Fortif AI (ASX:FTI) The Agentic AI “Data Plane” Behind the Models

The Real AI Bottleneck Is Data, Not Models Fortif AI

Fortif AI surged 14% to 32c this morning after announcing that its newly acquired business, Nol8, is building an agentic AI data plane.

We started writing on Fortif a couple of months ago around its acquisition strategy, and more recently the company delivered its first profitable quarter. Now we are starting to look deeper under the hood, because we think there are some genuinely interesting things developing here.

In simple terms, Fortif AI is trying to build an infrastructure layer that sits between AI models (the inference layer) and real world execution tools, like databases and workflows. If they can pull this off, it becomes less about “another AI product” and more about owning the plumbing that makes AI useful inside an enterprise environment.

What are the Best ASX Stocks to invest in right now?

Check our buy/sell tips

From Demo to Enterprise Proof, The Next Phase Matters

Fortif AI pitch here is that one of the real bottlenecks for AI is data, not models.

Enterprises are sitting on expanding datasets that are messy, fragmented, and poorly governed, which drives a lot of infrastructure spend just to process, move, and store unorganised information. In other words, the cost to make data usable is becoming the constraint.

Fortif AI’s solution is to move parts of data processing off large CPU clusters and into FPGA accelerated hardware. This is where their IP sits: neural network driven algorithms that can classify and route data while it is in motion, rather than forcing everything through heavyweight, centralised compute.

On the roadmap, management is presenting a relatively clear path to commercial revenue.

So far, Fortif AI has demonstrated the platform in demo form and is now working toward systems proven at demonstration scale, with the goal of showing high throughput and meaningful latency improvements.

The next step is less about building and more about proving. The company is developing the tooling so enterprise partners can ingest real world datasets and benchmark specific AI data pipeline scenarios. They say they are already in active partner conversations across multiple AI verticals to design and run tests in high performance environments.

This is the validation phase that matters most, because it is where demo results either translate to real enterprise workloads, or they do not.

The company is targeting a revenue ready platform by the end of CY26. Stage 3 is the commercial platform (target end CY2026), where the first contracts can be signed once benchmarking is complete.

5,000 CPUs to One FPGA, Hype or Breakthrough?

The headline claim is pretty punchy, management says the demo engine can take a workload that might otherwise require around 5,000 CPUs and run it on a single FPGA, while cutting latency from roughly 500 ms to 3 ms, which is about a 160x improvement.

The reason that pitch resonates is the unstructured data problem. A huge share of enterprise data is unstructured, think text, documents, images, audio, and video, and it is generally far more compute intensive to process than neat, structured tables. Some widely cited industry estimates put unstructured data at roughly 80% to 90% of enterprise data, and that gap is a big part of why AI infrastructure bills can balloon fast.

The key investor question is not whether the demo looks great, it is whether the performance holds up in real enterprise environments, with messy datasets, edge cases, and sustained throughput. If the benchmarking phase validates those latency and throughput claims in production like conditions, that is where the story starts to earn credibility.