Three blocs. Three ideologies. One outcome.

In the past 14 days, the US, the UK, and China have released their visions for the future of AI.

On the surface, this looks like just another policy debate.

But it’s far more.

The geopolitical heavyweights are in a battle to set the rules for the most important technology of the century. And while Washington, Brussels, and Beijing are positioning themselves as the fighters, the actual fighting will be done in Jakarta, Lagos, São Paulo, and beyond.

Why?

Because the Global South is the real prize.

To the victor goes a path out of stagnation and a chance to write the rules for the next century’s most powerful technology. And the loser? A slow bleed of relevance. A loss of control. Possibly a progressive decline into global side character status.

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 #55

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.

The UK, China Release Visions for AI’s Future

Two major announcements this week signal just how deep the global AI governance strategy divide really is.

The UK AI Security Institute, in partnership with its Canadian counterpart, Amazon, Anthropic, and civil society, launched The Alignment Project, broadly aligning with the EU’s risk-based regulatory approach. The purpose? Bring together like-minded partners to “tackle issues around safety, security, and human control to protect citizens.” How? “Fund cross-disciplinary research into AI control and oversight, in combination with cloud credits, grants, and venture backing.” Why? “Position the UK as a convening hub for pro-safety, pro-commercial alignment work.”

Meanwhile, China unveiled an action plan for global AI governance and proposed a new international AI cooperation organisation, headquartered in Shanghai, at the World AI Conference. Think IAEA. Premier Li Qiang framed AI as a “public good” and pledged to support Global South countries through shared infrastructure, open-source tools, and technical exchange. Why the Global South? Because the US hyperscalers haven’t yet monopolised those markets in all but name. Given the rising tide of distrust, the opportunity is there. And now is the time.

Why this matters:

  • The UK’s proposal is a valiant idea in principle. Safety, control, and social stability are the top-of-mind concerns that everyone is worried about. But there’s only £15m (~$20m) available. And when Meta is making $1b offers, that’s not even table stakes. It’s embarrassing.

  • China is continuing to play the infrastructure diplomacy game. Considering how successful this strategy has been with the Belt and Road Initiative, aspects of the model are proven. The question will likely be whether partners share in China’s vision for the future.

  • Considering the relative geopolitical void the Trump administration has left in the wake of recent actions, and subsequent decline in the global perception of the US (falling below China), it’s all to play for.

Download China’s Action Plan below:

China Action Plan on Global Governance of AI.pdf

China Action Plan on Global Governance of AI.pdf

Full Text for China's Action Plan for Governance of Artificial Intelligence

154.59 KBPDF File

Groq, Anthropic Target New Funding Rounds

Two challengers are closing in on massive funding rounds.

Groq is finalising a $600 million round led by Disruptive, doubling its valuation to $6B. Despite trimming its 2025 revenue guidance, insiders say the company expects to catch up in 2026. The funding extends Groq’s war chest to over $2B as it scales a global inference network and builds out data centre capacity in Europe.

At the same time, Anthropic is close to raising up to $5B in new funding at a $170B valuation. That’s up from $61.5B earlier this year. Iconiq is leading the round with a $1B check, with participation expected from Amazon, GIC, and the Qatar Investment Authority. Annual recurring revenue hit $5B in July and could climb to $9B by year-end.

Why this matters:

  • Capital is consolidating around inference challengers with real-world traction, and Groq is one of the breakaways.

  • Anthropic’s valuation tripled in under six months, faster than OpenAI’s last jump, and now it’s chasing infrastructure at a similar scale.

  • Valuations keep soaring across the sector, and with NASDAQ indicating there’s no slowdown in sight, these companies likely have much further left to run.

Nscale Brings OpenAI’s Stargate to Europe

Europe just got its first OpenAI-backed data centre.

Nscale, OpenAI, and Aker have unveiled Stargate Norway. The 100% renewable 230MW AI gigafactory in Narvik, Northern Norway, is targeting 100,000 NVIDIA GPUs by the end of 2026. The project is located in a hydropower-rich region with low local demand and sub-market electricity prices, and has the capacity to scale to 520MW. and serve as the first European site under OpenAI’s “OpenAI for Countries” programme. The first 20MW phase is backed by $1B in capital, with $250M in equity split evenly between Aker and Nscale.

Why this matters:

  • Landing a customer of this size for a project of this scale immediately catapults Nscale into the big leagues.

  • This is OpenAI’s first direct infrastructure venture in Europe, offering proximity to customers and access to sovereign contracts.

  • A name like OpenAI on the Nscale’s customer list is massive for overall credibility. Expect more big names to follow.

Google Signs EU’s AI Code, Voices Concerns

Google will sign the EU’s General Purpose AI Code of Practice, but isn’t thrilled about it.

The Code is the European Commission’s attempt to pre-wire compliance ahead of the AI Act’s enforcement. It offers a “voluntary” set of commitments on transparency, safety, and copyright, designed to shape model behaviour and pre-empt regulation. Google joins OpenAI and Anthropic in signing, but warned the Code could “chill” AI development if trade secrets and approval timelines aren’t handled properly. Meta, notably, refused to sign.

Why this matters:

  • Google’s signing signals support, but the wording also leaves future wiggle room.

  • The timing, just days before Article 51 enforcement begins on August 2, is deliberate. Better to show willingness before developers of “systemic risk” models are forced to toe the line or face penalties.

  • As such, Google is likely hedging: align today, garner some political goodwill at Meta’s expense, and litigate tomorrow if the final rules have the wrong kind of impact.

FuriosaAI Raises $125M, Cadence Gets Fined

Alternative semiconductors are having a moment.

Korea’s FuriosaAI just banked a $125M Series C bridge to ramp up production of its RNGD inference chip and design a next-gen processor for agentic AI. The raise, with participation from Korea Development Bank, Industrial Bank of Korea, Keistone Partners, PI Partners, and Kakao Investment, follows a 2.25x performance-per-watt win over GPUs in LG AI Research’s EXAONE workloads.

At the other end of the spectrum, Cadence Design Systems has been hit with a fine.

Why? They pleaded guilty to violating US export controls by selling chip design software to Chinese military-linked institutions. Including one suspected of simulating nuclear explosions. Which is obviously a totally normal workload to help customers run. And while the Trump administration recently walked back NVIDIA’s H20 ban, the $140M fine shows that supply chain compliance is no joke.

Why this matters:

  • FuriosaAI landed at LG last week and announced the raise this week. That’s some serious traction for a relatively unknown NVIDIA challenger.

  • The participation of so many local investors in this round, plus buy-in from a domestic juggernaut, could indicate that Korea is positioning Furiosa as a sovereign compute player.

  • Cadence’s guilty plea shows how fragile the tooling layer is. It’s not just chips under scrutiny. Every line of code that touches silicon is now geopolitically sensitive, and while export restrictions with China are in a state of relative flux, there are still serious penalties for falling foul of the rules.

Microsoft and OpenAI Renegotiate AGI Access

The deepest partnership in AI is now under renegotiation.

Microsoft and OpenAI are in talks to redefine their relationship, specifically around AGI development, model access rights, and intellectual property. The trigger? OpenAI’s commercial direction is diverging from Microsoft’s enterprise focus. The o3 model family, still unreleased to Azure customers, has pushed Microsoft to reassert its strategic priorities. The outcome of these talks could reshape who gets access to the most powerful AI systems, and under what terms.

Why this matters:

  • Microsoft has invested over $13 billion in OpenAI, but still doesn’t have access to GPT-o3.

  • Not having access to the latest frontier models undermines Azure’s AI lead and raises questions from enterprise customers, and likely shareholders, given the scale of the investment.

  • If OpenAI achieves AGI, the partnership is contractually over. The scale of the investment, the apparent progress in model capability, the restructuring, and the fact that OpenAI is increasingly acting like a standalone tech company and cutting deals with Google mean it’s probably about time for a relationship reset.

Superconducting Grid to Power AI DCs in Korea

KEPCO and LS Group will build the world’s first superconducting transmission grid, wired directly into a hyperscale AI campus.

By 2028, Korea Electric Power Corporation (KEPCO), LS Cable, and LS Electric plan to deliver a live 100MW superconducting grid. The cable design ditches copper for high-temperature superconductors cooled by liquid nitrogen, shrinking land footprint, cutting power loss, and enabling ultra-high throughput. The site? Gapyeong, Gyeonggi Province. The goal? 100x efficiency, near-zero loss, and enough headroom to handle surging AI workloads.

Why this matters:

  • Smaller substations. Higher density. Lower resistance. Superconductors mean a fundamentally different capex/opex profile for AI data centres.

  • Korea is moving into grid advantage, building specialised, low-loss delivery pipelines for AI-specific workloads.

  • In a decade where power is the bottleneck, success here could become the model for AI-driven grid upgrades the world over.

The Rundown

Money still matters.

But in AI, it’s not the differentiator. You can’t just outspend the competition. You have to outbuild them.

That’s what this week shows.

Nscale’s OpenAI-backed site in Norway. Korea’s superconducting AI grid. China’s attempt to bundle governance with infrastructure.

Different strategies, same goal:

Distribution.

Not just of compute, but of influence. Power. Rules. And ideology.

Because once you win ideological buy-in, the rest follows.

That’s the overarching play here.

Not hardware with a generous serving of hype.

Just control. By way of belief. Wrapped in infrastructure.

See you next week.

Reply

or to participate