In partnership with
Half a trillion dollars. Six of Wall Street's biggest. And the company selling the chips, quietly agreeing to cover the losses.
They are calling it independent capital, a brand-new asset class at arm's length from the vendor whose hardware it pays for.
Arm's length, give or take a quarter.
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 and GPU-management layer beneath all of this, so if you build, buy or fund the physical stack of AI, they are worth a look.
AI Infrastructure Leadership Summit
Hosted by Rafay, join fellow tech executives for this two-day summit 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. *Registrations subject to approval
NVIDIA Enlists Wall Street to Mobilise $500B for AI Infrastructure
The biggest chipmaker alive just made itself the bank.
NVIDIA has signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to build independent financing platforms designed to mobilise over $500 billion of third-party capital for AI infrastructure over time. Jensen Huang pitched it as a shift from buying chips project by project to financing AI factories as "productive infrastructure," with the money coming from the institutions rather than NVIDIA and each deal underwritten independently on its demand, utilisation and residual value. He is careful to say the $500 billion is capital the platforms are designed to mobilise, separate from NVIDIA's revenue or any single fund. His framing of the thesis is blunter: "in AI, compute is revenue."
Most of a LinkedIn essay went to getting ahead of the obvious charge, that NVIDIA is financing demand for its own chips. The defence is that the chips hold their value: CUDA keeps old silicon earning, and NVIDIA cites H100 rental rates rising from about $1.70 to $2.35 per GPU-hour between October and March. Lower down, though, NVIDIA may provide "residual-value support" for up to 25% of a given deal, so the independent capital arrives with the vendor backstopping the resale price of the very chips it sells.
Why this matters:
The 25% backstop is the tell. NVIDIA is putting its balance sheet behind a quarter of each deal to make a depreciating asset financeable, and it is exposed the day the residual-value case cracks. American Compute's read of 77,000 sales already puts H100 residuals in single digits by year five.
Every deal this fortnight turns on one question: who underwrites the buildout. Volta leaned on JPMorgan in Issue #118, Fermi still needs its guaranties, Riot has twenty years to fund. NVIDIA just centralised the answer.
It is memorandums rather than money: nothing committed, no facility closed. A $500 billion target announced while its own deals are still conditional is a headline first and an asset class second.
Anthropic Signs a $9.1B, 20-Year Compute Deal With Bitcoin Miner Riot Platforms
Two weeks, two ten-figure compute deals, and a second Bitcoin miner turned landlord for Anthropic.
Anthropic has signed a $9.1 billion, 20-year deal for AI data-centre capacity with Riot Platforms, the Bitcoin miner, whose shares jumped about 25% after hours as it turns hash power and Texas sites into AI hosting. Riot follows Bitdeer in repricing off Bitcoin onto contracted AI compute, taking its place behind SpaceX, Amazon and last week's $10 billion Volta deal on Anthropic's supplier bench. It has not said how it will fund the construction, only that Anthropic will pay to use it once built, which is exactly the gap NVIDIA's new platform exists to fill.
Why this matters:
Anthropic hedges suppliers the way a utility hedges fuel: the one thing a frontier lab cannot do is run short. Its offtake is now the demand signal that makes a supplier's buildout bankable.
Two weeks running, an Anthropic deal has turned a Bitcoin miner into an AI landlord, Bitdeer through Volta and now Riot direct. The lab brings the offtake, the miner brings the power, and someone else still has to fund the build.
It is a 20-year term for mostly-unbuilt capacity, worth $9.1 billion only if Riot can finance it. The same conditionality hangs over Fermi below.
CoreWeave Doubles Revenue to $2.6B as Interest Drives a $626M Loss
CoreWeave's scorecard leads with a $128 million non-GAAP operating profit; the filing shows a $49 million operating loss.
CoreWeave reported Q2 revenue of $2.575 billion, up 112% and more than double the $1.212 billion it booked a year ago. Adjusted EBITDA was $1.51 billion, a 59% margin, and the GAAP operating loss was just $49 million. Then the interest bill lands. Net interest expense hit $640 million for the quarter, and that is what turned a thin operating loss into a $626 million net loss, up from $290 million a year ago.
Why this matters:
CoreWeave now carries about $35 billion of debt, $31.4 billion recourse and $3.7 billion non-recourse, against $6.4 billion of capital spending in the quarter alone, per its Q2 8-K.
What the market bought was the backlog: revenue commitments reached about $104 billion, up 246%, with more than $25 billion of net new customer commitments added in early Q3.
It is the same wager NVIDIA is institutionalising in the top story, and CoreWeave is the live test of it at $35 billion in.
Meta Ships Muse Glimmer, a 30B Open Model Built to Run on One GPU
Meta's newest agent does not need a data centre; it runs on the laptop you already own.
Meta has open-sourced Muse Glimmer, a 30-billion-parameter agent model from its Superintelligence Labs, under a permissive Apache 2.0 licence. Quantised to roughly 4-bit it shrinks from 55GB to under 20GB, fitting a high-end consumer GPU or an M-series Mac, with a speculative-decoding drafter to stay responsive for real-time agent use. Meta built it by distilling its larger, closed Muse Spark model into a student that calls tools, recovers from failed calls, reads screenshots and slots into OpenClaw and other scaffolds. It benchmarks well for its size against Gemma4-31B and Qwen3.6-27B, which is the honest frame: a strong small model, and Meta claims nothing more.
Why this matters:
A capable 30B agent on a laptop pulls a slice of the fastest-growing inference workload off the metered cloud. That is the counterpoint to the top story: not every token has to come from a data centre someone borrowed half a trillion to build.
The compute did not disappear, it moved. Meta trained on the closed Muse Spark teacher and distilled a student cheap enough to run anywhere, the expensive GPUs staying in the lab while the cheap ones serve at the edge.
Meta gave away the student and kept the teacher, extending the down-market slide we tracked with DeepSeek and Qwen in Issue #118. The open gesture is real; the frontier stays behind glass.
An AI Memory Shortage Just Made CXMT China's Most Valuable Company
The industry lined up half a trillion for compute; the bottleneck that bit this week was memory.
Surging AI-infrastructure demand has made memory the scarce input, and the makers are pressing the advantage. SK Hynix spent the week scouting US fab sites as customers demand local supply, and HBM's manufacturing complexity keeps general-purpose DRAM tight enough that one estimate has Micron's margins reaching 95% by 2027. Three firms set the price of AI memory, and the squeeze lets them hold it.
CXMT (ChangXin Memory Technologies), a DRAM maker, passed Tencent on 13 August to become China's most valuable listed company, at about RMB3.54 trillion against Tencent's RMB3.45 trillion. A memory maker is now the most valuable company in China, which tells you where the scarcity, and the pricing power, has moved.
Why this matters:
Compute got the headlines and the capital; memory is the physical part you cannot conjure on a term sheet. HBM and DRAM sit on every accelerator, a fab takes years, and the squeeze lands as cost on every GPU buyer downstream, the neoclouds included.
Three firms set the price of AI memory, and the shortage is handing them the margins to prove it. That cost does not stay with them; it flows straight into the GPU build-out the rest of this issue is about.
The China section below is the demand side, home-grown compute for home-grown models. This is the supply side: a DRAM maker just became China's most valuable company while the West hunts for memory it cannot get.
China Assembles a Home-Grown AI Stack Around a New 100,000-Card Supercluster
Cut off from NVIDIA's best, China spent the week shoring up the domestic capacity to keep serving AI without it.
China switched on its first fully domestic 100,000-card AI supercluster this month, built on home-grown accelerators and split between the National Supercomputing Internet's Zhengzhou node and the Greater Bay Area. Above the metal, Alibaba Cloud made its M890 supernode generally available, 64-card units built to serve mixture-of-experts models up to 10 trillion parameters, with Qwen3.8-Max and Kimi K3 already running on it, and said it will more than double its global modular data-centre capacity.
DeepSeek is hiring to build its own data centres, Zhipu switched on more than 50,000 home-grown AI chips to serve nearly seven million API users, and Moore Threads is filing to list in Hong Kong to fund the next round of domestic silicon.
Why this matters:
Export controls were meant to slow China down. Instead they are forcing the domestic build, and the gap that matters, whether China can serve AI without the chips it is denied, keeps closing.
The M890 serves 10-trillion-parameter models like Qwen3.8-Max from Issue #118, and Zhipu's seven million users are real inference load. China is making the chips and the demand that keeps a home-grown stack running.
The awkward part for NVIDIA: the export rules are building its competitor, and China has run this playbook before with solar, batteries and EVs, each time coming out the low-cost supplier. The same script here shrinks the market the top story's $500 billion is chasing.
Fermi Finally Lands a Tenant at Project Matador: TensorWave's $6.5B Lease
The Fermi mega-campus we twice flagged for having no tenant just signed one.
Fermi (NASDAQ:FRMI), the power-first campus developer co-founded by former US Energy Secretary Rick Perry, has signed the first binding customer lease at Project Matador in Carson County, Texas. The tenant is TensorWave, an all-AMD neocloud fresh off a $350M Series B in June, taking 222MW worth about $6.5 billion over a 15-year term, with rights to expand past 650MW, for a campus built to run tens of thousands of AMD Instinct GPUs from the second half of 2027.
Why this matters:
This is the offtake it needed, signed by a neocloud rather than the hyperscaler these sites are built for. Read the conditions, though: the lease binds only once Fermi lands the "requisite project guaranties and financing" it has not yet raised, at a company whose board fired co-founder Toby Neugebauer in April.
With AMD as both TensorWave's supplier and a backer, this doubles as AMD's largest US reference campus. If the money closes, it is the biggest non-NVIDIA anchor on the board; if not, the second to fall through at Matador in a year.
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.