The Blackwell Ultra Surge: IREN’s $3.5 Billion Bet on 50,000 NVIDIA B300 GPUs

SYDNEY / NEW YORK — March 4, 2026 — In a move that cements its transition from a Bitcoin miner to a top-tier AI powerhouse, IREN Limited (formerly Iris Energy) has announced a massive purchase of over 50,000 NVIDIA B300 “Blackwell Ultra” GPUs. This single order—one of the largest in the sector’s history—aims to supercharge IREN’s AI Cloud fleet to a total of 150,000 units by the end of the year.

The announcement sent waves through the market, briefly pushing IREN’s stock up 13% before a subsequent $6 billion equity filing sparked dilution concerns. However, the underlying message remains clear: the race for “time-to-compute” has moved into the Blackwell Ultra era.

The B300: Why “Ultra” is the New Standard

NVIDIA’s B300 Blackwell Ultra, which began shipping in early 2026, has quickly become the primary target for hyperscalers like Microsoft and Google. The “Ultra” designation isn’t just marketing; it represents a major architectural leap over the original Blackwell (B200) series.

Massive VRAM: The B300 features 288GB of HBM3e memory, a 50% increase over the B200. This allows a single GPU to hold a full 70-billion-parameter model in high precision, eliminating the need for complex and slow “sharding” across multiple chips.
Reasoning Powerhouse: With 15 petaFLOPS of dense FP4 compute, the B300 is specifically designed for the “Reasoning” phase of AI—powering models that “think” through multiple steps before responding.
Power Density: This performance comes at a cost; each B300 chip draws a staggering 1,400W, requiring the advanced air-cooled and liquid-cooled infrastructure that IREN has spent the last year building.

Phased Deployment: Mackenzie & Childress

IREN plans to deploy the 50,000-unit fleet in phases throughout the second half of 2026. Rather than building new facilities from scratch, the company is leveraging its existing, vertically integrated “ready-to-plug” data centers.

Mackenzie, British Columbia: A major portion of the B300s will be installed in IREN’s Canadian flagship site, which is already optimized for high-density compute.
Childress, Texas: The massive Texas site will act as the primary hub for the expansion, utilizing its air-cooled infrastructure to handle the Blackwell Ultra’s thermal load.
Economic Scale: IREN projects that once the full 150,000-GPU fleet is operational, it could support an annualized run-rate revenue of over $3.7 billion.

Financing the $3.5 Billion Capex

Securing 50,000 state-of-the-art GPUs in a supply-constrained market requires immense capital. IREN has spent the last eight months building a $9.3 billion war chest through a mix of:

Customer Prepayments: Notably from its landmark deal with Microsoft (signed November 2025).
GPU Financing: Backed by major institutions like Goldman Sachs and JPMorgan.
Equity Management: A new $6 billion “at-the-market” (ATM) program ensures the company can fund the $3.5 billion in additional capex required for the B300 servers, storage, and networking.

Vertical Integration: The Competitive Edge

“Scaling to 150,000 GPUs positions IREN among the largest AI cloud infrastructure providers globally,” said Daniel Roberts, Co-Founder and Co-CEO of IREN.

By owning the land, the power grid connections (nearly 3 GW across its portfolio), and the data center buildings, IREN bypasses the 2–3 year lead times currently facing competitors who are waiting for new grid approvals. This “infrastructure-first” approach is why NVIDIA designated IREN as a Preferred Cloud Partner in early 2026.

IREN GPU Fleet Evolution

Date	Fleet Size	Primary Hardware	Annualized Revenue (Est)
Sept 2025	23,000	H100 / H200	$0.8 Billion
Jan 2026	100,000	B200 / H200	$2.3 Billion
Dec 2026 (Target)	150,000	B300 (Blackwell Ultra)	$3.7 Billion