Cracks might be showing in the UK’s AI crown jewel.
Nscale, the startup tasked with building Britain’s “Stargate” and a linchpin in the UK-US Tech Prosperity Deal, is caught in lost lawsuits, failed deals, and the ensuing political fallout.
Not ideal at the best of times.
Even worse when the story breaks the same week that your biggest customer signs multiple deals with multiple competitors.
The market giveth, the market taketh away.
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.
I’ve also got a new section for you in this issue: the appropriately named Everything Else.
Be sure to scroll all the way to the end to check it out. Especially if you’re looking for more regular company coverage or novel ways to get the right kind of attention.
But I digress.
Here’s what’s inside this week:
Let’s get into it.
The GPU Audio Companion Issue #72
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.
Nscale’s Infighting Raises UK AI Strategy Questions
Nscale is facing serious credibility questions.
New Sifted reporting reveals infighting, a failed acquisition, and a copyright lawsuit that expose cracks in one of the UK’s most high-profile AI projects. Court filings show Nscale lost a Dutch case against former contractors accused of stealing “trade secrets” for optimisation software called Paiton. Judges ruled the code was already public on GitHub before they joined, ordering Nscale to pay £7,000 in costs.
And that’s not all.
The company also falsely claimed it had acquired modular data centre firm Kontena.
That deal that never happened. Despite press coverage. And Nscale’s own announcement. The website has been updated and the announcement removed. Even so, it’s hard not to wonder if there’s anything else going on in the background.
While this is likely the biggest crisis Nscale’s PR team have ever faced, it’s an even worse look for a government that desperately needs a win.
Prime Ministerial popularity is dwindling, and tax rises loom on the horizon.
AI is a cornerstone of the Labour Government’s new industrial strategy, and the UK-US tech prosperity deal is the linchpin. With Nscale right at the centre, and with involvement from NVIDIA and Microsoft, one has to ask how all of the stakeholders involved in the deal missed this.
And, if they did, why?
Why this matters:
Stories like this raise serious doubts about governance and transparency in the UK’s sovereign AI strategy.
They also serve as a pretty stark reminder that multiple nationally significant projects the world over have come to rely on unproven private players with less than 2 years of trading history.
In an industry that runs on hype, this could be a watershed moment. Given the sums of money involved in all the deals currently being done, not just Nscale, it would likely come as no surprise to anyone if there’s more dirt to be found somewhere.
Neoclouds IREN, Lambda Land Microsoft Contracts
Two massive deals this week cement Microsoft’s position as the primary neocloud offtaker.
First, Iris Energy (IREN) signed a $9.7 billion, 10-year AI cloud contract with Microsoft - one of the largest in the company’s history. The agreement will see IREN supply NVIDIA GB300s from its 750MW Childress, Texas campus, and deliver 1.4GW of capacity across its US sites for Microsoft’s AI workloads. The deal builds directly on IREN’s existing 560 MW footprint, representing a scale-up that effectively turns the miner-turned-AI operator into a sovereign-scale hyperscale partner.
Then came Lambda.
The company announced a “multibillion-dollar, multi-year deal” with Microsoft to deploy “tens of thousands of NVIDIA GB300 GPUs”, including full NVL72 rack systems. This partnership will see Lambda operate high-density AI supercomputers for Microsoft’s expanding AI cloud, reinforcing its transformation from boutique GPU provider into one of the world’s most advanced AI infrastructure operators. Contract details are light, but it’s a huge win for Lambda.
Why this matters:
Microsoft is the central demand engine driving private AI cloud expansion across the neocloud space, striking long-term leases with emerging operators from Bitcoin-era miners to specialist GPU clouds.
The company’s approach mirrors traditional energy offtake: locking in multi-gigawatt capacity through multi-year contracts to secure compute supply ahead of its hyperscale needs.
By sourcing from new partners like IREN, Lambda, and others, such as CoreWeave, Nebius, and Nscale, Microsoft decentralises its AI infrastructure while maintaining control over capacity, compliance, and pricing.
Nebius Brings B300s to the UK, Launches Token Factory
Two updates from Nebius this week.
First, The Ultimate Cloud for AI Innovators switched on one of the UK’s first NVIDIA Blackwell Ultra deployments. The new London facility (hosted at Ark Data Centres) delivers B300 GPUs and Quantum-X800 InfiniBand networking. It’s part of Nebius’s global rollout across Europe, the US, and Israel, but with one subtle difference: it’s designed to meet the UK’s strictest data governance and compliance standards.
SOC2 Type II, HIPAA, and GDPR compliance come as standard from Day 0, enabling open-source enterprise AI deployment at scale.
That’s where Nebius’s new Token Factory comes in.
Per this week’s launch announcement, the platform turns open-source and custom models into production-ready inference systems. Guaranteed isolation, autoscaling throughput, and zero-retention inference in EU or US data centres ensure regulatory adherence for the most sensitive workloads. For healthcare, finance, and government users, this unlocks a middle ground between flexibility and compliance.
Think open infrastructure with enterprise-grade control seasoning.
Why this matters:
Nebius’s London rollout brings Blackwell-class compute onshore for critical sectors like healthcare and financial services which are major contributors to UK GDP.
The Token Factory signals a broader move toward open-model infrastructure, giving enterprises a way to deploy AI under strict governance frameworks without depending solely on closed APIs.
With outages from major hyperscalers recently knocking critical services offline, Nebius’s “sovereign-ready” open cloud model launch comes at the perfect time, offering resilience, auditability, and local control. All attributes that regulated industries increasingly demand.
Edison Scientific Launches “AI Scientist”: Kosmos
A new class of AI just entered the scientific arena.
Edison Scientific, the new commercial spinout from FutureHouse, has launched Kosmos. The pitch? An autonomous AI Scientist capable of running tens of thousands of experiments, reading over a thousand papers, and producing traceable, publish-ready discoveries in a single run.
Built to overcome the context limits of previous models like Robin, Kosmos uses a structured world model that lets it combine insights across hundreds of agent trajectories and maintain coherence over tens of millions of tokens.
Each 12-hour run executes around 42,000 lines of analysis code and processes 1,500 papers, producing reports where every conclusion is fully cited and auditable down to the original data or code line.
In trials, Kosmos replicated multiple unpublished human findings and made four novel discoveries across neuroscience, genetics, and materials science. These included identifying molecular mechanisms for Alzheimer’s progression and new pathways in Type 2 diabetes. Independent scientists verified that 79.4% of its findings were accurate, with collaborators estimating that a single Kosmos run delivered the equivalent of six months of postdoc-level research.
Why this matters:
If these claims continue to be verified by third parties, we could be seeing the start of AI-accelerated science, where models move beyond summarisation to autonomous hypothesis testing and discovery.
Demonstrates how structured agentic systems can sustain multi-hour, multi-million-token reasoning loops, a major leap from current LLM coherence limits.
That being said, PhD-level AI is a meme at this point, so until any results/data/findings/claims are subjected to rigorous peer-review in a journal of note, like human-researched science, maybe take any claims with two handfuls of salt.
Azure’s GB300 Racks Smash the Million-Token Barrier
Microsoft just benchmarked the new Azure ND GB300 v6 systems, and the numbers are staggering.
A single GB300 NVL72 rack hit 1.1 million tokens per second running Llama 2 70B, a 27% jump over the 865,000 tokens/sec record held by the previous GB200 NVL72. Each GB300 GPU pushes around 15,200 tokens/sec, roughly five times faster than an H100. The leap comes from 50% more GPU memory, a 16% higher power envelope, and massive bandwidth upgrades:
7.37 TB/s HBM throughput at 92% efficiency.
4× faster CPU-GPU transfers through NVLink C2C.
2.5× GEMM TFLOPS versus the H100-based ND v5.
In plain terms, the GB300 isn’t an iteration. It’s a new performance class.
Why this matters:
GB300’s efficiency crushes legacy economics, and shows everyone why it’s so difficult for XPUs to compete against NVIDIA.
Fivefold GPU-level gains make older inference architectures instantly obsolete.
With Azure deploying these racks at scale, Blackwell-era compute is rapidly becoming the new enterprise baseline for inference.
AWS Signs GPU Deal with OpenAI, Lease with Cipher
The contract will see AWS supply hundreds of thousands of NVIDIA GB200 and GB300 GPUs to the frontier AI lab, securing its biggest AI cloud win to date. It also builds on Amazon’s growing AI stack. Just last week, Anthropic’s new AWS complex went live, powered by hundreds of thousands of Trainium2 chips. That setup forms the backbone of Anthropic’s Claude models and gives Amazon two parallel bets: NVIDIA for Sam, Trainium for Dario.
And that’s not all.
Cipher Mining also signed a $5.5 billion, 15-year lease agreement with AWS this week.
Per the terms of the lease, Cipher will deliver 300MW of air and liquid cooled AI-ready capacity. The two-phase deployment is expected to begin in July 2026 and be completed in Q4 2026.
Why this matters:
AWS is now part of OpenAI’s compute network alongside Microsoft, Oracle, Google Cloud, CoreWeave, and Nscale (when it comes online).
OpenAI diversifies its compute supply amid a $1.4T infrastructure rollout, while becoming even more systemically important to the US AI buildout, edging closer to too big to fail territory.
As the AI buildout tends toward a multicloud strategy, the hyperscalers need ever-increasing power, whitespace, and experience. Ex-Bitcoin miners like Cipher have all three.
Space (Data Centres): The Final Frontier
Space is officially the next data centre frontier, and the race just got serious.
After Crusoe and Starcloud announced an orbital data centre partnership just weeks ago, and Starcloud put their first facility in orbit this week, SpaceX has now joined the fray. Elon Musk claims that the company will “scale up Starlink V3 satellites into full data centres.”
These next-generation satellites, equipped with laser interlinks and designed for gigabit throughput, could serve as the building blocks for orbital compute infrastructure. Using solar power, zero cooling costs, and global line-of-sight latency, a SpaceX orbital DC could deliver ultra-resilient compute networks.
Musk even suggested Starship could deliver 100GW of solar generation capacity to high orbit within five years, a figure that, if even partly accurate, would transform global compute economics.
Now Google has entered the picture.
Its new initiative, Project Suncatcher, aims to scale machine learning compute in space using solar-powered satellites equipped with its in-house-designed TPUs. Google, in partnership with Planet, plans to launch two prototype satellites in 2027. Initial launches will test TPU performance and radiation resilience in orbit, and be the first step toward what it calls “massively scaled computation in space.”
Why this matters:
With SpaceX, Google, and startups like Starcloud now all pursuing orbital infrastructure, space data centres are evolving from concept to competition.
Given power, land, and water constraints on earth, and falling launch costs, the incentives are aligning for space-based compute to become a significant market segment.
Everything Else
There’s so much going on in this industry that it’s impossible to fit it all into the seven-story+commentary structure.
That’s why I’m introducing the “Everything Else” section from this week.
These are the stories that are worth your attention, but didn’t quite make the cut for the full editorial treatment.
Going forward, I’ll do my best to aggregate all notable updates into this list.
To that end, if you’re looking to make sure your company updates are being seen by the people who matter, shoot me a message on LinkedIn or reply to this email.
I’m still not doing traditional PR, but I’m happy to help spread the word.
Are you interested in a deeper relationship with The GPU to ensure your news gets featured in this list?
The Rundown
First, the good.
IREN and Lambda landing Microsoft as a tenant is a huge deal for both of these companies. It’s not easy landing a hyperscaler as a customer (and it’s even harder still to meet stringent SLAs). Both companies have matured massively in the past 12 months, and that labour is clearly bearing fruit.
Congratulations to all involved.
The same can also be said of StarCloud.
Two years ago, the idea of putting data centres in space seemed absurd. Then two weeks ago, Crusoe announced a partnership with the space data centre builder. Bold. Now, two of the biggest names in technology, SpaceX and Google, are independently validating the same thesis by committing to their own projects. Not bad for a startup.
Nebius, technically, also still counts as a startup, and they continue to do what many in this industry struggle to do: deliver.
Switching on some of the first B300s in the UK to serve some of the premier biotech companies in London while simultaneously rolling out a compliant software suite that caters to the compliance requirements of the local market is no mean feat.
Then, there’s Edison Scientific’s Kosmos.
An AI researcher who can do six months of lab work in a day?
Maybe science fiction. Or maybe not, depending on peer review. If true, it changes R&D forever. If not, it’s the best marketing in the field.
This week, at least.
On the subject of marketing, OpenAI and AWS’s new partnership is almost predictable at this point.
Giant AI lab signs giant deal. Giant amount of money for a giant amount of GPUs is announced. Giant hyperscaler adds marquee name to giant customer list. Sama adds another provider to OpenAI’s giant list of partners.
Giant amount of press coverage, and thus the cycle begins anew.
And finally, there’s Nscale.
Six weeks ago, they closed the largest Series B in European history. This week? Retracting press releases of acquisitions that didn’t go through, losing lawsuits they brought against contractors they hired and, from the lack of response noted in Sifted’s exposé, avoiding questions fielded by reporters.
Maybe it’s nothing, or maybe it’s something.
Time will tell, as it always does.
If there’s a thread to be found that links this week’s stories together, it’s that speed continues to be everything in this industry. If you’re in control, that’s a blessing. But if you lose it?
If the spectre of past decisions makes its presence felt?
We all know what can happen next.
See you next week.