The numbers keep getting bigger.

OpenAI signing a $300B contract with Oracle. Nebius landing $17.4B with Microsoft. OpenAI and NVIDIA lining up billions in the UK with Nscale.

And all with NVIDIA’s next generation of GPUs peeking over the horizon.

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 #

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 and NVIDIA Line Up UK Data Centre Billions

OpenAI and NVIDIA are preparing to pledge billions into UK data centre infrastructure next week, in partnership with Nscale.

The announcement, according to Bloomberg, is expected to be made during President Trump’s visit. Jensen Huang, Sam Altman, Larry Fink (CEO of BlackRock), and Stephen Schwarzman (CEO of Blackstone) are also expected to attend, potentially rolling out tens of billions of dollars in new commitments.

For the UK, it’s a major endorsement of UK Prime Minister Keir Starmer’s “AI growth zones” agenda.

For Nscale, it’s something bigger: validation of its role as OpenAI’s preferred European partner.

The company has already locked OpenAI as the anchor for its Stargate Norway build with Aker ASA. Now, with OpenAI and NVIDIA backing, Nscale looks to be positioning itself as the bridge between European sovereignty goals and US capital.

Why this matters:

  • From launch in 2024 to delivering Stargate Norway with OpenAI, and now potentially the UK, Nscale is emerging as a serious contender in the neocloud space.

  • The Norway and UK deals show how OpenAI’s Stargate strategy works: partner locally, deploy NVIDIA hardware, and scale with sovereign-friendly operators.

  • The UK economy is stagnant, so this deal could be the boost Starmer’s government needs to get things moving in the right direction again.

Click here to read the full story from Bloomberg.

OpenAI Commits to $300B of Oracle Compute

OpenAI has signed a five-year, $300 billion compute contract with Oracle.

The contract, set to start in 2027, will require 4.5GW of new power capacity. To put 4.5GW into perspective, it’s the equivalent of more than two Hoover Dams. Or, roughly a quarter of the US’s current operational data centre capacity.

The scale is staggering.

But, so is the risk.

OpenAI currently generates ~$10B in revenue, less than a fifth of its average yearly commitment to Oracle under the deal. That makes this as much a gamble for Oracle as for OpenAI. Oracle is staking hundreds of billions of dollars in future revenue on a single customer and will likely need to borrow heavily to finance the chip and data centre buildout.

Why this matters:

  • Oracle already carries one of the highest debt loads among hyperscalers.

  • Building capacity at this scale risks stretching its balance sheet further, especially if OpenAI’s growth falters.

  • Deals of this magnitude echo the internet buildout of the early 2000s: massive spend before mass adoption. Survival hinges on whether the demand materialises, and who has the longest runway.

Click here to read the full story in the Wall Street Journal.

Nebius Signs $17.4B Deal with Microsoft

Nebius just landed its biggest contract yet: a $17.4B, five-year agreement to deliver AI infrastructure directly to Microsoft.

Capacity will come from Nebius’s new Vineland, New Jersey data centre, with deployments starting later this year. Financing for the build will be covered through cash flows from the deal and debt secured against the contract, with Nebius also exploring additional options to accelerate growth beyond its original 2026 plans.

CEO Arkady Volozh called it “the first of many” long-term contracts with major labs and big tech players, positioning Nebius as more than just another AI cloud: a supplier trusted by the very hyperscalers it competes with.

Why this matters:

  • This deal puts Nebius in the same group as CoreWeave, Crusoe, and Fluidstack, who are also delivering for hyperscale offtake

  • Aligning with hyperscalers is the fastest way for neoclouds to grow, due to the bankability of the underlying contracts and the creditworthiness of the end customer.

  • A five-year contract with a marquee name like Microsoft goes a long way in bolstering Nebius’ growth plans, while also serving as a powerful tool for closing additional contracts with customers of a similar scale.

Click here to read the full announcement from Nebius.

MBZUAI and G42 Drop K2 Think

MBZUAI and G42 have launched K2 Think, an open-source, 32B parameter reasoning model with frontier performance.

Built on long-chain-of-thought fine-tuning, reinforcement learning with verifiable rewards, and agentic planning, the model is already topping charts. Benchmarks already place it at the top of open-source reasoning models in math competitions like AIME, HMMT, and OMNI-Math-HARD.

K2 Think will also run on Cerebras wafer-scale hardware with speculative decoding, pushing throughput to 2,000 tokens per second.

Why this matters:

  • Unlike most “open” models, K2 Think releases everything: training data, weights, and deployment code, making it one of the most transparent systems in the field.

  • Benchmark performance shows that “smarter, not bigger” (20x smaller than peers) reasoning systems can compete at the frontier.

  • K2 Think’s performance is huge for both MBZUAI and the wider Abu Dhabi AI ecosystem, positioning them as a credible hub for sovereign, open-source AI research with global reach.

Click here to read the full announcement from MBZUAI.

NVIDIA Unveils Next-Gen Rubin CPX

Blackwell is out. Rubin is in.

NVIDIA just announced Rubin CPX, a new class of GPU designed specifically for massive-context workloads. How massive? Think million-token coding and generative video. The flagship Vera Rubin NVL144 CPX platform supposedly delivers 8 exaflops of compute, 100TB of memory, and 1.7PB/s of bandwidth in a single rack.

Or, 7.5x the performance of today’s GB300 NVL72 systems.

NVIDIA says the economics scale aggressively: $5B in token revenue potential for every $100M invested.

Why this matters:

  • Training was yesterday’s bottleneck; inference at million-token scale is tomorrow’s. Rubin CPX is NVIDIA’s answer.

  • A $5B return on a $100M investment builds on the theme of “spend $1, get $5 back.” 50x projected returns, however, is pretty bold. Even for Team Green.

  • If the generational performance jump is anything like they say, who knows what Rubin will do to Blackwell GPU/h pricing and the neocloud market at large? Lepton, it seems, can’t come soon enough.

Click here to read the full announcement from NVIDIA.

Alibaba Raises $3.2B to Fuel International Expansion

Alibaba is issuing a zero-coupon convertible bond worth $3.2B, the largest of its kind this year, to bankroll international cloud growth.

Nearly 80% of proceeds will go to expanding data centres, upgrading technology, and scaling AI services. The rest will boost e-commerce ventures. The raise comes alongside Alibaba’s pledge to invest ¥380B ($53B) in AI over three years, with CEO Eddie Wu emphasising that AI is now central to revenue growth.

Why this matters:

  • Hyperscaler cloud expansion demands upfront cash, and Alibaba is turning to creative financing just as US peers lean on debt markets to fund their own AI buildouts.

  • Washington has frozen China out of cutting-edge GPUs and advanced memory. That makes Alibaba’s cloud not just a growth story, but also a distribution channel for new domestic hardware, such as Huawei’s Ascend.

  • Expect new Chinese chips, optimised for local open-source models, to find their first real route to market inside Alibaba’s data centres, wrapped in a China-steered regulatory environment built to favour sovereign AI stacks.

Click here to read the full story on Reuters.

Vantage Raises $1.6B, Acquires Yondr’s Johor Campus

Vantage Data Centers has raised $1.6B to accelerate its Asia-Pacific expansion, beginning with the acquisition of Yondr Group’s Johor campus in Malaysia.

The Johor site, located in Sedenak Tech Park, currently has 72.5MW of IT capacity with plans to scale to 300MW. At full buildout, Vantage’s APAC footprint will hit 1GW. The deal also simplifies DigitalBridge’s portfolio, since both Vantage and Yondr are DB-backed. The investment was led by GIC and ADIA, both existing stakeholders in Vantage.

Why this matters:

  • APAC’s demand curve is steep, and the Johor campus, already delivering facilities ahead of schedule, gives Vantage a boost in keeping up.

  • With GIC and ADIA doubling down, sovereign capital is continuing to underwrite the growth of AI data centres in Southeast Asia.

  • Consolidating Yondr’s Malaysian assets under Vantage signals DigitalBridge’s intent to concentrate execution in fewer, larger platforms with the capacity to scale.

Click here to read the full story on DCD.

The Rundown

Big names this week. And even bigger numbers.

The OpenAI x Oracle deal announcement briefly made Larry Ellison (Oracle CEO) the richest man in the world. The Nebius x Microsoft announcement sent the stock soaring 40%+. The fact that one man’s personal wealth can balloon $101B, or a Nasdaq-listed stock can behave like crypto, because of just one deal, should give you an idea of just how valuable these compute contracts actually are.

And just how much of a boon the alleged Nscale deal could be for both the neocloud and the UK.

Especially given current economic prospects.

NVIDIA, as ever, sits firmly in the middle of all this activity. Both Nebius and Nscale are NVIDIA-only neoclouds. We’ve seen nothing to indicate Oracle will be deploying alternative hardware at this stage (though OpenAI is developing its own hardware with Broadcom). And now, we’ve had the first look at Rubin, which, if Blackwell is anything to go by, will change the game. Again.

The neocloud market stratification continues.

And the capex is only moving in one direction.

See you next week.

Reply

or to participate