Hikari Blue

Enterprise AI architecture · Cloud capacity

Microsoft pushed delayed Azure GPU capacity into 2027. Quota does not reserve yours.

Microsoft told investors on October 28, 2026 that Azure demand exceeds supply and that third-party GPU deliveries moved to January through June 2027. Its supplier owes it delay credits. We map what protects an enterprise workload in the meantime.

On October 28, 2026, Microsoft reported $45.2 billion of quarterly capital expenditures and said some third-party capacity deliveries expected in the quarter were delayed. Its CFO then gave buyers the sentence that matters: customer demand continues to exceed available supply.

We read the call next to the filings of one named GPU supplier and Microsoft's own deployment documentation. The 3 documents describe one supply chain. Microsoft is protected against its supplier's delay by contract. An enterprise running AI on Azure is protected mainly by the deployment type it chose.

What Microsoft disclosed

On the fiscal first-quarter call, Amy Hood put capex at $45.2 billion, including $7.4 billion of finance leases. She guided December-quarter capex above $50 billion and calendar 2026 to about $170 billion, "due to the 3P capacity delays." CNBC reports the prior forecast at $175 billion.

The delayed deliveries "shifted to H2." Microsoft's fiscal year ends on June 30, so H2 means January to June 2027. The Azure guide of 49% to 50% growth assumes efficiency gains offset "some 3P GPU as a service contracts that have been delayed to H2."

Two further lines describe how scarce capacity is shared. Satya Nadella described "a fungible fleet that allows us to dynamically allocate capacity to the highest-value opportunities across compute, data, Foundry, and Copilot." Hood said efficiency gains let Microsoft "put more allocation into Foundry," and that token pricing came in higher than expected. Commercial remaining performance obligation grew 73% to $677 billion, about 30% of it due within 12 months.

What the supplier filed

IREN signed a GPU services agreement with Microsoft in November 2025. Its Form 8‑K gives a contract value of about $9.7 billion through 2031, NVIDIA GB300 GPUs across 4 facilities in Childress, Texas, and tranches "targeted for deployment during 2026."

The same filing lists what Microsoft bought besides compute: 20% of each tranche paid before delivery, service level commitments, delay credits, and a right to terminate, after a cure period, if agreed delivery dates are missed.

On November 5, 2026, IREN reported Horizon 1 delivered to Microsoft and Horizons 2 to 4 in late-stage construction. Its target of more than $4 billion of annualized revenue moved from year-end 2026 to January 2027, and "may extend to February 2027," after "a customer-requested change and supplier timing."

An expert objects, and correctly. IREN is one supplier among several; CNBC also names CoreWeave and Nebius. IREN's target covers a 0.3 GW program, larger than the Microsoft contract, about 200 MW. We do not know IREN's share of Microsoft's shortfall. We know the dates moved in the same direction, and that Microsoft's contract prices the delay.

Microsoft priced its supplier's delay into a contract. A workload on a shared Azure deployment absorbs the same delay as variable throughput. Hikari Blue · operator note

Where the value sits

Microsoft's provisioned throughput documentation states the split. On standard deployments, "inference capacity is shared across customers and throughput can vary with demand." A provisioned deployment holds a fixed amount of capacity exclusively for its owner.

The same page adds that "having PTU quota doesn't guarantee that capacity is available." If a region is short, the deployment fails. When spillover is configured, overflow from a full provisioned deployment goes to a standard deployment, which is the shared pool.

Our reading, not Microsoft's statement: when demand exceeds supply and the fleet is allocated to the highest-value opportunities, uncommitted shared demand is served last. The value of reserved capacity is the delay it avoids on workflows where a slow answer costs money: claims intake, a service queue, fraud review.

What management decides now

First, inventory. List every production workflow that calls an Azure-hosted model, by deployment type and region.

Second, rank by cost of delay. Where it is material, hold provisioned capacity on a reservation and keep it. Microsoft's billing page warns that a scale-down and scale-up strategy "can leave you without capacity when you need it most."

Third, test a second route before it is needed: another region, or another model on another provider. A fallback model is a different model, so it passes the same evaluation before it carries production traffic. The NIST AI RMF, GOVERN 6.2, asks for contingency processes for failures in high-risk third-party AI systems.

Fourth, the contract. Microsoft secured delivery dates, delay credits and a termination right upstream. Ask the account team, in writing, what throughput commitment your agreement carries, in which region, and what the remedy is when it is missed.

What could prevent value capture

Idle cost comes first. Provisioned capacity is billed hourly whether or not requests are made. An oversized reservation converts a supply risk into a margin cost.

Capacity may not be for sale. Microsoft's guidance is to "always verify capacity availability before planning a deployment or purchasing a reservation." In a short region, the plan fails at deployment, not at purchase.

Residency limits the second route. A regulated firm's fallback region must sit inside the data boundary its compliance team signed. A fallback outside it is not one.

Supply can return. If the delayed capacity lands between January and June 2027, scarcity eases. A reservation sized on today's shortage outlives it.

If our shared Azure deployments slowed by half for a week, which workflow would miss its service level first, and does it run on reserved capacity today?

Track 2 internal numbers: the share of production AI requests served from reserved capacity, and the throttling rate (HTTP 429) per workflow. Track 2 external ones: IREN's January 31, 2027 target, and whether Microsoft's December-quarter capex clears $50 billion.

Thesis. When a provider rations capacity, an enterprise keeps its throughput through reserved capacity and a tested second route, both built into the architecture. The 4 tests of an AI operating layer

  • Microsoft Corporation (October 28, 2026). FY27 First Quarter Earnings Conference Call, transcript published by Microsoft Investor Relations. Primary source for: capex of $45.2 billion, finance leases of $7.4 billion, cash paid for property and equipment of $39.4 billion; third-party capacity delays; December-quarter capex above $50 billion and calendar-year capex of about $170 billion; Azure growth of 48% and guide of 49% to 50%; "Customer demand continues to exceed available supply"; fungible fleet allocated to the highest-value opportunities; allocation into Foundry; token pricing; commercial RPO of $677 billion, up 73%, about 30% within 12 months. microsoft.com, call transcript
  • CNBC (October 29, 2026). AI buildout hits snag as Google and Microsoft put off raising capex forecasts. Press source for: prior Microsoft forecast of $175 billion; third-party GPU providers named as CoreWeave, IREN and Nebius. cnbc.com
  • IREN Limited (November 3, 2025). Form 8‑K, agreement with Microsoft dated November 2, 2025. Primary source for: contract value of about $9.7 billion through 2031; NVIDIA GB300 GPUs; Horizon 1 to 4 facilities in Childress, Texas, about 200 MW of IT load; tranches targeted for 2026; 20% prepayment per tranche; service level commitments, delay credits, termination right on missed delivery dates. sec.gov, Form 8‑K
  • IREN Limited (November 5, 2026). First quarter fiscal 2027 results, Exhibit 99.1. Primary source for: Horizon 1 delivered to Microsoft, Horizons 2 to 4 in late-stage construction; target of more than $4 billion of annualized revenue moved from year-end 2026 to January 31, 2027, possibly February 2027; 0.3 GW program; revenue subject to commissioning, testing and customer acceptance. sec.gov, Exhibit 99.1
  • Microsoft Learn (last updated July 15, 2026). What is provisioned throughput for Foundry Models? Primary source for: shared capacity on standard deployments; exclusive capacity on provisioned deployments; quota does not guarantee capacity; spillover to standard deployments; verify capacity before purchasing a reservation. learn.microsoft.com, provisioned throughput
  • Microsoft Learn (last updated May 22, 2026). Provisioned throughput billing. Primary source for: hourly billing of reserved capacity; capacity risk of scale-down and scale-up strategies. learn.microsoft.com, PTU billing
  • National Institute of Standards and Technology (January 2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100‑1, GOVERN 6.1, GOVERN 6.2 and MANAGE 3.1. nvlpubs.nist.gov, NIST AI 100‑1
  • Our arithmetic, on reported inputs. Microsoft fiscal 2027 first half capex: $45.2 billion + more than $50 billion = more than $95.2 billion. Commercial RPO due within 12 months: 30% × $677 billion, about $203 billion. IREN prepayment if every tranche delivers: 20% × $9.7 billion, about $1.94 billion. Fiscal H2 2027: January 1 to June 30, 2027.

The Hikari Blue team · Austin, November 2026

Map your exposure before the board does.

30 minutes with an operator. No slides.

Direct call with one of the partners. We listen, we structure, and we tell you what the arbitration layer would look like (model routing, prompt sovereignty, audit trail) for your specific stack and regulatory perimeter.

Book a 30‑min call

A named partner replies quickly, in writing. If it fits, the next step is an Operating Assessment: 1 week, from $18,000.

The engagement behind this briefing: AI governance that answers to a regulator