75GW.

That’s how much new data centre capacity Brookfield sees being added by 2034. To put that number into perspective, it’s roughly the same as the power generation capacity of the UK in 2024.

Or, enough power for ~56m average homes.

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 #56

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.

2 Models, 1 Week: GPT-5 & GPT-oss Arrive

Power moves from OpenAI this week.

First up, OpenAI’s first opensource models since GPT-2. gpt-oss-120B and gpt-oss-20B ship under Apache 2.0, tuned for reasoning, tool use, and agentic workflows. The 120B runs near o4-mini on a single 80GB GPU; the 20B targets 16GB edge devices. Both have 128k context, structured outputs, and full chain-of-thought. Launch includes tokeniser, runtimes, integrations, and a $500k red-teaming challenge. Aim: own the open-model narrative and developer mindshare.

Next, GPT-5. This is the big one, and it’s not just another incremental model release. The new flagship is the step change many have been waiting for, and brings with it a unified system with a live router, a “thinking mode” for harder problems, and GPT-5 Pro for extended reasoning. It posts record scores across math, coding, multimodal, and health benchmarks, with fewer hallucinations and safer completions. Rollout is already live in ChatGPT, with the router choosing depth unless prompted to “think hard.”

Why this matters:

  • GPT-5 keeps OpenAI in the frontier race; gpt-oss opens doors to sovereign, regulated, and on-prem markets.

  • Apache 2.0 release pressures Anthropic, Google, and Mistral on capability and accessibility.

  • Owning both hosted and self-hosted lanes hedges against regulation, geopolitics, and hyperscaler risk.

Click here to read OpenAI’s GPT-5 announcement.

Click here to read the GPT-oss announcement.

CapitalG & Nvidia Circle Vast Data at $30B

The next most valuable AI infrastructure company might not be the one you expect.

Vast Data is reportedly fundraising at a potential $30B valuation, with Alphabet’s CapitalG and Nvidia in the mix. Why are they raising now? Because the AI storage player booked $200M ARR in January, expects $600M next year, and is free cash flow positive. With former Shopify CFO Amy Shapero recently joining the team, this level of investor interest, and the apparent thawing of public markets, an IPO may well soon be on the horizon.

Why this matters:

  • Closing a round at $30B would make Vast one of the highest-valued AI companies out there.

  • The last funding round was at a $9.1B valuation, so this uplift indicates just how important data processing and storage have become to the industry.

  • Competitors Weka and DDN are pursuing similar efforts, but at this stage, Vast Data is way out in front.

Click here to read the full story on Reuters.

Crusoe Taps Google Cloud AI Exec

Crusoe has hired Erwan Menard, former Head of Product for Generative AI at Google Cloud, as SVP of Product.

Menard previously led product and go-to-market strategy for Google’s AI stack, including model APIs and infrastructure services. At Crusoe, he’ll be tasked with refining the company’s GPU cloud platform and shaping its roadmap across inference, training, and sustainable compute delivery. Coupled with this week’s rebrand, it’s clear that Crusoe is stepping things up a gear.

Why this matters:

  • Senior-level hires like Menard indicate Crusoe is likely preparing to move up the platform stack, and go after the enterprise AI market.

  • The rebrand supports this, with a clear move away from the flare gas/Bitcoin mining heritage to warmer colours, high-energy yellow accents, a cleaner visual identity, and an enterprise feel.

  • Expect tighter product strategy, better developer tooling, and targeted enterprise attack vectors in the months ahead.

Click here to read the hiring announcement from Crusoe (and to check out their new website).

Apple’s ‘Answers’ team: ChatGPT, the Apple way

After years of saying it didn’t need a chatbot, Apple is building one anyway.

The new Answers, Knowledge and Information (AKI) group is working on an in-house “answer engine”: a lightweight, Apple-native rival to ChatGPT. Think conversational search built directly into Siri, Spotlight, and Safari, with Apple UI polish.

Why now?

Because Siri can’t reliably answer complex queries and often punts to Google, Apple pays Google ~$20B/year to be the default search option (a deal now under antitrust threat), and Generative AI is shifting how people search to an interface that Apple does not yet own.

AKI, led by former Siri boss Robby Walker, is hiring search engineers, exploring a standalone app, and building backend infrastructure for Apple-wide integration. Underneath, Apple is ramping Baltra, a cloud AI server chip, and standing up a Houston AI facility to run it. Partnerships, including with Perplexity, are still on the table, but the direction is now unmistakable: Apple wants its own search stack.

Why this matters:

  • Apple’s late entry may pay off if it ships a secure, ecosystem-first alternative.

  • A native answer engine could threaten Google Search for the first time in 20 years.

  • Control over search means control over user attention — and revenue.

Click here to read the full story on Bloomberg.

Cerebras Weighs $1B Raise, IPO on Ice

Wafer-scale chipmaker Cerebras Systems is in talks to raise up to $1 billion in private funding, pushing its IPO back yet again.

The company confidentially filed for a public listing in August 2024, announcing the filing in September, but the process has stalled under scrutiny over its reliance on G42. The Abu Dhabi-based tech conglomerate accounted for 87% of Cerebras’ revenue in both 2023 and H1 2024, and has already invested $900M to build a US-based AI platform on Cerebras hardware. CEO Andrew Feldman had said they were on track to list this year, until the private round talks began. So, will the IPO happen this year? Who knows.

Why this matters:

  • The inference market is booming, and Cerebras has been inking partnerships left, right, and centre.

  • Heavy revenue concentration from G42 could be seen as a risk in an IPO, though it wasn’t a problem for CoreWeave.

  • $1B in late-stage capital will therefore likely be used to extend runway, diversify revenue streams, and strengthen Cerebras’ position going into a public offering.

Click here to read the full story on DCD.

Meta Doubles Down on Custom Silicon

Meta is taking back control.

The company has placed orders for its next-gen AI servers, powered by the company’s in-house Artemis ASIC, with manufacturing partners Broadcom and Quanta Computer. The servers will back Meta’s next phase of AI infrastructure build-out, aimed at reducing reliance on third-party accelerators. The Artemis chip is optimised for Meta’s AI workloads, from recommendation engines to LLM inference, and is part of a broader push to cut costs and control supply.

Why this matters:

  • Custom silicon lets Meta sidestep GPU shortages, pricing volatility, and dependency on an ascendant potential competitor in NVIDIA.

  • Tight integration between hardware and Meta’s AI stack could boost efficiency at scale.

  • Broadcom and Quanta secure another multi-billion-dollar pipeline from hyperscaler AI demand.

Click here to read the full story on DCD.

Brookfield Sees 75GW New DC Capacity by 2034

Brookfield’s new AI infrastructure report makes some BIG predictions.

How big? $7 trillion in global AI infrastructure spend over the next decade, split roughly into:

  • $4T on chips (fabs + supply chain)

  • $2T on AI data centres

  • $0.5T on power & transmission

  • $0.5T on connectivity, cooling, and robotics

On capacity, Brookfield sees AI-focused data centres growing from ~15GW today to 82GW by 2034. That’s 6-8GW of new build every year, and a tenfold jump in a decade. GPU demand is expected to 7x from 7M to 45M units over the same period, with inference workloads making up ~75% of AI compute demand by 2030. They also warn that the rise of complex AI agents, chaining dozens of model calls for a single task, will further spike inference requirements, reshaping facility design around high-volume inference traffic rather than just large training jobs. GPUaaS revenues are also projected to grow from $30B in 2025 to $250B by 2034, driven by enterprises seeking AI capacity without the capex burden.

Why this matters:

  • This kind of buildout changes everything we think we know about how data centres could/should be built.

  • Brookfield highlights modularity as essential in avoiding obsolescence and enabling rapid power/cooling upgrades as chips evolve.

  • If these predictions are true, we haven’t seen anything yet.

Click here to read the full report on Brookfield’s website.

The Rundown

If Brookfield’s right, we’re about to add ~10% of the UK’s entire grid supply to data centre capacity, every year, for a decade.

That scale challenges everything we think we know about deployment. Not just for the builders, but for financing, power, cooling, and operations. And it all needs to happen at the same time. That means control is everything. And it’s also why this week’s stories all point the same way:

  • OpenAI is locking down both closed and open-weight AI to control the lanes.

  • Vast, Crusoe, and Cerebras are sharpening their plays to capture enterprise spend.

  • Apple and Meta are taking the stack in-house to cut reliance on anyone else’s silicon or search.

What’s the theme? Control.

Managing $7T of aggregate spend is impossible without it.

The winners will be the ones with it. The losers? The ones who give it up without even realising.

See you next week.

Reply

or to participate