Cohere Takes Sovereign AI Into Japan With Local Compute and a Government Test
Cohere has announced plans to establish a Japanese legal entity and to expand its sovereign AI investments in the country, alongside an expanding partner ecosystem and a training and certification programme for local partners. The technical core is Model Vault, which Cohere describes as proprietary inference infrastructure backed by local compute clusters that lets customers run fully in-region inferencing, with prompts and model responses remaining within Japan. The most concrete evidence in the release is that Takane, a Japanese-market language model Cohere built with Fujitsu, is among the domestic models Japan’s Digital Agency selected for evaluation in GENAI, the generative AI environment developed under the Agency’s Government AI efforts. Cohere’s Series E, which the release describes as targeting up to CAD $7 billion at a CAD $35.7 billion valuation, is not yet closed and the release does not say that it is.
What Model Vault actually guarantees
The distinction Model Vault draws is the one that separates a real sovereignty claim from a marketing one, and it is worth stating precisely.
Data residency usually means where information sits at rest. For a model API that covers half the question, because generating a response requires something to read the request and run it through the model. If that computation happens abroad, the information leaves the country for processing even when a copy stays locally, and it returns as an answer. A regulator examining where data travels would see a round trip rather than a local operation.
Inference residency closes that loop, and local compute clusters are what make it possible. Cohere’s description is specific on this point: prompts and model responses remain within Japan, and the deployment runs on infrastructure in the country rather than on a distant region. That is a materially stronger guarantee than storage, and it is the one a government buyer will test first.
The deployment options named go further than most vendors offer. Cohere lists public cloud, private cloud, on-premises and fully air-gapped environments. Air-gapped means the system runs with no connection to external networks at all, which is the configuration defence and intelligence customers require and the one that no cloud-hosted inference can provide. Very few model providers offer it, because it requires shipping the whole stack rather than exposing an endpoint.
The Takane selection is the proof point
Sovereignty claims are easy to make and hard to verify. The Takane detail is the part of this release that a reader can check.
Takane is a large language model optimised for the Japanese market, built with Fujitsu. Japan’s Digital Agency has selected it among the domestic LLMs for evaluation in GENAI, the generative AI environment the Agency developed under its Government AI efforts. That is a government body putting a specific model into an evaluation environment for public-sector use, which is a more meaningful signal than any partnership announcement.
It also explains how the two companies built the model. Cohere supplies the model and the platform; Fujitsu supplies the integration and the local presence. That division is the standard way a foreign AI company enters a regulated market, and it is the arrangement that lets a company describe a model as domestic without the underlying research being domestic.
The release states that organisations in Japan’s public sector and advanced manufacturing deploy Cohere’s products. It names no customer, which is the largest gap in the document and the one thing that would convert a description into evidence.
Why Japan, and why now
The release describes Japan as emerging as one of the world’s leading markets for secure, sovereign AI, and the framing holds up. Three things make the country unusual.
The first is language. Japanese is poorly served by models trained predominantly on English, and a model optimised for it has a defensible advantage that a general frontier model cannot simply outscale. The second is industrial structure: advanced manufacturing generates proprietary process data that firms will not send abroad, which creates demand for in-region inference independent of any regulation. The third is policy: a Digital Agency running a structured evaluation of domestic models is a government actively shaping the market rather than waiting for it.
That third factor deserves emphasis because it is the one a foreign vendor cannot manufacture. A government that evaluates domestic models has decided domestic capability is a strategic requirement, and that decision creates a market for exactly this kind of arrangement rather than merely permitting it. It also creates an incentive for foreign model builders to partner locally rather than sell remotely, which is the path Cohere has taken. Andrew Chang, Cohere’s Vice President of APAC, describes organisations across public and private sectors moving beyond experimentation into mission-critical deployment. That is the claim every vendor makes about every market, and in Japan’s case the Digital Agency evaluation gives it something to rest on.
How this differs from a standard model API
The contrast with a standard model API is worth drawing because it explains why the positioning exists at all. A typical frontier model is reached through an endpoint, and the compute behind that endpoint sits wherever the provider put it. A customer can usually choose a region for storage and sometimes for processing, and cannot choose who operates the facility, what jurisdiction governs it, or whether the model weights are present in the country at all.
A deployment where the customer runs the model on infrastructure inside their own jurisdiction changes all three. It is a heavier commercial arrangement and a smaller market, and it is the only one that satisfies a buyer whose requirement is that the model itself never leaves.
The funding round and the acquisition
Two corporate facts appear in a single paragraph and both need careful reading.
The release describes the Series E as targeting up to CAD $7 billion at a CAD $35.7 billion valuation, and the word targeting matters. A round in progress is not a round complete, and the release does not say the funding has closed. The figure is large enough to be the headline on its own and a reader should treat it as an intention until Cohere confirms a close.
The second is that the acquisition of AISLE establishes Cohere Cyber, adding security and threat detection for organisations in sensitive environments. That is a new product line created by acquisition, described in one sentence with no detail on what AISLE’s technology does or how it integrates with the North platform. It is the least explained item in the release and the one with the most obvious bearing on the sovereign AI proposition, because a customer running models in an air-gapped environment needs threat detection that also runs there.
The partner model is how this scales
One operational detail deserves separating. Cohere is investing in a training and certification programme designed to help local Japanese partners build expertise deploying secure enterprise AI.
That is the mechanism by which a company with a Tokyo office serves a national market. Sovereign AI deployments are not self-service; they involve on-premises installation, integration with existing systems, security review and ongoing operation. A vendor cannot staff that from Toronto, and a partner programme is how capacity gets built locally while the model stays with Cohere. It also means the quality of the customer experience depends on the partners Cohere certifies rather than on Cohere directly, which is a risk the programme is designed to manage.
Why the air-gapped option is different in kind
Air-gapped deployment deserves a paragraph of its own, because it is not a stronger version of a private cloud. It is a different product.
A private cloud deployment still assumes a network path, a support arrangement and a way for the vendor to deliver updates. An air-gapped environment has no connection to external networks at all, which means the customer takes delivery of the whole stack and operates it themselves. There is no remote telemetry, no automatic model update and no vendor access when something breaks.
That is what defence and intelligence customers require, and it is why few model providers offer it: it inverts the usual commercial model. A vendor selling an endpoint keeps control of the software; a vendor shipping into an air-gapped environment gives that control away and depends on documentation, certification and trained local staff instead. The partner certification programme named in the release is the mechanism that makes this workable, because the people who can operate an air-gapped deployment have to exist locally.
What would complete the picture
Four things would let an enterprise or a government buyer plan against this rather than admire it.
A named customer is the first, and it is the gap a reader notices. Public sector and advanced manufacturing describe two categories without identifying a single organisation in either.
Model Vault’s capacity is the second. The release names local compute clusters without saying how much compute, where it sits, or what latency a customer should expect. Those figures determine whether in-region inference is practical for interactive workloads or only for batch.
The third is what Takane’s evaluation in GENAI actually involves and when results appear. An evaluation in progress is a genuine signal and a completed one with published findings would be considerably stronger.
And the fourth is detail on Cohere Cyber. A product line created by acquisition, aimed at the most sensitive customers, described in one sentence, is the item most in need of a technical briefing.

What to watch
Whether the Series E closes at the stated figure or below it, because a round targeting seven billion and a round closing at seven billion are different announcements. Whether Cohere names a Japanese customer, since that is what converts a partnership structure into evidence. And, whether the GENAI evaluation produces published findings, because a government assessment of a foreign-built model would be a genuinely useful document. And whether Cohere Cyber ships into the air-gapped deployments the platform already supports, because sovereign inference without in-region threat detection answers half the question a defence customer asks.

