• The GPU
  • Posts
  • Issue #34: Llama 4 Issues, A Neocloud's $45m Raise, & TSMC's $1B Fine

Issue #34: Llama 4 Issues, A Neocloud's $45m Raise, & TSMC's $1B Fine

Feat. Meta, Nvidia, NexGen Cloud, Google, Oracle, and TSMC

Author

Ben Baldieri
April 12, 2025

It’s official. Llama 4 is out.

Meta says it’s their best model yet. The open-source crowd is (semi) thrilled. But there’s a catch (isn’t there always?): they might have trained it with data they weren’t supposed to touch.

I’m shocked. Shocked, I tell you.

Well…not that shocked.

Meanwhile, Trump’s team walk back their planned H20 crackdown after a cosy chat at Mar-a-Lago, and NexGen Cloud raises $45M for sovereign AI infra in Europe.

Plus: Google commits to data centre capex, expands its multicloud pact with Oracle, and TSMC might be staring down a billion-dollar US fine.

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.

The GPU Audio Companion Issue #34

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.

Meta Drops Llama 4… And Gets Caught In Controversy

Meta just released the Llama 4 family of open-weight large language models.

The new herd gives a strong showing on benchmarks like LMArena and tough-to-crack academic tasks. Multimodal support is in, performance looks solid, and the "Maverick" chat-tuned variant was already second only to Gemini 2.5 on LMArena.

But there’s a catch.

The model Meta submitted to LMArena, LLAMA-4-Maverick-03-26-Experimental, wasn’t the same one it released publicly. It was a tuned, private build. Optimised for stylistic appeal, verbosity, and even emoji use.

After widespread confusion and backlash, LMArena updated its leaderboard policies, confirming the model wasn’t representative of what devs could actually deploy.

Meta didn’t deny it

In fact, they called it an “experimental chat-optimised version” designed for that kind of contest. Users were less forgiving, reporting a big quality gap between what was promised and what shipped.

Why this matters:

  • Meta’s top benchmark placement came from a custom model that’s not actually available, which raises questions as to how good the massively delayed Llama 4 family actually is.

  • This kind of behaviour, especially from a supposed leader, erodes trust in crowd-ranked benchmarks and raises questions about benchmark gaming more broadly.

  • User pushback was enough for Meta’s own GenAI boss to step in and (try to) blame inconsistent inference platform support, which doesn’t land particularly well in a post-DeepSeek world. Especially when engineers are allegedly resigning over internal actions.

A pause before we resume our usual programming. I’d love to understand more about you - my audience!

If you could answer the question below, that would really help me out.

What best describes your role?

Login or Subscribe to participate in polls.

Now, back to the news.

Trump Backs Off H20 Crackdown

In a sudden reversal, the Trump administration has backed off its planned crackdown on Nvidia’s H20 AI chips.

The H20 is the most advanced processor legally allowed for export to China under current US restrictions. The shift reportedly followed a dinner between Nvidia CEO Jensen Huang and Trump at Mar-a-Lago. The administration had been preparing to restrict exports of the chips as recently as this week. But after the meeting, and Nvidia’s promise to ramp up US investment in AI data centres, the plan was dropped.

Why this matters:

  • The H20 has been a hot seller for Nvidia in China, with ByteDance, Alibaba, Tencent, and DeepSeek placing over $16 billion in orders just in Q1.

  • Nvidia’s diplomatic playbook appears to work - corporate promises now directly influence national trade policy.

  • With the rest of the US tech firms navigating a volatile mix of politics, tariffs, and geopolitics, Nvidia likely just dodged a major hit.

NexGen Cloud Raises $45M For Sovereign AI Infrastructure

Big news from Europe - NexGen Cloud just secured $45 million in Series A funding.

The round was led by Adit Ventures, and supports NGCs plans to deploy GPUs in new facilities focused on data sovereignty and enterprise compliance. It’s a strong move from a challenger neocloud looking to carve out ground in the post-GDPR AI infrastructure landscape.

Why this matters:

  • NexGen’s strategy mirrors growing demand for Europe-first AI infrastructure, especially in regulated sectors.

  • This brings them closer to competition with Fluidstack, Sesterce, and Nebius for European AI infra dominance.

  • Sovereignty is no longer a buzzword - it’s becoming a capital-backed competitive advantage.

Google Pushes Agent2Agent Interoperability

In software land, it’s been a busy week.

Announcing the Agent2Agent Protocol (A2A)- Google Developers Blog

Explore A2A, Google's new open protocol empowering developers to build interoperable AI solutions.

developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability

Alongside Llama 4, Google announced A2A, a new initiative to improve agent interoperability. The protocol aims to let AI agents talk to each other, making autonomous workflows more collaborative and standardised. It’s part of Google’s bigger push to dominate not just models, but also the infrastructure AI agents rely on.

Why this matters:

  • Agents are where much of the overall industry thinks the value creation is going to happen for enterprise, which means ROI on the ever increasingly massive CapEx.

  • Google’s A2A protocol could become the TCP/IP layer for agents, if adoption sticks.

  • Orchestration and coordination across agents, not just model performance, is rapidly becoming the next battleground.

TSMC Faces $1B+ Fine Over Huawei Linked Chips

A fresh headache for the world’s top foundry.

TSMC is under investigation by the U.S. Commerce Department and could face fines exceeding $1 billion over alleged violations of export controls. The case centres on AI chips produced for China-based Sophgo, which match the design of Huawei’s Ascend 910B.

That’s a chip the U.S. deems off-limits without a licence.

Nearly 3 million of these chips may have ended up with Huawei, despite restrictions barring U.S.-tech-based chipmaking for sanctioned firms.

TSMC denies supplying Huawei since 2020 and says it’s cooperating fully with regulators. But with Trump eyeing new 32% tariffs on Taiwan and the CHIPS Act money still flowing, the timing couldn’t be worse.

Why this matters:

  • A billion-dollar penalty would be the largest ever for an export control breach in semiconductors.

  • Sophgo’s role raises more questions about proxy supply chains skirting U.S. chip bans.

  • With TSMC investing $100B into U.S. fabs, the fallout could complicate its strategic positioning and Washington’s reliance on it.

Google & Oracle Deepen Multicloud Partnership

Two giants, one increasingly entangled stack.

Oracle and Google are expanding their multicloud partnership, adding 11 new global regions for their Oracle Database@Google Cloud service. That includes new coverage across Europe, North America, Asia, and Australia. Alongside this rollout, they’re launching the Oracle Base Database Service, offering lifecycle automation, Exadata support, and pay-as-you-go pricing. Even government workloads now have a home on this stack.

Why this matters:

  • Google and Oracle are quietly building one of the most credible multicloud stories in the market, and doing it on each other’s metal.

  • Oracle gets the reach and developer mindshare of Google; Google gets sticky enterprise workloads that aren’t going anywhere.

  • There’s also a new partner reseller program planned, which moves beyond mere strategic alliance and into joint go-to-market engine territory.

Alphabet Sticks To $75B Data Centre Spend

Alphabet isn’t blinking: $75B is still going into AI infrastructure.

Despite tariff chaos and rising build costs, Alphabet is sticking with its $75 billion capex plan for 2025. That figure includes data centres, AI chips, and cloud infrastructure to support products like Search and Gemini. Google Cloud execs said demand from enterprise clients is too strong to justify slowing down, even with such massive global uncertainty.

Why this matters:

  • Alphabet’s aggressive spend is a signal to investors: AI infrastructure is still a high-conviction bet, even in uncertain markets in the face of market uncertainty.

  • With $75B locked in, Google is doubling down on vertical integration of chips, software, and hyperscale delivery.

  • This keeps pressure on AWS, Microsoft, and Meta, who are now chasing similar AI ROI at even higher capex levels.

The Rundown

This week’s plotline:

Two tech giants flex on the open model front. Sovereign AI gets a funding boost. TSMC’s probe. Trump’s chip diplomacy. Meta’s benchmark gaming. New strategies and commitments from hyperscalers.

Big moves aplenty.

And all of this against a backdrop of global trade turmoil and the shredding of long-held assumptions about international norms.

Where are we heading is no longer the only question.

What state will we be in when we get there, and what will it cost, is becoming just as important.

The answer?

Join me next week to find out.

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.

Ben Baldieri

Independent insights on GPU compute, AI infrastructure, and data centres. No fluff. Just sharp analysis on the trends shaping AI and digital infrastructure.

buymeacoffee.com/bbaldieri

Reply

or to participate.