Seven shifts. One week.

Non-GPU AI infrastructure. Mega mergers. B300s. Apple relinquishing control. Space fabs. Oracle landing the biggest cloud contract ever. Meta playing private credit games.

And, of course, billions of dollars of capital flowing.

I’m Ben Baldieri. 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:

Let’s get into it.

The GPU Audio Companion Issue #50

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.

OpenAI Rents Google TPUs

Nvidia’s grip may not last forever.

OpenAI has started renting Google Cloud’s TPUs to run inference for ChatGPT and other products. It’s the company’s first meaningful move beyond Nvidia chips and Microsoft’s data centres. OpenAI is still one of the world’s largest GPU buyers for training, but TPUs could help lower inference costs as workloads scale. Google, for its part, isn’t handing over its most powerful TPUs but is reportedly treating the deal as proof that its custom silicon can compete.

Why this matters:

  • This is the first time OpenAI has deployed non-Nvidia silicon at meaningful scale.

  • With growing pressure on cost, power, and supply, Anthropic, Apple, and SSI are all starting to chart their own paths beyond the GPU.

  • If the results of this POC are positive, the inference hardware landscape might be about to change very rapidly.

Read the full story on Reuters here.

HPE x Juniper Deal Cleared

The $14B HPE-Juniper merger is officially a done deal.

With regulatory approval now locked, HPE can begin integrating Juniper’s AI networking stack across its enterprise and cloud offerings. CEO Antonio Neri says the combined company will focus on “AI-native” infrastructure, leveraging Juniper’s Mist AI and data centre switching to compete more effectively against Cisco and Arista. The real test now lies in execution: aligning go-to-market strategies, stitching together product teams, and proving out cross-sell capabilities across edge, cloud, and AI workloads.

Why this matters:

  • Networking is moving to the heart of the AI infrastructure conversation, and simplicity is what the enterprise AI market wants.

  • HPE is betting on a future where compute, interconnect, and orchestration are sold as a single platform.

  • Any player with ambitions in enterprise AI now face a new heavyweight with vertical integration, sovereign AI ambitions, and a sharpened focus on end-to-end infrastructure control.

Read the full story on SiliconAngle here.

Neoclouds Kick Off the B300 Era

The B300 wave has officially landed: one on sovereign soil, one in the AI cloud.

Hydra Host just delivered the first-ever sovereign deployment of NVIDIA B300s for El Salvador’s National AI Lab. President Bukele wants compute he can control. Hydra made it happen.

Bare metal, in-country, for a nation-state.

On the other end of the stack, CoreWeave is now the first cloud provider to deploy GB300 NVL72 systems.

These aren’t the flagship GB200s for training. GB300s are tuned for inference. 50x higher output. 10x better user responsiveness. 5x higher throughput per watt than Hopper.

Why this matters:

  • With both the next iteration of Blackwell now live, H100/H200 infrastructure is looking increasingly outdated.

  • New generations of hardware mean massive downward pressure on GPU/h pricing. Good for users. Bad for margins.

  • Expect massive changes in Hopper pricing dynamics and a flood of second-hand H100s and H200s to hit the market before the end of the year.

Read the announcement from Hydra Host here and CoreWeave here.

Apple flirts with outsourcing Siri’s brain

After years of downplaying LLMs, Apple may finally be conceding.

Bloomberg reports Apple is testing Claude and GPT models on its own infra to power the next-gen Siri. This would move Siri beyond just calling ChatGPT, baking third-party LLMs into Apple’s cloud stack. Why now? Myriad delays and technical dead ends have reportedly pushed completion of the “LLM Siri” project to 2026 or later.

Why this matters:

  • Letting OpenAI or Anthropic in the door means handing over at least partial control of the thinking layer of Apple’s future interfaces and products.

  • The world’s most vertically integrated company relinquishing any control at all of the UX is a stark departure from the Apple ethos of old.

  • Embedding tech that you continue to dismiss at the centre of your product stack looks an awful lot like capitulation.

Read the full story on Bloomberg here.

Space Forge Launches Orbital Fab

This isn’t your average fab.

Last week, SpaceX launched a payload designed to kickstart semiconductor production in space. The satellite, built by Space Forge, will attempt to manufacture chips in low-Earth orbit. Why? Zero gravity, extreme cold, and a vacuum environment enable the production of materials that are impossible on Earth.

Why this matters:

  • Space Forge is targeting high-performance materials like gallium nitride, sapphire, and other exotic substrates.

  • These materials require extreme purity or precision in production that space-based manufacturing could unlock.

  • If successful, space fabs may yet become the default for the advanced materials of the future.

Read the full story on Tom’s Hardware here.

Oracle Lands $30B Cloud Deal

Oracle just signed one of the largest cloud contracts in history.

In a regulatory filing, the company revealed a $30 billion-a-year cloud services agreement, set to kick in from fiscal year 2028. The customer wasn’t disclosed, but the scale dwarfs Oracle’s current cloud revenue (~$10.3B over the last year). The deal builds on Oracle’s push into AI workloads and follows its Stargate JV with OpenAI, which was already one of the most aggressive GPU buildouts announced to date.

Why this matters:

  • This is 3x Oracle’s existing cloud infra business, and could vault the company ahead of other second-tier hyperscalers.

  • A deal of this size likely requires multi-billion-dollar annual investments, so expect significant capital expenditures.

  • From OpenAI to unnamed mega-clients, Oracle’s AI infrastructure pivot is real and accelerating.

Read the full story on Bloomberg here.

Meta Wants $29B for AI Data Centre Buildout

Meta is in talks with Apollo, KKR, Brookfield, Pimco, and Carlyle to raise a staggering $29 billion.

The raise, with $3B in equity and $26B in debt, is to bankroll its next wave of AI data centres. The goal? Lock in the capacity needed to train and run Meta’s Llama models and future AGI ambitions, without overloading its own balance sheet. Meta has already raised its 2025 capex forecast to up to $72B, citing rising infra and GPU costs. This new raise could push that number even higher.

Why this matters:

  • This raise follows a string of aggressive AI moves: a $15B investment in ScaleAI, multiple nuclear and clean power deals, and sky-high sign-on bonuses to lure OpenAI talent.

  • Meta wants to control the infrastructure it needs for growth without further increasing its balance sheet weight.

  • Off-book structures like these enable Meta to scale without compromising its credit profile.

Read the full story in the FT here.

The Rundown

The giants are slipping, and they know it.

Apple’s been forced to delay its own LLM project and is now testing third-party models to power Siri, after years of dismissing them. Meta’s scrambling to secure $29B in off-balance-sheet financing just to fund its next wave of AI data centres. OpenAI, the poster child for Nvidia loyalty, is now running inference on Google TPUs to cut costs.

The pattern is hard to ignore:

Cost pressures are rising, technical setbacks are real, and the old models aren’t scaling cleanly.

From infrastructure spend to foundational model control, the strongest players are having to rethink how they build. That’s why the next wave of dominance won’t come from who’s biggest. It’ll come from who adapts fastest and can make their runway last the longest.

See you next week.

p.s. If you refer a friend using your unique code below, you get access to The GPU Premium for a month…

Go on.

You know you want to.

Reply

or to participate