Over the weekend, Anthropic has leaked news that it was looking for 1.4GW of compute in Australia.

in my opinion is likely to land that deal, if it hasn’t already found a buyer for its compute. Regardless, Anthropic’s entry into Australia will raise prices for compute.

In this post, i will walk you through how much is likely to earn.

From the announcement, we know that is deploying 40,000 GB300 GPUs from NVIDIA.

From data points, we can see that GB200s (an older version) had deals cut at $11.6 (SPCX-GOOG) to $6 (deals done here) - GB300s should trade at a premium to this pricing. For the sake of being conservative, we will stick to the GB200 pricing.

For 40,000 GPUs, we are looking at roughly an ARR of 2-4B depending on the price.

currently trades at 2B market cap.

In my neocloud walk through, i have already gone through the economics of how this translates down into the bottomline and to achieve a payback period of 1-2 years.

Given that is under Nvidia’s partnership program, this reduces the upfront capex for the buildout tremendously (GPU costs are ~44% of the total buildout in this case)

Again, the valuation of a neocloud is the NPV of all the projects that it has done. In this 1 singular project, it is likely to earn the PV of its entire market cap currently.

If we were to apply industry ARR multiples to $SHAZ, this would command a range between 6-10B in market cap.

I of course, argue that valuation by ARR multiples is a terrible way to look at neoclouds.

Tldr

has 40,000 NVIDIA GB300s that is likely to be able to earn it 2-4B a year in gross revenues while its market cap sits at 2.5B.

“Meta signaling that they are open to selling excess computing infrastructure”

“Softbank plans to offer AI Compute in US at 10GWs”

Everywhere you go turn, it looks like everyone wants to become a neocloud and start selling compute. Why is this the case?

Let’s start off by examining a real life case scenario as i walk through the economics of a neocloud with you.

On the 5th June 26, Google announced a deal to pay Space X 920million a month for compute capacity at XAI data centers.

This was for 110,000 NVDA GPUs as well as CPUs and other memory components i.e a fully furnished datacenter (Colossus 2)

The GPUs in question are the NVDA GB200s NVL72.

For that amount, we can see that each GPU fetched 920M/110,000/720 = 11.6 per hour

Let us construct a fictitious 100MW datacenter filling it with GB200s

My estimates land us at about 50B per GW. In this case, our 100MW buildout will cost us about 5B.

Which references well with industry estimates. Note, this is for GB200 buildout. As you get more advanced, the chips are more expensive. For example, the GB300 buildout, the compute cost is 20% higher.

Jenson saying that 100B per GW is actually realistic when you factor in that compute costs for Rubin will be twice as much as the Blackwell costing. The cost of power etc, are also rising.

We also know that each GB200 draws about 1200Ws, hence 100MW converts to about 83,333 GPUs.

On the XAI deal economics, this would land us at 83333*11.6*365*24 = 8467M/year …

On the following cost structures

0.74M/MW per year for electricity
0.25M/MW per year for infrastructure maintenance
0.3M/MW per year for staffing

We have this insane looking scenario where your payback period is less than 1 year.

This is where i would come in and say that the Space X deal was remarkably generous. Was Google facing a massive compute shortage?

If we were to run this on less insane and more normal looking assumptions like a price of ~6/hr

Our payback period stretches to about 2 years.

I am estimating that on long term deals that the neoclouds sign with the hyperscalers, it is done at about $4/ equivalent. The hyperscalers in turn are able to monetize the compute out at much higher rates, like $12+ (AWS).

Just for fun, if we do a pro-forma of a singular build out - it would look like this

Now, of course - we can dispute that GB200s will have no value after 5 or 6 years. But A100s came out in 2020 - they are still being rented out at $1 - $2/hr. This was ~ $3 in 2021. The hyperscalers are still renting them out at anywhere from $2 - 3.5/hr

Gavin Baker made the case for the GPU depreciation timeline to run for about 10 years. I would agree with him especially as agentic AI and inference usage pick ups.

In essence - current compute markets are pricing a severe shortage.

Anyone that tells you otherwise is covering their eyes.

Space X is renting out compute, because at those rates that they rent out, they are making a ton of cash (also because they face issues with connecting Colossus 1 and 2). Musk also made the deal with the right to cut the lease short saying that “If compute gets super tight I said we might need it back at some point”

Tldr: Everyone wants to be a neocloud now because compute is super tight and it mints cash.

From the facts of the matter, being able to construct a datacenter and to sell compute is the best business right now.

This is why the hyperscalers are willing to prepay neoclouds for compute availability.

Again, a neocloud’s valuation should be the summation of the NPV of all the compute deals that it can do. Hence the ability to obtain financing, the exact depreciation schedule and the ability to execute are the 3 most important levers impacting it.

For those, who then cry that it is all circular financing etc (insert whatever phrase), know that Anthropic now is making more than 70% gross margins on serving inference. The frontier labs that are the biggest users of compute can switch over to become cash flow positive if they want to.

As part of my journaling, I would like to look back on my calls on June to assess what i did well and where i can do better.

Reviewing through - Here is the list of calls i made

On the 3rd of June - I made the call to “gross down exposure and shore up cash….probably the best times to be bidding protection”.

On 4th of June - during the mkt open - i called for “the selling pressure into the open quickly reverses and we would end up green on the day”

The market ended up V - reversing with NBIS continuing to be strong.

On the 5th of Jun 26 - I said that “today the picture right now looks alot more murkier….would recommend grossing down exposures”

On the 9th of Jun 26. I called for the “start of the new wave of the rally in semis”

And i continued to bang the drum for the rally on the 11 of Jun and there after.

Until the 15th of Jun 26 - where on my telegram channel i posted that these are good areas to trim before going into FOMC (sorry got too busy, didnt post on X - this is why you should be in Tele. I get too busy sometimes)

On the 22th of June 26 - I then made the case to “reduce down leverage going into end of June ….flows wise this current period will be met with headwinds”

On the 26th of June - I told everyone that “I see the current pain and sell off in the sector as technically driven….that these are the levels you should be bidding…areas of interest for me NBIS 240s, ALAB 360s, CRDO 240s, SMH 605s”

I also posted some of my public purchases….because fortune favors the bold.

Lastly in my weekly review , i explained that and puked because of “mechanical flows” and that “they can offer some good opportunities”. Since my post, BE is now up 25% and CRDO up 12%.

I also have been calling for Compute and Memory pooling to be leading the next rally instead of everyone’s favorite memory. Thus far this has been playing out.

All my posts on X are time stamped. Likewise on my telegram channel. Everything is free. I do not intend to charge for my posts nor knowledge. If you enjoyed it, i am happy.

I want to say that in terms of reading flows for June, i have done pretty well for myself, calling every single move.

Where can i do better? I think i can become even more patient. My execution last week could have been better for sure. I also could have managed my cash position a lot better.

Alright - that’s all for the June review.

Before we touched on Asia markets today. A quick recap yesterday on the quarterly rebalance. KOSPI was said to have a huge rebalancing yesterday, however price action remained firm towards the close - this was a good indication to me that a great deal of rebalancing has already happened and this would likely be the case for the US markets.

Indeed, US equities were extremely firm - while memory was stagnant in performance, we saw NBIS leading in the neocloud sector before the entire neo sector rallied into the close.

This has shown up in today’s Asia markets. KOSPI was relatively weak - down 2% while NKY was barely up. Again, KOSPI has a huge memory weightage, hence it is more reflective of the memory beta than any other risk factor.

One post that has caught some attention is by Andrew Curran that “there has been a significant break through in architecture - specifically around memory efficiency”

This has been picked up by Citrini that DRAM prices “are not going to keep going up the way”

From a trading perspective, i can see the flows softening into memory. On 25th of June, I said that “I think the forward looking traders are seeing this and will rotate out of MU which is why i believe that memory will NOT be the next leader for the rally ahead. Humans are smart and will try to find workarounds via expensive solutions.”

Now - from a fundamental perspective, this probably does NOT dent FY28 earnings. We are still in a structural shortage that will last through to 2030 especially if robotics comes into play. However, the market as you know is deeply irrational and trades more on vibes in the short term.

Again - if you dont understand nuance. Let me be simpler, memory will not be leading the next rally.

On neoclouds and compute. Dylan of Semianalysis has revealed that “Anthropic’s margin on an Opus 4.8 API token is north of 80%”

Hey - this is actually crazy. Please. Re read it again.

The frontier labs are no longer ponzis. Anyone that says that is not seeing reality.

The frontiers labs have basically shaped themselves into a position where they are earning 80% margins. “Anthropic excluding SBC was net income profitable in Q2 2026. Potentially profitable including SBC by Q3.”

This came alot earlier than many people are expecting. I thought it would come far later.

The market needs more compute.

At this point Situational Awareness portfolio is mostly in neoclouds. Why? People need to really understand that a neocloud bet is NOT a bet on current revenues.

It is more like a bank - the valuation of a neocloud is the NPV of ALL the projects it will take on in the future.

Currently, any compute will be taken up. Given that Anthropic is earning 80% margin on tokens, I can see the compute market to stay heated for MUCH LONGER.

For readers who come from crypto - this is slightly similar to a DAT. Except instead of taking your cash to buy useless tokens, this is taking your cash to construct +EV, +NPV projects.

One lever that nobody talks about is that people are arguing that 5 years depreciation is too lenient. Like Burry.

But A100s are still being rented out, these came out 6 years ago. If a GPU life is much longer than 5 years, this is an additional lottery windfall for neoclouds. This is NOT being priced in at the moment.

Of course, no 2 neos are alike. Execution is key.

It is like running a bank. You can have 2 banks with the same deposit base. But they can be valued differently based off how well they are run. See ICBC vs JPM.

I wrote about this in my note on yesterday.

You can read more about it there.

Trading wise, this is a short week. Seasonally, this week is characterized by an upwards drift.

Things i am looking for - NBIS breaking 300.

Good luck!

I am writing this as a follow up to document what i have learnt while digging into the company.

First off, Situational Awareness filed it yesterday

What strikes me immediately is that Situational Awareness stake ballooned to 19.9%. This is significant for many reasons. 20% is the threshold where the stake crosses a major accounting and regulatory level.

Under both IFRS and GAAP, owning 20% or more presumes “significant influence” and the fund has to use the “equity method” of accounting. (goosebumps, thank you accounting class)

At 20%, the stake can hardly be called passive anymore. It becomes one of the largest singular shareholders and have extreme leverage in getting a seat on the board.

Situational Awareness bought all the way till the limit.

Next, in order to acquire such a stake, this means for the entire equity deal, Situational Awareness took 35%.

Linking into the first quarter financial statements, we can see that Sharon AI only had ~164M of cash on the balance sheet.

Post this deal, will have an additional 1.6B of cash on its balance sheet to fund the build out.

Also, back in January 26, USDai ($CHIP) approved a 500m debt facility for

This answers one of the most important question for - Can this company build what it is promising to build. With access to about 2.1B of cash and debt - the build out is now financially realistic.

Next, the revenue stream for cannot be discounted anymore. It has a 2.2B backlog in a take or pay contract.

Also, is making a strategic collaborator in the 72MW deal in Australia. This secures the GPUs.

In any neocloud buildout, the three questions i will ask is

1 - Can the company fund this in an equity accretive manner?

2 - Can the company deliver the compute? (Sourcing GPUs etc is a big part of the problem)

3 - Can the company continue to sign good contracts that are suitable for its capital structure

Thus far, i see that has significantly derisked the first 2 concerns i have for a starting neocloud.

This brings the total capacity for SHAZ to 132MW. Which indicates about ~1.3B in ARR for

A neocloud is more like a bank for compute. It raises capital and deploys it into compute. The valuation of a neocloud is therefore the NPV of all the projects that it will ever deploy into.

Currently, any neocloud operator will tell you that compute is sold out. Every neocloud project is +EV.

Don’t believe me? Go rent out some GPUs yourself. Tell me what is the payback period.

Non believers of neos often dont understand the real business of a neocloud and hence anchor valuation to current deployment. That misleads people.

If it was any other anchor investor besides Situational Awareness, i would not put too much weight on the ability of SHAZ to execute.

However, Leopold lending his name and weighting this deal heavily biases me towards that the execution for SHAZ will probably be good.

He in all likelihood knows a great deal more than me.

To recap, the latest funding deal has answered 2 crucial questions for SHAZ, whether it has the ability to raise capital for the build out and to procure the GPUs required.

Leopold lending his name to SHAZ also heavily reduces any reputational taint that SHAZ might have.

Hope you enjoyed this!

Subscribers to my telegram channel would note that i commented on the MOC orders at 3:58, 2minutes before the closing. - I comment on telegram much more in real time.

Why such a reaction?

Let’s go over what happened on (chart below)

Into the closing auction, SPY suddenly nose dived all the way down to 717s (these prints are real). Orders were executed at those prices. Congrats to those who were filled below, those with stops that were triggered? Not so great.

SPY as a tradable asset sets its clearing price via the closing auction. How this works is that, there a bunch of participants that know they want to trade “at the close” and get the price at the close - think people who are tracking some benchmarks etc.

Hence they submit these MOC (Market On Close) and LOC (Limit On Close) orders. These get collected before 3:58PM E.T. Just before 4PM, the exchange has a view of all those orders and then searches for maximizing the number of shares that can trade.

Its a straight forward process - however when there are large imbalances, the closing price can possibly diverge from “fair value”

What is the fair value of SPY? It is simply the sum of all the value of basket of stocks that it holds in the ETF - each undergoing their own closing auction mechanics.

This mirrors the closing for SPX Index - Because this is an index, it is not tradeable. It is calculated instead by taking the official closing price of each of the constituent stock multiplied by its weight.

Because, the closing price of each stock is done via their respective auctions, SPX prices are only published after the close.

Given that SPY is one of the largest products, such a “flash crash” is actually quite unusual.

The effects ripple out much larger in single stocks.

was down ~16.5% on Friday. It was trading sluggishly, dribbling down before puking almost 10% in the last hour (the last 5minute candle was a 5% wide candle puke)

This effect can be due to the fact that BE was removed from the Russell 2000 to the large cap Russell 1000 index. The R2000 index has more AUM than the R1000 AUM. eg, which tracks the R2000 has about 84B while the which tracks the R1000 only has about 48B tracking it…

This mechanical rebalancing sell pressure gave it a good push into the close. At the closing price of $253 - I think this gives a good value zone.

also toppled by ~11% on Friday. No results for guessing. Yes - was also rebalanced out from the Russell 2000 into the Russell 1000 index.

In my daily notes, you would realize i often reference my trading around flows. This is because while fundamentals drive the stock in the long run, the short term performance of a stock is mainly driven by flows.

Some of these flows are mechanical and not very intelligent.

They can offer some good opportunities. Of course, having a good knowledge of these will also allow you to understand why certain things are moving “weirdly”.

Extending beyond the week, we have the quarterly rebalancing coming up on Tuesday. JPM estimates that ~165billion of equities will be sold into the close on Tuesday.

The net effect of the rebalancing will be between how much gets sold into the close and the amount covered by hedge funds who anticipate and trade against such flows.

Given the fragility of markets at this point, i would say that expect volatility to be high.

Good luck and enjoy the rest of your weekend!