The US government giveth, the US government taketh away.

On one hand, Nvidia is back in the Chinese market after a Trump intervention, clearing $4.5B of stranded H20 inventory for sale. On the other, a high-stakes GPU deal with the UAE just hit pause because Washington isn’t sure G42 should get a slice.

One decision breathes life into Nvidia’s Asia strategy.

Another throws a wrench in America’s most ambitious AI export plan.

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 #52

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 Wins China Reprieve After Trump Meeting

Nvidia just got the green light to resume H20 sales to China.

After a $4.5B write-down tied to unsold China inventory, Jensen Huang met with Donald Trump last week to argue for licensing exemptions. It worked. The US will now grant export licenses for the H20, Nvidia’s previously restricted, lower-performance AI chip designed for China. Deliveries begin soon. Nvidia has also developed another export-compliant model, further de-risking its position in the world’s second-largest AI market.

Why this matters:

  • China is a massive market for Nvidia, and losing access would have serious long-term consequences for the shape of the global AI market and US dominance.

  • Even with $4.5B in stranded inventory, Nvidia’s last earnings report was eye-wateringly positive.

  • With this approval, the doors are now back open, so a substantial balance sheet impact is likely in the next quarter.

Read the full story in The Wall Street Journal here.

UAE-Nvidia Chip Deal Hits a Wall

The multi-billion-dollar AI chip deal between the US and UAE has stalled.

After Trump’s May visit, the agreement to supply hundreds of thousands of Nvidia GPUs, earmarked for sovereign compute buildouts, looked like a done deal. But behind the scenes, US officials are now debating whether G42 should be excluded from the arrangement. The reason? National security concerns. The chips were slated to support large-scale data centres operated by US firms in the Emirates, including Microsoft and OpenAI, with G42 expected to receive ~20% of the supply. That share is now in question.

Why this matters:

  • US officials are concerned that China could gain access to Nvidia’s most advanced hardware as a result of this deal.

  • At the same time, White House AI Czar David Sacks is concerned that if the US fails to meet market demand in the UAE, China will.

  • Delays will likely complicate timelines for the Stargate Campus, OpenAI’s potential regional dominance, and the broader rollout of AI infrastructure across the UAE.

Read the full story in the Wall Street Journal here.

TensorWave Deploys 8,192 AMD MI325X GPUs

TensorWave has fired up North America’s largest AMD training cluster.

They’ve brought a whopping 8,192 MI325X GPUs online in a custom, liquid-cooled setup in Arizona. Each GPU is paired with 256GB of HBM3e, and the system delivers an estimated 21 exaFLOPS of FP8 throughput. The racks are a dense stack of OCP Yosemite v3 trays, cooled via direct-to-chip cold plates.

And this isn’t some proof-of-concept.

TensorWave says the cluster is fully operational and available now, with CEO David Marvin calling it “phase one.”

Why this matters:

  • AMD’s ROCm has finally matured enough to power real workloads.

  • With a cluster this size, TensorWave is proving there’s life outside the green ecosystem.

  • Liquid-cooled MI325Xs at this scale are no joke, and will likely open the door to similar deployments of the MI350X later this year.

Read the full story on Tom’s Hardware here.

Claude Gets Suited for Wall Street

Anthropic is going after financial services.

The AI-safety-focused lab has launched a version of Claude fine-tuned for the money men. The new offering includes tools tailored for banks, insurers, and asset managers, covering risk modelling, client communication, document summarisation, compliance support, and code generation for quant teams. It runs on Claude 3.5 Sonnet and can operate via API, secure chat, or within private cloud deployments.

Why this matters:

  • OpenAI and Google are ahead in enterprise deals, so it looks like Claude for Finance is a bid to change that.

  • Financial firms are lucrative but demanding clients; reliability, privacy, and compliance are critical.

  • This is the OpenAI challenger’s clearest move yet into enterprise verticals, with Anthropic pushing Claude beyond a mere chatbot and towards a multi-agent analyst layer.

Read the full announcement from Anthorpic here.

AWS Is Building Its Own Liquid Cooling

Amazon Web Services is done waiting around for cooling vendors to catch up.

In true AWS fashion, it took the “if you want something done right, do it yourself” route, designing, prototyping, and deploying a fully custom liquid cooling system in just 11 months. Rather than go off-the-shelf, AWS engineers built a direct-to-chip, closed-loop system that routes hot-tub-temp fluid across cold plates to cool next-gen AI chips without adding to water consumption. The internal coolant distribution unit is also homegrown, because (naturally) AWS decided everyone else’s was inefficient and overpriced.

Why this matters:

  • Cooling is core to performance, cost, density, and margins at scale.

  • Custom solutions like this give AWS even more control over its hardware ecosystem.

  • Given the concurrent push into custom silicon, expect more custom solutions in the future.

Read the full announcement from Amazon here.

Google Wins $200M Pentagon AI Cloud Contract

The US DoD just handed a $200 million contract to Google to support advanced AI infrastructure and applications for national security.

The contract, under Google Public Sector, means the DoD now has full access to Google’s US-based cloud infrastructure, including TPUs and air-gapped Distributed Cloud (GDC) systems. The Pentagon says it’s part of a broader effort to support “agentic AI workloads” across warfighting, intelligence, and enterprise domains. OpenAI, Anthropic, and xAI also won similar contracts, and will contribute their own secure AI models and offerings like Claude Gov and Grok for Government.

Why this matters:

  • Google has already worked with the US Navy, Air Force, and Defense Innovation Unit.

  • In addition to this deal, Google Distributed Cloud (GDC), and the air-gapped version, recently received DoD Impact Level 6 authorisation.

  • With TPU access as part of the deal, GPUs are increasingly no longer the only hardware game in town.

Read the full story in DCD here.

Meta is Deploying “Titan-Scale” AI Clusters

Meta is redefining data centre campus scale in the US.

The company plans to invest hundreds of billions of dollars into AI infrastructure, with several multi-gigawatt superclusters under development. The flagship 5GW Hyperion cluster will “cover a significant part of the footprint of Manhattan.” Two names have already surfaced: Hyperion, the 5GW flagship, and Prometheus, its counterpart. Zuckerberg also confirmed there are further “titan clusters” in the pipeline.

Why this matters:

  • Hyperion, in Greek mythology, was the Titan of heavenly light. Prometheus, his son, stole fire (ie knowledge) from the gods to give it to humans.

  • These names, and confirmation of the Titan naming convention, are a pretty clear statement of Meta’s intent.

  • With these announcements, along with the recent hires and acquisitions, it looks like Meta has gone full gloves off.

Read the full story on Data Center Dynamics here.

The Rundown

Nvidia’s still wearing the crown, but there’s something in the air.

The $4.5B China write-down may have turned into a handshake and a green light for shipping, but US officials hitting pause on Nvidia’s GPU deal with the UAE means the same risk still exists: loss of market share to Chinese competitors like Huawei. So, while the Department of Commerce worries about G42’s access being a backdoor for China, it risks handing the Middle Kingdom a regional foothold - something far more valuable.

Couple this with TensorWave scaling AMD, Google winning DoD work with TPUs, AWS is building its support infrastructure, and Meta channelling godlike power in its hiring and messaging, and you’d be forgiven for thinking that light breeze may yet develop into something else.

Headwinds?

Only time will tell.

See you next week.

Reply

or to participate