The era of massive, compute-hungry enterprise AI is fading, making way for much more efficient, scalable solutions.

“It’s not about unlimited resources, but about smart, efficient solutions.” – Aidan Gomez, Co-Founder and CEO, Cohere

This month we launched our most efficient and performant model, Command A.

Command A’s capabilities are a powerful confirmation of our view that innovation and efficiency, rather than brute compute power, will be the key to deliver enterprise value from AI.

Command A runs privately on just two GPUs (versus eight for DeepSeek-V3) while matching or beating the performance of DeepSeek-V3 and GPT-4o across a range of agentic and multilingual metrics.

Efficient, secure AI matters

Companies face a difficult trade-off between accelerating secure AI adoption and preventing costs from spiraling. In fact, 74% of companies say they are struggling to achieve and scale AI’s value to their organization.

It’s a challenge that’s been particularly acute for Fortune 500 enterprises in tightly regulated industries like healthcare and banking. These sectors overwhelmingly need private AI deployments to ensure data security and compliance, but that involves a series of increasingly heavy costs:

Specialized GPU chips, costing around $40,000 each and that are thirsty for energy
Computing costs, which rose an estimated 89% from 2023 to 2025, overwhelmingly due to generative AI
Customization costs, which can “skyrocket when it comes to training custom models with everything from data prep, along with tools and infrastructure hitting the bottom line.” (AI Infrastructure Alliance Report, 2024)

Efficiency is set to be the defining factor in scaling private AI deployments. Large enterprises that optimize for low-cost, high-performance secure AI are poised to lead the next phase of AI transformation.

Smarter, scalable agentic AI is here

Much of the coming wave of productivity for big enterprises is set to come from agentic AI applications, like our secure AI workspace North, powered by Command A.

Consulting firm Deloitte claims that a quarter of companies using GenAI will try out agentic AI this year, a figure that will swell to 50% by 2027.

Yet, running a system with that kind of power can quickly become costly.

100 times more compute for agentic AI

One hundred times more computing resources are needed to run agentic systems compared with conventional large language models, according to Nvidia’s VP of product management for AI, Kari Briski.

In the healthcare setting, for example, a multi-agent AI system could scan incoming patients’ electronic health records, update them, flag any omissions and next steps, and use the information to optimize staff scheduling and bed assignments. It would shorten processing times, reduce errors, and free up staff for more direct patient care.

“As the volume of data continues to grow and become more complex, the need for intelligent, adaptive systems will increase.” – Barry Mostert, Senior Director, Artificial Intelligence and Analytics, Oracle

The constant feedback loops and coordination needed to deploy these AI agents — akin to a neural or “reasoning” network — demand increased computational resources.

North, with Command A, lets you build and deploy custom AI agents using just a fraction of the GPU power required by less efficient models.

Customized secure AI at a fraction of the cost

Secure AI customization has become essential, especially for companies in highly regulated sectors, because off-the-shelf models rarely satisfy the complex, domain-specific language required in these industries and data privacy is paramount.

According to a recent MIT report, over half of companies surveyed say that employing fine-tuned models on their proprietary data and workflows is key for businesses to unlock unique insights and create more personalized service for their customers. Of these, 86% say that security is critical as they customize more actively.

Yet, the larger a model, the more complexity and the more computational resources are needed to train it. Customizing more efficient models can dramatically reduce costs while still delivering strong performance for targeted business use cases. By fine-tuning these leaner models in private deployments, organizations can achieve faster deployment, lower infrastructure demands, and cheaper inference at scale in a secure environment (on-premises or VPC).

As agentic and deeply customized secure AI solutions start to reshape entire industries in the coming years, enterprises that are quickest to embrace lean, efficient, and secure models and platforms will outpace and outlast the competition.

Building enterprise AI agents

A comprehensive blueprint for developing scalable AI agents in regulated industries.

Download the eBook

Taming AI costs: Leaner models for smarter agentic use