Some truths feel like gravity. Unquestioned. Unshakeable.

That enterprises would always default to the big three. That public cloud was the final form. That Europe would grumble, regulate, but stay put.

Not anymore.

Private conversations are making their way into the public forum.

At KubeCon EU, vendors said European companies are actively fleeing US hyperscalers. Tariffs. Espionage fears. Regulatory overreach.

And Microsoft has responded.

With a €4.8B commitment.

But, as ever, that’s not all that’s going on in the market.

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:

• Nscale Goes Big, Fast

• Meta Finally Becomes a Cloud

• Nebius Joins Forces With DDN

• Qwen3 vs. Regulatory Capture

• Intel’s Foundry Gamble

• Amazon Launches Kuiper Satellites

• Microsoft Commits €4.8B to Europe

Let’s get into it.

The GPU Audio Companion Issue #38

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.

Nscale Goes Big, Fast

Nscale is 11 months old.

It’s now seeking $2.7 billion in capital. $1.8B in private credit. $900M in equity.

The goal?

Build out AI data centres across the globe.

Supplying ByteDance with Nvidia GB200s at scale and hosting them in Norway is a big part of this vision. However, ByteDance actively denies the rumoured 27,000 GPU scale. With Goldman Sachs already lined up to lead the debt, this could be a big one.

Why this matters:

• ByteDance can’t buy the chips because of the chip restrictions. But it can rent them.

• While the geopolitical risk is real, so is the demand, and a contract with ByteDance isn’t one many would turn down.

• This model echoes that of CoreWeave: fast capital, chip-as-collateral, hyperscaler-style ambition. The race to build a European counterweight continues.

Read the full story here.

Meta Finally Becomes a Cloud

Meta can finally be considered a cloud.

The pseudo-hyperscaler just launched hosted APIs for Llama 3, turning its models into a fully managed service. Developers can now plug into Meta Cloud directly, a late but decisive move into inference-as-a-service. The platform includes integration tools, safety features, and partnerships with AWS, Azure, and Oracle for infra.

Why this matters:

• The cloud incumbents just picked up a not-insignificant new competitor, as did the inference providers.

• Meta is moving up the stack, offering infra, not just open weights.

• This move also gives it a monetisation path outside advertising…or a novel way to serve yet more ads. And if recent comments out of the Perplexity camp are anything to go by, the hypertargeted ads are a case of when, not if.

Read the full story here.

Nebius Joins Forces With DDN

DDN and Nebius are collaborating to power next-gen enterprise AI infrastructure.

The idea? Seamless orchestration across compute, storage, and model deployment, optimised for sovereignty and hybrid deployments. The collaboration targets mid-sized enterprises that don’t want the full Azure/AWS lock-in but still need something enterprise-ready. Nebius brings the cloud-native platform. DDN brings storage muscle.

Why this matters:

• Enterprise AI infra is still a fragmented mess, and Nebius wants to fix that.

• DDN’s hardware credibility could be a differentiator as workloads scale.

• Nebius is increasingly positioning itself as a full-stack neocloud alternative.

Read the full story here.

Qwen3 vs. Regulatory Capture

Alibaba Cloud has launched Qwen 3, its new open-weight model suite.

Qwen 3.5, the flagship variant, outperforms Mixtral, LLaMA 2, and even Claude 2.1 on MT-Bench and MMLU. The full release includes models from 0.5B up to 72B parameters. And all trained with context windows up to 128K.

But the real story might be the meta one.

Anthropic recently published a national policy position paper calling for export controls on open weights. It’s part of a growing trend where the largest model providers are actively lobbying the US government to regulate their smaller competitors under the banner of national security. Which sounds an awful lot like regulatory capture masquerading as a “free world vs. the rest” story.

Why this matters:

• Alibaba’s Qwen3 is yet another indicator of declining Western dominance in the open-weight arena.

• Regulatory capture is bad for everyone.

• Calling for state control of distribution (and indirect state support as a second-order consequence) under the guise of national interest is a slightly ironic line of argumentation to take.

Read the full story here.

Intel’s Foundry Gamble

Intel’s foundry business is rebranding, restructuring, and rebooting trust.

This week, it rolled out a broader product portfolio, new ecosystem partnerships, and updated timelines for its five-process-node-in-four-years plan. The effort includes formalising foundry services for chiplets, packaging, and advanced nodes, and expanding design support for both x86 and Arm. Intel also announced new customers for its IFS (Intel Foundry Services) arm and hinted at an expanded ecosystem with cloud, EDA, and packaging partners onboard.

Why this matters:

• Foundry customers don’t forget. Intel lost credibility by failing to deliver consistently.

• Winning back trust means showing volume, not just PowerPoint.

• With US–China tensions still high, Intel could become the go-to foundry for Western firms seeking supply chain insulation.

Read the full story here.

Amazon Launches Kuiper Satellites

Amazon just launched the first 27 satellites for Project Kuiper, its answer to Starlink.

The launch, via ULA’s Atlas V rocket, marks the beginning of a planned 3,236-satellite constellation designed to provide global broadband internet. And compete directly with Starlink. The goal: deliver connectivity to underserved regions, support AWS edge workloads, and control the orbital layer of AI deployment.

Why this matters:

• Just as AWS controls its own silicon, Amazon now wants to own the delivery layer for internet access.

• Ground infrastructure is already in progress, and customer terminals have been demoed with peak speeds of up to 1 Gbps.

• With those speeds, Amazon can serve cloud + AI workloads in remote areas. Something even AWS regions can’t cover.

Read the full story here.

Microsoft Commits €4.8B to Europe

Microsoft has announced €4.8 billion in AI infrastructure investment across Europe.

This new push for dominance includes sovereign cloud regions in Spain and Germany, and plans for 600,000 GPUs by 2026. It’s also part of a broader strategy that includes training programmes, responsible AI frameworks, and new regional HQs in London and Brussels.

But there’s an elephant in the data centre:

While Microsoft wants Europe, Europe is quietly drifting away.

At KubeCon EU, vendors like Civo and Nextcloud reported a wave of clients seeking alternatives to US hyperscalers. Concerns include tariffs, data residency, and espionage risk. Even Vultr confirmed it’s seeing rising demand for sovereign deployments.

Why this matters:

• We’re entering an era where true data sovereignty becomes a procurement requirement, not a niche feature.

• Microsoft is pre-emptively embedding itself into European AI infrastructure before regulators force their hand.

• This is the most explicit hyperscaler commitment to European AI so far, and it probably won’t be the last.

Read the full story here.

The Rundown

A year ago, Nscale didn’t exist. Not publicly, at least.

Today, it’s pitching $2.7B to host Nvidia’s best chips for ByteDance. Meta is now a cloud. Alibaba’s open-weight stack is beating GPT-3.5. And Anthropic wants Uncle Sam to build more walls.

Meanwhile, Nebius and DDN are building sovereign infra, Intel is trying to win back the foundry crowd, Amazon is going orbital, and Microsoft is doubling down on Europe… just as Europe looks for a way out.

Everything’s in motion.

See you next week.

Keep The GPU Sharp and Independent

Good analysis isn’t free. And bad analysis? That’s usually paid for.

I want The GPU to stay sharp, independent, and free from corporate fluff. That means digging deeper, asking harder questions, and breaking down the world of GPU compute without a filter.

If you’ve found value in The GPU, consider upgrading to a paid subscription or supporting it below:

☕ Buy Me A Coffee → https://buymeacoffee.com/bbaldieri

It helps keep this newsletter unfiltered and worth reading.