On 5 November 2026 DIGITIMES reported that South Korea’s Moreh claims AMD Instinct MI355X GPUs running its inference software delivered more than twice the throughput of Nvidia B200 GPUs in company tests. The claim sits within South Korea’s plans for AMD-inclusive AI infrastructure. Results are Moreh’s own measurements and are not presented as an independent third-party benchmark in the same report.

What it means: For Greek firms planning private inference clusters, the point is not “buy chip X” but that stack software can swing tokens-per-second and cost-per-request across GPU vendors. Run tests on your own models, batch sizes and latency targets before CAPEX locks in. Vendor-reported figures raise overstatement risk; demand reproducible scripts and comparable workloads. Put performance acceptance criteria and remedies into procurement contracts. Korea’s push to diversify AI hardware also echoes European interest in avoiding single-supplier lock-in. FinOps teams should track cost per 1,000 tokens alongside power and EU support coverage. ISVs selling inference in Greece reduce customer risk by supporting multiple backends, not only one GPU ecosystem. In practice, teams should name an owner, update the risk register and agree success criteria within 30 days. Where suppliers or customers are affected, send a short notice that points to the primary source and avoids overstated claims. Compliance, cybersecurity and payments operations are linked; a change in one area often requires a pass over contracts, access rights and continuity plans. Finally, keep a review date so the item becomes concrete actions for the Greek business rather than newsletter noise alone.

Read the original on DIGITIMES

Related service: Custom software

Related stories