Here's what's on our plate today:

• 🧪 Rakuten's state-backed flagship model runs on NVIDIA chips and DeepSeek's design.

• 📰 Boards plan for digital shocks; ASML backs 12-inch masks; Apple's foldable arrives today.

• 🧠 Brain Snack: name the layer of your stack you don't actually control.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Europe built an aircraft it could not fly without American export licenses, and Asia is now building AI the same way.

• Why the sums force the choice: training modern AI costs more than most national markets can repay, so governments build the layers they can afford and buy the rest.

• How far it reaches: Japan's flagship state-funded model runs on American chips and a Chinese base design, and India, Korea, and their neighbors have made the same trade at different weights.

• What it costs: NVIDIA booked over $30B from sovereign programs in fiscal 2026, and expects the line to keep growing.

• Who pays: Korea cut Motif Technologies from its sovereign model contest in August despite the highest benchmark score in the field.

• What could reverse it: BCG and Brookings argue full independence was never available, and buying what you cannot build is the disciplined answer.

• What remains unresolved: the C919 halt lasted five weeks, and nobody knows what an assembled stack does when a supplier stops for an extended period.

Asia's sovereign AI is assembled from American and Chinese parts

In the late 1960s, Britain, France, and West Germany each concluded separately that none of them could build a passenger jet capable of beating Boeing. The obstacle was money rather than engineering, since designing an airliner cost more than any one European market could earn back on its own. Splitting that bill three ways turned an impossible project into a feasible one, which meant dividing the aircraft itself between the three countries. Britain took the wings while Germany built fuselage sections, and the pieces came together on an assembly line in Toulouse. The arrangement worked well enough that Airbus now accounts for roughly half of the global market for large commercial aircraft.

One part of the aircraft never came under European control, and it happened to be the part that turns fuel into thrust. Much of the engine supply is American or comes from US-French joint ventures, which means an Airbus stops being fully European the moment Washington decides who may buy what.

Sovereignty has become something a country can order

Artificial intelligence has now arrived at the same arithmetic, though the sums involved are considerably larger than anything Airbus faced. Training a modern AI system requires machines costing more than most national markets could ever repay, forcing any government seeking its own capability to choose among layers. It can build the layers it can afford while buying the ones it cannot, and the resulting mixture gets announced as 'sovereign AI'. The phrase refers to the systems, data, and computing infrastructure that a country controls within its borders.

How those choices are resolved in practice first shows up in the accounts of the company selling the equipment. NVIDIA reported quarterly revenue of $96.2B on August 26, 2026, and used the release to list Japan's national AI infrastructure and Korea's sovereign buildout among its own business highlights. Sovereignty has become a segment on the supplier's income statement, worth more than $30B in fiscal 2026 by the company's own count. That figure more than tripled year over year, and management expects it to keep growing as countries spend in proportion to the size of their economies. Every government that decides it cannot depend on foreign computing power turns into a new customer for foreign computing power, which is the circle none of the announcements close.

What Japan actually bought

Japan has spent more on closing that circle than any of its neighbors, committing more than ¥10T (roughly $65B) in public support for AI and semiconductors through fiscal 2030, and its trade ministry nearly quadrupled its own annual chips-and-AI budget to about ¥1.23T (roughly $7.9B) for the year starting April 2026. The largest thing the money bought appeared the following summer, when NVIDIA and a Japanese consortium called Noetra announced a 140-megawatt AI factory on July 16, 2026. An AI factory is a data center built for the single job of training and running models, and this one holds 27,5k NVIDIA Rubin GPUs, the chips that do the bulk of AI computation.

Japan's Ministry of Economy, Trade and Industry stands behind the project as its sponsor, and the AI factory supplies the computing foundation for METI's FRONTia program in robotics and physical AI. NVIDIA calls the facility the world's first national AI infrastructure for physical AI, and the description holds at the level of ownership, since a Japanese company established it and Japanese institutions get the resulting model weights. Every processor inside the building was designed by an American firm and fabricated in Taiwan.

The model layer repeats the arrangement

The same division runs one layer higher, where Japan's trade ministry pays part of the training cost for domestic models through an accelerator program called GENIAC. Rakuten released Rakuten AI 3.0 on March 17, 2026, calling it Japan's largest high-performance model and thanking the open-source community without naming what the model was built on. The naming happens instead in Rakuten's own public model repository, which carries an architecture tag reading 'deepseek_v3' at the top of the listing. That listing gives 671B parameters, the adjustable numbers a model learns during training, of which 37B are used to answer any single query. DeepSeek-V3 carries precisely those specifications, and the Chinese lab DeepSeek published it as a free download that anyone is entitled to build on.

Building on a published model is legal, cheap, and entirely ordinary, which is why so many national models are assembled the same way. Community developers found the architecture in Rakuten's own configuration file within hours of the launch, and the specifications sitting in that repository are public and unchanged. The record establishes a flagship subsidized model drawing its computing power from California and its underlying design from Hangzhou.

Buying keeps winning on speed

Chip manufacturing is the one layer Japan is trying to rebuild rather than buy, through a state-backed foundry called Rapidus, meaning a factory that makes chips to other companies' designs. Rapidus raised ¥267.6B (about $1.7B) in February 2026 and targets production in the second half of fiscal 2027, with full output the year after that. By the time Rapidus starts, TSMC and Samsung will have been shipping the same generation of chips for roughly two years, and no high-volume customer has been publicly named.

Put that timetable in front of an official working to a budget, and the real choice is between chips in 2028 and chips next quarter. The purchase order wins every time, which is the mechanism converting sovereignty programs into export sales for the suppliers they were meant to reduce reliance on.

The same trade at different weights

India set the dial toward breadth, building a shared national pool of computing power that BCG puts at about 62k chips as of March 2026 and renting it cheaply to startups and researchers. The same analysis records Microsoft buying around 485k chips of an earlier generation during 2024, which leaves India's entire national base below one company's annual order. That gap determines what the pool can be used for, since a shared resource of that size supports research and small products rather than frontier models.

South Korea moved the dial the other way and bought at the largest scale available to it, agreeing with NVIDIA at the APEC summit in Gyeongju on October 31, 2025, to take more than 260k chips. Korea also owns a layer outright, because Samsung and SK hynix make most of the fast memory that NVIDIA's processors need in order to run at all. That combination should make Korea the exception in the region, and the exception dissolves at the model layer. Korea's science ministry ran a national competition to pick its sovereign model champions and published the second-round scores on August 27, 2026, one day after NVIDIA's earnings release. SK Telecom led the field and holds equity in Anthropic from a $100 million investment made in 2023.

The case that this is the right answer

A serious argument holds that the goal was miscategorized from the beginning, and BCG makes that case in a March 25, 2026 analysis of national AI strategies across a dozen countries. Its authors conclude that owning a finished model is not the same as owning the ability to keep making one, since a competitive system depends on hardware partners and foreign training data that no single country supplies. Brookings reaches the same conclusion from a different angle, calling full-stack sovereignty structurally infeasible and proposing managed interdependence through alliances instead.

Read through either paper, and Japan looks disciplined rather than compromised, because it is refusing to spend public money reproducing what it can already buy. BCG attaches one warning to its own recommendation, since the approach bets on today's compute-heavy methods continuing to dominate, and a change in method would strand everything already committed.

The real test of sovereign AI comes when the parts stop arriving

Both papers rest on an assumption about behavior rather than engineering: that these dependencies remain governable through partnership. China's COMAC C919 has already tested that assumption, since the airframe is Chinese while the engines ship under American export licenses. Washington suspended the engine export licenses in late May 2025 and production stopped, though the licenses were restored roughly five weeks afterward.

The same kind of interruption has now happened once inside AI itself, when Anthropic pulled its most capable models from all users worldwide in June 2026 to comply with US export controls. Access came back on July 1 after 19 days, though only the more restricted of the two models returned worldwide, while the other went back to a short list of approved US organizations and stayed there. Neither halt lasted long enough to test the stacks Asia has spent the past few years assembling. What none of these programs has answered is what an assembled stack does when a supplier stops shipping and the licenses do not come back after a month.

Brain Snack (for Builders)

Quick Bits, No Fluff

• Boards start rehearsing digital shocks: A new survey finds boards moving digital infrastructure failure into formal scenario planning rather than leaving it with IT.

• ASML wins the industry over to bigger masks: TSMC, Samsung, and Intel backed the shift to 12-inch photomasks with a pilot line targeted for 2031, alongside High NA machines running about $400M each.

• Apple's foldable lands today: The iPhone Ultra is expected to open to a 7.8-inch display, swap Face ID for Touch ID, and cost more than $2k.

Wednesday Poll

Meme Of The Day

The Toolkit

• Modal: Serverless cloud for Python and AI workloads, spins up GPUs in seconds with no infrastructure work.

• Leonardo AI: Image and video generator with fine-grained controls, built for teams that need consistent style at scale.

• Sourcegraph: Code intelligence platform whose assistant reads your entire codebase before answering questions or writing anything.

Rate This Edition