Constantin

Builder & Investor

[Antonio Linares @alc2022]

Good morning

Pressure is always highest right before you win

Don’t tap out

[Joshua Kushner @JoshuaKushner]

C.S. Lewis: The Inner Ring

[Constantin @ConsJMJ]

Pure software businesses with no significant lock in / data will slowly diminish over the coming years.

Players like eg $HIMS that combine a physical product / hardware with a AI-powered software layer will shine.

[Constantin @ConsJMJ]

In Cape Town right now building something new in the health tech space.
Such a great place to focus and get stuff done.

[Antonio Linares @alc2022]

I just wouldn’t go short Luis

[Joshua Kushner @JoshuaKushner]

in moments of euphoria, people often misjudge the ordinary as extraordinary

[Constantin @ConsJMJ]

@mjkayum Nano

[Jim Liu @jiahanjimliu]

$IREN $MSFT Deal Breakdown

Although the price action is not the 50% candle that $NBIS $MSFT saw on Sept 9, topline is in the same range and bottomline will be better for IREN because IREN doesn't need to pay colocation to DataOne.

Topline Comparison
1. NBIS MSFT = 17.4B/5yr/300MW = $11.6m/MW-yr
2. IREN MSFT = 9.7B/5yr/200MW = $9.7m/MW-yr
3. CIFR AWS = 5.5B/15yr/400MW = 2.75m/MW-yr
Once you break down the per yer and per MW, you will see that IaaS contracts are much better than Colocation.

I do love CIFR's deal as it raises the market price of colocation from 1.8m/MW-yr (WULF/CIFR previous deals) to 2.75m/MW-yr. This will increase cost for CRWV/NBIS and at the same time raises the value of the 2GWs at SW 1 & 2.

I suspect IREN got negotiated down by Microsoft due to establishing HS credibility. We saw both CIFR/WULF take a below market deals and catch up on subsequent deals. In exchange, IREN got the 2B prepayment which is quite valuable as NBIS did a combination of senior convertible notes and ATM offering for 4.2B total funding for it's 300MW deal with MSFT.

This difference between NBIS/IREN's top line (1.9m/MW-yr) is fully made up by colocation as NBIS is likely playing between 2-3m/MW-yr for behind the meter power colocation with DataOne.

Costs
The all in cost $14-16m/MW seem much higher than $7-8m/MW projected but as usually IREN is transparent about ALL costs. For an apple to apple comparison, the datacenter infrastructure itself is $9-11m which is still higher than the $7-8m/MW but not double.

The 3m for super cluster architecture cost means that IREN is developing this cluster for training. It may be reused for inference during downtime or later in the GPU lifetime but primary usage will be for training. As Infiniband and cabling is part of the 5.8B GPU capex, I don't know at this time what exactly this is for. Industry standard estimates for 11-15m/MW is likely for inference DCs so industry estimates for training DCs should be $14-18m/MW.

The 2m charge for acceleration means that there are bottlenecks in the supply chain and you have to bid to get components in time. I am tangentially aware of this dynamic during the covid supply chain crunch when many electronic component suppliers had customers bid to get priority. This highlights the multiple bottlenecks in GPU DCs, not just power and makes existing DC build outs more valuable. I had called out that DC inflation was gooing to be much higher than CPI and this is basically DC "temporary" inflation charges. This dynamic is likely industry wide and industry estimates for training DCs available in 2026 should be $16-20m/MW.

I was definitely not ready for higher capex but IREN is fully transparent on where the extra cost are coming from and they are industry wide not IREN specific. DC inflation is real.

Bottom Line
We estiamte net profit from (Topline ARR - GPU+Networking/5yrs - DC/15yrs). All-in DC is 200MW*15m/MW = $3B.

Yearly profit = 9.7B/5 - 5.8B/5 - 3B/15 = 580m Net Profit.

Comparison to $NBIS = 17.4/5 - 5.8B GPU+Networking * (300MW/200MW)/5 - 2.5m/MW-yr Colocation * 300MW = -748m Net Profit. Unless NBIS is getting significantly better price for GPU+Networking and below market rate on DataOne colocation, NBIS is losing money on their deal with MSFT. Maybe a combination of both will help NBIS into profitability but IREN clearly has higher bottom line than NBIS.

$CRWV is probably getting by on profit from ripping off CORZ for colocation but is making less than IREN.

[Nicholas Fabiano, MD @NTFabiano]

You're not depressed, you just lost your quest.

[Constantin @ConsJMJ]

@andreibreha CRO

[Constantin @ConsJMJ]

Network effects are the moat behind the most successful businesses of the digital economy.

[Daniel Roberts @danroberts0101]

The Fourth Industrial Revolution is defined by GPUs as the new workers. i.e. millions of computers inside AI factories, turning raw data into intelligence.

Everyone talks about Time to Power.
An even bigger bottleneck is Time to Data Center… or more simply, Time to Compute. That’s the race.

[Constantin @ConsJMJ]

Invest in companies that profitably provide a marginally better user experience and watch them grow into industry giants.

[Constantin @ConsJMJ]

@teamburkisoy 52

[Antonio Linares @alc2022]

In theory, this entire market could be collapsed into a single vertically integrated peptide platform powered by an AI doctor.

A.k.a $HIMS in a year or two.

[andrewdudum @AndrewDudum]

One of the most under appreciated aspects of personalized medicine is often people end up needing meaningfully less medicine to achieve the targeted health benefit.

[Joshua Kushner @JoshuaKushner]

the goal of life is to be excited to go to work and excited to go home

[Constantin @ConsJMJ]

Doing sport daily is the key to happiness

[Constantin @ConsJMJ]

The best investments happen when you get in before two things: earnings growth and multiple expansion. Catching both is where real outsized returns are made