There are many routes to AI ROI, and some are less savoury than others.

Option one is the slow march towards enterprise compliance. Option two is Stargate-as-a-Service. Option three consists of deals with the companies you once sought to dethrone. And option four?

Hyper personalised, ad-optimised, TikTok-style, AI-generated porn feeds.

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 #69

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, click here to fix that.

OpenAI’s Solution to a $1 Trillion Problem: Porn?

OpenAI has a plan to pay for everything, but can barely afford what it’s already promised.

The Financial Times revealed this week that the company has mapped out a five-year plan to cover more than $1 trillion in commitments to partners like Oracle, NVIDIA, AMD, and Broadcom. But there’s a problem. OpenAI’s annual recurring revenue sits around $13 billion, with only 5% of ChatGPT’s 800 million users paying.

The maths doesn’t work unless the business model changes.

And it already is.

From Pulse (its social layer) to Sora’s upcoming video features, to Altman’s confirmation that adult content will return for verified users, OpenAI is slowly inching toward the same direction every attention platform eventually goes: advertising and engagement economics, with a potentially NSFW twist. Why does this make sense?

Because the “AI girlfriend” industry is projected to hit a $15.8 billion market value by 2032.

Pair that with Sora’s real-time video capability, and the next frontier is easy to imagine: algorithmic, NSFW, TikTok-style companion feeds, optimised for user preferences/fetishes, watch-time and ad spend.

Why this matters:

  • OpenAI wanted to build AGI.

  • Now, it has opened the door to the world’s most addictive content platform to pay for it.

  • If the next growth phase of AI runs on ad-supported intimacy, the question isn’t whether it works. It’s what it costs. Socially, economically, and psychologically.

Click here to read the full FT article.

Nscale Signs 200,000-GPU Deal with Microsoft

Nscale are on a roll.

Following an aggregate $1.1b + $433m fundraise across two rounds, we finally have details on the big deployments with Microsoft. Top level? 200,000 NVIDIA GB300 GPUs, spanning four countries and two continents, in collaboration with Dell Technologies. Per this week’s announcement, the split is as follows:

  • Texas, USA: ~104,000 GB300 GPUs across a 234MW AI campus, leased from Ionic Digital. Microsoft holds an option to expand by another 700MW, potentially scaling the site to 1.2GW by late 2027.

  • Sines, Portugal: 12,600 GB300 GPUs hosted at the Start Campus data centre, providing sovereign EU-based AI infrastructure for Microsoft and enterprise customers.

  • Loughton, UK: 23,000 GB300 GPUs at Nscale’s 50MW AI campus, scalable to 90MW, delivering the UK’s largest NVIDIA AI supercomputer for Microsoft Azure from Q1 2027.

  • Narvik, Norway: 52,000 GB300 GPUs through the Aker-Nscale joint venture, already under a multi-year agreement with Microsoft to expand Nordic hyperscale capacity.

Together, these projects catapult Nscale into the multi-gigawatt global deployment range, putting them in the same league (at least on paper) as Oracle and CoreWeave.

Why this matters:

  • If Nscale can pull this off, they’ll have earned a place as Europe’s first sovereign hyperscaler, rivalling US incumbents.

  • That’s a big if, not necessarily due to any lack of ability from Nscale, but rather because of the sheer scale (and resultant complexity) of the planned deployments.

  • Execution risk is real, especially with kit as allegedly finicky as the GB300s, so all eyes are on Josh and the Nscale team for the foreseeable future. No pressure.

Click here to read the full announcement from Nscale.

Nebius Rolls Out “Aether” AI Cloud 3.0

Nebius has rolled out AI Cloud 3.0 “Aether”.

The major platform update is designed to provide enterprises with tighter control, stronger compliance, and smoother scaling for production-grade AI. Unlike many updates or announcements focused on capacity, Aether zeroes in on the operational friction points that slow enterprise adoption: governance, observability, and secure multi-team collaboration.

The release brings new enterprise-grade compliance certifications, including SOC 2 Type II, ISO 27001, and HIPAA, alongside alignment with NIS2 and DORA, clearing regulatory barriers for customers in healthcare, finance, and the public sector.

Other key feature updates include:

  • Granular IAM and self-service tenant creation, allowing IT leaders to enforce access controls without slowing development.

  • Advanced observability tools, with Grafana-based dashboards for Nebius’ Managed Soperator (Slurm) clusters, giving teams real-time visibility into performance, uptime, and power usage.

  • Cilium support and static routing options for more flexible networking, and a new built-in secrets manager (MysteryBox) to eliminate API key exposure risks.

  • Reliability upgrades, including active health checks, self-healing nodes, and major file storage throughput improvements — now up to 12 GB/s read and 8 GB/s write per 8-GPU VM.

  • Simplified developer experience, with refreshed navigation, auto-allocated resources, unified GPU SKUs, and the ability to launch containerised AI apps like Jupyter and ComfyUI directly on VMs.

Why this matters:

  • Regulated enterprise customers cannot and will not engage with a platform that doesn’t meet stringent compliance requirements.

  • Few neoclouds meet the necessary standard, and “enterprise-ready” is little more than a marketing slogan. Many fall woefully short. Nebius no longer has that problem.

  • With compliance, control, and developer velocity now all integrated into one stack, Nebius is quietly positioning itself as a legitimately enterprise-ready alternative to traditional hyperscalers.

Click here to read the full announcement from Nebius.

Microsoft Launches MAI-Image-1

Microsoft AI has unveiled MAI-Image-1, its first fully in-house text-to-image model.

MAI-Image-1 was designed for speed, realism, and creative control, avoiding the repetitive or stylised look typical of many diffusion systems. Training emphasised rigorous data curation and real-world evaluation. Professional designers and photographers then shaped how it handles lighting, reflection, and composition.

The result?

Photorealistic images, from landscapes and interiors to product shots, at speeds fast enough to support interactive workflows in Copilot and Microsoft’s creative tools.

Why this matters:

Click here to read the full announcement from Microsoft.

Intel Unveils First Inference-Only GPU

Intel just announced Crescent Island, a new data centre GPU optimised purely for AI inference.

Built on the upcoming Xe3P architecture, a higher-performance version of the Xe3 cores found in next-gen Panther Lake CPUs, Crescent Island carries 160GB of LPDDR5X memory.

Beyond that, however, the design remains partially mysterious.

Intel hasn’t confirmed specs, but analysts note that fitting 160GB of LPDDR5X likely requires either a 640-bit interface for a single massive GPU or dual GPUs with 320-bit buses, hinting at a multi-die approach. Sampling begins in 2H 2026, targeting air-cooled enterprise servers and cloud inference providers seeking energy efficiency and memory bandwidth for large language model deployment.

Why this matters:

  • Like NVIDIA with the Rubin CPX series, Intel too are pivoting toward dedicated inference silicon.

  • With LPDDR5X memory and Xe3P efficiency, Crescent Island could become Intel’s first viable data-centre GPU for LLM serving.

  • Success would be a boon for both Intel and the market, introducing yet more competition to the burgeoning inference-specific hardware space.

Click here to read the story from Tom’s Hardware.

Oracle Reveals Zettascale10 at AI World

Oracle just launched its most ambitious infrastructure project yet: the OCI Zettascale10, revealed this week at Oracle AI World in Las Vegas.

The system is built to deliver multi-gigawatt AI capacity, connecting up to 800,000 NVIDIA GPUs across tightly clustered data centres. Oracle claims up to 16 zettaFLOPS of peak performance, making Zettascale10 the largest AI supercomputer in the cloud. Orders are already open, with availability expected in the second half of 2026, marking the start of Oracle’s next phase in the hyperscale AI race.

Why this matters:

  • Zettascale10 moves AI infrastructure beyond megawatt clusters into gigawatt-scale AI factories.

  • It’s a direct challenge to hyperscalers like AWS and Microsoft, offering Stargate-class compute as a service.

  • Oracle is betting that scale, not just software, will define who controls the next wave of AI model training.

Click here to read the full announcement from Oracle.

NVIDIA’s 800V DC Vision: The 1MW Rack Era Begins

The future of data centres is dense. Very dense.

Today, a typical hyperscale rack draws between 30-60kW. NVIDIA’s Kyber rack design, built on its new 800V distribution standard, pushes that an order of magnitude higher. By distributing high voltage DC directly to the GPU nodes and eliminating multiple conversion steps, NVIDIA claims the design delivers 3.4x higher rack density than Hopper-class systems while maintaining efficiency and reliability. That shift completely rewrites the economics of the data centre floor.

A 1MW rack isn’t just a hardware milestone.

It’s a financial one.

At a hypothetical market rate of around $100 per kW per month, a single rack could generate $100,000 in monthly recurring revenue, or $1.2 million a year, before power and cooling costs. Multiply that across a 100MW campus, and you’re looking at a $120 million annual lease base from compute alone.

Why this matters:

  • Data centres are shifting from megawatts per hall to megawatts per rack, meaning the overwhelming majority will be unable to support this rack density.

  • If MW racks become the industry standard, expect a lot of stranded assets. Equally, if MW racks don’t catch on, we’ll have the same problem in the other direction.

  • Whatever happens, many DC operators are going to have to make some pretty aggressive bets to position themselves for what’s to come.

Click here to read the full story from DCD.

The Rundown

“Our mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity.”

That’s OpenAI’s mission statement.

It’s the mission that’s given rise to every single thing in this week's issue.

European startups landing deals with US hyperscalers at less than two years of age. Relentless platform updates in the pursuit of enterprise adoption. Image models and silicon that are the stuff of sci-fi. Zettascale compute clusters. Racks providing the power equivalent of 164 homes. Every single one of these stories would have been impossible just three short years ago.

They are only possible now through the thousands of hours of collective effort an entire industry has spent in support of OpenAI’s mission.

If one could travel back in time to the beginning of the buildout to let the industry know that the pinnacle of their collective efforts would be the organisation seeking to herald the arrival of AI, quietly laying the groundwork for AI-porn generation at scale, would we have seen the same buy-in?

That’s the paradox now facing the industry.

The infrastructure built to accelerate intelligence may end up optimising for attention. The same racks that power the frontier of science could soon power algorithmic intimacy feeds.

And somewhere between trillion-dollar debt plans, NSFW engagement loops, and sovereign-grade compliance frameworks, the question has shifted.

Not can we build AGI (and that’s debatable with LLMs), but what will it be built for?

See you next week.

Reply

or to participate