Neoclouds were meant to eat the hyperscalers’ lunch.

Faster builds. No price gouging. None of the procurement theatre that slows the big three down.

For a while, that pitch worked.

Especially when AI demand was vertical and “build it and they will come” was a viable business model.

But a lot of them built the exact same product. Pointed it at the same customers. Used the same messaging and colour schemes. And financed it all with expensive debt.

And now?

H100 rental rates are in freefall.

H200s? Not far behind. Margins? Gone. Lease payments? Missed. First mover advantage? A distant memory. And the companies that sought to break the hyperscaler monopoly by deploying thousands of systems into a commodity market under the illusion that they’d achieved product-market fit?

They’re taking their calls and signing their term sheets.

And it looks like consolidation has begun.

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.

Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.

Northern Data is selling off business units while weighing a future under new ownership.

The company has signed a non-binding, non-exclusive term sheet to sell Peak Mining, its crypto unit, to Elektron Energy in a deal valued at $235M. It includes $175M upfront, with the rest tied to performance metrics and the transfer of a deposit agreement with a mining hardware supplier. At the same time, video platform Rumble has made a $1.17B all-share offer to acquire the rest of Northern Data, folding in its Taiga Cloud GPU business and Ardent Data Centres.

So, what’s going on?

Northern Data operates more than 22,000 GPUs across Europe.

Those systems are predominantly collateralised H100s, deployed with leverage on the cash flows they’re meant to generate. But H100 rental rates have plummeted. And cash flows have likely dried up.

Given how painful that leverage is likely feeling, asset sales and partnership deals with last year’s competition are all you have left.

Why this matters:

Click here to read the Peak Mining announcement and Rumble announcement from Northern Data.

The US has given NVIDIA and AMD the green light to sell AI chips to China, but only if they hand over 15% of the revenue to Washington.

The deal will see NVIDIA share proceeds from H20 sales, while AMD will do the same for its MI308 chips. This comes after the Trump administration briefly banned those exports in April, only to reverse course when NVIDIA pledged up to $500B in US data centre investments. Licenses are now being issued for the sales, with the deal reportedly tied to broader trade talks over rare earths.

Shakedown? Or the cost of doing business?

You decide.

Why this matters:

Click here to read the full story on TechCrunch.

Core42 is expanding its European sovereign AI footprint through a new strategic deal with Northern Data.

The partnership gives Core42 access to up to 10,000 GPUs from Northern Data’s Taiga Cloud estate, boosting its ability to deliver compliant, sovereign-grade AI infrastructure across Europe. It builds on Core42’s recent European moves in France and Italy, and aligns with its sovereign strategy: blending its own deployments with trusted local partner capacity to scale quickly.

For Northern Data, the deal comes as the company faces falling H100 pricing, senior exec departures, and takeover offers.

Why this matters:

Click here to read the announcement from Core42.

OpenAI’s long-teased GPT-5 has become a case study in over-promising and under-delivering.

Hyped as a “PhD-level expert” that would revolutionise interaction, the new model landed with an immediate backlash. Reddit threads with thousands of upvotes called it “horrible,” with so many users demanding the return of GPT-4o that OpenAI caved. Benchmarks show only incremental gains, and in some areas, users feel performance has slipped. A scathing piece from Gary Marcus accused the release of being rushed, overhyped, and riddled with the same hallucination and logic failures that have plagued LLMs for years.

Why this matters:

Click here to read Gary Marcus’ Substack article.

The quantum race is no longer just physics experiments in a lab.

IBM and Google both claim they can deliver industrial-scale quantum computers by 2030, pushing the field from sub-200 qubit prototypes toward the million-qubit systems needed for practical use. That leap means overcoming brutal scaling challenges: qubit instability, interference, manufacturing limits, and the staggering cost of error correction. Google is betting on surface code error correction, requiring millions of physical qubits but already showing scaling gains. IBM is pursuing low-density parity-check codes to slash qubit requirements by up to 90%, though manufacturing proof is pending.

AWS and others, however, remain sceptical, suggesting “useful” quantum may still be decades out.

Why this matters:

Click here to read the full story in the FT.

Microsoft is investigating reports that units from Israel’s Intelligence Corps are using Azure to run airstrike targeting decision workloads.

Unit 8200 is allegedly using Azure to store and process vast intelligence datasets on Palestinians and high-priority targets. This isn’t the first time the company has answered questions about their engagement with Unit 8200. Prior reviews found “no evidence” that Microsoft’s tech had been used to harm people in Gaza. But new reporting from The Guardian, +972 Magazine, and Local Call suggests some employees may have known more.

Why this matters:

Click here to read the full story on DCD.

TeraWulf has signed two 10-year HPC colocation agreements with Fluidstack for 200+MW of capacity.

The deal is worth $3.7B in contracted revenue, with two 5-year extensions potentially pushing that to $8.7B. And to make it even sweeter, Google is backstopping $1.8B of Fluidstack’s lease obligations while taking an 8% equity stake in TeraWulf via warrants. Phase one (40MW) is due online at the Lake Mariner campus in 1H 2026, with the rest deployed by year-end. TeraWulf also granted Fluidstack a 30-day exclusivity window for another 160MW at the site.

With this new agreement, Lake Mariner is now firmly cemented as the anchor site for AI hosting in New York State.

Why this matters:

Click here to read the full TeraWulf announcement.

At Christmas, there were 145+ neoclouds and GPU providers.

I wrote that the market was splitting into three tiers:

Eight months later, the lines are sharper, and the stakes are higher.

Tier 1 are still the leaders, and they’re way out in front. They’ve carved out their niche, dictate terms, own the relationships that matter, and are selective about the partnerships they sign.

Tier 2 are the survivors and dealmakers. They’re pragmatic enough to cut capacity deals, swap workloads, or align with sovereign operators if it keeps the lights on.

And Tier 3? The left behind. They’re still in the market, but who knows for how much longer.

Because in August 2025, you’re either overextended and living with the consequences of questionable operational decisions. Or you’re not. And that changes everything.

The calculus has shifted from independence vs growth to two simple questions:

Can you make it through the next twelve months?

And if you can’t, what are you willing to do to survive?

See you next week.