In partnership with
Partridges and pear trees remain conspicuously absent this week, though we seem to have just about everything else.
Multi-GW deployments. New regional HQs. Sovereign frontier-grade GPU deployments. A new family of models. 9-figure funding rounds. A credible threat to NVIDIA’s CUDA moat.
Oh, and two stargates.
Because who wants a slowdown towards the end of the year?
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.
The GPU Audio Companion Issue #79
Want the GPU breakdown without the reading (or the ads)? The Audio Companion does it for you, but only if you’re a paid subscriber. If you can’t see it below, click here to fix that.
Effortless Tutorial Video Creation with Guidde
Transform your team’s static training materials into dynamic, engaging video guides with Guidde.
Here’s what you’ll love about Guidde:
1️⃣ Easy to Create: Turn PDFs or manuals into stunning video tutorials with a single click.
2️⃣ Easy to Update: Update video content in seconds to keep your training materials relevant.
3️⃣ Easy to Localize: Generate multilingual guides to ensure accessibility for global teams.
Empower your teammates with interactive learning.
And the best part? The browser extension is 100% free.
Hut 8, Anthropic, Fluidstack Ink Multi-GW Partnership
Hut 8 are partnering with Anthropic and Fluidstack to build between 245 MW and 2,295 MW of AI data centre capacity in the US.
The build starts at Hut 8’s River Bend campus in Louisiana, where the first 245 MW of IT load will be delivered, backstopped by Google. Anthropic also receives a ROFO for a further 1,000 MW at the same site, contingent on securing more power. A third tranche opens the door to an additional 1,050 MW across Hut 8’s broader development pipeline.
Fluidstack will operate the clusters, extending its position as Anthropic’s primary infrastructure partner.
Why this matters:
Anthropic is locking in long-term power and land positions, signalling more vertical integration in AI infrastructure, and Fluidstack cements its role as Anthropic’s primary operating partner, repeating the same deal st
Hut 8 joins the top tier of AI infrastructure builders by securing a pipeline that rivals hyperscaler-scale expansions.
With Google stepping in to backstop the financing obligations, this deal is broadly similar to the TeraWulf and Cipher Mining contracts from earlier in the year.
Carbon3.ai Announces UK Blackwell Ultra Deployment
Carbon3.ai has announced plans to deploy one of the UK’s first large-scale NVIDIA Blackwell Ultra clusters.
The setup pairs HPE’s liquid-cooled XD685 servers with Spectrum-X networking, BlueField-3 DPUs and HPE’s modular AI Mod PODs. The wider stack includes VAST Data and NVIDIA AI Enterprise, targeting high-performance workloads in finance, manufacturing and healthcare. Carbon3.ai says the design can cut energy use by up to 30% and increase density, backed by renewable power across its sites.
The systems aren’t live yet, but the pace of announcements shows Carbon3.ai positioning itself aggressively in the UK sovereign-compute race.
Why this matters:
Another UK operator is pushing into Blackwell-class deployments, increasing domestic competition.
DLC and renewables give Carbon3.ai a credible path to high-density, lower-cost sovereign compute once deployed.
Carbon3.ai’s rapid cadence of announcements signals a company trying to scale into a top-tier UK AI infra player.
Core42 Sets Up European HQ in Dublin
Core42 has picked Dublin for its European headquarters.
The HQ will serve as the hub for delivery, engineering, regulatory work, and partnerships. It builds on Core42’s recent national deployments in France and Italy, where the company is helping to stand up some of Europe’s highest-density sovereign clusters. Dublin offers a neutral regulatory environment, strong cloud talent, and proximity to EU institutions at a time when sovereignty and trusted AI infrastructure have become board priorities.
Why this matters:
Demand for sovereign infrastructure is rising rapidly across the EU. Especially in the wake of the recent Azure and AWS outages.
This move will likely increase pressure on domestic providers and local neoclouds.
Dublin is emerging as a strategic nerve centre for sovereign compute, with both Core42 and Crusoe anchoring European operations there.
NVIDIA Launches Nemotron 3 Model Family
NVIDIA has released the Nemotron 3 family of open models.
The new family, including Nano, Super, and Ultra, is built around a new hybrid MoE architecture designed for multi-agent systems. Per NVIDIA’s announcement, Nemotron 3 Nano, a 30B-parameter model with 3B active, delivers 4x the throughput of Nemotron 2 Nano, cuts reasoning-token generation costs by up to 60%, and supports a 1M-token context window. Artificial Analysis ranks it as the most efficient open model in its class.
Nemotron 3 Super (≈100B parameters, 10B active) and Ultra (≈500B parameters, 50B active) target high-accuracy reasoning and multi-agent workflows, trained using NVIDIA’s 4-bit NVFP4 format on Blackwell systems. NVIDIA also released 3T tokens of new training and RL datasets, plus NeMo Gym, NeMo RL, and NeMo Evaluator for post-training and safety testing.
Why this matters:
Meta’s potential step back from open releases with Avocado leaves a clear space for a heavyweight open model provider. NVIDIA is moving to fill it.
NVFP4 training lets teams push larger models onto current Blackwell fleets, widening what counts as “frontier-class” infrastructure.
NVIDIA now owns both layers of the agentic stack: Blackwell hardware and the open model ecosystem running on top of it.
Mythic Raises $125m in Analogue Chip Push
Mythic has raised $125m to scale its analogue compute-in-memory chips.
The company claims up to 120 TOPS per watt, roughly 100x the efficiency of current GPUs once memory movement is included, by merging compute and memory into the same analog plane. The chips are already in customer hands across defence, automotive, and robotics. Internal benchmarks point to large-model advantages too, with Mythic reporting up to 750x more tokens per second per watt than high-end GPUs on trillion-parameter workloads.
Why this matters:
Energy efficiency, not FLOPs, is now the binding constraint on AI scale, and analog compute is emerging as one of the few viable escapes.
If Mythic’s gains hold in data-centre settings, inference economics could tilt toward architectures that run cooler, cheaper, and on mature fabs.
The hardware frontier is widening beyond GPUs: the next breakthroughs may come from companies with no legacy in digital accelerators.
Google targets CUDA with “TorchTPU”
Google is working on TorchTPU, a major push to make its TPUs fully PyTorch-native.
Until now, Google tuned TPUs almost exclusively for its internal stack (Jax + XLA). This forced external customers into extra engineering work to hit acceptable performance. TorchTPU is designed to remove that barrier and erode Nvidia’s lock on the AI software layer by providing developers with a drop-in alternative that doesn’t require rewriting workloads in Jax. If and when TorchTPU lands, switching costs fall sharply for any team already standardised on PyTorch, which is almost everyone.
That means greater optionality for customers and renewed hardware competition.
Why this matters:
Successfully weakening that link will likely change the competitive landscape, though even efforts from the likes of AMD have only had limited success.
If successful, we could see hyperscalers, neoclouds, and model labs reassessing their hardware stacks at scale.
Vantage Breaks Ground on Two Stargates
Vantage has broken ground on two Stargate campuses in Texas and Wisconsin.
First up in Shackelford County, Texas, work has begun on the 1.4GW Frontier campus. The 1,200-acre site will house ten single-story data centres, support rack densities above 250kW, and deliver its first building in H2 2026. Next, in Port Washington, Wisconsin, Vantage has also started construction on the $15bn Lighthouse campus. The 672-acre site will provide up to 902MW of zero-emission capacity for Oracle and OpenAI.
Why this matters:
Vantage is now the core delivery arm for Stargate, with multi-GW campuses moving ahead in parallel.
The Texas site sits near Abilene, where Crusoe is building another Oracle–OpenAI facility, tightening the regional cluster.
The Abilene corridor is rapidly becoming one of the largest concentrations of AI power demand in the US.
Introducing…Ads
As The GPU has grown, so has the time it takes to research, write, and ship a weekly issue.
That time comes straight out of paid client work, and the trade-off is starting to pinch. So from this week, I’m turning on ads from the beehiiv ad network. If you’ve read this far, you’ve already seen one. This is not a decision I have taken lightly, but needs must!
A few things to be clear about:
I’ll endeavour to only run ads that are broadly relevant to my audience (that means no AG1)
These are separate from direct sponsors. Direct sponsors are still hand-picked and clearly labelled.
If you’d rather keep the issues clean, the £5/month subscription keeps The GPU ad-free from beehiiv network ads (you’ll still see direct sponsors, because they directly support the work).
This change keeps the newsletter sustainable, keeps the weekly cadence intact, and lets me keep spending time on the parts that matter:
The analysis, the sourcing, and the calls with operators who actually build this stuff.
If you’ve got strong feelings on what should or shouldn’t be allowed as an ad here, reply and tell me.