Iren Ltd.'s Q4 2025: Contradictions in GPU Financing, Horizon One Capacity, and CapEx Allocation Strategies

The above is the analysis of the conflicting points in this earnings call

Date of Call: None provided

Financials Results

• Revenue: $187M, up $42M sequentially

Guidance:

• Scaling AI cloud to ~10,900 GPUs; commissioning over coming months; targeting $200M–$250M in annualized AI cloud revenue by December.
• Plan to maintain ~50 EH/s by reallocating ASICs; mining generates ~$1B annualized revenue at current economics.
• Horizon 1 (Childress) on schedule for Q4 2025; Tier III-like redundancy; flexible rack densities.
• Early works and long-lead procurement started for Horizon 2 (50MW liquid-cooled).
• Sweetwater 1 energization targeted for April 2026; construction and substation upgrades progressing.
• Secured 100% GPU purchase-price financing at single-digit rates; 2–3 year paybacks; near-term CapEx funded.
• Expect strong AI cloud margins given low power costs and fully owned data centers.

Business Commentary:

* Record Financial Performance: - Iron delivered 10 times EBITDA growth year-on-year, with strong net income, attributed to a 33% increase in contracted grid-connected power to nearly 3 gigawatts. - The growth was driven by significant increases in Bitcoin mining capacity and expansion into AI cloud services.

• AI Cloud Expansion:
• The company plans to scale its GPU deployments to 10,000, expecting at least $2,000,000,000 to $2,500,000,000 of annualized revenue by December.

• This expansion is supported by non-dilutive financing for GPU purchases and significant demand for AI cloud services due to an industry-wide supply constraint.

• Bitcoin Mining Profitability:

• Iron's Bitcoin mining generated over $1,000,000,000 in annualized revenue, with a cash cost of $36 per bitcoin mined versus an average realized price of $99,000.

• The profitability is attributed to best-in-class fleet efficiency and low net power costs, which support the company's AI cloud expansion.

• Data Center Capacity Growth:

• The company increased its operating data center capacity by over 300% to 8 ten megawatts, driven by projects like the Sweetwater two gigawatt data center hub and Horizon 1.
• This growth aligns with the company's strategy of scaling across the full AI infrastructure stack and capturing a broad addressable market in AI services.

Sentiment Analysis:

• Management cited “record results,” “10x EBITDA growth YOY,” and a “strong bottom line.” They’re “scaling to more than 10,000 GPUs,” became an “NVIDIA preferred partner,” secured “100% GPU financing at single-digit rates,” kept Horizon 1 “on schedule for Q4 2025,” and confirmed Sweetwater 1 “energization in April 2026.” They’re “targeting $200M–$250M annualized AI cloud revenue by December.”

Q&A:

• Question from Paul Golden (Macquarie): How do PUE differences across sites influence GPU rollout prioritization, and what is your approach to backup generation/UPS given customer requirements?
  Response: BC sites run ~1.1 PUE (air) and <1.2 with liquid cooling; Horizon 1 averages ~1.2 (peak <1.4). Deployment is driven by customer demand, and Iron is adding gens/UPS across the GPU fleet to meet redundancy asks.

• Question from Paul Golden (Macquarie): What did the GB300 NVL72 flexibility entail at Horizon 1, and any implications for financing or Rubin readiness?
  Response: They reworked electrical/mechanical to support lower rack densities while retaining 200kW rack capability; GB300s (~135kW/rack) and future Rubins remain supported; no change to Rubin readiness.

• Question from John Todaro (Needham): What is the average duration of AI cloud contracts relative to GPU paybacks?
  Response: Contracts range from 1-month rolling to 3 years; newer-gen (e.g., B200/GB300) are trending multiyear, aligning with ~2-year GPU and ~3–4-year GPU+infra paybacks.

• Question from John Todaro (Needham): Given cloud traction, how are you weighing cloud vs. colocation and leverage targets?
  Response: They’ll optimize risk-adjusted returns: cloud offers higher margins/shorter paybacks; colo is longer-term/lower returns. Leverage will scale prudently with cash flow visibility and offtake quality.

• Question from Darren Aftali (Roth): How will Horizon 1 and 2 be utilized (cloud vs. colocation), and will you borrow power from Bitcoin operations? Also, color on FluidStack/neo cloud demand.
  Response: Horizon 1 is GPU-only (liquid-cooled) with flexibility to monetize via cloud or colo; lower-density support added. They may reallocate mining MW as needed and keep building. Cloud paybacks (~3 years for GPU+infra) currently look most compelling.

• Question from Joseph Baffi (Canaccord Genuity): What are the end-of-lease options for Blackwell financing, and how much financing capacity exists for GPUs vs. colocation?
  Response: They use both full-payout and FMV leases with end-of-term options to retain or rotate. Financing terms depend on customer mix, credit, and duration; both asset-level (cloud/colo) and corporate markets (convertibles/debt) remain available.

• Question from Reggie Smith (JPMorgan): What key hires have you made for cloud/colo, and how are you winning AI clients?
  Response: They’re hiring across operations, networking/InfiniBand, software, sales, solutions architects, and marketing. Demand comes via inbound/outbound, conferences, and word-of-mouth; vertical integration, capacity, and performance/cost advantages are differentiators.

• Question from Brett Knobelk (Cantor Fitzgerald): Do you order GPUs ahead of signed customers, and what are the power-per-GPU dynamics?
  Response: They buy ahead to meet on-demand needs; capacity sells quickly, and ordering pace follows traction. Power per GPU is rising; ~20k B200s fit in 50MW (Prince George) vs. ~19k GB300s per 50MW (Horizon 1).

• Question from Nick Giles (B. Riley): How will Horizon 1 be split between your own GPUs vs. colocation, and could you mix both?
  Response: They can blend cloud and colocation within Horizon 1 (and Horizon 2) and will choose the mix that maximizes risk-adjusted returns as market conditions evolve.

• Question from Nick Giles (B. Riley): You focus on bare metal today; when would you layer software and what’s the upside?
  Response: Large, sophisticated customers prefer bare metal and bring their own orchestration; software may be layered later to serve smaller startups/enterprises but isn’t required for core demand.

• Question from Steven Gagola (Jones Trading): What are the benefits/economics of participating in NVIDIA’s Lepton marketplace, and does cloud offering conflict with colocation?
  Response: Lepton can broaden reach and simplify onboarding; it’s an additional demand channel (still early). Offering cloud doesn’t hinder colo; customers value Iron’s proven ops at scale.

• Question from Ben Summers (BTIG): Why begin Horizon 2 early works now, and how does customer scale at Sweetwater differ?
  Response: They’re placing long-lead orders to preserve fast time-to-power given strong demand signals; full CapEx remains optional. Sweetwater targets very large-scale customers with flexible structures.