Every decision a company makes is a financial decision. So finance isn't just the function that reports on the business; it's where the business gets decided.

Which makes it worth asking where your finance team's hours actually go. Reconciling last month. Chasing receipts. Policing a $30 subscription. All of it backward-looking.

That's why I use Brex: corporate cards plus AI-native workflows that absorb the record-keeping. Zero-touch receipt matching, real-time policy enforcement, continuous reconciliation. Controls fire before money leaves the account. Close takes days, not weeks.

Defense handled. Go play offense. Thousands of companies — including Anthropic, Coinbase, and DoorDash — already run on Brex. Head to brex.com/metrics to join them.

I’m making a rare trip across the pond. Next month I’m hosting a happy hour for our London readers. If you want to have a cheeky pint with ya boy, RSVP below.

It’s on Thursday October 15th.

The dusty ass Lenovo ThinkPad I write this on (X1 Carbon btw, not to brag) is older than Nscale, the two year old UK neocloud that has somehow signed more than $100 billion in take or pay contracts with the likes of Anthropic and MSFT.

They're also gross margin negative, a recovering crypto miner, and participate in the circle j of Nvidia vendor financing.

Welcome to September of 2026!

Nscale filed to list on the New York Stock Exchange under the ticker NSCL (which I can’t help reading as Npopsicle). The price range is still blank, but the Financial Times reports the company is targeting a valuation of about $35 billion. In March, its last private round valued it at roughly $14.6 billion, according to The Information.

If you like Florida swampland, I've got a negative gross margin neocloud to sell you!

To be fair, the contracts are real. Microsoft has signed up for as much as $43.8 billion through 2033, and Anthropic for as much as $44.6 billion (OMG), which pushes total contracted value to $103.4 billion (not a typo). Only $2.6 billion of that is live today, and most of the rest depends on data centers, power plants and GPUs that haven't been built or fully financed yet.

This is a balance sheet business.

The whole deal comes down to three questions:

But first… WTF is a neocloud?

A neocloud is a company whose in the business of renting out GPUs for AI. It buys chips from Nvidia, installs them in data centers with enough power and cooling to keep them from melting, and charges customers by the hour to use them. Nscale prices its product "on a per-GPU-hour basis," the same way CoreWeave, Nebius, Crusoe and Lambda do. Those are the four companies Nscale names as its primary competitors (along with SpaceX, which I'll get to).

The easiest way to place a neocloud within the extremely confusing AI ecosystem is by what it sits between.

Microsoft is one of Nscale's two biggest customers, even though Azure also shows up in the filing as a competitor.

Hyperscalers rent from neoclouds because the bottleneck in AI has moved from chips to electricity.

If you already have a site with a few hundred megawatts hooked up, you can bring a GPU cluster online faster than a hyperscaler waiting in line for a grid connection. In the short term, a neocloud is selling speed to power as much as it's selling compute.

It helps to think of the AI stack as five layers:

Nscale's pitch is that it owns layers one through four end to end, what it calls "power to token." The Anyscale acquisition, which closes with the IPO, is its bet on pushing further into layer four with software.

However, the filing is more modest about “ownership” once you get to the risk factors:

So the chips come from Nvidia, the generators from Caterpillar, and a lot of the buildings from someone else who does building stuff.

Of Nscale's 17 data center sites, seven are wholly owned, nine are colocation (renting space in someone else's facility) and one is leased.

Measured by contracted capacity, about 1 GW of the 1.37 GW sits at owned sites, but the capacity that is live and generating revenue today is mostly in colocation and leased space. For now, the vertical integration is more assembled than owned.

Nscale sells two products:

Nscale Infrastructure is the landlord business. A giant customer like Microsoft or Anthropic reserves an entire dedicated cluster, often a whole building's worth of GPUs, for about six years at a fixed price. Nscale handles the land, the power, the building and the hardware, and the tenant runs its own workloads on top.

Nscale Cloud is the hotel business. Smaller customers such as AI startups and enterprises check in for fine-tuning, inference, or a few hours of on-demand GPU time, and Nscale handles everything up to the model. Anyscale fits here. In theory this is where the higher margins are, because you're selling software and convenience on top of the hardware.

Almost all of the money today is in the landlord biz. Infrastructure makes up the "vast majority" of the $103.4 billion in contracted value, on-demand usage is "not currently a material component" of revenue, and the filing never breaks out how much revenue Nscale Cloud brings in.

They are very much selling you the dream on where the cloud biz goes.

This is an admission that their margins suck right now and they are very much overly concentrated amongst their customers. But it can change if things go to plan and their second act goes well.

Take-or-pay contracts come from the natural gas and pipeline world, and the concept carries over really well for companies like Coreweave and Nscale who are also selling access to refined power and capacity. The customer reserves a set amount of capacity at a fixed price per GPU-hour and gets billed monthly for that reservation, whether they use it or not. Nscale's version usually includes a big prepayment at signing, which averages 23% of contract value, and that cash helps fund the build.

The catch is when the meter starts. Under Nscale's contracts, the term begins only once a GPU cluster is delivered and accepted. Until then, a signed contract is a promise on both sides.

What do I mean by that?

That's why only $2.6 billion of the $103.4 billion is “live”.

OK, now that we’re on the same page we can go through the numbers.

What's a contract worth?

Ask ten people (or auditors) and you'll get ten answers.

It depends on how likely you think they are to fulfill it.

Nscale's headline number is $103.4 billion of total contract value, meaning revenue over the full term of every signed deal. That's a ceiling. It's what Nscale collects if every cluster gets built, delivered on time and kept running for six years.

If, if, if…

Two customers make up most of that ceiling. Microsoft's contracts provide for payments of "up to approximately $43.8 billion" through 2033, and Anthropic's for "up to approximately $44.6 billion." Together that's about 85% of the total.

Up to, up to, up to…

It's also brand new. Total contract value was $38.0 billion in December, so more than half of the backlog was signed in the eight months before the filing.

Each Microsoft GPU term runs five to six years from when Microsoft accepts the cluster. The filing only describes termination for cause, like an uncured breach or insolvency. There's nothing about walking away for convenience.

The Anthropic deal was signed August 25, less than a month before the filing, and all of it runs out of one place:

We're big in West Virginia! Cue the John Denver!

Anthropic can terminate a tranche without liability if it's delivered late, and supply chain delays only count as an excuse in narrow cases. It can also walk if uptime falls below set thresholds. And Nscale still has to go find the money to build it:

Not to be a pessimist, but the single largest contract in the backlog runs through a campus that isn't built, on power that isn't powering yet, funded by money that hasn't been raised.

Every GPU Nscale deploys is made by Nvidia and bought through third-party suppliers. Then there's this:

OK, you have contracts to buy it if you have it, but you don't have contracts to secure that you will have it!

That's what makes the Anthropic delivery clauses bite. The customer can leave if the chips show up late, and nothing on the supply side guarantees they'll show up on time.

(Sees Nvidia is an investor… Conspiracy Carl pulls tinfoil hat out of desk drawer)

Vendor financing is when a supplier puts money into a customer so the customer can keep buying its shit. Nvidia shows up on every side of Nscale. It makes the GPUs, it has invested about $2.2 billion (including $1.0 billion that closes in November), it guaranteed up to $860 million of rent on a Texas data center, and it rents $1.2 billion of GPU capacity back from Nscale as a customer.

Nvidia vouched for the rent on a building that will be filled with its own chips (like your dad co-signing for your college apartment), and Nscale paid for the favor in warrants. Those warrants have since gone up in value, which is a big part of the $457 million non-cash loss this year.

Still, Nvidia's $2.2 billion is small next to the $24.0 billion Nscale has committed to spend on equipment. Customer prepayments and debt are doing most of the funding.

And in Game of Thrones often your sister is your lover who's also your cousin.

When a customer signs, Nscale usually collects about 23% of the contract up front. That's why deferred revenue went from $2.0 billion to $6.5 billion in six months, and why operating cash flow was a positive $1.7 billion while the company lost $1 billion.

GAAP treats a prepayment that big as a loan from the customer, so Nscale books interest on it. Of the $95.1 million of interest expense in the first half, only about $12 million was on actual debt.

Real debt was just $137 million at June 30, but it's arriving quickly. Since then, Nscale has drawn $830 million on a revolver, signed $4.17 billion of new loans, started $2.54 billion of equipment leases with Dell, and raised $3.1 billion of convertible notes. The early loans from Sandton cost 13% to 15%, while the newest are investment-grade at SOFR plus 2.375%.

That's good!

There's still $2.6 billion of lease liabilities spread across 17 sites. Thoughts and prayers to whomever has to do the lease accounting for this thing.

Right now, demand for AI compute is bigger than the supply of powered data centers. A company that shows up with land and power during a shortage can sign a lot of contracts fast, which is how the backlog nearly tripled in eight months. The real question is what's left when the shortage eases, and the filing addresses it:

Their plan for that world is to be the cheapest producer.

That's a clear mission statement. Walmart!

Most of that edge is power. Nscale says its sites often buy electricity about 70% cheaper than major US markets, and new grid connections and generation take years to line up.

Being cheap doesn't lock anyone in, though. The risk factors admit customers "may encounter relatively low switching costs at the end of their contractual commitments." Nscale's answer is software, through Nscale Cloud and the $1.65 billion Anyscale acquisition, but almost none of the backlog is tied to that yet.

At least the contracts match the hardware. The average contract runs 5.7 years, and Nscale depreciates its GPUs over five to six years, with no change to that schedule disclosed. The customer pays for roughly the whole useful life of the chip, and anything after that depends on signing someone again.

1. There's going concern language in the notes.
Management initially flagged "substantial doubt" about the company surviving, because the plan depends on financing it hasn't secured yet. Its fallback is to cut capex, which would mean late deliveries, which would let Anthropic cancel tranches.

2. The accounting controls have material weaknesses.
Nscale doesn't have enough accounting staff or documented controls yet, and the fix runs through 2027. The auditor, KPMG, only started this year and had to disclose an independence issue of its own.

3. Two customers are about 85% of contracted value.
Microsoft is up to $43.8 billion, Anthropic is up to $44.6 billion, and the Anthropic buildout has no committed financing.

4. A lot of the growth was paid for in stock.
The West Virginia campus, the Norway joint venture buyout, Anyscale and the convertible notes add up to roughly $7.3 billion of equity issued or promised this year.

5. Some neighbors don't want a data center.
The filing admits Nscale has "faced local community opposition" at some sites, and New York just paused big data center approvals.

From the founder's letter:

Good news though! They get the greenlight on the following:

Nscale depreciates its GPU equipment straight-line over five to six years and hasn't touched that schedule, which is more than you can say for some of the other companies in this space that have played it fast and loose with how long a chip lasts.

The ownership percentages are still blank, so the filing tells us who the big holders are but not how much each one owns.

Three names clear the 5% bar:

The old parent is in here too. Shareholders of Arkon Energy, the Australian bitcoin data center company Nscale was spun out of, swapped their Arkon stock for Nscale stock in May.

Payne holds options on 7.4 million shares at an exercise price of $0.000167 a share, which makes them essentially free. Half vest over time, and the other half vest as the company's valuation passes $5 billion, $10 billion and $20 billion. In August the board let him exercise early, so he took 2.5 million shares in September.

He also got an IPO award of 19.1 million shares, worth about $350 million at grant and equal to roughly 2.5% of the company. It vests from 2028 through 2032, and only if he hits targets:

Almost half of the CEO's IPO award depends on how many megawatts he gets built (which, like, makes a lot of sense bc that’s the only way the contracts get fulfilled and the company makes $).

The outside directors include Sheryl Sandberg, Nick Clegg, Susan Decker and Fidji Simo, who joined in September after leaving her operating role at OpenAI. At the Series C price of $20.39, Sandberg's restricted shares are worth about $39 million, Decker's about $23 million and Clegg's about $19 million, each vesting over four years.

Nscale names CoreWeave as its first competitor, and CoreWeave is the obvious comp. It's a neocloud that also started out in crypto, rents Nvidia GPUs by the hour on long take-or-pay contracts, and leaned hard on Microsoft, which was 62% of its revenue the year before it went public. It also went out at almost exactly the valuation Nscale is reportedly targeting.

Nscale is asking for the same price with about a twelfth of the revenue CoreWeave had, and a backlog about seven times bigger. The backlog numbers aren't measured exactly the same way. CoreWeave's is remaining performance obligations, and Nscale's is total contract value, which includes deals signed but not yet started. On the same basis as CoreWeave, Nscale's June 30 number was $56.4 billion. CoreWeave trades at 4.6x NTM revenue today and Nebius at 9.3x, per this week's Mostly Multiples.

So what does $35 billion buy? Today the business is running at about $400 million a year, based on the second quarter. Revenue over the next 12 months should be at least $1.2 billion, because that much prepaid cash is already scheduled to turn into revenue. A year from now, if the contracts signed before December go live on schedule, the run rate could be somewhere around $3 billion to $4 billion. That estimate comes from the December backlog, which had about $5.5 billion set to be recognized across 2026 and 2027, most of it in 2027. It doesn't count a dollar from Anthropic.

At $35 billion, Nscale would trade at about:

CoreWeave marketed its IPO at $35 billion and priced at about $23 billion, roughly 12 times its trailing revenue. It then did $5.1 billion of revenue in 2025, so IPO buyers paid about 4.5 times what its first year as a public company actually delivered.

1. There's a table comparing Wales and Delaware corporate law.
If you can't sleep tonight they literally have a table comparing the differences between Wales vs Delaware corporate law. It's an absolute banger. Among the highlights: English shareholders get preemptive rights on new stock, amending the articles takes a 75% vote, and a takeover needs 90% acceptance before the buyer can squeeze out the rest. IYKYK.

2. SpaceX is listed as a competitor.
Nscale names "new large-scale compute providers such as SpaceX" alongside CoreWeave, Nebius, Crusoe and Lambda. If you read our SpaceX S-1 breakdown, that's the xAI side of the house: the COLOSSUS supercomputer in Memphis and the plan to put data centers in orbit.

3. The fish farm.
Nscale's data center in Glomfjord, Norway pipes its waste heat into the water supply of a nearby fish farm.

4. Two banks share the top spot.
Goldman Sachs and J.P. Morgan sit at the top of the cover with an asterisk noting they're "listed in alphabetical order," which means they share lead-left. Morgan Stanley is third, out of 23 banks total. Goldman ran the Series B placement and co-ran the Series C with J.P. Morgan, and the CFO spent more than two decades at J.P. Morgan before joining last December.

5. UK retail investors get a slice.
The shares list only on the NYSE, but part of the offering goes to UK retail investors through depositary interests that settle in the UK system. Nscale also got the UK tax authority to confirm the setup won't trigger stamp duty.

6. It's filed as prepackaged software.
The SIC code on the cover is 7372, "prepackaged software," for a company whose biggest assets are land, power generation and GPUs. LOL.

7. They hate pictures.
This filing has very few images. No splash pages.

8. It still counts as an emerging growth company.
Nscale has $103.4 billion of contracted value, but its annual revenue is under the $1.235 billion threshold, so it qualifies for the lighter disclosure rules. One of the perks is an exemption from having the auditor attest to its internal controls, which is worth knowing given item 2 in Red Flags.

Sometimes I feel like I just don't know anything anymore.

I've spent the last few hours learning what a neocloud is instead of watching the end of Troy. I had it teed up on the TV until I saw this filing drop. ("It's too early in the day to be killing princes!" What a line.) I digress.

Also, sometimes it's better to be lucky than good. Stumbling from crypto mining into a generational supply and demand imbalance for AI is better than hitting the Megabucks.

Is this opportunism or a durable company?

My spidey senses do go up anytime I see references to "low cost provider" and "customer concentration," but hey, what do I know. As long as they don't build it in my backyard, I could use a cheaper LLM.

None of this is investment advice. I wrote this instead of watching Troy, and neither I nor my dog Walter (editor in chief) has ever been to a "neocloud." Do your own homework. For information and entertainment purposes only.