$IREN In-depth review of Q3 Earnings Call, MSFT Deal, Profitability
First I want to acknowledge where my previous posts were wrong:
1. I stated 4.3m capex for IREN's liquid cooled DC which was way off (3). Written in Sept, I had used total load so if I used IT load it still would have been 6-7m. The estimation was based on this Data Center Dynamics Article which had stated Childress buildout was 300-350m (1) which had likely taken the figure IREN's prev for non T-3 datacenters. The actual is 14-16m. 9-11m of it is the data center itself. It's reasonable the extra 3-4m was for a T3 datacenter as the previous 6-7m figure did not state T3 datacenter. Furthermore, training clusters cost more than inference clusters which is a 3m charge. I now think both clusters are training and each training cluster is 100MW so there are two training clusters with each Horizon making up one training cluster.
2. I had wrongly mentioned that no ATM dilution is needed on my first take of the slides. I had over looked the part from the earnings where focus is on maintaining a "prudent balance of debt and equity" (4).
@FransBakker9812 had found in footnote to confirm that dilution via ATM not convertibles are coming (2). This is warranted because we don't want to be like $CRWV who is issuing overloaded debt which makes their dilution look good but their balance sheet is saddle which makes it harder for them to issue debt in the future. IREN will want a healthy balance sheet to be ready for SW1.
3. I always that IREN could do IaaS at a high level which the MSFT contract did validate that IREN can serve hyperscales but I did not anticipate that the rate would be much lower than Nebius. This is the main reason we are not seeing $IREN stock hit $90+.
4. Finally I did participate in some degen speculation on Wednesday night from a bridge 84 picture posting and took it as a hint for a possible Meta deal. That did not materialize.
Now that we have more data, I will finish this post with more accurate net earnings project and share price for Canada + Childress + SW1 than last projection.
Earnings Call Notes
IREN uses IRR to compare IaaS deals to colocation. They subtract off a $130k/kW-month colocation.
Colocation Net Profit
This is equivalent to 1.56m/MW-yr. IREN has stated that the DC itself is 14m-16m/MW and can be used for 20 years. If we take 15m/MW and divide by 20 we get that the cost for the colo is .75m/MW-yr.
First we calculate (Revenue - Depreciation) which is 1.56m/MW-yr - .75m/MW-yr so that's .81m/MW-yr. If we used 85% project margins, then we get a Net Profit of 688.5m/MW from colocation or 137.7m for 200MW. Note, we don't know what composition of margin is colo and which margin is IaaS we just know the entire project averages out to 85% margins.
Note that the 85 margins includes power, repairs, maintenance but not SGA, financing (5). SGA was high this earnings report due to key employees getting paid in RSUs which did a 10x. Their refresher RSU contracts will be based upon market price. I acknowledge that SGA may be higher in a hot industry but I assert it won't anywhere near 10x higher RSUs.
IaaS IRR
IRR (internal rate of return) and not ARR is the best metric for Neoclouds. IRR is used in real estate where the value of the property is very high compared to the yearly cashflows. This makes absolutely sense because value of GPU depreciation is very high and the elephant in the room. I do like that IREN is very transparent and uses IRR to contextualize their ARR. Every Neocloud should do this.
Thanks to @GlobalCollapse for original IRR calculations. I have calculated IREN's unleveled IRR to be 12.47% using Wolfram Alpha as shown in picture 1 (calculations available in source 6 for modification). Wolfram Alpha was ChatGPT for math before LLMs were cool and is classical algorithm based so it doesn't hallucinate and is has calculator precision. I check with @MarkosAAIG and this is inline with DC projects.
In real estate, leveraged is often used in the form of debt. Note that debt is secured by the asset which in real estate has stable value so this is not degen margin - every single DC project is levered whether it's GCP, Meta, Microsoft, NBIS, CRWV, or IREN. I have calculated IREN's leveraged rate of return to be 25.76% subtracting off the prepayment and then using a 7% interest rate shown in picture 2 (calculations available in source 7 for modification). I don't have the exact interest rate they are getting from Dell, but the earnings call explicitly stated with 2.5B of leverage, they are getting 25-30% IRR on 2.5B of leverage and if they use 3B of leverage then IRR is 35%-40%.
The above IRR assumes 0 residual value on GB300s after 5 years. With 20% residual value, Unleavered IRR becomes high teens and with 3B leverage IRR becomes 50%.
IaaS Net Profit
I will subtract off deprecation of GPUs and then multiply by 85% project EBDITA margins to get EBITA since we already subtract off depreciation of GPUs. Then I will subtract of interest to get EBTA. There's no significant amortization from a cost of a software team so EBTA = EBT. So I will get EBT which is just missing taxes a proxy for Net Profit. Note IREN has tax abatements on everything except the land which they bought for cheap.
We have 9.7B - 1.56B colocation = 8.14B IaaS revenue. 5.8B - 1.94B prepayment = 3.86B which over 5 years at 7% is 3.86 * 0.7 + 4/5 * 3.86 * 0.7 -> etc. Note that 4/5 is the remaining principal on year 4, etc. This gives us 810.6m of interest payment.
(8.14B IaaS Revenue - 5.8B GPU+Networking) * .85 = 1.989 - .8106 = 1.1784B over 5 years which comes out to 235m for 200MW IT per year.
Topline/GPU Comparison to NBIS, Nscale
Now that earnings call clarify 76k GPUs, we can compare topline per GB300 which is the most accurate. Let's do per 100k GPU for round numbers.
NBIS = 17.4B per 100k GB300
IREN = 9.7B * 100/76 = 12.76B per 100k GB300
Nscale = 14B * 100/104 = 13.46B per 100k GB300 (10)
Thus IREN is earning 73% of NBIS's topline/GPU which I attribute to uptime or credibility of uptime (9). Note that this is Nscale's second deal with MSFT where they were previously getting 6.2B for unknown amount of GPUs (10).
Notes from Earnings Calls
1. Dell has financed IREN for GPUs. Payment is due 30 days after shipping -> I'm assuming it will cover a portion with prepayment and cash from IREN covering the rest.
2. Other miners are giving up equity for colocation deals, while IREN gives up zero equity and gets huge a pre-payment worth 1/3 of the entire GPU+Networking+Cables cost.
3. Microsoft started shifting to negotiating for IaaS instead of Colocation 6 weeks prior to earnings.
4. Other HS more interested in colo SW but open for combination of both.
5. Other HS have interest in Horizon 5-10.
6. Contract obligations cannot be revealed.
7. Horizon 1-4 is 76k GPUs total.
8. Superclusters makes 2m extra charge for training only, not required for inference.
9. For Canada, 11k contracted from news released, 1k more contracted for 12k total contracted. Non-contracted ones are because of delivery timeline from Nvidia is far out.
10. Both AI native + enterprise customers in BC.
11. Hyperscalers and AI natives only want bare metal GPUs and want to bring their own orchestration layer.
12. Only smaller, less sophisticated customers want orchestration layer.
13. IREN is only focused on bare metal offerings because that’s the best way to scale their offerings to utilize their power assets.
Net Profits for Canada + Childress + SW1
Assumptions: Canada IaaS + H1-4 at 73% of NBIS Topline + H5-10 at % of NBIS Topline + 1.4GW SW Colocation
Canada IaaS Net Profits
Capex for Prince George is as follows for 500m ARR then we will subtract by capex and multiply by .85% margins. 7.1k NVIDIA B300s, 4.2k NVIDIA B200s and 1.1k AMD MI350Xs for approximately $674m (11). 1.2k B300 and 1.2k GB300 for $168m (12). 4.2k B200s for 193m (13). 1.2k B200, 1.2k for 130m (14). Total is 1.165B, over 5 years lifespan that's 233m depreciation. (500m - 233m) * .85 = 226.95m. Canada has two other sites projected for 1.5B net profit so we total net profits for all Canada 3 sites is 680.85m.
Childress
H1-4: 235m + 137.7m = 372.7m
H5-10
Note IREN got 73% of NBIS's Topline/GPU for H1-4. Let's assume that IREN gets 86% of NBIS's Topline/GPU for H5-10. This is very important because the 13% is pure profit if we assume same rate of cost of GPUs and DC. Then H5-8 would be 200MW for 11.42B instead of 9.7B for H1-4. H1-4 has 372.7m profit while for H5-8 an additional 11.42B - 9.7B = 1.72B would go bottom line so net profit for 5-8 would be 2.0927B. Now that's for H5-8, we can multiply by 1.5 to include H9-10 so H5-10 = 3.139B.
Childress: 3.5117B
Sweetwater
For colocation of 1.56m/MW-yr, we get a Net Profit of 688.5m/MW-yr. However, CIFR got 1.83m/MW-yr from MSFT while IREN is conservatively giving internal colocation numbers. If we take a 1.83m/MW-yr and apply the delta as profit then we get an additional 270m/MW-yr of profits for a total of 958.5m/MW-yr. For 1.4GW that's 1.3419B for Colocation. Now I know why Dan says colocation deals are bad.
Canada + Childress + SW1 Share Price
Adding all 3 sites, total net profit is 5.53B. Take whatever P/E you want but I'll use conservative 20 P/E. Note the 5.53B doesn't include the 2-3GW confirmed pipeline. That's 110.6B market cap. We have 281m diluted shares today. Assuming more 40% more dilution for finishing SW1 and H5-10 that gets us to 393.4m diluted shares or a share price of $281 by 2027.
$IREN is pleased to announce the signing of a $9.7bn AI Cloud contract with @Microsoft
Key details of the transaction:
- $9.7bn AI Cloud contract value
- 5-year average term
- 20% prepayment
- 200MW (IT load) data centers
- NVIDIA GB300 GPU deployments
Refer to the press release and accompanying presentation below for further information
Press Release: https://t.co/3pG9N2HfkF
Presentation: https://t.co/uyysnL9ffS
$IREN is pleased to provide an update on its AI Cloud business.
- New multi-year AI Cloud contracts with leading AI companies
- Contracts secured for 11k GPUs, representing ~$225m AI Cloud ARR by end of 2025
- $500m AI Cloud ARR from 23k GPUs on track for Q1 2026
- NVIDIA Blackwell GPUs contracting ahead of delivery on an average term of 2 years
- Capacity for >100k GPUs across British Columbia campuses and Horizon 1 & 2
- Customer workstreams progressing through site tours, technical diligence and commercial negotiations
Press Release: https://t.co/nNABJL8qQm
I just put 70% of my portfolio into one stock.
It's a secret smallcap that will 10x because it powers Anduril's subsea drones.
Here's why I'm not out of my mind.
And why the ocean is the next $1T defense spending supercycle 🧵
Cipher Mining Signs 168 MW, 10-Year AI Hosting Agreement with Fluidstack
In this transaction, Cipher will deliver 168 MW of Critical IT Load to Fluidstack at its Barber Lake site, securing ~$3B in contracted revenue over the initial 10-year term. The agreement also includes two 5-year extension options which, if exercised, would bring the total contract revenue to ~$7B.
To support the buildout, @Google will backstop $1.4B of @fluidstack's lease obligations and will receive warrants to acquire shares of $CIFR common stock, equating to a ~5.4% pro forma equity ownership stake.
“We are thrilled to be working with Fluidstack to develop HPC data centers, and we look forward to welcoming Google as an investor in Cipher. This transformative transaction reinforces our HPC momentum as we continue to attract attention for our large and growing pipeline of sites,” said Tyler Page, Cipher’s CEO. “We believe this transaction represents the first of several in the HPC space as we continue to scale our capabilities and strengthen our position in this rapidly growing sector.”
Full Press Release here: https://t.co/hoRq53vWWy
$BDP's multi-agent arsenal kicks off with Agent 008 (Crypto Discretionary)
Soon to be storming ALL TradFi & DeFi frontiers: credit risk, macro, and market-neutral arbitrage.
Alpha insights: live on X. Buckle up.
Advertisement and malware blocking is coming soon to the Anyone iOS Browser app!
Currently being tested in the community, and due for imminent release. Get the most secure browser for iPhone now:
🔹 https://t.co/TQiquWSGPd
What if AI agents could work together, securely, privately, and autonomously?
At ICML 2025, we propose AgentaNet, a decentralized swarm network for AI agents to collaborate securely, efficiently, and with aligned incentives.
It tackles core challenges in interoperability, security and privacy, laying the groundwork for scalable, privacy-preserving agent collaboration.
The future of autonomous AI doesn’t just scale, it swarms.
We are now live on @Base!
Our mission aligns with Base: bringing the next billion users onchain by building an ecosystem of products that truly matter.
This expansion opens up new pathways for participation in the Anyone ecosystem, making privacy and DePIN more accessible than ever.
Transfer crosschain now via @StargateFinance, and access liquidity on @AerodromeFi. 🧵
✨ Platform UI Upgrade: Trading Page v2 is LIVE
We’ve just launched the first phase of our v2 UI upgrade—focusing first on a streamlined, user-friendly trading experience.
- Updated layouts for Bridge, Swap, and TWAP pages
- Improved trading charts & notifications
Combined with our recently revamped sidebar, https://t.co/mNQvMMkKFc has never looked better. 💎
More UI overhauls arriving this week, making it easier and more intuitive than ever to navigate and trade the powerful Bittensor $TAO token ecosystem.
Trading subnet tokens should be easy—let’s make it happen. 🦾
The Anyone multichain expansion continues.
We’re thrilled to bring Anyone to @Base, expanding liquidity and unlocking key functionality for our growing DePIN and privacy ecosystem.
Look out for new partnerships with projects building on Base, and onchain integrations for our protocol!
As part of this rollout, we’re working with @StargateFinance to enable seamless bridging, and preparing for liquidity on @AerodromeFi.
❕ Stay tuned for update on when we are officially live
We're building the first full-stack decentralized AI infrastructure on the fastest EVM chain today - @base.
As Base brings millions of people onchain, FLock brings millions of people to decentralized AI training.
Train, earn, own. Not your models, not your AI.
We’re exploring what it takes to train healthcare AI models across continents — without moving a single datapoint.
Our latest research shows how blockchain + federated learning (FL) can solve healthcare’s biggest challenge: data sharing.
(we won a prestigious award for it) 🧵
1/ Thread on what I think it’s one of the most under-the-radar projects in the market: $FLOCK 🧵
>> @flock_io became the first & only decentralized AI integrated by @Alibaba_Qwen, Alibaba’s LLM (the GPT to OpenAI, the Claude to Anthropic).
This is not an empty partnership, like Flock simply using Alibaba Cloud or integrating Qwen models.
Instead, Qwen is building models and training on Flock’s infrastructure for decentralized privacy-first AI.
Some alpha:
>> Qwen’s 2.5 on-edge version is being trained with FLock, and will sell directly to industry use cases, like hospitals, financial institutions, and other sectors where data privacy is of outmost importance.
In their own words: "This marks the first deployment of a decentralized AI agent model within a major centralized AI ecosystem".
Synaptogenix has begun to generate revenue through $TAO staking from qualified custody with BitGo, enabling the company to generate yield while maintaining the highest standards of security and regulatory compliance. Read more: https://t.co/GHKp5uhnkV $SNPX #TAO#Crypto
The strength of the Anyone ecosystem comes from the builders around it - from dedicated organizations deploying relays to new apps built with the Anyone SDK.
The spirit of activism runs through our entire ecosystem, and there's never been a more important time for it.