Last Updated on March 13, 2026 by Jason Andrew
In 2013, two Aussie brothers bought some Bitcoin, only to sell thinking the market had ‘topped’.
“Yeah nah, mate, that’s probably as high as she goes.”
But five years later, they took another look at Bitcoin, betting that Satoshi’s invention was the future of digital currency by starting a crypto mining project in Canada.
That business is IREN.
They went public in 2021, only to have their stock crushed by 95% in the 2022 Bitcoin crash.
Despite almost losing the company to debtholders, the company survived, and in 2023 pivoted the business to go all in on the AI thesis.
Today, IREN is a US$12bn company and would be the 37th largest company on the ASX if not for a US listing. Up over 300% this year, IREN is ripping on the AI thesis.
But are we in a bubble? Is the AI hype real?
This is the story of how two Aussie brothers built a $10bn empire with IREN, and a breakdown of the state of the data centre market.
IREN Story
IREN started in 2018 when brothers Daniel and Will Roberts saw an opportunity to use renewable energy to mine Bitcoin at attractive returns.
In the words of the IREN prospectus:
“We are a Bitcoin mining company. We build, own and operate data centres and electrical infrastructure for the mining of Bitcoin primarily powered by renewable energy.”
Will and Daniel were uniquely positioned to build a business around this thesis, having previously worked as renewable energy infrastructure investors at Macquarie and Palisade Investment Partners. The business went public on the NASDAQ in November 2021, raising ~$232m at a $1.5bn valuation, positioned as a clean Bitcoin miner (EEA Partners).
The business historically traded poorly given the volatility inherent in Bitcoin’s price, an external factor outside the control of the company, which has a huge impact on the viability of the business. This discount is not unreasonable, given that the company becomes unprofitable (on an EBITDA basis) at a $45k Bitcoin price.
IREN, like many neoclouds, was conveniently positioned on large data centre assets which could be repurposed to serve AI workloads after the recent language model takeoff. In the last 12–18 months, the business has started to undergo this transition, swapping out ASICs for NVIDIA GPUs in many of their data centres, given the huge spike in demand for these services among the hyperscalers.
Over this period, IREN has also generated significant positive free cashflow (blue line below), whilst maintaining heavy capital investments in order to fund their future growth, showing management’s ability to execute on a thesis of being an opportunistic digital infrastructure provider.
“All Tesla has done in relative terms over the past decade is figure out how to produce factories that churn out increasingly autonomous vehicles. Similarly, IREN is successfully becoming the machine that prints data centres.” — Antonia Linares
Financials & Operating Advantages
IREN has a number of advantages over other neocloud players.
- Full Stack Operation – the business is ‘full stack’, controlling everything from the land to data centre construction and operation. This gives them unique expertise in building and operating data centres at scale that other players do not yet have.
- No Colocation Fees – Because IREN is full stack, they pay no colocation fees, unlike many of the other neoclouds on offer (in other words, they own the data centre and do not rent capacity in existing data centres in order to provide cloud services).
- Power Procurement – IREN has procured ~3 GW of grid-connected power capacity with access to cheap renewable power across British Columbia and West Texas at $0.03 – $0.04/kWh.
- Land Control – The company owns 1,800 acres of land across four active sites (Canal Flats, Mackenzie, Prince George, Childress) and one development site (Sweetwater), all of which have been de-risked from a permitting and grid-access standpoint (810 MW of which are currently active, with 2.2 GW set to come online in the coming years).
Let’s talk numbers, then.
Financially, the business is divided into two parts:
- Bitcoin Mining Business ($485m rev FY25)
- AI Cloud Business ($16m rev FY25)
Despite the Bitcoin mining business being significantly larger today, the AI Cloud business is ramping much faster moving forward. From an enterprise value perspective, it also offers the ability to lock in contractual revenue instead of being indexed to the Bitcoin price. This should allow IREN to trade at a higher multiple moving forward, something that CEO Daniel Roberts has commented on.
Numbers
Financially, IREN’s revenue has ramped significantly over the last few years as their infrastructure projects have slowly come online, with revenue ramping significantly over the last year.
The business generated ~$270m in EBITDA last year and had operating cash flow of $245m (excluding capex). That being said, on an EBITDA less capex basis, IREN generated a loss of ($1.1bn). Wowzers.
This was split between upgrades to the Childress and Sweetwater facilities, as well as increasing spend on Bitcoin ASICs. IREN is still investing heavily into future growth, understandable given the size of the prize and the potential to serve huge cloud contracts for the hyperscalers.
It’s difficult to tell the split between maintenance and growth capex here, so it’s hard to say how profitable IREN’s operations have been to date, but on a cash basis, the business doesn’t look great.
On top of their recent strong performance, on 3 November, IREN announced a $9.7bn 5-year GPU deal with Microsoft, sending the stock surging 20%.
But what is the contract actually worth to the business?
The deal involves:
- $9.7bn of total contract value
- $5.8bn GPU capex, including ancillary equipment
- 5-year average term targeting phased starts in 2026
- 20% prepayment, credited to years 3–5
- NVIDIA GB300 GPU deployments
- 85% EBITDA margins (excl GPU capex), $1.94bn ARR deal
Doing some rough back-of-the-napkin math, this will add ~$490m per year in EBITDA less capex to the business over the five years. Customer concentration is obviously a big risk for IREN here, given that the Microsoft deal will account for roughly ~55% of their 2026 AI Cloud revenue target.
That being said, from a compute capacity perspective, this contract and the subsequent cloud revenue IREN expects to generate only account for 16% of grid-connected power (460 MW), showing the potential to diversify across more of the hyperscalers as they secure access to more GPUs in the coming years (more on this later).
Sweetwater Development
Nothing gets me fired up more than a big construction site. And IREN’s got a big one in the works.
Project Sweetwater is set to come online at the end of 2027, with ~2000 MW of capacity (vs 810 MW across their existing portfolio). Notably, Sweetwater already has contractual access to renewable, grid-connected power, unlike many other data centre projects in development, with power shortage being cited as one of the biggest barriers to growth, as data centres are expected to grow to roughly ~10% of US electricity consumption by 2030 from a mere 3% today.
CEO Daniel Roberts has spoken extensively about this, arguing that “The word pipeline implies made-up spreadsheets … (it’s) really hard to get grid-connected power. You’ve got to get that signed contract” when speaking of other players in the data centre market.
Relative to current data centre capacity on the market, Sweetwater would be the largest data centre in the world, although larger projects like Stargate (10 GW) are currently under development.
So what will IREN do with all of this capacity?
Signing more long-dated AI cloud contracts needs to be top of mind for IREN given the growing supply of capacity amongst other neoclouds. Similar to the glut in dark fibre following the Telco Bubble of 1999/2000, the adage that “supply creates demand” is a risky one to live by.
The narrative to date has been that selling “picks and shovels” in the age of AI has been a good business because it’s difficult to tell exactly how the AI models will be used. But what happens when you are selling shovels and the gold mines are
Data Centre and NeoCloud Market
The AI data centre market has exploded.
Much of this spend has been driven directly by the hyperscalers, with ~$400bn in capex in 2025 between Meta, AWS, Alphabet and Microsoft alone (Benedict Evans).
Among the neoclouds, there was significant growth this year, with leader Coreweave posting 134% YoY revenue growth and a 270% increase in their revenue backlog to over $55bn.
Although estimates range widely, the AI Data Centre Market is to grow somewhere in the region of 2–3x by 2030, hitting 150–200 GW of global capacity.
But what is the risk here?
Compute is a commodity product. Whilst IREN and the other neoclouds have contracts with the hyperscalers today, there is a risk that a supply glut will render their technology obsolete, and demand in the future won’t match current supply.
Michael Frazis puts it best:
“As with equipment leasers to massive mining companies – who can also afford their own equipment but sometimes choose to rent – they can scale up fast and generate outsize returns in shortages. But even a minor oversupply (which the massive sums of capital in neoclouds now all but guarantee) can lead to disastrous pricing dynamics.”
This brings us to the notorious GPU depreciation schedules.
IREN CEO Daniel Roberts has publicly stated that the business is underwriting the equipment to have 5–6 useful years.
This trend is broadly consistent with the accounting policies of AWS, Google, Azure and Coreweave, but more aggressive than Nebius and Lambda. Is a 6-year depreciation schedule reasonable? Notably, this 6-year depreciation schedule adopted by the hyperscalers (and IREN) is a relatively new phenomenon, only adopted in 2023 by Google and Microsoft, and in 2024 by Amazon, where previously they had 4–5-year depreciation schedules in 2021–24, and before that, 3-year depreciation schedules.
Frontier chips are only necessary for training, not inference, so the chips can be useful for inference, and a longer tail of tasks related to accelerated computing, or speeding up tasks traditionally performed on CPUs. MBI deep dives expressed a similar view that a 6–7-year depreciation schedule may be reasonable, assuming that the workload the chips are being used for transitions over the life of the chip, raising the example that Azure only recently retired their original NC, NCv2 and ND-Series VMs (powered by Nvidia K80, P100 and P40 GPUs) for August/September 2023, launched between 2014 and 2016, and implying a useful service life of 7–9 years.
That being said, it still feels aggressive. Just because the hyperscalers are adopting this depreciation schedule doesn’t mean it’s reasonable, and it’s very likely that this was employed as a tactic to inflate near-term earnings by dropping R&D expense, rather than being reflective of a long-term increase in the useful life of the hardware.
Notably, Nebius adopts a 4-year depreciation schedule, far more conservative than other players in the market. In addition, NVIDIA has just recently shifted to a 1-year chip iteration cycle, increasing the cadence of their chip rollouts, which will likely reduce the time to obsolescence. Jensen himself has hinted at this, noting at NVIDIA’s AI conference in March that:
“When Blackwell starts shipping in volume, you couldn’t give Hoppers away,” Huang said in March at Nvidia’s AI conference. “There are circumstances where Hopper is fine,” he continued. “Not many.” (CNBC)
For IREN, signing long-dated upfront contracts similar to the Microsoft deal mitigates the risks associated with ageing hardware, although it will be top of mind as IREN builds out their AI Cloud across a broader range of clients.
From a performance standpoint, there are also questions about how smaller neoclouds like IREN will fare relative to the larger players in Coreweave, Lambda, Crusoe or Nebius.
In Dylan Patel’s recent ClusterMax GPU cloud rankings, IREN was ranked in the underperforming category, showing some fragility in the IREN business model (ClusterMax serves as a measure of data centre quality, including measuring 10 factors including security, lifecycle, orchestration, storage, networking, reliability monitoring, pricing, partnerships and availability).
Assessing GPU Market Demand
So how will demand for chips fare for neocloud players like IREN?
The interesting thing about the current state of the AI market is that demand weakness is difficult to observe.
In particular, John Coogan from TBPN raises a great point that it’s actually about the “customer’s customer”. In other words, NVIDIA has a resilient business selling GPUs, with strong demand from the hyperscalers and leading AI labs. The hyperscalers still remain equity financed with free cashflow on the AI side, so there is not an inherent blowup risk.
But what about their customers? i.e., who is actually using these tokens, and are the direct users of tokens seeing a strong return on investment?
On the demand side, token growth has been promising, with OpenRouter showing token usage up from 330bn/week to 5 trillion/week over the last year, a good sign that demand will hold up. But is it enough? What about in dollar terms?
It’s difficult to assess the revenue generation from the hyperscalers themselves, but one useful datapoint is the total sum of the revenue generated by application layer AI startups. Taking this basket of 10 of the most valuable AI startups (I have tried my best), with the most recent revenue figures available can serve as a proxy.
ChatGPT: ~$13bn
Anthropic: ~$9bn
Cursor: ~$1bn
Midjourney: ~$500m
Harvey: ~$100m
Cognition: ~$70m
Sierra: ~$100m
Perplexity: ~$150m
Glean: ~$100m
OpenEvidence: ~$50m
Total Revenue: ~$24bn
If you assume that these companies account for ~70% of total token usage today, that puts the total AI market revenue at roughly ~$34bn in revenue. Certainly a far cry from the revenue figures required to make the maths work with hundreds of billions of dollars pumping into data centres annually.
In other words, whilst there is currently strong demand for tokens, the suppliers of tokens (model providers, applications) do not yet have a monetisation model that can provide this value in a way which makes economic sense.
Looking Ahead - What’s Next For IREN
So what does this mean for IREN?
Well, the data centre market is certainly looking frothy. But will IREN and the other players generate strong returns? Only time will tell.
In the meantime, IREN needs to focus on continuing to sign hyperscalers to lease their data centres on long-term contracts.
If you’re analysing companies building AI infrastructure — from neoclouds to hyperscaler suppliers — and want to separate durable businesses from hype:
👉Book a call with SBO to break down infrastructure economics, capital allocation, and the long-term winners in the AI compute race.
