Most neoclouds are token farms on rented land: that's what they own.
How they sell is closer to airlines. $IREN is Ryanair. $CRWV is Emirates on borrowed money.
So $IREN must prove its planes leave on time. Next test: Microsoft's H2 to H4, due early 2027. CoreWeave needs high season to last.
I own $IREN, no $CRWV. Not advice.
Most neoclouds are token farms on rented land: that's what they own.
How they sell is closer to airlines. $IREN is Ryanair. $CRWV is Emirates on borrowed money.
Where it breaks: passengers can't wait, a lot of compute can. Training and batch jobs run whenever a GPU is free: OpenAI, Anthropic and Google sell batch at half price. That's a battery. It fills empty seats and flattens the peaks.
$CRWV = Emirates. Best service: Platinum, SemiAnalysis' top grade.
But it leases its hubs and flies on debt:
$ 47bn at ~8%, and $ 4bn of equity under $ 91bn of assets. On Nov 2, its 2031 bonds yielded ~16%.
$IREN = Ryanair. It owns its land, grid connections and data centers, in cheap remote spots: its secondary airports. No debt on the buildings.
Weak spot: service. SemiAnalysis rated it "Underperforming" in 2025 and reported multi-day outages at two old mining sites in October.
A GPU-hour is a plane seat: unsold, it's gone, and almost all the cost is fixed. So neoclouds sell most seats early under contract, the rest last-minute: pricey when full, cheap when empty.
It's high season: H100 rental is up ~50% since last December (Silicon Data).
Agents are here to stay. At work and at home, there's no going back on usage.
But look at GPU rental prices as an indicator. The same pattern shows up twice. H100 rents have fallen to $2.00/hr on Dec 9, 2025, then rose 20% by the end of February, in the wake of Opus 4.5 and 4.6. This summer they dipped again. Since September they're up 14%, and B200 rents are up 15% in the last four weeks alone. That's when Grok Bot and Muse scaled.
The catch: by their makers' own words, they're expensive to run. Musk says the cost of serving Grok Bot was "unsustainably high relative to revenue." Zuckerberg says Muse is "very compute-intensive to run".
So for agents to scale economically, compute has to get much cheaper. How do we get there? Efficiency alone? Or a crack first: an oversupply that makes GPUs and memory cheap? I don't know.
What do you think comes first?
@OInvests Best-case scenario for $IREN is Microsoft pulling out of the contract, letting IREN immediately sell the 200MW IT for $5bn instead of Microsoft's $2bn.
@RamaswmySridhar Better analogy than airlines: farming. GPUs are the trees, tokens the fruit, the model the seed. No compute, no harvest; no model, no harvest either. Whoever owns the seed takes the 85% margin, so open weights decide the winners.
@RamaswmySridhar Anthropic S-1 breaks the 6/16 math. Cost of revenue, substantially all serving compute, was $2.4B on $16.3B of H1 revenue. That's ~$15 of compute per $100 of tokens, an 85% gross margin, not $100 for $150.
@DanielTNiles@elonmusk Agree on hyperscalers, power and construction. It's one trade: own the land, the power and the walls. A neocloud that owns its site sits closer to your list than to the 200+. The rest are token farms on rented land.
@davelee Read this list as a credit memo. Google, Amazon, Microsoft and Broadcom get ~$431B whether Anthropic uses the compute or not. SpaceXAI gets 90 days' notice from early 2027. Lambda's biggest piece dies if it can't secure the building. Contracts and walls get financed.
@amitisinvesting Item 1 is the real story. Even SpaceX, investment grade with Starlink revenue, financed $40B of GPUs by letting Apollo own them and lease them to SpaceX. That's the new rule of AI credit: the lender holds the title. Most neoclouds can't offer that, they don't own the building.
@chamath Energized land is also the best collateral in this market.
A lender to a neocloud renting someone else's building gets chips it has to pull out past a landlord. A lender to the owner of the land, the building and the GPUs gets a whole factory. 2027 will price that gap.
Starlink and launch bring in ~$28B a year. Enough for a $40B lease, not for the ~$500B that 10 GW of new compute costs in 2027. And compute rented near spot gives lenders no long contract to lean on. In the end, SpaceX borrows against SpaceX.
(I love @SpaceX and I’m a happy investor btw)
Counter-intuitive: the neocloud with the biggest funding gap is $SPCX
Musk's plan: from ~3 GW at the end of this year to ~13 GW at the end of 2027. At ~$50B per GW, that's ~$500B to fund in 14 months.
What SpaceX has: $90B of cash and securities, the $40B Apollo lease, and ~$38B a year of operating cash flow at Q3's pace. About a third of the bill.
Musk says lenders want revenue outside AI. SpaceX has it: Starlink and launch, ~$28B a year, now less than half its revenue. That backs a $40B lease. It doesn't back $500B.
And it can't borrow against customers. Its compute rents near spot, and renters can leave SpaceX on 90 days' notice from early 2027.
Neoclouds with 5-year contracts borrow against Microsoft, Google or Meta. SpaceX borrows against SpaceX.
Most neoclouds are token farms on rented land.
Rented land is a problem for lenders. Being a token farm is a problem for equity investors. My thinking on Musk's "guppy-level death" post:
Rented land.
I learned this on our palm farm in Morocco. The trees, the irrigation and the power lines are most of the investment. The land is on a 40-year lease from the state. The bank won't lend us a dime without our personal guarantees, because the company has nothing to mortgage without land ownership.
Neoclouds have the same problem with GPUs. If one stops paying, the lender has to pull its chips out of a building someone else owns, and the landlord, with rent to collect and lenders of its own, controls the door. So the best GPU loans come with the customer contract attached: $CRWV's is rated A3, while its unsecured bonds yield over 15%.
Own the land, the building and the GPUs, and the lender can seize the whole factory, running, and re-lease it. That's IREN at Childress.
Token farms.
We call them AI factories. Most are farms: GPUs are the trees, tokens are the fruit. The model is what makes a token intelligent and gives it a price. A neocloud is a factory only if it owns the intelligence: its own model, or open source, which anyone can run as its own. Otherwise it grows tokens for someone else's model.
No open weights: the value goes to the closed labs, and compute (chips and data centers included) becomes a commodity crop.
Open weights: every token farm can become an AI factory, the value moves to compute, and the labs are the ones under pressure.
$IREN's full-ownership model will prove more resilient in a challenging funding environment.
And the battle over open weights will decide who the AI winners are.
Pretty much.
As of the past few weeks, debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers.
SpaceX closed $40B of debt for GPUs, but we’re investment grade and have Starlink & launch revenue.
That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many.