Nscale's $103.4B Backlog Has $2.6B Live as IREN Delays Build

What happened

Two Texas data center disclosures inside two days showed the gap between signed AI compute backlogs and live capacity. On 6 Nov 2026, Reuters reported that AI data center operator IREN delayed some infrastructure deliveries for Microsoft in Texas after Microsoft requested network-design changes, while laying out another $15 billion in planned capital expenditure over the rest of its fiscal year.

IREN, valued at about $7 billion, is building 200 MW of data center capacity for Microsoft in Texas. The company pushed the operational timeline for its more than $4 billion annualized run-rate revenue target from year-end to January, and said timing could extend to the end of February. Its quarterly filing showed capital commitments of $18.95 billion at the end of September, up from $13.81 billion a quarter earlier.

One day earlier, on 5 Nov 2026, Supercomputing Online reported that Microsoft declined an option to more than double its AI computing commitment with British infrastructure provider Nscale at Nscale's Cedarvale campus in Ward County, Texas. Microsoft keeps its base agreement for about 234 MW using roughly 104,000 NVIDIA GB300 GPUs, but passed on an expansion option described at roughly 500 MW to 700 MW.

Nscale's September IPO filing had disclosed approximately $103.4 billion in total contract value, led by up to $43.8 billion with Microsoft through December 2033 and $44.6 billion with Anthropic. As of August 31, 2026, only $2.6 billion of that contract value was operational, and 25,000 of 461,000 contracted GPUs were active.

What the chart shows

The first chart breaks down Nscale's $103.4 billion reported backlog against what was running at the August filing cutoff. Anthropic accounts for $44.6 billion and Microsoft for $43.8 billion, or 85.5% of the entire backlog across two customers. Operational contract value stood at $2.6 billion, or 2.5% of the headline figure.

The physical GPU count tells the same story. With 25,000 active GPUs out of 461,000 contracted, only 5.4% of the contracted fleet was live at the filing date. Even Figure's new initial computing commitment of $3.5 billion is targeted for the second half of 2027.

The second chart tracks IREN's balance sheet over the three months to September. Capital commitments rose by $5.14 billion to $18.95 billion, while GPU hardware on the balance sheet grew 2.9x, from $2.04 billion to $5.82 billion, with another $4.61 billion sitting in construction in progress.

What it means for founders

A hundred-billion-dollar compute backlog is an option book, not energized capacity. Hyperscalers sign multi-year take-or-pay schedules tied to delivery milestones, network topology and cooling specs, and they use change orders and expansion options to pace when megawatts actually come online. When Microsoft asks IREN for network-design changes or passes on Nscale's Texas expansion option, the capital burden stays on the infrastructure builder's balance sheet first.

Notice where the bottleneck moved. IREN already holds $5.82 billion of GPU hardware on its balance sheet and $4.61 billion of construction in progress, yet a fabric-design change from one customer shifts a $4 billion run-rate milestone into the new year. Buying GPUs is no longer what separates a live cluster from a delayed one; optical interconnect layout, substation commissioning and thermal integration decide when billing starts.

For AI infrastructure and software founders, the customer with the sharpest pain is no longer the model lab shopping for raw GPU hours. It is the neo-cloud operator sitting on $18.95 billion of capital commitments or 436,000 uncommissioned GPUs that do not generate revenue until cluster acceptance tests pass.

What would change our mind

We read the Texas delays and dropped options as evidence that hyperscalers are pruning unbuilt capacity while forcing neo-clouds to carry the construction and network-redesign risk. Two data points would change that view. If IREN brings its full $4 billion annualized run-rate online in January without further slippage, the Microsoft network change was a routine engineering revision rather than a demand pause.

At Nscale, the test is the initial delivery of its 234 MW Cedarvale cluster with 104,000 NVIDIA GB300 GPUs. If Nscale converts that block into operational revenue on schedule and backfills the dropped Texas expansion option with another anchor buyer, then a 2.5% operational-to-backlog ratio was simply the starting line of a steep ramp.

A $100 billion compute backlog is an option book; network commissioning and power delivery decide when billing actually starts.

What we would back: Software and hardware teams that compress GPU cluster commissioning, optical fabric testing and grid-to-rack bring-up for neo-cloud operators.