Don’t Build Your Enterprise AI on Borrowed Access

All AI applications and enterprise deployments have been built until now with the assumption that access to frontier AI models will remain possible and almost unlimited. That was the right mindset, until now. The trend was that frontier models’ new capabilities were being commoditized or open-sourced within 6 months and therefore the focus was only about understanding future capabilities as early as possible. The companies, teams or people who could build products by grasping and adapting to new frontier capabilities as early as possible to build applications have been winning. However, several signals might be indicating that this era will stop soon. The U.S. Commerce Department’s decision to restrict access to Anthropic’s Fable and Mythos models to U.S. nationals was only a first sign. However, that restriction was later lifted after coordination with Anthropic, showing just how quickly access conditions can shift and the need for adaptable enterprise AI systems.

The four reasons why frontier model access is not guaranteed

The mindset of building for frontier models only might stop being the right one if their availability stops being guaranteed. There are at least four reasons why that might happen.First, the economics of frontier AI are built on subsidized pricing that will not last. The recent IPO preparations from OpenAI and Anthropic have forced them to open the books and clearly revealed the real cost of current sponsored AI services, which may not be economically viable post-IPO. Rumors are that the $200 private subscription offered by Anthropic actually costs several thousands to the company. Therefore, token prices are expected to keep rising, and the era of artificially cheap frontier model access will end.

Second, frontier models are not getting cheaper to run. Unlike classical computing, where Moore’s Law steadily reduced the cost of any given operation over time, AI does not seem to follow this curve. The models are getting bigger but also more compute-intensive. All of our engineers have exhausted their Claude credits within 10 minutes after the (temporary) release of fable 5. This might hint that the next frontier models will require more tokens, more compute, and more infrastructure with every iteration.

Third, it seems that fewer and fewer players are actually staying in the race for the true frontier models. Mistral, once Europe’s hope to compete on the Frontier models global arena, has revised its ambitions to be a modest follower. Chinese models are fast followers but continue to lag behind the leading labs’ best releases.

Finally, frontier models might just be too dangerous to be used in every application. We are not using miniaturized nuclear reactors in cars, partly because this field is highly regulated and it’s dangerous. Anthropic’s own experiments show that it is very difficult to keep a frontier model on leash to ensure it remains harmless while preserving its capabilities.  This may be a long-term real lesson and underlying rationale behind U.S. regulation of Fable 5. In any case, it is very likely that frontier models will be treated as strategic assets by states and therefore will not be as freely available as they are now.

Therefore, we may be entering a world where token costs matter and where frontier models are not available to all applications in all countries.

Enters the application layer

In a world where not every model call goes to the best available model, how can we still guarantee that every task is solved reliably and with the best possible level of performance? The answer may lie in the so-called “application layer”.

NVIDIA’s CEO Jensen Huang described AI as a 5 layer cake: energy, chips, data centres, models and applications. He argues that the application layer is both the most important one and the one that is still largely missing.

For some applications, such as chatbots, this layer is extremely thin—essentially just a chat interface. In others, such as advanced industrial and engineering applications, it is far more important and critical. It is the determining layer in the value chain. As access to a specific frontier model will not be guaranteed anymore, the application layer’s role will be to uniformize outcome across models and to keep token costs under control.

Concretely, if I want to perform a complex, mission critical task like redesigning my manufacturing chain for a new aircraft model, I will ask a powerful frontier model to help me. The model might be able to figure out how to do it alone, build a planning tool for me and communicate the outcome to all of my teams. However, if I switch models to a more capable one, a less expensive one or one from a different provider, I might lose performance or lose consistency in my final output. In many practical cases in enterprise setups, this unpredictability is not acceptable. This is where the application layer’s role is critical: rather than letting the model operate in an open-ended space and relying on its raw capabilities to produce the outcome, the application layer defines a clear canvas, a constrained set of available actions, and a defined perimeter within which it can operate. The model is no longer asked to figure everything out from scratch, it is given a well-defined canvas that will greatly improve the likelihood that any model will perform well.

Cost-wise, my model might also decide to perform very low-level actions to reach its goal. It might rewrite a planning algorithm from scratch, build a new interface for me to visualize its work or even develop a new optimization solver. But, this might be incredibly token expensive, whereas good guidelines or pre-built tools may help the model achieve the same outcome with fewer tokens. This is the role of the application layer. The good news is that models tend to be lazy and look for shorter paths, so we might not need to constrain them too much once it has access to the right set of skills, tools and compute resources.

In short, a good application layer will provide clear skills, tools and guidelines for the model to operate reliably and cost effectively. In addition to improving performance and controlling token costs, this also means dependency on a specific model decreases. if you embed your intelligence in the application layer rather than in the model itself, the underlying model becomes interchangeable. In many cases, it might be what makes the difference between experimental and scalable enterprise AI.

Call for action for enterprise leaders

Across industries, we are meeting everyday with enterprise leaders working to establish their AI stack. Most companies are starting with one or several frontiers model providers to establish their necessary model layer in their stack. For the application layer, they either rely on home-built tools, fully purchased capabilities, or a mix of both.

The Anthropic ban was a reminder that this stack can be disrupted overnight. Companies’ architectures must be ready for when it will happen again. To be ready, companies must consider speed of deployment, but also build solutions that will mitigate the effects of model changes as well as growing token costs. It’s not an easy path, but my advice would be:

• Invest in your application strategically and early. Companies that are missing the train will always have a gap compared to early adopters
• Make model-agnosticism a design principle, not a workaround
• The cost of switching models should not be an existential question for your operations
• Partner with specialized companies that can benchmark models, validate performance across multiple setups and help ensure long-term performance guarantees.

The companies that will win the game in the next enterprise AI phase are not necessarily the ones with access to the best models, they are those that will have built the right layer around them.