Nvidia just beat out 190 sovereign states.

Only five countries have a nominal GDP higher than Team Green’s market capitalisation: the US, China, Germany, India, and Japan. Considering that the worst AI will ever be is right now, Jensen may yet be on the podium.

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 #51

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.

Nvidia Flirts with $4 Trillion

Nvidia just became the first company to cross a $4T market cap, and the most valuable public company in history.

That’s an astronomical number. Larger than both Apple and Microsoft. And there’s probably still room for Team Green to run.

Why?

Because AI is currently the worst it will ever be.

If you believe it improves from here, in capability, adoption, or commercial application, then Nvidia’s growth is only heading in one direction. Every new AI use case, from chips in hospitals to voice agents for lawyers, adds demand. And that’s before you even start to think about robotics.

Why this matters:

  • $4T looks unsustainable until you remember how early we are.

  • Even if AI workloads shift to alternative hardware, Nvidia still owns the developer stack.

  • The pie is likely to get much bigger, and even if Nvidia loses market share, the size of their slice will probably still be unprecedented - never mind Jensen’s net worth.

G42-Led Group Proposes $2B AI DC in Vietnam

Vietnam may soon become home to Southeast Asia’s most ambitious AI data centre.

A consortium led by G42, FPT, VinaCapital, and Viet Thai Investment Group has submitted plans for a $2 billion hyperscale AI campus in Ho Chi Minh City. GDP growth, tech upskilling, and foreign capital are all on the table. But with investors flagging Vietnam’s restrictive data localisation rules and monitoring requirements as risks, the funds may yet flow elsewhere. The group is now requesting “special mechanisms” from the Prime Minister, mirroring Singapore’s digital economy policies.

Why this matters:

  • G42’s ambitions are stretching well beyond the Middle East and Europe.

  • This would be its first major play in Southeast Asia, and could make Vietnam a serious contender in Asia’s AI infrastructure race.

  • With $2B on the table, governments may need to modernise data rules or risk losing a seat at the table in this buildout for the ages.

CoreWeave is Buying Core Scientific

CoreWeave is buying Core Scientific in an all-stock deal worth $9B.

The acquisition locks in 500MW of power across Core Scientific’s US campuses, plus 200MW more in the pipeline. It’s a major vertical integration move, linking CoreWeave’s AI platform with physical sites already optimised for AI workloads. The deal also gives Core Scientific a path out of bankruptcy, repurposing its scale from crypto mining to AI infrastructure.

Why this matters:

  • In a commodity market like compute, you need every advantage you can get.

  • Vertical integration, or moving “down the stack”, leads to massive cost savings, better margins, and competitive advantage.

  • With this deal, CoreWeave gains direct control of the most valuable resource in the AI infrastructure race: power.

Cerebras Brings Ultra-Fast Inference to Agentic AI

Hugging Face, DataRobot, and Docker just integrated with Cerebras Systems.

Hugging Face’s SmolAgents now run on Cerebras inference infrastructure, offering near-instant interaction on Hugging Face Spaces. DataRobot’s new “syftr” framework is powered by Cerebras, too, enabling enterprise-grade agent workflows out of the box. And with Docker integration, developers can now deploy full-stack, multi-agent systems using a single Compose file.

Why this matters:

  • Cerebras’ inference stack is screaming fast and increasingly accessible.

  • Higher token throughput is needed for real-time AI use-cases, and inference speed is rapidly becoming the competitive differentiator, not just training throughput.

  • These integrations give developers from across the AI stack access to the hardware they need to bring those use cases from theory to deployment.

AMD’s MI600 Chip Quietly Confirmed

The MI600 is real, even though we didn’t hear it from AMD.

The director of Germany’s HLRS supercomputing centre casually confirmed the existence of AMD’s next-gen MI600 accelerator in a press Q&A. Despite AMD having made no formal announcement. HLRS plans to use MI600s in its future Jupiter system, where they’ll be paired with SiPearl’s Rhea CPUs. Specs remain under wraps, but the MI600 will likely replace the MI300A and feature a denser chiplet layout and architectural refinements to boost inference performance.

Why this matters:

  • While Nvidia dominates the GPU market as a whole, compute for science-specific applications is an AMD stronghold.

  • AMD skipped the fanfare, likely to avoid cannibalising MI300 demand too early (or to avoid a repeat of Jensen’s Chief Revenue Destroyer antics at GTC).

  • With the rack-scale MI450 system on the horizon, growing hyperscale support, and rapidly improving software capabilities, the MI600 is yet another arrow in AMD’s quiver for a shot at meaningful inference market share.

Google’s New AI Cable Will Land in Bermuda

Bermuda is officially part of the AI infrastructure map.

Google has announced SOL, a new private transatlantic subsea cable connecting South Carolina, Bermuda, Portugal, and Spain. It’s the first hyperscaler cable to land on the island, and part of Google’s growing global fibre network powering its AI and cloud services. SOL uses space-division multiplexing (SDM) to deliver higher capacity and lower latency, with expected completion in 2026. Once operational, it will support the growth of AI workloads and cross-continental compute flows.

Why this matters:

  • Subsea cables are the arteries of global AI.

  • Owning or landing one gives you geopolitical and economic leverage.

  • Bermuda just went from tourist destination to fibre landing point, opening doors for AI-hosting, low-latency routing, and digital economy growth.

Firmus to Cool AI Infra with Seawater

Firmus just signed an MoU with Singapore’s Port Authority to deploy seawater-cooled data centres.

The agreement will enable Firmus to utilise the port’s deep-water access to cool high-density AI compute workloads, specifically those running on NVIDIA H100 and GB200 platforms. The project targets 20MW of AI compute with a PUE under 1.1. Seawater cooling isn’t new in theory, but it has rarely been deployed at this scale. Singapore’s land and power constraints have prompted operators to reassess their fundamentals, and water-based thermal transfer presents a significant efficiency upside.

Why this matters:

  • PUE under 1.1 is elite-level efficiency, especially for H100 and GB200-class systems.

  • Better PUE means lower cooling costs, improved environmental impact, and healthier margins.

  • They say necessity is the mother of invention, and Asia is rapidly becoming a proving ground for alternative data centre architectures as a result.

The Rundown

AI infrastructure isn’t just compute.

It’s energy, heat, capital, and bandwidth. All multiplied at scale.

And each one is either a cost.

Or a moat.

That’s why CoreWeave bought power, Firmus is building on seawater, and Google is routing fibre through Bermuda.

The infrastructure is being shaped by physical incentives, not just AI models.

And it’s dragging every industry along for the ride.

Whether they’re ready or not.

See you next week.

p.s. I’ve put together a survey for existing subscribers so I can make sure I’m giving you exactly what you want to see.

It’ll take you less than a minute to complete (and it’ll net you an invite to The GPU LinkedIn group).

Reply

or to participate