VORTIQ-X challenges NVIDIA: New benchmark stops 89 percent of AI calls

European AI governance layer wants to change how companies build AI by stopping unnecessary model calls before they reach the GPU. According to the company, new benchmark results show that up to 89 percent of AI calls can be avoided even before execution.

The most expensive AI is the one that should never have been run. That is the starting point for the benchmark results that VORTIQ-X Consilium The company says the results show that its AI governance platform can stop, hold back, or completely avoid unnecessary AI executions before they reach production, reducing model calls, token consumption, and GPU power consumption while allowing legitimate workflows to continue without unnecessary delays.

The publication also raises a larger question about the development of the AI market. If the greatest cost, energy consumption and security risk lies in AI processes that should never have been started, why does the industry continue to focus on faster computing power instead of controlling which AI decisions are allowed to be implemented in the first place?

From AI acceleration to AI control

In recent years, the AI industry has been dominated by investments in more powerful GPUs, larger language models, and higher performance. NVIDIA has become the world's most valuable technology company by delivering the infrastructure behind this development, while solutions like NeMo Guardrails focuses on the certainty surrounding the models' responses.

VORTIQ-X, however, believes that the real challenge in regulated businesses is no longer about how fast AI can work.

The crucial question is instead whether AI should be allowed to act at all before it moves money, opens sensitive information, changes business systems or affects critical operations. According to the company, that decision is made in a separate governance layer that sits on top of models, cloud platforms and hardware and is not owned by any single vendor.

“Acceleration drives AI. Authority governs AI. The strategically important question for every organization is no longer which GPU is used, but who controls the decision itself on top of the infrastructure. That layer works on top of both NVIDIA and other accelerator platforms.”

Benchmark shows massive reduction in AI calls

In a two-hour benchmark test that included 19,676 AI events and 18,682 protected AI actions VORTIQ-X reports the following results:

• 1,890 model integrity-related events was stopped or held back before any AI model was executed, corresponding to a stop rate of 38.78 percent.
• 0 unauthorized approvals.
• 0 incorrect routing decisions.
• 0 unexpected endpoint errors.
• 0 retention policy violations.
• Replay verification was carried out without deviations and simulated manipulation attempts were correctly identified.
• False positives are reported to be below one percent. The company's internal tests across eight different language models report no false positives.

The benchmark also shows that:

• 6,480 out of 7,200 model calls could be avoided, corresponding to 90 percent.
• About 604,080 tokens never needed to be generated, a reduction of 90.1 percent.

Separate benchmark shows reduced GPU power consumption

In a standalone benchmark focusing on energy consumption, VORTIQ-X reports:

• 23,930 avoided model calls
• About 4.1 million tokens avoided
• 14.42 percent reduction in total measured GPU power consumption.
• 65.27 percent reduction of dynamic inference energy when idling consumption is excluded.

The company emphasizes that results vary depending on workload, models and policy configurations, but describes the benchmark as evidence that between about 38 and 90 percent of unnecessary AI execution can be avoided depending on the application.

Stopping AI before the costs arise

According to VORTIQ-X The big win isn't about making every single AI call cheaper.

In modern agent-based AI environments, a single incorrect or unconfirmed response can lead to new model calls, more AI agents, more tool calls, and increased costs throughout the infrastructure.

The company's strategy is therefore to stop these chain reactions before they even start.

According to VORTIQ-X, the cheapest, most energy-efficient and safest AI execution is the one that never needs to run.

Real-time control without noticeable delay

A common objection to pre-execution control is the risk of higher response times. According to VORTIQ-X, the benchmark results show the following p95 latencies under full load:

According to the company, this means that AI control can occur in real time without becoming a noticeable burden on performance.

Reduced costs and lower risk

For organizations using AI, every avoided model call means:

• lower costs for tokens,
• less GPU usage,
• lower energy consumption,
• reduced carbon footprint,
• reduced regulatory and legal risk.

VORTIQ-X therefore believes that AI governance is evolving from being a security function to becoming a tool for both cost control and more efficient AI infrastructure.

Built for multiple models and platforms

The company describes its solution as a complementary governance layer rather than a replacement for existing AI platforms.

VORTIQ-X is designed to work on top of various language models, cloud platforms and accelerators and can be used in private AI environments, on-premises installations, air-gapped systems, bare metal infrastructure, sovereign clouds, multi-agent systems, RAG architectures, robotics and automated workflows.

The company also states that the technology is based on Swedish intellectual property law with patent applications filed through the Swedish Patent and Registration Office (PRV). The platform is designed to allow customers to choose where data is stored and which jurisdiction applies, without dependence on a single cloud provider.

The benchmark is open for review

VORTIQ-X emphasizes that benchmark results should not be accepted uncritically.

Methodology, test setup and results are available for review and the company encourages customers, researchers and industry players to evaluate the solution themselves in their own AI environments.

“We are not afraid of being scrutinized. On the contrary, we want to be tested. The only AI that belongs in banks, hospitals, government agencies or critical infrastructure is AI whose decisions can be proven to have been guided before it acted. Anyone selling AI acceleration should ask themselves who actually owns that proof. The answer is not the chip manufacturer.”

Raymond Steen, CTO and Co-Founder, VORTIQ-X Consilium

About VORTIQ-X Consilium

VORTIQ-X Consilium is an AI technology company with Swedish intellectual property rights that develops real-time deterministic AI governance for regulated and mission-critical environments. The company focuses on verifiable governance, auditability, cybersecurity and sovereign AI deployments for organizations in Europe, the Middle East and Africa.