The shift is real.

Nebius brought Blackwell to the UK. Together AI is bringing 2GW. And NVIDIA’s everywhere, GTC Paris, London Tech Week, shaking hands with Starmer and Macron.

But behind the fanfare, AMD made its most serious move yet.

An open ecosystem. A rack-scale system. A real CUDA alternative.

I’m Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.

Here’s what’s inside this week:

Let’s get into it.

(Plus, I have some extras for you at the end of this issue - be sure to scroll to the end to find them!)

The GPU Audio Companion Issue #46

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, fix that here.

AMD Unveils Vision for an Open AI Ecosystem

AMD just laid out its blueprint for an open AI stack to break CUDA’s grip.

At its June keynote, AMD pitched a full-stack, plug-and-play platform built around Instinct GPUs, bolstered by the Brium acquisition. The stack knits together Triton, SHARK, WAVE, and open compilers, aiming to deliver frictionless performance without hand-tuned configs. It’s designed for enterprises that want to move fast, stay compliant, and avoid vendor lock-in. AMD is betting that time-to-value matters more than theoretical FLOPs.

Why it matters:

  • Hardware doesn’t win by itself, and Nvidia’s ecosystem is a big part of why it dominates, even though it’s closed à la Apple.

  • Enterprises want open, auditable stacks they can tweak, so a turnkey open platform lowers friction for pilot-to-production rollouts.

  • Competition is good for markets, and Lord knows does the AI infra market need it.

Read the announcement from AMD here.

Jensen’s European Tour

This week, NVIDIA went on a charm offensive across the continent.

At London Tech Week, Jensen launched sandboxes with the FCA and pledged deeper support for the UK’s compute stack. Over in Paris, GTC Europe featured Grace Blackwell systems in full production, industrial AI clouds, and regional tech hubs from Finland to Spain.

Why it matters:

  • Europe’s AI spending is accelerating, and so is the appetite for local control.

  • NVIDIA’s moving beyond just hardware, and playing soft power games with research centres, curriculum, and national AI strategies.

  • With AMD increasingly pushing an alternative ecosystem view, and hardware/software releases on the horizon that might actually provide a meaningful challenge, the stakes have never been higher.

Read all of NVIDIA’s myriad blog posts from Jensen’s tour here.

Nebius Lands Blackwell in the UK

Nebius just became the first Euro neocloud to offer Blackwell Ultras in Britain.

Following its launch in mainland Europe, Nebius is now bringing Blackwell Ultra to the UK. And they’re starting with 4000 GPUs. Big. With a footprint spanning Amsterdam to London, Nebius is framing itself as the compelling alternative to CoreWeave, Lambda, and the Big 3 hyperscalers.

Why it matters:

  • Sovereign infra is scaling fast to meet regional AI demand, but not in the UK.

  • This move finally puts the UK on the map for the latest generation/bleeding edge of hardware.

  • Nebius are execution first, everything else second.

Read the announcement from Nebius here.

Meta Backs Scale AI at $29B Valuation

Meta just took a ~$14B swing at the most important data pipeline in AI.

It’s Meta’s biggest AI investment yet: a 49% stake in Scale AI, which annotates and labels data for most top-tier labs. Founder Alexandr Wang joins Meta to steer its “superintelligence” push, but Scale stays operationally independent. The deal funds growth and shareholder exits, while giving Meta direct influence over the raw material that trains frontier models.

Why it matters:

  • Labelled data is still the lifeblood of large-scale training.

  • Meta is locking in its upstream data supply, likely to fix issues with the lacklustre Llama-4 family and beyond.

  • Scale’s independence preserves its role in the wider AI stack (and likely avoids adding further fuel to the antitrust fire).

Read the full story here.

AMD Unveils Helios, Rack-Scale Solutions Are Go

72 MI400 GPUs, 2.9 exaFLOPS, and 31TBps bandwidth. All in a double-wide rack designed to challenge Vera Rubin.

AMD pulled the curtain back on Helios at Advancing AI in San Jose. And it’s a monster. Slated for 2026, Helios is AMD’s direct answer to NVIDIA’s Vera Rubin NVL144, packing 72 MI400 accelerators connected via UALink, the open alternative to NVLink. The system pools memory and compute into one logical “super-GPU,” with 432GB of HBM4 per GPU and 2.9 exaFLOPS of FP4 inference throughput.

It’s also paired with 6th-gen Venice EPYC CPUs and Pensando Vulcano superNICs, offering 50% more memory bandwidth and capacity than NVIDIA’s NVL144, but requires double-wide racks and a presently unknown power draw, potentially limiting its footprint in dense, energy-constrained sites.

Why it matters:

  • NVLink has real competition: UALink is the first serious challenger, and it’s open.

  • Helios offers 31TBps and 1.4PB of pooled memory. That means massive models.

  • High power and oversized racks could slow adoption in legacy data centres, but that’s not a problem unique to AMD.

Read the full breakdown of Helios on The Register, here.

Google Breaks Internet

Google ruined everyone’s day this week.

Google Cloud reported an IAM issue on June 12 at 11:46 PDT, affecting 40+ regions and 26 services, including the Cloud Console. Cloudflare, which relies partly on Google infra, went dark before Google posted its first alert. By mid-day, services began recovering—but not before retry storms and cache failures triggered knock-on outages across major platforms. It was a reminder that hyperscaler incidents don’t stay contained.

Why it matters:

  • A not insignificant portion of the internet went down on the 12th.

  • Many consumer-facing services, like Spotify, Discord, and more, got caught in the blast radius.

  • Given growing pushback against the hyperscalers, while this probably won’t change anything, it certainly won’t help their case with the sceptics.

Read about the outage here.

Together AI to Deploy 2GW of European Capacity

Together AI is bringing 100K GPUs and 2GW of AI infra capacity to Europe with Hypertec and 5C by 2029.

Bringing 100,000 GPUs to Europe

www.together.ai/blog/together-ai-expands-in-europe

Together’s rollout starts late 2025, spanning France, the UK, Italy, and Portugal. The stack includes NVIDIA Blackwell and next-gen chips, powering training, inference, and fine-tuning of 200+ open-source models. A new EMEA team, anchored in London and Amsterdam, will handle compliance, dev support, and infra integration. The play? Meet Europe’s AI needs with fast, sovereign, sustainable deployments.

Why it matters:

  • Sovereign demand is going to be the regional driver.

  • Together’s inference first approach brings a different approach.

  • Europe loves local infra from a compliance and data control perspective, and localised strategy is increasingly the way to go.

Read about the full rollout plan here.

The Rundown

Europe just arrived.

This week wasn’t about vague future ambitions. It was real capacity, real hardware, and real investment. Together AI’s mega buildout. Nebius’ UK launch (plus G42 and Crusoe), all committing billions to continental infrastructure.

And AMD?

It didn’t just show up - it delivered a roadmap.

Its open AI stack is polished, complete, and built for production. Helios brings serious rack-scale performance, memory, and bandwidth. This means the market may yet have a credible alternative.

Nvidia still dominates 90%+ of the market.

But for the first time in years, that dominance looks…challengeable.

The next 12 months will test whether openness, flexibility, and speed-to-deploy can beat raw scale and lock-in.

Europe’s building fast. AMD’s betting bold. NVIDIA isn’t sleeping.

But the board is shifting.

And the pieces are moving.

See you next week.

p.s. I’m working on some new subscriber onboarding flows, and lifting the lid on how I’m running The GPU.

To that end, I have three things for you:

First, here’s my Spotify Hot List.

Because I am, in fact, a real human. This is what I’m listening to during long research and editing sessions. A wild mix of metal, electronica…and salsa.

Second, I’ve created a private community on LinkedIn, only for subscribers.

For your invite, complete the survey below so I can give you more of what you want, and less of what you don’t.

It’s only eight questions long, and it’ll take you less than a minute, so if you get a minute (see what I did there?), click below.

Complete Survey

Once you complete it, you’ll automatically receive an email with an invite.

Third, I’ve finally figured out a referral program.

Share your personal link below. Refer one person. Get a free month of premium. Simple.

Reply

or to participate