Issue #21: Jensen Prints, A Neocloud Eyes Going Public, & G42 Heads to Italy

Nvidia just reminded the world of who dominates AI infrastructure.

Its latest earnings blew past expectations, with data centre revenue hitting $18.4B. A 409% YoY increase. Meanwhile, CoreWeave is reportedly closing in on a $35B IPO, Starcloud just raised $10M to build space-based data centres, and OpenAI’s latest model update comes with a catch: GPT-4.5 is expensive to run and already facing competition. Elsewhere, Nvidia’s RTX 50 series is shipping with missing ROPs, Microsoft is quietly cancelling data centre leases (but it might not mean much), and ENI, MGX, and G42 are backing a 1GW AI data centre buildout in Italy.

I’m Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centers that power it all.

Here’s what’s inside this week:

• Nvidia’s AI dominance grows as it posts (more) record-breaking earnings.

• CoreWeave’s $35B IPO could make it the first publicly traded AI cloud startup.

• Lumen Orbit and Starcloud land $10M to build space-based data centers.

• OpenAI’s GPT-4.5 and Perplexity’s Comet take different bets on the future of AI.

• Nvidia’s RTX 50 series ships with defects. But how bad is it?

• Microsoft pulls back on data centre leases, but does it actually matter?

• ENI, MGX, and G42 are backing a 1GW AI data centre buildout in Italy.

Let’s get into it.

The GPU Audio Companion Issue #21

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

Nvidia Prints $39.3B in Q4

Nvidia’s Q4 earnings left no doubt: AI infrastructure is still on fire.

Revenue hit $39.3 billion, up 12% QoQ and 78% YoY. Data centre sales made up $35.6 billion, surging 93% YoY. Full-year revenue? $130.5 billion, more than doubling YoY.

Jensen Huang didn’t hold back, calling the demand for Blackwell “amazing” and teasing a future where reasoning AI creates a new scaling law where more compute = better answers. Nvidia is already ramping up massive-scale production of its latest AI supercomputers, claiming “billions” in sales in their first quarter.

It’s Jensen’s world. We’re all just living in it.

Why this matters:

• Nvidia’s data centre revenue alone is larger than most Fortune 500 companies.

• Despite smashing expectations, gross margins slipped, which Nvidia blamed on “more complex and higher-cost systems” in its data centre business.

• More expensive hardware, higher R&D costs, and increasing competition from other manufacturers producing cheaper tin likely means Nvidia won’t be able to print money forever.

Read the full story here.

CoreWeave Eyes $4 Billion IPO at $35 Billion+ Valuation

CoreWeave, one of the fastest-growing AI cloud firms, is preparing to file for an IPO within the next week.

Targeting a $4 billion raise at a valuation north of $35 billion, the Nvidia, Cisco, Magnetar Capital, Coatue, Jane Street, and Fidelity-backed neocloud are at the forefront of AI infrastructure. CoreWeave’s public filing will likely be closely watched as investors gauge whether AI infrastructure providers can sustain sky-high valuations. If successful, this could open the floodgates for other AI infrastructure IPOs.

Why this matters:

• CoreWeave’s public debut will set a benchmark for other neoclouds eyeing the public markets.

• While AWS, Azure, and GCP dominate cloud infrastructure, CoreWeave is proving that specialised providers can carve out market share.

• A $35 billion+ IPO valuation suggests that investors still see massive growth potential in AI cloud providers despite recent volatility in AI stocks.

Read the full story here.

Space-Based Cloud is No Longer Sci-Fi

Two companies are pushing data storage and compute beyond Earth.

Starcloud (formerly Lumen Orbit) just raised $10M to develop LEO-based data centres, leveraging solar power and space’s natural cooling to offer ultra-secure, high-speed storage beyond Earth’s constraints.

Meanwhile, Lonestar and Phison have launched the first-ever lunar data centre aboard a SpaceX Falcon 9, set to land on March 4. The goal? Petabyte-scale off-world storage, protecting critical data from cyber threats, climate risks, and geopolitical instability.

Why this matters:

• Space-based storage and compute are no longer sci-fi. They’re real, funded, and happening now.

• This move makes sense given the massive geopolitical uncertainty—storing critical data off-Earth makes nation-state cyberattacks far more difficult.

• Given the near-unlimited solar power and cooling available, space could legitimately become a major market for AI inference and high-performance computing.

Read the full stories here:

Starcloud.

Lonestar and Phison.

OpenAI Unveils GPT-4.5, Perplexity Builds a Browser

OpenAI just dropped GPT-4.5 ‘Orion’, its biggest AI model yet.

Trained with more compute and data than ever, it boasts "higher emotional intelligence" and better factual accuracy. But it’s not a frontier model. And it lags behind DeperepSeek and Anthropic’s latest reasoning models on key benchmarks. And it’s by far and away the most expensive model to access via an API out there.

Meanwhile, Perplexity is moving beyond search, teasing Comet, an AI-powered browser designed to reshape how users interact with the web. Think of an AI-first experience, prioritising curated, conversational search results over (ad-riddled) traditional browsing.

Why this matters:

• The lacklustre performance on key benchmarks is a pretty good indication that pre-training scaling is over.

• The astronomical API costs relative to competitors for mid-level performance will likely make some users ask, “What for?”

• Perplexity is taking the fight to Google. With Comet, it’s going beyond search and attacking how people consume the web.

Read the full stories here:

OpenAI.

Comet.

NVIDIA’s RTX 50 Series Already Has Defective Silicon

Nvidia’s upcoming RTX 50 series is shipping with silicon defects.

The issue? Some RTX 5090, 5080, and 5070 Ti GPUs are missing Render Output Pipelines (ROPs). Nvidia blames manufacturing defects. But at ~10M+ units per series, even a fraction of failures is a lot of bad chips. And with the data centre Blackwell GPUs launching later this year and margins already under pressure, we may yet have a perfect storm for stock performance.

Why this matters:

• Missing ROPs means lower performance in some workloads, even if most users won’t notice.

• Even at its scale, defects slip through, raising questions about Blackwell’s production quality.

• If this is happening in consumer GPUs, high-end AI silicon could have similar issues.

Read the full story here.

Microsoft Pulls Back on Data Center Leases. But Does It Matter?

Microsoft is quietly cancelling data centre leases, leading to speculation about its AI infrastructure strategy.

Analysts flagged the move as a sign of slowing AI demand, but hyperscalers cancel lease options all the time. The pullback was also on only 200MW of capacity, which is nothing in the grand game of AI. With land, power, and permits in short supply, locking in excess capacity is just smart business, whether or not you ever use it.

Why this matters:

• Hyperscalers grab space early to keep competitors out, not just to build.

• Losing an option premium makes more financial sense than committing to sites you don’t need, and even trillion-dollar firms won’t waste billions on unnecessary capacity.

• This probably isn’t Microsoft pulling back. They’re just playing the long game.

Read the full story here.

G42 & ENI Team Up for a 1GW AI Data Center in Italy

G42 and ENI are partnering to build a 1GW AI data centre cluster in Italy.

The project taps ENI’s power assets to secure stable energy for AI workloads, while G42 likely handles the compute stack. This announcement follows G42’s deal with AMD and DataOne in France, cementing its European AI cloud expansion. Phase one launches in 2026, and full capacity will scale to 1GW over the next several years.

Why this matters:

• Following its France deal, the UAE is clearly going to play a major role in shaping European AI infrastructure.

• AI data centres are now energy projects, and the tie-up with ENI ensures long-term power stability, a key challenge for high-density AI computing.

• Sovereign AI capability is what everyone is targeting, given the shifting geopolitical landscape, and G42 are keen to enable it.

Read the full story here.

The Rundown

I thought “To the moon” was meant to be tongue-in-cheek.

2025 clearly has other ideas.

NVIDIA keeps smashing records, CoreWeave is gearing up for a blockbuster IPO, and OpenAI’s latest release is making waves (if not quite a tsunami). Meanwhile, Perplexity is creeping up on Google, Microsoft is trimming some data centre fat, and G42 is locking down AI infrastructure across Europe. And to top it off, we have data centres in space.

Ridiculous.

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.