In partnership with
The company building the most AI infrastructure this decade just decided it would rather not own most of it.
It sold most of a fourteen-billion-dollar campus to an asset manager and leased back the compute it runs on its own land. The spending left its balance sheet; the risk did not leave the building.
And they say AI will kill creativity.
I'm Ben Baldieri. Every week, I break down what's moving in GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
Today's issue is brought to you by Rafay
The GPU stays independent because sponsors cover the bills, not because they shape the copy. Rafay works the orchestration layer this whole issue keeps returning to, so if you build, buy or fund the physical stack of AI, they are worth a look.
AI Infrastructure Leadership Summit
Join fellow tech executives for a two-day summit hosted by Rafay bringing together leaders from the world's most innovative neoclouds, telcos, NVIDIA Cloud Partners, sovereign AI initiatives, and enterprise AI platforms to share real-world strategies for commercializing AI infrastructure, expanding margins, and building the next generation of AI businesses.
Meta Sells 80% of Its $14B El Paso Data Centre to BlackRock
Meta just handed most of a $14 billion data centre to Wall Street.
Meta and BlackRock formed a joint venture to build and run a 1GW El Paso campus at roughly $14 billion. BlackRock's GIP and HPS funds take 80%; Meta keeps 20% and contributes the land and $2.3 billion of construction. It then leases the campus back and guarantees the residual value, up to around $13 billion. That guarantee is why 20% understates it: Meta carries most of the economic risk, off its own balance sheet. Compute lands in 2028.
Why this matters:
The asset-light hyperscaler, out in the open. Meta keeps operational control and first call on the compute while an asset manager owns the shell. The other three copy this the moment their capex lines hurt.
We watched infrastructure capital buy the finished hall in Issue #116, MGX and BlackRock's GIP taking Aligned for $40 billion. This is the earlier entry point: funding the campus to own it from day one.
Utilisation is the risk that stays with Meta. The structure only pays off if the campus keeps earning after the lease, and it is Meta behind the residual guarantee if it does not.
Nvidia Puts a Reported $5B Into Sutskever's Safe Superintelligence
The one frontier lab that stayed off Nvidia's chips just let Nvidia in.
Nvidia and Safe Superintelligence, Ilya Sutskever's lab, announced a partnership giving SSI access to Nvidia's Vera Rubin platform and, in Nvidia's words, an order of magnitude more compute. The release names an investment but no figure; Bloomberg put it near $5 billion, so treat it as reported. SSI had leaned on Google TPUs, and this pulls its next training generation onto Nvidia silicon. The lab has raised about $6 billion at a reported $32 billion valuation.
Why this matters:
Nvidia is funding the customer that spends the money back on its chips. The investment and the compute commitment are the same dollars in a circle; it only counts as new demand for the share Sutskever would not have bought anyway.
We flagged the identical loop with AMD in Issue #116, a reported $5 billion Anthropic stake alongside a 2GW order. Equity in, GPUs out is now the standard way a chipmaker locks a frontier lab.
The prize is allocation, not revenue alone. A lab this closely held signals who gets Vera Rubin first, and every rival reads that as where 2027 supply is already spoken for.
Nscale Moves to Buy Anyscale for $1.65B
Nscale spent $1.65 billion to stop being just a place you rent GPUs.
.png)
Nscale, the UK-based GPU neocloud, agreed to acquire Anyscale, the company behind the open-source Ray framework, at a reported $1.65 billion. It is an agreement, not a close: completion is expected in the second half of 2026. Ray was donated to the PyTorch Foundation in 2025, so Nscale is buying the company and its commercial layer, not the framework. It marches Nscale up the stack, from bare metal to scheduling what runs on it.
Why this matters:
Neoclouds are done fighting on price per GPU-hour. Owning the scheduling layer raises switching costs: a customer built on your orchestration will not re-plumb to save a few cents an hour.
We tracked the neocloud consolidation wave in Issue #114, where Nscale already featured. Buying software instead of concrete is the land grab turning into a stack grab.
Neutrality cuts both ways. Anyscale won by running anywhere, and the moment a Nscale-native path looks favoured, customers on rival clouds go hunting for an owner-agnostic scheduler.
Moonshot Opens Kimi K3's Weights and Lands on Azure and Dell
China's loudest open model just landed inside Microsoft and Dell.
Moonshot AI released the weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window, downloadable and self-hostable. Open-weight, not open source: any product past 100 million users or $20 million in monthly revenue must show "Kimi K3" in its interface, and larger hosted-model businesses need a separate deal. It is already served through Fireworks on Microsoft's Foundry and listed on Dell's Enterprise Hub. K3 rivals the top open systems rather than topping any US frontier benchmark.
Why this matters:
The licence is the strategy. Give the weights away, then meter where the money is: consumer scale and hosted inference. Open enough to spread, closed enough to bill.
We covered Microsoft testing Kimi K3 to shave its Copilot bill in Issue #115. Weights in the open, plus Foundry and Dell, turns that experiment into something any enterprise can run without a China contract.
Watch enterprise adoption, not the leaderboard. A frontier-class open model inside Azure procurement and on Dell hardware is a pricing problem for every closed lab.
SK Group and Nvidia Pledge a Company-Stated $500B and 2GW of AI Factories
SK Group and Nvidia's $500 billion partnership binds neither of them to anything.
SK Group and Nvidia announced a partnership they put at more than $500 billion across AI factories and memory, signed as letters of intent: non-binding, spread across years. The concrete piece is a 2GW AI factory from SK Telecom on Nvidia's Vera Rubin accelerators, powered by SK hynix HBM4, targeted for 2027. The real story is memory: HBM4 is the bottleneck, and this ties a large slice to one national build-out.
Why this matters:
The scarce input is shifting from GPUs to the memory beside them. Whoever secures HBM4 secures the accelerator, and this claims two gigawatts of it for Korea before the racks exist.
We traced SK hynix going to the capital markets in Issue #114. This is the demand side: the memory maker and the chip designer building the factories that will swallow what SK hynix ships.
Queue position is what a deal like this buys. A national programme at the front of the HBM4 line pushes every other buyer back for 2027, whatever the headline number means.
Amazon's Capex Guidance Climbs to $220B as Cash Flow Turns Negative
Amazon beat its quarter and turned cash-flow negative doing it.
Amazon reported second-quarter revenue of $200.6 billion, a company first, with AWS up 37% to $42.2 billion. What moved sentiment was spending: Andy Jassy lifted full-year 2026 capex guidance to about $220 billion, up from $200 billion, blaming memory costs. The trailing accounts already sting: property and equipment hit $173 billion over twelve months, up 64%, and free cash flow swung to an outflow of $7.6 billion from an $18.2 billion inflow a year earlier. The first negative print of this capex cycle.
Why this matters:
The spending line now runs the story at every hyperscaler. A revenue beat no longer carries the quarter; the market wants capex turned into contracted demand, and a negative cash-flow print is where that patience gets tested.
We logged the same capex vertigo in Issue #115, Amazon and Microsoft both climbing. Guidance at this level says the cycle is still accelerating, and one company resets what counts as normal.
Read the memory line as the tell. When the biggest buyer blames its own forecast on the price of memory, it confirms the squeeze the chipmakers are building around.
Local Opposition Has Killed 50 Data Centre Projects This Year
Fifty data centre projects have died this year, and none of them ran short of money or chips.
Heatmap Pro counts more than 530 counties and municipalities moving to restrict or block new data centres, nearly 190 since June. More than 50 projects were cancelled in 2026, double the 2025 total, and the kill rate on challenged projects has climbed from about 40% to roughly half. New York went furthest with a one-year permit moratorium, and campuses tied to Amazon, Microsoft and Google have all fallen. Demand is not in doubt: BNEF models US load at up to 207GW by 2033. Whether it gets permitted is the new variable.
Why this matters:
Money and chips were never the binding constraint the models assumed. Local consent is, and it does not scale with more capital. The asset-light deals in this issue move who owns the risk, not whether the site clears its planning board.
We have chased the power scramble since Issue #110, always on the supply side. This is the other blocker: line up the generation and a county can still vote the campus down.
Track cancellations, not the announcements. If half of challenged projects keep dying, built gigawatts diverge from announced, and whoever locks permitted sites now holds the scarce asset.
Everything Else
p.s. The GPU Daily now goes out every morning. Opt in below if you want the physical layer of AI in your inbox daily.