NVIDIA Minerva targets GPU downtime, with up to $40m a year per uptime point at IREN
NVIDIA promises less GPU downtime with Minerva but gives no measured gain or price. Meta's figures put the potential recovery near ten uptime points, worth up to $40m a year each at IREN if it reaches its $4bn target.
NVIDIA launched Minerva on Thursday, software that aims to win back GPU hours lost to failures. The potential recovery is capped near ten percentage points of uptime, and NVIDIA has published no measured gain.
Each recovered percentage point adds output without new chips, power or floor space. For IREN, the cloud provider NVIDIA names as a user, one point is worth up to $40m a year if it reaches its $4bn revenue target.
NVIDIA calls Minerva an AIOps platform, software that applies purpose-built AI models to infrastructure and workload data. It pulls telemetry from GPUs, networking, storage, schedulers and jobs into one map. It groups an event storm into one incident with a likely root cause. NVIDIA says this works across a factory with hundreds of thousands of GPUs.
The target is what NVIDIA calls AI factory yield. That is the useful work a cluster produces through GPU hours, goodput, uptime and token output. NVIDIA describes the problem plainly: a single degraded component, such as a GPU running hotter than its peers, can slow an entire job.
What Meta's training run says about downtime
Meta's training data give a measure of the downtime problem. Its Llama 3 paper covers 54 days on up to 16,384 H100 GPUs. A single GPU failure may restart the entire job.
Over those days the job was interrupted 466 times. Of those stops, 419 were unexpected. That is one unexpected stop every 3.1 hours across the whole cluster.
About 78% of the unexpected stops traced to confirmed or suspected hardware faults. GPU problems alone were 58.7% of all unexpected issues. Meta says significant manual intervention was needed only three times in that period, and automation handled the rest.
Meta still reported "higher than 90% effective training time". So it lost under ten points. That includes 47 planned stops that no monitoring tool removes, so the slice faster diagnosis can win back is smaller.
Meta also says it built "tools for fast diagnosis and problem resolution" to reach that figure. The 90% therefore already includes in house tooling. It is the best public analogue, using 2024 H100 numbers from Meta's own cluster.
What recovered GPU time could be worth
IREN's FY26 results show what a point of uptime could be worth. Its annualised revenue is GPU hourly pricing for commissioned GPUs under contract, multiplied by 8,760 hours a year, plus storage and ancillaries. One point of uptime is 1% of that figure.
IREN reports $4bn of contracted annualised revenue for 2026 capacity, but the $4bn is a target to have running, not yet earned. It aims to have $4bn running by 31 December, on internal assumptions. Only $1bn was running on 26 August. At $4bn a point is worth $40m a year, and at $1bn it is worth about $10m.
Ten points would be $400m a year at full scale, the most the Meta figure allows. These are run rate figures. IREN says recognised revenue may be materially lower than annualised revenue.
IREN also says recent three year contracts earn more than $20m of revenue per megawatt of IT load, with payback of about two years. A point of uptime also avoids buying roughly one percent more GPUs.
NVIDIA has not published a measured gain or price
NVIDIA has published no measured improvement. Its blog promises "longer time between interruptions", "faster time to repair" and "greater insight", with no figure for any of them.
The product page says the software is "preventing hundreds of lost GPU hours", without naming a cluster or a period. It says predictive maintenance "increases mean time between interrupts". It gives no percentage.
NVIDIA lists no price. A buyer cannot yet set cost against benefit, and nobody outside NVIDIA can calculate a payback.
The product page promises that "actionable intelligence recovers GPU hours that would otherwise be lost to undetected failures". It adds that the machine learning analyses billions of data points an hour to learn normal behaviour and predict anomalies.
The one named user, IREN, offers no number either. Its chief technology officer, Denis Skrinnikoff, says the tool lets operations "focus more on proactive maintenance rather than troubleshooting". Tesla, IBM, Lambda, Jio, Aolani, Bitdeer AI and Quake AI "plan to deploy" it. Dell, Lenovo and Supermicro "will offer" it.
NVIDIA also controls the measurement. It calls Minerva "the primary observability solution for NVIDIA AI factories", and the tool reads NVIDIA's own infrastructure, network, scheduler and workload telemetry.
The collector will be open sourced in DSX OS, and NVIDIA says key capabilities will follow it. Customers can run Minerva as a data collector only, or as a full turnkey platform.
Generic tools and in-house dashboards already cover much of this work; buying more GPUs may have a larger effect. NVIDIA's FAQ answers that its full stack data helps operators isolate root causes that "generic tools and siloed point solutions may miss".
Meta's log shows one unexpected stop every 3.1 hours on 16,384 GPUs. Minerva may also help sell Vera Rubin, since NVIDIA presents it as part of that platform.
IREN's financing costs and next report
IREN pays financing costs even when its GPUs are idle. IREN's $2.4bn GPU loan carries a 9.0% fixed rate, about $216m a year of interest if fully drawn. The loan funds the Mackenzie air cooled expansion and covers 90% of the GPU spending there. Its $3.6bn Microsoft financing costs 6.0%.
Each point of uptime increases the revenue available to pay that unchanged interest bill.
IREN reports the quarter ended 30 September on Thursday 5 November, with a call at 5:00 pm Eastern. The June quarter brought $70.5m of AI Cloud revenue, up from $33.6m in March. That quarter's revenue will show how fast capacity came online.
Annualised revenue does not measure uptime, because it counts every commissioned GPU for all 8,760 hours. A figure for recovered hours would have to come from IREN or NVIDIA. If neither gives one, the $40m a point stays a ceiling with no measured gain behind it.