CIFR vs IREN vs NBIS: Three AI Infrastructure Models and Why the Market Keeps Confusing Them

The 43% Decline That Was Supposed to Happen

Cipher Digital’s revenue fell 43% year over year in Q2 2026 and the stock dropped roughly 32%. The revenue decline was deliberate, not a business failure. It is the arithmetic of a company switching off one revenue engine before the replacement engine starts billing.

The reported numbers were poor by any conventional reading. Q2 2026 revenue came in at $24.8 million, below the $31.9 million consensus and down from $35 million in Q1 2026. The GAAP net loss widened to $268 million from $114 million in Q1. Loss per share was $0.65 against an expected $0.24 loss. Adjusted EBITDA swung to negative $29.99 million. On a screen, those four lines look like deterioration.

The critical detail most coverage skipped is what sits inside the net loss. Approximately $150.5 million of the $268 million loss was a warrant remeasurement charge. That is a non-cash accounting adjustment tied to the fair value of outstanding warrants, not an operating outflow and not money leaving the business. Strip that single line out and the loss picture changes materially. An investor comparing the headline loss to the cash position without separating the warrant charge is comparing two different things.

The reason revenue fell is the reason to own the stock or avoid it, depending on the thesis. Cipher deliberately powered down legacy bitcoin mining machines to free physical capacity for hyperscale AI tenants. Mining revenue was being traded for contracted lease revenue. A company converting from spot-priced bitcoin production to fifteen-year contracted rent should show exactly this pattern during the transition, because the old revenue stops before the new revenue starts. There is no version of that conversion where the income statement looks clean in the middle of it.

The offsetting news arrived in the same release and received a fraction of the attention. An amendment to the Black Pearl campus lease with an investment-grade hyperscale tenant accelerated initial capacity delivery by two months, with rent commencing in early August 2026. In a business where the entire model is contracted rent, pulling forward the start of that rent is the single most valuable operational outcome available. It arrived on the same day as the weak revenue print.

The identity change matters for anyone screening the name. The company renamed from Cipher Mining to Cipher Digital in February 2026, reflecting the pivot from bitcoin mining to high performance computing data center development. The Nasdaq ticker CIFR was unchanged. Investors still classifying it as a bitcoin miner are modeling a business the company is actively dismantling. For ongoing coverage and price data, see the Cipher (CIFR) stock analysis on CleaRank.

“The market read a transition quarter as a collapse. When a business is converting from spot revenue to contracted revenue, the reported line goes down before it goes up, because you have to switch off the old machines to make room for the new tenant. Add a $150.5 million non-cash warrant remeasurement on top and the print looks far worse than the cash reality. The contracts signed during that quarter were better than the quarter itself.”

Shaun David, CleaRank Co-Founder

Cipher Digital: The Landlord Model

Cipher’s business is real estate with electricity attached. It secures cheap power and land, builds the shell and the cooling, and rents the result to hyperscalers on long-term contracts. It does not sell compute, it does not run a software platform, and it does not buy accelerators.

The anchor deal defines the model. Cipher signed a 15-year lease with Amazon Web Services covering 300MW of gross capacity at the Black Pearl campus, worth approximately $5.5 billion in contracted revenue over the initial term. Delivery runs in two phases, beginning July 2026 and completing in Q4 2026, with rent commencing in August 2026. The capacity includes both air cooling and liquid cooling for AI workloads. The company laid out the terms in its business update covering the 15-year AWS lease at Black Pearl, and the structure was reported in detail as AWS signing a 300MW hosting deal to lease capacity from Cipher.

The power position is the moat. Cipher secured long-term fixed-price power supporting costs near 2.8 cents per kilowatt-hour, among the lowest in the industry. Its Odessa facility operated at roughly 11.6 exahash per second on that power base during the mining era, which is how the company proved out the energy infrastructure before repurposing it. In a business where electricity is the largest recurring input, a locked low rate for a decade is not a marginal advantage. It is the entire spread between a profitable lease and an unprofitable one.

The water and permitting angle is underrated. Conventional data centers consume very large volumes of water for evaporative cooling and draw against municipal allocations, which slows them down in planning committees and permitting processes. Cipher’s approach uses prefabricated structures and advanced liquid cooling, which reduces that friction substantially. That is a real part of why it delivered the AWS capacity two months early. Speed to energized capacity is currently the scarcest commodity in the sector, and a developer who can shorten the path from land to powered rack is selling something hyperscalers cannot buy elsewhere at will.

Liquid cooling is not a specification detail. Modern NVIDIA accelerators generate heat densities that conventional air cooling cannot keep pace with. Building liquid cooling into the shell from the beginning, rather than retrofitting it into a facility designed for a previous generation of hardware, is what makes a site viable for current-generation AI equipment at all. Retrofits are slow, expensive, and frequently capped below the density the tenant actually needs.

The risk transfer is the defining feature of the model. Cipher does not buy GPUs. Tenants bring their own accelerators. Hardware obsolescence risk, which is the single largest uncertainty in AI infrastructure, sits with the tenant. Amazon absorbs the question of what a Blackwell-generation GPU is worth in 2030. Cipher’s exposure is counterparty credit on the lease and construction execution on the delivery schedule. Those are real risks, but they are underwriting risks rather than technology risks, and they are the kind institutional investors have priced for decades.

The backlog gives the model scale. Cipher reports roughly $11.4 billion in contracted pipeline, with AWS attached to two separate sites and Fluidstack, backed by Google credit support, at a third. Google backstopped $1.4 billion of Fluidstack’s lease obligations and received warrants for roughly 24 million Cipher shares, an approximately 5.4% pro forma stake. A hyperscaler taking equity exposure to its landlord is a signal about how scarce energized capacity has become.

IREN: The Vertically Integrated Operator

IREN takes the opposite side of the hardware question. It develops its own sites and buys the NVIDIA accelerators itself, then sells compute by the hour. Where Cipher hands the depreciation problem to its tenant, IREN keeps it and charges for keeping it.

One accuracy point deserves stating clearly, because it is frequently misreported. IREN is not simply a GPU buyer sitting on someone else’s infrastructure. It develops and owns its own data centers, with roughly 480MW on track for 2026 and 1,210MW in build for 2027. What distinguishes it from Cipher is that it also owns the accelerators on top of that infrastructure. It is the vertically integrated model, covering power, shell and silicon in a single stack.

The NVIDIA relationship anchors the AI cloud business. IREN signed a $3.4 billion five-year AI cloud contract with NVIDIA, alongside a strategic 5GW partnership that includes rights to purchase up to 30 million shares, worth roughly $2.1 billion. The arrangement covers Blackwell GPUs at the Childress site with ramp targeted from early 2027. The details were set out in the company’s business update and Q3 FY26 results covering the $3.4 billion NVIDIA AI cloud contract.

The financial picture reflects a company mid-ramp. Q3 FY26 revenue was $144.8 million, with a net loss of $247.8 million and adjusted EBITDA of $59.5 million. The company raised its year-end 2026 AI cloud annual recurring revenue target from $3.7 billion to more than $4 billion, with approximately 85% of that target under contract. The gap between $144.8 million of quarterly revenue and a $4 billion ARR target is the entire investment case, and it closes only if the buildout lands on schedule.

The economics of ownership are straightforward. By owning the GPUs, IREN captures the full spread between the cost of compute and the price it charges per hour. That is a materially higher margin opportunity than collecting rent on the building those GPUs sit in. A landlord earns a lease yield. An operator earns the compute margin, which in the current pricing environment is considerably wider. It also means IREN, not a tenant, absorbs the obsolescence cycle. When a new accelerator generation ships, IREN owns the previous generation.

The honest framing of that risk is that nobody knows the answer yet. GPU depreciation schedules are one of the most contested assumptions in the entire AI trade. If accelerators hold economic value for five or six years, IREN’s model produces exceptional returns and the current losses look like a financing bridge to a very profitable asset base. If the useful life is closer to three years, the math changes considerably and the same capital produces a much thinner return. No operator has run a full accelerator cycle from purchase to retirement under current demand conditions, so both camps are arguing from theory.

The bull case and the bear case on IREN are the same fact viewed from two angles: it owns the hardware. For price data, contract history and ongoing coverage, see the IREN Limited (IREN) stock page on CleaRank.

“Hardware ownership is the real fork in the road, and most comparisons skip straight past it. Cipher hands the depreciation risk to Amazon and collects rent for fifteen years. IREN keeps that risk and gets paid a wider margin for keeping it. Both can be correct positions in the same market. What an investor cannot afford is to buy one while believing they own the other.”

Shaun David, CleaRank Co-Founder

Nebius Group: The Platform at the Top of the Stack

Nebius sits highest in the value chain. It is a full cloud platform with a large engineering organization, and it funds its GPU purchases through structured debt secured against its own deployed infrastructure. It owns hardware like IREN, and then builds a software business on top of it.

The Q2 2026 results describe a company scaling faster than almost anything else in the sector. Group revenue reached $582 million, up 454% year over year. Annual recurring revenue reached $3 billion, up 598%. The core AI cloud business generated $575 million, up 514%. Adjusted EBITDA swung to a positive $236 million from a $21 million loss a year earlier, with margin at 41%. The move from an EBITDA loss to a 41% margin inside four quarters is the number that separates Nebius from most of the AI infrastructure cohort.

The contract book explains the durability behind those growth rates. Nebius signed four landmark mid-term contracts each averaging over $1 billion, with customers including Reflection, Cohere, a US-scale AI lab, and a large quantitative trading firm. Deals in the core AI cloud averaged above $20 million in revenue per megawatt, which is a far higher monetization rate per unit of power than a lease model can produce. Pricing on older-generation GPUs ran more than 30% higher than in Q1, which is a notable signal about supply tightness: when previous-generation hardware reprices upward, the shortage is real rather than anticipated.

The financing structure is the genuinely unusual part and deserves attention beyond the growth figures. Nebius entered its first senior secured debt facility of approximately $775 million, backed by deployed GPU infrastructure and contracted cash flows from an investment-grade customer. The facility matures October 31, 2030 and is priced at SOFR plus 2.50%. Financing accelerator purchases against the contracted revenue those accelerators generate is a structural advance for the sector. It converts GPUs from a pure equity-funded capital expense into a financeable asset class, which changes how fast a platform can scale without diluting shareholders.

The balance sheet position reinforces the point. Total deferred revenue stands near $6 billion, up from roughly $4.8 billion at the end of Q1, and management expects customer prepayments to deliver more than $9 billion of upfront funding across 2026. Customers paying in advance are effectively financing the buildout at zero cost. Full-year guidance is $7 to $9 billion in ARR, $3 to $3.4 billion in revenue, and roughly 40% adjusted EBITDA margin.

Execution is harder at this layer, and that is the trade. Running a full platform means orchestration, scheduling, storage, networking, developer tooling and support, not just powered racks with hardware in them. That requires hundreds of engineers and creates far more ways to fail than a lease or an hourly rental. It also creates the strongest customer lock-in of the three models once it works, because customers build workflows, pipelines and internal tooling on the platform rather than simply renting capacity they can move elsewhere next quarter. For the full financial history and ongoing coverage, see the Nebius Group (NBIS) stock page on CleaRank.

The Comparison Table

The table below places the three models side by side on the dimensions that actually differentiate them. Read the hardware ownership row and the primary risk row together, because those two lines explain most of the difference in how the three stocks behave.

Which Risk Do You Actually Want?

The right question is not which company is best. It is which risk an investor is willing to underwrite. Each of the three models is a different answer to the same question about who absorbs uncertainty in the AI buildout, and each is priced accordingly.

Cipher’s risk is concentrated in counterparty credit and construction execution. If Amazon pays for fifteen years, the model works almost regardless of what happens to GPU generations, chip architectures or compute pricing. The downside case is a tenant default or a build that runs late and blows its delivery schedule, pushing rent commencement to the right while interest and construction costs keep accruing. It is the most defensive of the three, and correspondingly the least leveraged to an AI compute boom. Rent does not accelerate when demand spikes.

IREN’s risk is depreciation. It gets paid a premium for absorbing hardware obsolescence, and that premium is real and visible in the margin structure. The question is whether the useful economic life of an accelerator supports the return assumptions embedded in the buildout. Investors comfortable with hardware cycles, who have watched servers and networking gear earn returns well past their nominal replacement dates, will find the risk-reward attractive. Investors who believe refresh cycles are compressing as each NVIDIA generation widens the performance gap should be cautious, because that view breaks the model directly.

Nebius’s risk is execution across a much wider surface area. The reward is the widest margin and the strongest customer stickiness, evidenced by roughly $6 billion in deferred revenue and prepayments funding the business before a single invoice comes due. The risk is that platforms are hard. Scheduling, storage performance, network reliability and support quality all have to work simultaneously, and a platform that underdelivers on reliability loses customers who have credible alternatives at AWS, Microsoft Azure, Google Cloud and CoreWeave. Growth of 454% invites scrutiny of whether the operational organization scaled with the revenue.

A note on correlation is necessary before anyone treats this as a diversification exercise. All three depend on sustained AI infrastructure demand. Spreading capital across the three does not remove that exposure. It only changes which failure mode dominates in each position. A broad slowdown in AI capital spending damages all three, in different ways and on different timelines, but it damages all three.

What Could Break Each Thesis

Cipher’s vulnerability starts with concentration. The model depends on a small number of tenants, and AWS plus Fluidstack account for the overwhelming majority of the $11.4 billion backlog. Construction delays convert directly into delayed rent, and a fifteen-year lease that starts six months late is a fifteen-year lease with a six-month hole in its financing schedule. The bitcoin mining transition means reported financials will look poor throughout the changeover, which can pressure the stock even when the underlying contracts are fully intact, as Q2 2026 demonstrated. Regulatory and permitting risk on power procurement remains, and the cheap power position that defines the model is only as durable as the contracts underneath it.

IREN’s vulnerability is the depreciation question, which is genuinely unresolved rather than merely debated. Large net losses continue while the AI cloud ramps, and a $247.8 million quarterly net loss requires either continued capital raising or a fast path to cash generation. Contracted ARR is a target, not collected revenue, and roughly 85% contracted means 15% is not, on a figure that itself sits well above current run-rate revenue. Concentration around the NVIDIA relationship cuts both ways: it delivers hardware access and a marquee customer contract, but it also means a single commercial relationship shapes an outsized share of the forward story.

Nebius’s vulnerability begins with the arithmetic of its own growth. Growth of 454% is not repeatable indefinitely, and the comparison base hardens with every quarter, which means the reported growth rate will decelerate sharply even if the business performs perfectly. Debt secured against GPU infrastructure is a new structure that has not been tested through a downturn in compute pricing, and the collateral value in that structure is exactly the asset whose useful life the market cannot agree on. Customer concentration among a handful of billion-dollar contracts means a single non-renewal is material to both revenue and the ARR narrative that supports the valuation.

One Buildout, Three Different Places to Take Risk

The habit of grouping these three as AI data center stocks obscures the only thing that matters, which is where in the stack each one sits and what it is paid to absorb. They appear on the same screens, move on the same headlines, and get discussed in the same sentence, but they are not the same business and they do not fail for the same reasons.

The recap is simple once the models are separated. Cipher is real estate with cheap electricity and liquid cooling, renting to Amazon on fifteen-year terms without touching a chip. IREN owns the silicon and sells the compute, capturing the full spread and the full depreciation risk that comes with it. Nebius runs the platform above the hardware, borrowing against its own deployed GPUs and its customers’ prepayments to fund the next batch. Landlord, operator, platform. Three different jobs in one supply chain.

The Q2 reaction to Cipher illustrates the cost of the confusion better than any argument. A 43% revenue decline read as deterioration when it was a deliberate conversion from spot bitcoin revenue to contracted rent, amplified by a $150.5 million non-cash warrant charge that never touched the cash account. Investors who understood the model saw a transition quarter with an early delivery bonus attached. Investors who did not saw a collapse and sold into it.

The AI buildout will not reward every participant equally, and it will not punish them for the same reasons. Choosing among these three is less a forecast about AI demand than a decision about which specific risk an investor is best positioned to hold: a tenant’s creditworthiness, a chip’s useful life, or an engineering organization’s ability to keep a platform running while it grows five times over.

Frequently Asked Questions

Disclaimer: This comparative analysis of Cipher Digital (CIFR), IREN Limited (IREN), and Nebius Group (NBIS) is for informational purposes only and does not constitute financial, investment, or legal advice. All three companies depend on sustained AI infrastructure demand, and none is diversified against a broad slowdown in AI capital spending. Cipher’s reported results will appear weak through the transition from bitcoin mining to high performance computing, and the company depends on a small number of tenants for the majority of its contracted backlog. IREN carries GPU depreciation risk whose useful-life assumptions remain unresolved across the industry, and it continues to report large net losses while its AI cloud business ramps. Nebius’s growth rates are not repeatable indefinitely as its comparison base hardens, and its debt structure secured against GPU infrastructure is new and untested through a downturn in compute pricing. Contracted ARR targets, lease backlogs, and pipeline figures cited in this article are commitments rather than collected revenue, and delivery schedules can move. All three stocks have shown high volatility, and position sizing should account for that. CleaRank and its analysts may hold positions in securities mentioned in this article. Past performance is not indicative of future results. Always consult with a licensed financial advisor before making investment decisions.