A benchmark moves out of the lab and onto a live network

For six months FortifAI has asked investors to take its performance numbers on trust. Recently in August 2026 it tried to change that. The company, which trades on the ASX under the ticker FTI, said its NOL8 “AI data plane” had been run across live infrastructure in three continents — Virginia, Tokyo and Sydney — over Megaport’s private network, rather than on an in-house rig. FortifAI (ASX:FTI) reported that NOL8 processed 34 terabytes of governed data a day against 2.6 terabytes for legacy CPU-based software, a claimed cost advantage of “more than 10 times”. It is the firmest evidence yet behind a stock that has swung from 11 cents to well above a dollar in a year. It is still, however, a company claim — and one that has yet to earn a dollar of NOL8 revenue.

 

What just happened

FortifAI is the former Mighty Kingdom, an Adelaide mobile-games studio that pivoted hard. It announced the Acquisition of the NOL8 technology (via FastAI Pty Ltd) in December 2025, completed the deal in February 2026, and rebranded. The underlying research traces back to Israel’s Technion institute.

The commercialisation cadence since has been brisk. NOL8 version 1.0 launched recently in late June 2026. A day later, the company joined Megaport’s AI Exchange as a distribution and proof-of-concept channel. The recent multi-region test result is the latest step, and it lifted the shares modestly to around 65 cents in early August.

 

Why it matters

The pitch rests on a simple bottleneck. Modern AI clusters are built around expensive GPUs, yet those chips often sit idle waiting for data to be cleaned, filtered and governed by conventional CPUs. FortifAI’s argument is that if you do that data-wrangling in dedicated hardware — a field-programmable gate array, or FPGA — before the data reaches the GPU, you keep the costly silicon busy and throw away what it never needed to see.

In the recent test, NOL8 was said to strip out up to 64 per cent of data before it hit the GPUs, letting “each megawatt of GPU capacity deliver up to 2.8 times more useful AI work”. If that holds outside a design-partner setting, it speaks directly to the Economics that keep data-centre operators awake.

 

The numbers that matter

Treat the performance figures as company-commissioned benchmarks; FortifAI itself concedes they have not been independently verified by a third party.

The headline claims are striking. Across 2026 the company has variously reported a 200,000-fold throughput edge over Google’s RE2 pattern-matching engine at high complexity, latency cut from roughly 500 milliseconds to 3 milliseconds, and throughput scaling from 5,000 to two million events a second. Earlier in the year it claimed a single accelerated server could replace the compute of around 60,000 CPUs. In the recent run, it said policy complexity rose 400 per cent while throughput fell less than 1 per cent, and in-region latency held at 3 milliseconds.

The financials are more sobering. FortifAI raised roughly $20 million in 2026 — $5 million at 30 cents earlier in the year and $15 million at 71.5 cents later — to fund development and working capital. Legacy games Revenue was just $1.23 million in the half to 31 December 2025, down 52 per cent year on year. NOL8 has generated no recurring revenue. The 52-week share-price range spans about 11 cents to $1.26, a measure of how speculative the market still regards the story.

 

The bigger industry picture

FortifAI is chasing a genuine shift. The industry’s centre of gravity is moving from Training giant models to running “agentic” AI — software agents that act continuously in the background, generating relentless streams of data that must be moved, filtered and governed in real time. The company cites its own market work sizing the data-movement layer at roughly $120 billion by 2030, and points to the fast-growing “NeoCloud” providers as its route to market.

Whatever the precise number, the thesis is mainstream: interconnects and data plumbing, not just raw GPU horsepower, are becoming the constraint. That is why the Megaport tie-up matters — it puts NOL8 next to the private-network layer where hyperscale data actually travels.

 

The bull case

If even a fraction of the claimed efficiency survives independent scrutiny, FortifAI is selling something operators want: more output from GPU fleets they have already paid for. The FPGA approach is hardware-differentiated rather than another software wrapper, and the company frames its moat as the full stack — proprietary algorithms plus accelerated hardware — rather than any single layer. The Megaport channel offers distribution without a large direct sales force, and the multi-region test at least demonstrates the product runs on real infrastructure across geographies with consistent governance outputs.

 

The risks

The gap between demonstration and dollars is the whole Investment case here. There are no signed commercial contracts, no NOL8 revenue, and no independently verified benchmarks. The competitive backdrop is formidable: Intel (via Altera) and AMD (via Xilinx) own the FPGA silicon, and hyperscalers build their own data-movement stacks. A company that raised $20 million this year and burns cash will likely need more capital, and the current cash balance was not clearly disclosed in the material reviewed for this article. The shares have already round-tripped violently, and “proof-of-concept” is doing a lot of work in the disclosures.

 

What happens next

Watch for a handful of concrete catalysts. First, the quarterly Appendix 4C, which will reveal cash burn and runway. Second — the decisive one — any conversion of a Megaport design partner into a paying, named customer. Third, independent or third-party validation of the benchmark claims. Fourth, the production-grade commercial engine the company has targeted for the end of 2026. Fifth, any further capital raising, which would signal how long the runway really is.

 

Conclusion

FortifAI has done what a pre-revenue technology story is supposed to do at this stage: it has produced a sharper, more credible demonstration and lined up a distribution partner in Megaport. The recent multi-region test is a step up from a lab bench, and the underlying problem — feeding hungry GPUs efficiently — is real and expensive. But investors are still being asked to price a set of unverified performance claims and a market opportunity the company has largely sized itself. The next two quarters should tell whether NOL8 is a genuine piece of AI infrastructure or an elegant benchmark in search of a buyer. Until a paying customer appears, the revenue meter reads zero, and that is the number that ultimately matters.