';
Back
A briefing for Senior Executives in European Banking and Asset Management
For two years, the working assumption across European financial services has been simple: if you want frontier AI, you buy it from a handful of US labs, you access it through their cloud endpoints, and you accept the bill. That assumption is now breaking down and the implications for banks and asset managers are strategic, not merely technical.
A market that has quietly flipped
According to recent reporting by the Financial Times, companies from Silicon Valley to Europe including DoorDash, Siemens and Airbnb have begun adopting Chinese AI models from labs such as DeepSeek, Z.ai, Alibaba and Moonshot AI. The drivers are familiar to any COO reviewing a technology budget: these models are dramatically cheaper, increasingly capable and critically many are open weight, meaning the model itself can be downloaded, hosted on your own infrastructure and fine-tuned on your own data.
The economics are hard to ignore. Together AI, a cloud provider serving enterprises in this space estimates that the best open-weight models cost between 10 and 60 times less than proprietary equivalents. Meanwhile, the leading US labs have been moving enterprise services from flat subscriptions to usage-based billing, sharply inflating costs just as institutions scale up AI-assisted workflows. Anthropic, for example, has announced a 50 per cent price increase for its Opus 4.8 model from September.
The capability gap, once assumed to be a year or more is closing fast. The June release of Z.ai’s GLM-5.2 prompted prominent Silicon Valley figures including Marc Andreessen to argue publicly that Chinese models now match and in some cases beat the public models from major American labs. Practitioners are voting with their workloads: DoorDash now routes routine tasks to Moonshot’s Kimi K2.6 and reserves premium US models only for the most demanding work, reporting better performance at lower cost. Some firms, such as San Francisco start-up Lindy have switched to Chinese models entirely and report savings in the millions.
Kimi K3: the model to watch
The next inflection point may arrive within days. Moonshot AI is preparing to release Kimi K3, reported to be China’s largest model to date at an estimated 2 to 3 trillion parameters a scale comparable to and possibly exceeding, the speculated size of Anthropic’s Opus 4.8. According to people familiar with the launch cited by the FT, K3 is expected to outperform Opus 4.8 on mainstream benchmarks.
What makes K3 potentially transformative for regulated financial institutions is not the benchmark scores. It is the distribution model. K3 will be released as an open-weight model, freely downloadable. For the first time, a bank or asset manager could run a model of genuinely frontier scale entirely on its own premises inside its own data centres, behind its own firewalls, under its own security regime.
For an industry governed by GDPR, DORA, outsourcing rules, banking secrecy obligations and client confidentiality duties, this changes the calculus fundamentally. On-premises deployment means client data never leaves the institution. There is no third-party processor to diligence for the inference layer, no cross-border transfer analysis for every use case, no dependency on a vendor’s uptime, terms of service or pricing decisions. The model becomes infrastructure you own, not a service you rent.
Why this matters specifically for banks and asset managers
Client centricity at scale.
The most valuable AI applications in financial services hyper-personalised advice, portfolio commentary tailored to each client, intelligent servicing across every channel, relationship-manager copilots that actually know the client all require processing sensitive client data. That is precisely where cloud-hosted proprietary models create friction with compliance and precisely where self-hosted open-weight models remove it. Institutions can fine-tune a model on their own product shelf, investment philosophy, house views and tone of voice, producing an assistant that speaks the institution’s language rather than a generic one.
Cost efficiency that compounds.
AI in a large bank is not one application; it is thousands of daily inference calls across research summarisation, KYC document review, code generation, regulatory reporting drafts and client communications. At usage-based pricing from proprietary vendors, these costs balloon. A tiered architecture capable open-weight models handling the bulk of routine work, with premium closed models reserved for the hardest problems is exactly the pattern early adopters describe, and it can reduce run costs by an order of magnitude.
Sovereignty and resilience.
European executives no longer need this argument made in the abstract. The Trump administration’s export controls on Anthropic’s Mythos and Fable models since overturned but not forgotten demonstrated that access to US frontier AI can be revoked by a foreign government’s decision. As one UK consultancy AI chief put it in the FT’s reporting, the ban may have been lifted, but the market’s perception has permanently changed. Cohere’s chief executive Aidan Gomez made the broader point: relying on any single entity for critical workloads is now visibly a risk. Several European firms have already begun offloading work to Chinese open-weight models explicitly to reduce dependence on US labs. Notably, venture investors quoted by the FT observed that for European companies, a self-hosted model, whatever its origin is increasingly viewed as the more secure choice, because control sits with the institution.
Under DORA, boards are explicitly accountable for ICT third-party concentration risk. A frontier-class model running on owned infrastructure is arguably the cleanest answer available to that mandate in the AI domain.
A pragmatic playbook
Institutions moving now are following a recognisable sequence. First, benchmark: run open-weight candidates such as Kimi K2.6, DeepSeek and GLM-5.2 against your actual workloads document summarisation, client-email drafting, code assistance and rather than public leaderboards. Second, tier: keep premium closed models for the small share of tasks that genuinely require them, and migrate high-volume routine work to self-hosted alternatives. Third, harden: deploy through vetted infrastructure with independent security review of model weights, strict network isolation, and full audit logging Airbnb’s approach of running China-origin models only through approved, controlled environments is instructive. Fourth, prepare for K3: if the release meets expectations, an on-premises frontier-class deployment moves from aspiration to a budgetable 2026-27 programme.
The counterarguments, honestly stated
Executives should weigh the case against as seriously as the case for. Security and reputational scrutiny of China-origin models will be intense: supervisors, clients and boards will ask hard questions, and institutions must be able to demonstrate that self-hosted weights have been independently evaluated and that no data flows to the model’s originators. US labs still hold the edge on the most complex reasoning tasks, and Anthropic has alleged that some Chinese labs built their capabilities partly by distilling US models, a dispute with potential legal and ethical ramifications.
Running trillion-parameter models on-premises requires serious GPU infrastructure and scarce MLOps talent, costs that partially offset the per-token savings. And the geopolitical wind can shift in both directions: European regulators could yet impose their own restrictions on China-origin AI, just as Washington did on its own exports. Prudent institutions will treat open-weight adoption as a diversification strategy reducing single-vendor dependence rather than swapping one concentration risk for another.
The bottom line
The AI market has moved from a monopoly on capability to a genuine choice of operating models. For European banks and asset managers, open-weight models offer a rare alignment of three goals that usually pull against each other: better client experience, lower cost and stronger data control. The imminent arrival of Kimi K3 frontier-scale performance, runnable in your own data centre makes this the right quarter to put the question on the board agenda: not whether to use AI, but who controls the AI you use.