# AI capex — X 热门讨论 (2026-10-09 12:25 UTC)
## @FransBakker9812 (Frans Bakker) · 10-09 11:26 · ♥72 ↻2 💬6 Part 1: Construction speed as a moat.
We have heard the phrase "time to compute" far too many times. But what does it mean for $IREN's greenfield sites?
The below 2 SW1 images are taken exactly 30 days apart.
On a complete shell basis, the results of 30 days progress: 1. 25% of the roof completed. 2. 25% additional steel has been erected, bringing the current steel frame of the building to nearly 50% of the total. 3. 45MW IT load in terms of dry-coolers have been installed.
And this is just the bulk of the visible work, there's so much more going on which I will cover in a separate post.
If IREN can keep this pace up, before the end of Q1 2027, the complete 200MW IT load shell will be up with a roof, and all the external equipment for cooling will be positioned for commissioning.
Based on extensive AI analysis of my available aerial footage, I have concluded there are currently no more than 350 workers on site.
By contrast, at present, IREN has approximately 3000 workers active in Childress.
Can you imagine what will happen in Sweetwater if IREN ramps up the workforce there during the first half of 2027?
Subcontractors are already preparing for this. I know that, because I talk to them.
The market is placing a considerable amount of execution risk on IREN's valuation. And we know the company has guided incredibly conservative for liquid-cooled capacity to come online from Q4 of CY2027.
When I look at the situation at hand, I see a different story.
IREN is going fast at Sweetwater 1, and by my calculations, the first part of this 200MW IT load building could be ready to receive GPUs in Q2 of next year, but only if there is a phased delivery of 25-50MW increments.
From what I understand from IREN, is that there will in fact be 4 segments of 50MW IT load in this hall, and that means there is a strong possibility of a much earlier initial delivery than Q4, and thus a higher likelyhood of earlier contracting, and subsequently billing, of additional FY2028 revenue.
So if you wonder why IREN hasn't contracted any capacity at this site yet, it's not because they cannot find any customers, it's simply that by holding out a few more months, they will be able to:
1. Sell that capacity to a much larger group of customers, including non-IG, 2. at a higher price per MW, 3. without risking having to factor in inflationary components of the buildout, and associated hardware, because by then it's already done and procured.
So why is construction speed a moat? By being your own general contractor with a large team of subcontractors active on schedule today, you can factor in the exact time-to-data center, and negotiate GPU delivery timelines, lock in prepayments, and have an edge in locking in favorable GPU debt.
It sounds like a simple 1,2,3, but in this AI industry, things change almost every day.
Being predictable in your buildout, with a demonstrated pace of construction, having all your supply chains in sync, and all your contractors locked in: that's when you have the strongest negotiation position with your customers.
IREN in 2027 and beyond, is all about unit economics, EBITDA margins, and profitability as a result of true vertical integration. Being able to build data centers fast, on multiple sites at once, with a standardized design, and repeatable CAPEX cycle, the only thing that is variable is your ability to attract debt, with on the other hand the uptrend in demand, and GPU hour economics.
I will address debt in another part of this post, but clearly the demand is here to stay, and GPU hour rates for Blackwells are making all time highs.
IREN has 3 energized primary substations in Sweetwater 1, and 6 in Childress.
Currently, the early earthworks for 2 more are showing at Sweetwater 2.
All of this is Batch Zero base load, energized, or approved to energize.
Power not a moat? Okay, but power is also not a concern.
If you can build fast, you can capitalize on a relatively larger part of this supply-constrained environment, and monetize a larger share of the overwhelming demand, at the best economics per MW on average.
It's really that simple.
Even if you believe that the demand is temporary, having the ability to construct these data centers fast, without having to worry about your source of power, gives IREN an edge in locking in the most favorable revenue mix, across the most strategically selected customer mix.
This is why we track the construction at OnlyFrans. Not to look at pretty pictures, but to establish data points, and quantify the construction cadence.
If execution risk is the main concern for this stock, zooming in on just that, is exactly what gives investors the information they need to position themselves accordingly. https://x.com/FransBakker9812/status/2108519530687566120
## @KrisPatel99 (Kris Patel 🇺🇸) · 10-09 08:00 · ♥47 ↻1 💬11 $SOFI and long-duration Treasuries via $TLT began to rally as concerns around AI capex overcapacity started to work their way into the market yesterday. This was accelerated by an FT report that, in my view, misunderstood OpenAI’s revenue-recognition framework.
That reaction may offer a useful blueprint for how investors should think about positioning if they believe the AI capex cycle will eventually slow—or even contract—as the market works through the underlying supply-and-demand dynamics.
Right now, AI compute demand still materially exceeds available supply. But a meaningful portion of frontier-AI consumption is effectively being subsidized by venture-capital and strategic-investor funding.
That dynamic could change once OpenAI and Anthropic become public companies.
At that point, access to incremental capital may become more constrained, and both businesses will likely face greater pressure to optimize for sustainable economics rather than maximizing growth at almost any cost.
At the same time, increasingly capable open-source models and lower-cost, “good enough” alternatives should put pressure on frontier-model pricing power. If that happens, the economics of continuously scaling massive training runs become less attractive.
The result could be a shift in priorities:
Less emphasis on ever-larger training runs.
More emphasis on inference efficiency.
Lower cost per token.
Higher utilization.
Better gross margins.
And ultimately, a much stronger focus on generating acceptable returns on the enormous amount of capital already invested in AI infrastructure.
That is where $AVGO becomes particularly interesting.
The purpose of custom silicon is not simply to build faster AI chips. It is to fundamentally change the cost curve of inference. OpenAI’s work with Broadcom is part of that transition—from relying heavily on expensive general-purpose accelerators toward purpose-built infrastructure designed around the economics of serving models at enormous scale.
If the AI industry moves from a “build capacity at any cost” phase toward a “monetize the installed capacity” phase, the winners—and the market leadership—could look very different.
*Wrote this and had AI correct grammar and spelling* https://x.com/KrisPatel99/status/2108467565253013761
## @FransBakker9812 (Frans Bakker) · 10-09 11:26 · ♥53 ↻1 💬3 Part 2: Cost of Capital.
The current credit market environment is showing increasing signs of strain. Rising Treasury yields, persistent inflation concerns, and uncertainty over the Federal Reserve's next moves are pushing borrowing costs higher, while investors are becoming more selective about where they deploy capital.
At the same time, the extraordinary financing requirements of the AI infrastructure buildout are placing additional demands on credit markets. With hundreds of billions of dollars in GPU financing, data center asset-backed financing (ABF), and other AI-related debt being arranged or contemplated, the question is increasingly whether investor appetite can keep pace with the sheer volume of new debt issuance.
Capital remains available, but the combination of higher benchmark rates, growing financing needs, and tighter scrutiny of risk raises questions about how sustainable current funding conditions will be.
This is the increasingly cautious picture emerging from credit markets. However, a more challenging borrowing environment doesn't necessarily mean that all companies, assets, or financing structures will be affected equally.
What does this mean for $IREN?
What is the real concern about the current financing risk that's priced into the valuation of the stock?
Is the concern that IREN will have to service too high an amount of debt-related interest payments?
Is the market worried that IREN will not be able to attract additional debt to accommodate its FY2027 capex roadmap and beyond?
What about the cost of equity and concerns around dilution?
IREN has given a clear capex roadmap in its FY2026 earnings presentation. The numbers are there:
$8B in GPU debt and prepayments expected, and a $3–8B funding gap beyond that.
With a current market cap of $14B, it's hard to imagine that this capex roadmap is going to be an easy one.
Especially given that IREN's reported debt principal is now equivalent to almost 55% of its market cap, a ratio that has deteriorated significantly as the share price has declined.
But are the concerns valid? Or is this a snapshot in time, and will the situation look very different once IREN's $4B in CY2026 ARR becomes operational and starts generating meaningful cash flows?
According to the latest available data, $IREN pays just $59.4M annually in cash coupons on its convertible debt, while the Microsoft and Mackenzie GPU financing facilities would represent approximately $435M in annualized interest at full drawdown, before amortization.
Interest rate fluctuations should not impact this, as most of the debt carries fixed interest rates or is hedged through interest-rate swaps.
But as investors look forward, future borrowing and refinancing remain sensitive to the overall credit environment. But if the last 12 months have shown us anything:
1. 0.88% weighted cash coupon on the entire convertible debt position. 2. Fixed 9% interest on non-IG-backed GPU debt with a maturity date beyond the GPU payback period. 3. Blended ~6% interest on IG-backed GPU debt, hedged against rate fluctuations.
Sure, these achievements are not a guarantee of future convertible or GPU debt financing. But they do show that, within the last 12 months, the tools that the company utilized came with fantastic terms, ensuring that their maturities are in line with the associated contracts and buildout cycle.
Given that the current demand environment, where GPU availability is pushing MW economics up, is contributing to better projected EBITDA margins for vertically integrated AI clouds like IREN, I believe concerns about IREN's existing interest obligations should be considered a relatively low priority, while future borrowing costs deserve more attention as the company continues to scale.
With that said, the next concern around access to debt is perhaps a more important one.
Can IREN continue obtaining competitive corporate, GPU-backed, and data center ABF financing in a more selective credit market?
I attached a short clip from BlackRock's Rick Rieder. Yesterday, he was interviewed by Bloomberg TV, and the TL;DR is that he does not think there will be fewer GPUs bought because of a 25 bps rate increase.
Now, he does not explicitly talk about IREN here, but his overall point is interesting: a 25 bps increase in interest rates is unlikely to meaningfully change the economics and appetite driving AI infrastructure investments. Demand for GPUs and related debt isn't suddenly going to disappear because financing becomes marginally more expensive.
How can we narrow this down to IREN?
IREN, so far, has demonstrated continuous access to GPU debt despite a large amount of convertible debt being present on the balance sheet.
Furthermore, the company has demonstrated substantial customer prepayments, representing approximately 45–55% of associated GPU CapEx, providing an additional source of funding alongside its GPU-backed debt.
The early track record is solid, but for now, data center ABF has not yet been utilized, and since not many data centers have actually been completed and begun billing yet, there is room for execution risk here.
Overall, a more selective credit market could affect IREN's expansion pace and financing costs, but it has not yet invalidated management's funding model.
I believe the market sees the $8B of GPU debt and prepayments for FY2027 as a real risk, and since IREN was very conservative with this number, it would be smart to address this in its next earnings call to reflect a more accurate estimate.
I see the remaining $3–8B funding gap as a smaller concern, particularly because IREN has yet to utilize its unencumbered data center infrastructure as collateral for additional debt financing. With all four Microsoft Horizons targeted for delivery by the end of this year, I expect the company's ability to attract data center ABF to improve significantly as these assets become operational and start generating contracted revenues.
Given that Kent Draper has stated that IREN expects the vast majority of its funding gap to be covered by non-equity financing, I see little reason to believe management is anticipating substantial difficulties in attracting the necessary debt to finance the FY2027 buildout.
Just as with the physical infrastructure buildout, IREN will need to show the market that it is able to execute.
For now, this risk weighs on the stock, but it also shows what the delivery of the Microsoft contract could do for both the buildout execution risk and the financing execution risk.
Dilution risks are present. The company has issued $2.5B out of the $6B ATM facility so far, and with $3.5B remaining and a $14B market cap, there is a valid concern around dilution.
But going back to Kent's words, and considering that IREN has demonstrated the ability to raise substantial equity capital at attractive valuations, this concern should be assessed within the broader context of its capital requirements and management's expectation of a declining equity component.
If demand holds up and GPU-ready data center capacity remains scarce, $IREN could continue benefiting from increasingly attractive economics per MW.
It is precisely for that reason that unit economics will matter going forward, as project economics will determine the ability to service debt principal, beyond the interest payments.
Having a 30-month maturity term on your GPU debt only works if you can actually service the principal, so excellent project economics are essential to demonstrate a healthy capex cycle and attract appetite to finance the next project.
Currently, IREN's market cap is under pressure, but a declining share price does not automatically translate into a higher WACC. What matters is whether the sell-off reflects a higher required return on equity, changes in debt financing costs, or a deterioration in the company's underlying cash-flow expectations. I don't believe it does.
IREN has come a long way from tapping the ATM at $6 last year. Its ability to secure billions in convertible debt, GPU financing, and customer prepayments speaks for itself, and I am confident that the bankers from Sydney will continue to deliver. https://x.com/FransBakker9812/status/2108519536244978089