AI Agents Are the Future. Inference Costs Are Keeping 90% of Companies From Getting There.

In the past few weeks, AI agents have moved from the theoretical next big thing in AI to the clear next frontier. Peter Steinberger’s OpenClaw gained enough traction to lead to a high-profile acquihire by OpenAI, with promises to keep it open and build agents as a core part of their strategy. The message is clear: Agents are here to stay. But how companies will actually grow and pay for them is never quite that simple.

In DigitalOcean’s 2026 Currents research report, which surveyed over 1,100 developers, CTOs, and founders at technology companies around the world, 60% said that applications and agents represent the most long-term value in the AI stack. However, only 10% are actually scaling agents in production today. This gap between conviction and execution is fundamentally an infrastructure problem, and open source models and inference are at the center.

Open-Weighted Models Are Winning. So Is Its Infrastructure Bill.

The same open source momentum behind OpenClaw’s rapid rise is reshaping the open-weights model layer just as fast. The Currents data shows Meta’s Llama and DeepSeek are each used by 21% of respondents for AI agent development, with DeepSeek matching Llama’s adoption at a much quicker pace. OpenClaw’s spread in China made this concrete, as developers adapted it to run on DeepSeek, configured it for WeChat, and Baidu announced plans to give users of its smartphone app direct access. The open source architecture made localization easy, which is exactly why open source agents are spreading faster than any proprietary alternative could.

But the narrative glosses over that while open-weighted models are free to download, running costs for them are not free.

Our data shows that 44% of respondents report spending between 76% and 100% of their AI budget not on model training, but on inference. The cost of running models scale with usage and compounds with agent complexity. A single AI query is one inference call, but as tasks get more complex and autonomous, the number of queries to complete a workflow quickly balloons into the hundreds. DigitalOcean’s data shows companies are already thinking about this: 49% of respondents identified the high cost of inference at scale as their number one blocker to scaling AI.

Also Read: AiThority Interview With Arun Subramaniyan, Founder & CEO, Articul8 AI

The Adoption Numbers Tell a Different Story.

Directionally, companies agree that agents are the future, with 60% stating agents represent the most long-term value in AI. 37% say it’s their top area for budget growth in the next twelve months. Despite this, only 10% are scaling agents in production or treating them as core to their business strategy, while 33% are running pilots and 28% are still in exploration mode.

For companies that have overcome adoption challenges such as security concerns (cited by 41% of respondents) and tooling complexity (called out by 48%), the ROI is already evident. Fifty-three percent of companies using AI agents report measurable productivity and time savings. Seventeen percent report productivity gains of 26-50%. Nine percent report gains above 75%, which is not just incremental efficiency, but something closer to a step change in how work gets done.

The gap between first-movers and holdouts is real, but it’s not yet decisive. The organizations waiting for agents to be more mature, cheaper to run, and easier to orchestrate are right that those things are coming, but waiting for them is costly. Every quarter spent in pilot mode is a quarter the early adopters are using to build the institutional knowledge and compound their advantage over holdouts.

The Agent Era Runs on Inference.

Steinberger’s decision to join OpenAI and move OpenClaw to a foundation shows that  building the infrastructure to run agents reliably, securely, and economically at scale is beyond what any single developer (or a modest team) can sustain.

That’s the same problem businesses are sitting with right now. The models are good enough, and the use cases are clear. The gap is in the operational layer underneath. Inference infrastructure must be cost-predictable at scale, with orchestration that handles multi-agent complexity without a custom integration for every tool, and reliability that’s built into the architecture, without the human approval checkpoints that 40% of organizations currently rely on.

OpenClaw went viral because it made the agentic future feel tangible. It also exposed why most organizations aren’t ready to meet it: the infrastructure layer hasn’t kept pace with the ambition. The industry’s center of gravity is now moving away from model training, and towards inference as organizations begin to run models in production across autonomous workflows.

The organizations that treat inference optimization as a priority now are the ones that will be able to move quickly when agent use cases demand it. The ones still stitching together disconnected tools and absorbing unpredictable costs will find themselves rebuilding their stack at exactly the wrong moment. In 2026 and beyond, the companies that win will be the ones that made inference a priority architectural concern early on.

Also Read: Cheap and Fast: The Strategy of LLM Cascading (Frugal GPT)