Issue #36: Nvidia’s $5.5B Headache, Dolphins, & Hugging Face Robots

Spare a moment for Jensen’s neck.

Last week, Nvidia was good to go for continued H20 shipments to China. This week? It’s a no-go. Meanwhile, Hugging Face dives into open-source robotics, AMD hits 2nm, and nuclear power continues to hold the spotlight.

Oh, and Google?

They’re trying to talk to dolphins.

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:

• NVIDIA’s H20 Whiplash: One Week, Two Policies

• Hugging Face Buys a Robot Company

• CoreWeave Fires Up GB200 Systems At Scale

• Google Trains AI on Dolphin Sounds

• AMD Tapes Out Its First 2nm HPC Chip

• Google Goes Geothermal in Taiwan

• Nuclear Could Supply 10% of AI Infra Energy Demand

Let’s get into it.

The GPU Audio Companion Issue #36

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’s H20 Whiplash: One Week, Two Policies

Nvidia’s attempt to secure H20 chip exports to China just got undone.

After last week’s news that the Trump administration backed off a planned crackdown following a dinner with Jensen Huang, the dream is over. The US government has flipped again and imposed formal restrictions on Chinese access to the H20 chip. The estimated cost to Nvidia? $5.5 billion in lost sales. Ouch.

Why this matters:

• This is a sharp reversal that undermines the diplomatic optimism Nvidia spun just days ago.

• $5.5B in lost revenue is a major blow, especially given the looming threat of increased competition from AMD (and even Huawei).

• Policy reversals like this set the scene for an unknowingly volatile export regime, where one dinner can't override geopolitical consensus for long.

Hugging Face Buys a Robot Company

Open-source robots are now on the menu.

Hugging Face has acquired Pollen Robotics, makers of Reachy 2, a humanoid robot already used at places like Cornell and CMU. The move accelerates Hugging Face’s robotics play via its LeRobot framework. And kicks off sales of physical robots built for research, education, and embodied AI.

Why this matters:

• Robotics is the next compute frontier and open-source could win hearts (and budgets).

• HF is positioning as the platform for open-source physical AI as well as text and vision models.

• Their move aligns with Nvidia’s GR00T launch, signalling early convergence around open robotics stacks.

CoreWeave Fires Up GB200 Systems At Scale

CoreWeave is first out of the gate.

They’ve officially launched Nvidia Grace Blackwell GB200 systems. Thousands of them. This makes good on CoreWeave’s early access claim and strengthens their reputation as the launch partner of choice for Nvidia's bleeding-edge silicon. The GB200s will be deployed across CoreWeave’s network at scale.

Why this matters:

• Being the first mover in the compute market with the top-of-the-range hardware means you get to set the market price.

• From what we saw with the H100s, being first to market also means being first to pay off hardware debt.

• Given how limited the GB200 supply is in the market right now, CoreWeave are in a very enviable position right now.

Google Trains AI on Dolphin Sounds

This one’s different.

Google just released DolphinGemma, a new model trained on decades of wild dolphin vocalisations. Built with DeepMind and Georgia Tech, it’s designed to decode and eventually respond to dolphin communication. This paves the way for two-way human-animal interactions using AI.

Why this matters:

• This is large-scale language modelling for non-human language.

• It showcases Gemma’s flexibility and Google’s ambition to stretch open-source AI into new domains.

• Also, it runs inference on Pixel phones, highlighting both the power of small, efficient models and what could be possible with on-device intelligence in the near future.

AMD Tapes Out Its First 2nm HPC Chip

Zen 6 is real - and tiny.

AMD just taped out its first 2nm EPYC chip (“Venice”) on TSMC’s N2 process. It’s the industry’s first HPC CCD on a gate-all-around transistor node. At the same time, AMD confirmed some of its current EPYC chips are now being built in the US, at TSMC’s Arizona Fab.

Why this matters:

• AMD just leapfrogged Intel’s 18A delays, potentially adding to their competitor’s woes.

• US-made chips for US clients is a strategy the entire industry is adopting.

• Venice could anchor AMD’s next-gen AI server push, giving it a stronger footing in the AI data centre race.

Google Goes Geothermal in Taiwan

Google just signed its first geothermal power deal in Asia.

The 10MW agreement with Baseload Capital will help power local data centres in Taiwan and complements a larger 1GW solar pipeline. It’s also part of a broader strategy to invest in 24/7 clean power, alongside geothermal R&D with Fervo and Project Innerspace. Nuclear, it seems, is not the only game in town for the hyperscalers.

Why this matters:

• Hyperscalers are hunting new baseload options. Nuclear came first. Geothermal might be next.

• Geothermal runs 24/7, making it ideal for AI data centre loads.

• Taiwan’s grid constraints are real, and this unlocks both infra and optics for Google.

Nuclear Could Supply 10% of AI Infra Energy Demand

A new Deloitte report says new nuclear capacity could meet 10% of US data centre power demand growth over the next decade.

Energy demand from AI infrastructure could hit 176GW by 2035 - a 5x jump from today. This means we need more power generation capacity, and fast. 35–62GW of new nuclear capacity could fill this gap via SMRs, re-licensed plants, and even coal site retrofits.

Why this matters:

• SMRs are moving from buzzword to serious solution with AWS, Google, and Oracle already signing deals.

• Coal-to-nuclear retrofits alone could unlock up to 174GW of capacity.

• While we have the technology to do this today, time, cost, and a 3.5x increase in the nuclear workforce by 2050 are the real bottlenecks.

The Rundown

At the risk of sounding like I’m rehashing an old hook:

There’s never a dull moment.

Policy flip-flops. Hugging Face goes physical. CoreWeave accelerates GB200. Dolphins join the LLM party. AMD hits 2nm. Google digs deep for geothermal. The nuclear renaissance continues.

What’s the thread that links it all together?

Speed across multiple domains.

Geopolitics. Robotics. New hardware rollouts. Human animal communication. Semiconductor improvement. Power generation.

We live in interesting times.

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.