코인톡

My view on AI Infrastructure and $IREN

Before I start, I am not bearish or bullish on IREN. This is simply how I look at companies structurally. I have been a Iren shareholder in the past so I believe I have some understanding of the business. Two quick points:

1) Yes I did sell early
2) No I’m not bitter.

Genuinely happy for everyone who has made money in it.
Anyways, I look at investments the same way I look at systems, what goes in, what comes out, how fast it decays, and where it breaks. When I run that lens on data center names like IREN, the thesis gets difficult from a business perspective pretty fast.

The issue is not whether AI demand exists. It does. The problem is whether the asset itself can actually throw off durable, risk adjusted returns after you factor in physics, accounting, and who the customers are.

Depreciation is everything in this space. These are not traditional real estate plays. A warehouse holds value for decades because concrete lasts. A GPU based data center ages like a race car. The useful life of an H100 or B200 cluster is maybe 3 to 4 years before performance per watt makes it economically obsolete. IREN and others run 4 to 5 year straight line schedules (not going to debate anything longer), but that still understates how brutal the replacement cycle is. If a competitor stands up a site with next gen chips 18 months later, your $40,000 GPU might still sit on the books at $30,000, but the revenue it can command dropped 60% because nobody wants last gen FLOPS for training. So you end up on a capex treadmill. Every dollar of EBITDA gets plowed back into new chips just to keep revenue flat. That looks like growth on the top line, but it is really asset churn with a giant power bill. Free cash flow never shows up because the depreciation is real cash walking out the door each refresh.

The supply and demand setup is also lopsided. The bull case is that AI needs infinite compute, which is true. The part I worry about is that AI does not need *your* compute specifically. These economics only work above 85% utilization with long term, take or pay contracts from solid customers. IREN’s history has been Bitcoin mining shifting into AI and HPC, and that brings two headaches. First, customer concentration. You are not selling to thousands of software companies. You are pitching to five hyperscalers and a few model labs. They have all the leverage, they multi source, and they will build in house if your pricing creeps up. If Microsoft or Meta decides to own instead of lease, your demand can vanish overnight. Second, the revenue is lumpy. Without 5 to 10 year power purchase agreements, you are exposed to spot pricing. Training workloads are bursty. Inference can run on older gear. If model scaling takes a breather, utilization can swing from 90% to 50% and your EBITDA margin flips negative because power, staff, and debt service do not scale down. You build a $500M site assuming seven years of rent, but the customer’s roadmap changes every 18 months. That duration mismatch is brutal.

The unit economics get worse with scale unfortunately. People call data centers picks and shovels. I see negative convexity. Power is the biggest variable cost and you do not control it. A 100MW site at $0.04 per kWh is $35M per year just for electricity. If grid rates move or carbon rules change, margin evaporates. Bitcoin miners lived this in 2022. Cooling is another trap. Air cooling will not handle 1000W GPUs. Liquid cooling retrofits are capex you did not model. Every hardware generation shifts the thermal profile, so your so called future proof building is not. Geography can strand you too. Build in West Texas or British Columbia for cheap power, but if the customer needs low latency near Virginia, your asset is obsolete.

The terminal value assumption is also off. Real estate underwriting assumes the land and shell hold value without the tenant. With AI data centers, a shell without current gen GPUs, specialized power delivery, and liquid cooling is just an empty warehouse with a huge utility bill. The second your hardware is obsolete and you lack a contracted customer to fund the refresh, the asset value falls to industrial land minus demo costs. You bought a depreciating supercomputer with a lease attached.

So from a systems view, IREN and similar operators are capital intensive, customer concentrated, and exposed to tech obsolescence on a 3 year clock.

One more thing that hurts the equity case is dilution. These businesses raise constantly to fund the next build or the next GPU refresh. Share count goes up, and existing holders get watered down even if revenue grows. Until they can self fund from cash flow, which the capex treadmill makes hard, dilution is the default setting. Dilution can be normal in a high growth business but there needs to be a roadmap to normalization and FCF. I don’t necessarily see it here.

That said, there is potential upside if they pull it off. If IREN locks in long term, take or pay deals with a top tier hyperscaler at above market rates, and pairs that with low cost, fixed price power, the model flips. You get potentially contracted cash flow, utilization protection, and a path to real free cash flow after the first refresh cycle. Add in geographic scarcity near key fiber routes, and the asset starts looking like infrastructure instead of a depreciating chip home. It is a narrow path, but it exists i believe. The equity becomes a leveraged bet on execution and contracting which is possible this year with larger deals.

I wish the best to all investors and stay open to hearing about different viewpoints.