for expanded radio resources. In March 2026, a millimeter wave demonstration with high-speed moving vehicles ran in a full-scale tunnel facility. The Ministry of Land, Infrastructure, Transport and Tourism operates it.

None of that is a network. All of it is the two decades of groundwork that has to happen before a network can exist.

What This Means If You Build on the Edge

The honest reading of the CEATEC numbers is simple. The 6G era, whatever the marketing says, is an edge AI era. The network stops being the pipe and becomes the compute tier. That has consequences for people writing software now, and they arrive long before 2030.

• The first is latency budgets. A cloud round trip is hundreds of milliseconds. An edge inference path is tens of milliseconds. That is the difference between an interface that feels instant and one that feels sluggish. It is now a network design decision, not an application optimization. If your product has an interactive AI feature, the cheapest latency win available is not a better prompt. It is moving the inference closer to the user.
• The second is who captures the value. A carrier offering inference in its own closed network competes with cloud providers. The workload used to sit unambiguously in a data center. The radios are the easy part to understand and the hard part to build. The margin on inference is the part that will be fought over.
• The third is the accelerator timeline. 5G networks were built for phones. The 6G push is being shaped by AI workloads instead. The useful question for an Indonesian developer is not when 6G arrives. That answer is set by the ITU process and points to the 2030s. The useful question is when inference moves out of the cloud and into regional edge nodes. That is a question about data center capacity and electricity prices, and it can be answered much sooner.

And this point is often missed. The airship that crossed the Pacific carried a 4G radio, not a 6G one. Most of the world still runs on 4G. Which means the whole edge AI machine can be built on networks that already exist. The priority order is clear. What needs building first is not 6G. It is moving workloads from the cloud to edge nodes closer to the user.

What Is Still Open

Three questions do not have answers yet, and guessing at them would be a mistake.

Whether the token metrics actually replace the speed metrics, or just get added alongside them. Operators have every incentive to keep advertising peak speed, because buyers compare it. Expect both numbers on the same slide for a long time.

Whether 6G ever needs a new radio band at all. The claim that 7 GHz and the FR3 range will carry 6G assumes the WRC-27 outcome goes one way. Spectrum decisions are politics as much as engineering. A decision in 2027 can redraw the map these trials are betting on.

How much of the AI-RAN story is real. The numbers on the CEATEC slides came from the companies presenting them. Outside measurement does not exist yet. The same caveat applies here as to the depreciation schedules of the companies funding all of this. The honest position is that the direction is credible and the magnitude is unverified.

Sources

• SoftBank and Nokia 7 GHz outdoor trial in Ginza, reported by Robotstart, July 10, 2025
• SoftBank News, HAPS communication test over Japan, September 30, 2026
• Robotstart, AI-Native 6G at CEATEC 2026, October 16, 2026
• NICT, demonstration of many-connection performance for the 6G era, January 15, 2026
• NEC, Ministry research and development program for 6G communication performance, June 23, 2026
• Ericsson, 6G and the road to first commercial services
• ITU-R, IMT towards 2030 and beyond, the IMT-2030 framework
• JETRO, China's Ministry of Industry and Information Technology on accelerating 6G demonstration experiments, August 2026