In partnership with
The compute financing ouroboros continues.
This week a chipmaker took a five-billion-dollar stake in the lab that just agreed to buy its GPUs. Another chipmaker bought a slice of the cloud that rents out its own chips. And the biggest data-centre deal on record closed at forty billion dollars.
Everyone is now financing everyone.
It is a lovely machine, until one payment in the loop stops.
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.
Your prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.
AMD Takes a $5B Stake in Anthropic and Lands a 2GW Order
The clearest crack yet in Nvidia's monopoly came with a cheque going the other way.
AMD used its Advancing AI 2026 event to agree to invest up to $5 billion in Anthropic, with Anthropic committing to deploy up to 2GW of AMD's Instinct MI450 GPUs and the first gigawatt due in the first half of 2027. Anthropic was not an AMD customer at this scale before, which makes this the most significant non-Nvidia commitment a frontier lab has made. It also rhymes with the rest of the week: the chip supplier is taking equity in the customer that buys its chips, the same loop showing up at Nvidia and Nebius below. Anthropic already runs large fleets on Google TPUs and AWS Trainium, so read this as widening the field, not walking out on Nvidia. Nvidia is still the default. It just picked up a credible rival at gigawatt scale.
Why this matters:
The $5 billion and the 2GW order are the same money moving in a circle. AMD funds Anthropic, Anthropic buys MI450s, AMD books the revenue. The deal is only as real as the slice of that 2GW Anthropic would have bought without the cheque.
We flagged AMD's challenger push in Issue #113. The MI450 ships in 2027, so this locks Anthropic's next training generation, not this one.
The tell is what actually runs on the racks. Inference on AMD is routine. Frontier training on ROCm at CUDA-level utilisation is the milestone nobody has hit.
The Biggest Data-Centre Deal Ever: $40B for Aligned
The largest cheque ever written in data centres just cleared, and it came from the Gulf and Wall Street.
MGX, AIP and BlackRock's Global Infrastructure Partners completed a $40 billion acquisition of Aligned Data Centers, the largest deal the industry has seen. It pairs Gulf capital, through Abu Dhabi-backed MGX, with the biggest infrastructure investor on Wall Street to buy one of the largest hyperscale developers in the Americas. $40 billion for data centres is a bet that AI demand keeps the halls full for decades, at a scale once reserved for pipelines and airports. Aligned's campuses across the US and Latin America now sit inside the same vehicle that is building for the hyperscalers. Infrastructure capital has decided the data centre is the new toll road.
Why this matters:
Infrastructure funds are paying toll-road prices for assets that carry data-centre risk. Pipelines throw off contracted cash for 30 years. A GPU hall's lease and its hardware both age out far sooner, so the $40 billion works only if the second tenant pays as much as the first.
We watched private capital move into the physical layer in Issue #110. At $40 billion it stops being a trade and becomes the benchmark every other campus gets marked against.
The risk is duration. Debt this size has to be served for longer than any GPU lease yet signed, which means it is underwritten on demand no one has actually contracted.
Nvidia Takes 9.3% of Nebius, Then Nebius Borrows Another $775M
Nebius has a new shareholder.
Nvidia disclosed a 9.3% stake in the largest independent European neocloud in a regulatory filing, and days earlier Nebius secured $775 million in debt to fund its build-out. In May it lined up a Bloom Energy fuel-cell offtake backed by $1.7 billion from IDF and Oaktree. One tier down from the lead: Nvidia takes equity in the cloud that buys its chips, lenders finance the halls those chips sit in, and the whole thing runs on frontier-lab demand. Good business while demand compounds. It also means the chipmaker, the lender and the customer are increasingly the same handful of names.
Why this matters:
Nvidia holding 9.3% of its own customer is also an allocation signal. A seat on the cap table tells the market who gets Rubin first, which quietly disadvantages every neocloud Nvidia does not own.
We tracked the neocloud financing machine in Issue #112. The fresh $775 million is borrowed against GPUs that lose value faster than the loan amortises, so the collateral shrinks while the debt holds.
Concentration is the quiet risk. When one company is shareholder, supplier and demand-setter across the same neoclouds, a single downturn hits all three roles at once.
Microsoft Is Auditioning a Chinese Model to Cut Its Copilot Bill
The company that bankrolled OpenAI is testing whether a Chinese open model can do the job for $600 million less.
Microsoft is internally evaluating Moonshot's Kimi K3 as a replacement for some Copilot inference, with reported savings of up to $600 million. It follows the pattern from last week: Microsoft leaning on cheaper models to run more of its own inference and depend less on OpenAI. The complication is provenance. The White House has accused Moonshot of distilling Anthropic's Fable model and training on Nvidia GB300s routed through Thailand, and the Treasury has threatened sanctions over it. Microsoft testing that model inside its flagship product sharpens the bind: the cheapest inference on the board may also be the most politically radioactive.
Why this matters:
The $600 million is leverage as much as savings. Microsoft does not have to ship Kimi to use it. A credible cheaper option resets what it pays OpenAI, whether or not a single request ever routes to Beijing.
We covered the repatriation story in Issue #115. Reaching past Western models to a Chinese open-weight one crosses a political line the earlier cost-cutting never did.
The exposure is timing. If Microsoft wires Kimi into Copilot and sanctions land after, the cost is a forced rip-out mid-production, not just lost savings. Watch whether this leaves the test environment.
AMD Officially Launches Helios, and the Big Buyers Line Up
AMD's answer to Nvidia's rack finally shipped, and the deployment list reads like the whole industry.
AMD officially launched Helios, its rack-scale system built on MI455X GPUs, Venice EPYC CPUs and Pensando networking, a direct answer to Nvidia's NVL72. OpenAI is standing up its first Helios rack under the 6GW partnership it signed last October, Meta is scaling the 6GW AMD deal it struck in February, and Microsoft is deploying Helios for Azure inference. None of those commitments are new this week. The hardware to honour them is. A rack sells itself once your rival's biggest customers are already racking it.
Why this matters:
At rack scale the contest is the interconnect, not the die. Nvidia's NVL72 wins on NVLink, so Helios has to prove MI455X and Pensando scale as cleanly across a rack. Buyers purchase the fabric, not the chip.
We flagged AMD's move to the rack layer in Issue #113. OpenAI, Meta and Microsoft are already contracted, so the real test is the first big buyer that racks Helios with no prior AMD deal behind it.
A shipping rack makes AMD benchmarkable. Within two quarters the market will have production dollar-per-token figures for Helios against NVL72, and that number, not the keynote, sets the 2027 order book.
Alphabet Spends $45B in a Quarter and Guides Higher
The quarter beat. The spending line moved it.
Alphabet reported Q2 2026 with revenue up 24% to $119.8 billion and Google Cloud up 82% to $24.8 billion, then lifted its full-year capex guidance to $195-205 billion, up from $180-190 billion. Capital expenditure hit $44.9 billion in the quarter alone, roughly double a year earlier, with management pointing about 60% towards servers and 40% towards data centres and networking. That is the reported cash capex line; Alphabet does not fold finance leases into the headline figure, so the true commitment runs higher. The reaction was unease rather than applause: costs are now climbing faster than revenue, and investors are asking out loud whether the capex cycle has outrun near-term monetisation.
Why this matters:
A big share of that $44.9 billion is server spend on in-house TPUs, which skip the margin Nvidia charges everyone else. Alphabet turns capex into compute more cheaply than rivals buying merchant GPUs, so flat capex comparisons flatter the companies paying Nvidia.
We logged the same capex vertigo in Issue #115, where Amazon hit $44.2 billion and Microsoft $31.9 billion. Alphabet just matched them and guided higher, which makes this the sector's baseline, not one company's bet.
The real signal is the market's mood. For two years more capex lifted the stock; unease on a beat quarter says investors now want revenue, not just spend. Microsoft and AWS report within days: matching gaps make it a sector condition.
Crusoe Wants to Run 5GW of GPUs Without the Grid Noticing
The hardest part of an AI campus isn't the chips. It's not tripping the grid when they all spin up at once.
Crusoe and ON.energy are deploying 5GW of what they call AI UPS across multiple hyperscale campuses, with commissioning starting this year and running into 2027. The pitch addresses a problem few discuss: GPU training loads swing power draw hundreds to thousands of times a day, and utilities hate it. ON.energy's system sits between the cluster and the grid, absorbing those swings so a 100,000-GPU site does not destabilise the local network. Solve that, and sites a grid operator would otherwise reject become buildable. It is the logic driving behind-the-meter gas and dedicated nuclear: whoever controls the power interface gets to build where others can't.
Why this matters:
The scarce resource is shifting from megawatts to load quality. The market has raced to secure generation. Crusoe is attacking the other half: the violent swing of training load that makes utilities say no even when the power is there.
We covered the scramble for firm power in Issue #110. This is its complement. Winning the generation race still leaves you in the interconnection queue, and flattening the load is what clears it.
The buffer is more than insurance. Batteries sized to smooth those swings can arbitrage power prices and sell grid services, turning the power interface into a margin line rather than a cost.
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.