Issue #15: Jail Time for DeepSeek, Deep Research, & Housework Robots

The AI infrastructure race is honestly starting to feel a bit like a reality TV show.

Every week introduces a new character, sees a new storyline emerge, or a plot thickens. The pace is ridiculous, and that’s from inside the industry. I can only imagine how absurd it all must look from the outside.

Still, it makes for (hopefully) entertaining reading.

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:

• US Senator proposes prison time for downloading Chinese AI models.

• OpenAI co-founder’s new AI startup, SSI, seeks a $20B valuation.

• CoreWeave launches Nvidia GB200 NVL72 instances ahead of hyperscalers.

• OpenAI launches DeepResearch, and HuggingFace clones it within 24h.

• Cerebras Systems breaks inference speed records.

• The UAE commits billions to AI infrastructure in France.

• Meta funds AI-powered robotics research, testing human-AI collaboration.

Let’s dive in.

The GPU Audio Companion Issue #15

Want the GPU breakdown without the reading? The Audio Companion does it for you—but only if you’re subscribed. Fix that here.

US Senator Proposes Prison Time for Downloading Chinese AI Models

A new bill proposes up to 20 years in prison for Americans caught downloading or collaborating with Chinese AI companies.

Senator Josh Hawley (R-MO) has introduced a bill that would criminalise the transfer, research, or financial support of AI models from China. The penalties? Up to $1 million in fines plus 20 years in prison for individuals and $100 million in fines for companies. Yet given Sam Altman has come out and said OpenAI is “on the wrong side of history” with their open-source stance, and Mark Zuckerberg expressed concerns to Joe Rogan about an open-source Chinese model becoming the default, this move is an interesting one. Considering how open-source development works, it’s hard to see how that won’t happen if the US criminalises engagement with the open frontier models.

Why this matters:

• If the US closes itself off from the open-source AI community, there’s a good chance Chinese open-source will become dominant.

• Dominance in the domain of the future will actively harm global US tech hegemony.

• The battle over who controls AI infrastructure is getting political. Fast.

  Or perhaps it always was, and we’re only now getting to see it.

Read the full story on the Register here.

OpenAI Co-Founder’s New AI Startup Seeks $20B Valuation

Another player joins the game.

Ilya Sutskever’s Safe Superintelligence Inc. (SSI) is reportedly in talks to raise funding at a $20B valuation. Quadrupling its worth from last year. The former OpenAI co-founder is pitching investors on the idea of "scaling in peace", free from the commercial pressures that have shaped OpenAI and Anthropic. With AI valuations skyrocketing (OpenAI is targeting $300B, Anthropic $60B), SSI’s ability to attract investment shows that the race is far from over for players with the right pedigree.

Why this matters:

• The frontier AI labs still command massive valuations despite DeepSeek’s efficiency shift.

• Sutskever is betting on "safe AI", a response to concerns about uncontrolled AGI development, although what “safe” means in this context is anyone’s guess.

• Investors are still all-in on foundation models despite open-source disruption.

Read the full story on Reuters here.

CoreWeave Launches Nvidia GB200 NVL72 Before Hyperscalers, But Microsoft Got There First

CoreWeave is the first cloud provider to make the Nvidia GB200 NVL72 system generally available.

Despite Microsoft deploying Blackwell GPUs last year, the AI Hyperscaler has beaten AWS, Azure, and Google Cloud in offering GB200 NVL72 instances. Nvidia’s official blog even highlights CoreWeave’s early leadership in rolling out Blackwell, reinforcing how NeoCloud providers are now competing directly with hyperscalers at the bleeding edge of AI infrastructure.

Why this matters:

• The GB200 NVL72 system delivers a 30x performance increase over the same number of H100 chips.

• Combine this performance increase with the efficiency gains DeepSeek unlocked, and we’ll probably see at least one more massive jump in foundation model capability when the first training runs come to an end.

• CoreWeave being first to market with general instances will likely go a long way to further entrenching their dominance in the neocloud space.

Read the full press release from CoreWeave here.

OpenAI’s Deep Research: AI-Powered Autonomous Research

OpenAI has launched Deep Research, an AI agent within ChatGPT designed for autonomous, multi-step investigations.

Unlike standard chatbot responses, Deep Research independently browses the web for 5 to 30 minutes, analysing data, compiling structured reports, and citing sources. Initially available to Pro users, OpenAI plans to expand access. Powered by a specialised version of OpenAI’s upcoming o3 model, it outperforms prior AI models on deep contextual tasks, scoring 26.6% on Humanity’s Last Exam.

More than double previous OpenAI models.

But open-source AI isn’t far behind.

Hugging Face’s AI research agent nearly matched Deep Research’s performance after just a 24-hour hackathon, highlighting how community-driven AI can challenge proprietary models faster than ever.

Why this matters:

• Highly capable AI research assistants are here, automating complex information analysis.

• The quality of the outputs is such that even expert-level analysis is rapidly becoming commoditised along with previously specialised work like IPO filings.

• Hugging Face’s challenge shows open-source AI is catching up fast, meaning proprietary models may not hold the advantage for long as open-source innovation accelerates.

Read the full stories on the OpenAI blog and Ars Technica below:

• OpenAI – Deep Research

• Ars Technica – Hugging Face AI Agent

Cerebras Sets AI Inference Speed Record

Cerebras just launched the world’s fastest DeepSeek R1 Distill & LLaMA 70B inference.

The CS-3 system from Cerebras now claims the fastest inference speeds for DeepSeek R1 Distill and Meta’s LLaMA 70B. The AI chipmaker is positioning itself as the go-to alternative to Nvidia for high-performance AI workloads, particularly as the industry looks for non-GPU-based solutions to break Nvidia’s near-monopoly. This could be a significant shift in AI hardware strategy, especially as hyperscalers and AI labs look to diversify their compute stacks.

Why this matters:

• The market is slowly starting to shift away from just training towards inference as models start being deployed in the “real world”.

• Inference is where ROI will be made, with power cost and efficiency having a massive impact on return profiles for hardware investments.

• Alternative hardware manufacturers like Cerebras may yet have an edge over Nvidia with their niche focus and lower cost-to-performance ratio.

Read the press release on BusinessWire here.

UAE to Invest Billions in AI Data Centres in France

The UAE is pouring billions into France’s AI data centre industry, strengthening its European AI footprint.

In a major geopolitical move, the UAE has committed billions to develop AI data centres in France, reinforcing its strategy to expand beyond its own borders as a dominant AI infrastructure player. This investment positions France as a key AI hub in Europe, while also giving the UAE strategic control over compute resources outside of the Gulf region.

Meanwhile, in the UK, Chainergy, a biogas-powered AI data centre is being developed. While the UAE is throwing billions into European hyperscale AI, the UK is experimenting with cleaner, independent solutions to circumvent energy prices as a barrier to AI expansion.

Why this matters:

• The UAE is expanding its AI infrastructure dominance beyond the Middle East.

• France, with its predictably priced nuclear power, is becoming a key European AI infrastructure hub.

• Grid connections and power prices in the UK are a major barrier to AI deployment at scale, so alternative solutions like Chainergy will continue to emerge as the market looks for solutions.

Read the full story on TechCrunch here.

Meta Funds AI-Powered Robotics Research

Finally, the AI news we’ve all been waiting for.

Meta is funding robotics research focused on teaching AI models how to collaborate with humans on housework. The research aims to improve AI’s ability to understand physical environments, opening the door for future AI-powered robotic assistants. While this may sound like a consumer tech experiment, the real goal is to advance AI models that can interact in real-world environments.

The ironing robot is on the horizon.

Why this matters:

• Meta is pushing AI beyond text and image models, into real-world interactions.

• This signals a long-term bet on physical AI beyond just LLMs and Agentic AI in the virtual world.

• Human-AI collaboration research could accelerate robotics adoption in industry.

Read the paper from Meta here.

The Rundown

The saga (and drama) continues.

US lawmakers are pushing for prison time for downloading Chinese AI, and Ilya Sutskever’s new startup, SSI, is targeting a $20B valuation. Meanwhile, CoreWeave beat hyperscalers to launch Nvidia GB200 NVL72, and OpenAI launched Deep Research with HuggingFace responding in kind. Cerebras Systems broke records while the UAE is pouring billions into AI infrastructure in France. And finally, Meta is looking to bring you the Housework Bot 9000.

Tune in next week for another episode of Humanity’s Final Infrastructure Race.

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.