Technology procurement has changed more in the last year than in the previous 20 due to a combination of political, economic, and technological factors. Countries and organizations are realizing that dependence on a handful of platforms is no longer sustainable. For enterprises, this was brought home with the on-off-on saga of Anthropic’s Fable 5.
For those who hadn’t already spotted it, [building enterprise workflows and products around a single AI vendor is not good business](https://www.cio.com/article/4178779/5-things-cios-must-do-as-sovereignty-becomes-a-design-constraint.html?utm=hybrid_search). “An infrastructure whose models and computing power we don’t control is an infrastructure that others can unplug,” stated former French prime minister Édouard Philippe. So what can CIOs do to mitigate the risks of suppliers or governments cutting off essential services?
While simultaneously using multiple vendors for critical applications such as CRMs and ERPs isn’t viable, pulling in different AI models to optimize for cost and efficiency is. Model gateways such as OpenRouter, LiteLLM, and Portkey can route between the applications an enterprise runs, and are becoming key components in AI infrastructure plumbing. In the year to April 2026, OpenRouter disclosed a rise from 5 trillion to 20 trillion tokens per week its products coordinated. AI orchestrators complement this layer by chaining calls and managing agent loops.
Many organizations are already hedging their AI deployment strategies in this way. A June 2026 survey of 145 enterprises by VentureBeat showed two-thirds had already adopted a diversified strategy before the Fable 5 shutdown. Just over half of those surveyed were blending closed frontier models with open-weight ones that they run internally, and another 16% were taking core workflows off closed APIs completely.
We can expect this trend to continue and the commodification of models to accelerate as enterprises focus on reworking their business processes, adjusting business models and stripping out vulnerable points of failure.
Diversifying across models is a necessary start to building resilience. But when 90% of enterprise model spend goes to three American vendors, according to research from Menlo Ventures, it isn’t sufficient. The US government’s involvement in slowing down the launch of OpenAI’s GPT-5.6 models highlights that the Fable 5 saga was not a one-off.
Running open-weight models inhouse is the safest way to prevent disruptions to APIs. While the capability gap between frontier and open-weight models may be slowly growing, says research by Epoch AI, they’re only lagging by approximately four months, similar to the gap between GPT-5 and GPT-5.5. For many applications, this capability imbalance makes little or no difference, and the trade-off in performance is more than compensated for by the security it offers, not to mention potential cost savings.
While Anthropic and OpenAI’s terms and conditions state the possibility of immediate and uncompensated suspension of their services where the law requires it, we may be seeing an emerging divergence among some vendors. Aware of growing concerns from European customers regarding digital sovereignty, Microsoft announced last year it would uphold Europe’s digital resilience regardless of geopolitical and trade volatility. What this might mean in practice if pressure were put on them by a combative government remains to be seen, but it signals a growing awareness that business as usual is no longer the case for US digital services being sold abroad.
Since 2025, EU financial firms, for example, have been required under the Digital Operational Resilience Act (DORA) to maintain tested exit plans for any critical technology supplier. The UK has had a similar requirement with four US-owned cloud vendors designated as critical third parties: Microsoft Ireland Operations Limited, Google Cloud EMEA Limited, AWS EMEA SARL, and Oracle Corporation UK Limited. These firms will be supervised jointly by the Bank of England, the Prudential Regulation Authority, and the Financial Conduct Authority, and be required to undergo resilience testing and report major incidents. We can expect to see more vendors fall under this new regime, including AI frontier model providers.
As technology embeds further into enterprise workflows, government control extends into new realms, and points of failure multiply, so businesses need to adapt. Philosopher Luciano Floridi anticipated this six years ago when he said of digital sovereignty that control is the ability to influence something and its dynamics, and it comes in degrees and, above all, can be pooled and transferred.
The ability to pool and transfer AI control is being enabled by gateways and orchestrators, and we can expect to see power shift away from a small number of frontier model developers as customers spread their workflows across multiple models.
At a recent AI conference in New York, Brian Craig, senior director of architecture at Liberty IT, part of insurance firm Liberty Mutual, said you can’t lock in right now to one vendor or even one framework. “You need to keep being able to have the flexibility with that backbone to be able to hook into different models and vendors, depending not so much on who’s the flavor of the day, but on what you can feel confident about for the next six months,” he said. After all, a more fluid and uncertain world keeps emerging.