In partnership with
Seven billion dollars moved this week. None of it went to a model company.
A chipmaker put $2 billion behind a single neocloud and committed 5GW of next-generation systems to it by 2030. A year-old lab got a gigawatt of chips before shipping a single product. Europe's largest ever Series C closed. A hyperscaler reported half a trillion in contracted revenue. And somebody got permission to burn a lot of natural gas.
Different companies, different continents, same signal:
Physical infrastructure > everything else.
I'm Ben Baldieri, and 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 #96
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.
Today's issue is brought to you by Hyperstack
The GPU runs on sponsorships. If you value independent, no-fluff analysis of the AI infrastructure market, the sponsors are who make it possible. I'd appreciate you checking them out.
New region. New capacity. NVIDIA Blackwell available now in Europe.
Hyperstack's EU1 region (Sweden) is now online with NVIDIA RTX PRO™ 6000 Blackwell Server Edition — available on demand and via reservation.
EU1 has been built with NVIDIA B300 scale in mind: the power, cooling, and data centre capacity are already in place for next-generation cluster deployments.
Secure Private Cloud is coming. Early discussions are open now.
NVIDIA Invests $2 Billion in Nebius, Partners to Deploy 5GW by 2030
Last week, Jensen Huang said NVIDIA was pulling back from model company equity. This week, it put $2 billion into the neocloud that built its own full-stack AI cloud and is scaling it to 5GW.
NVIDIA and Nebius have announced a strategic partnership spanning AI factory design and engineering, inference software, fleet management, and multiple generations of NVIDIA hardware including Rubin, Vera CPUs, and BlueField storage. The $2 billion investment backs a target of 5GW+ deployed through Nebius by end of 2030. Huang framed it around agentic AI driving "incredible compute demand."
Why this matters:
5GW by 2030 through a single neocloud. We first profiled Nebius in Issue #20. The $17.4B Microsoft contract, the $4.2B raise, the 1.2GW Missouri approval, and now this. Ex-Yandex spinoff to one of the largest AI infrastructure operators on earth in roughly 18 months.
The inference and agentic AI focus is the forward signal. NVIDIA is co-developing Nebius's inference stack, providing early silicon access, and optimising fleet health. The model companies fight over Pentagon contracts. NVIDIA and Nebius are building the infrastructure layer underneath all of them.
NVIDIA Partners with Thinking Machines Lab for 1GW Vera Rubin Deployment
NVIDIA didn't stop at infrastructure.
Team Green and Thinking Machines Lab have announced a multi-year partnership to deploy at least 1GW of Vera Rubin systems, with a "significant investment" in the startup. Deployment targeted for early 2027. Thinking Machines, founded by former OpenAI CTO Mira Murati, raised $2 billion led by a16z with participation from NVIDIA, AMD, Cisco, and Jane Street.
Why this matters:
A gigawatt-scale NVIDIA partnership and direct investment puts Thinking Machines in the same infrastructure bracket as Meta's Indiana campus and Nebius's Missouri facility.
The company is building frontier models and platforms for "customizable AI at scale." "Customisable AI" is closer to the open-weight thesis we've tracked for months than to the closed model that the other major labs conform to.
NVIDIA backing that thesis with a gigawatt of Vera Rubin hardware is a bet on where inference demand goes next. Not one model serves everyone. Many models on many racks, customised per customer, running on NVIDIA silicon. The more models proliferate, the more GPUs get sold.
Nscale Raises $2 Billion in Europe's Largest Ever Series C
Nscale just raised the biggest Series C in European history.
$2 billion led by Aker ASA and 8090 Industries at a $14.6 billion valuation. The Aker-Nscale joint venture from July 2025 has been fully rolled into Nscale. The round builds on a $155 million raise in December 2024, a $1.1 billion round in September 2025, and a $1.4 billion GPU-backed term loan in February 2026.
The raise landed the same week as a Guardian investigation questioning UK AI investment commitments, reporting that Nscale's Loughton supercomputer site remains a scaffolding yard, and the government acknowledged it is "not playing an active role in auditing these commitments."
Why this matters:
Nscale was the benchmark Bloomberg named in last week's export control story: 200,000 GB300s for Microsoft across four sites. The capital to service that order is now in place. Whether Washington approves the hardware is the open question.
$14.6 billion valuation. CoreWeave went public at $23 billion. Nebius trades around $25 billion. At a time when the US government is drafting chip export controls and blacklisting AI companies, Nscale's European positioning becomes a source of differentiation, not a limitation.
The Guardian's reporting raises a broader question: whether capital commitments match physical reality on the ground. We flagged governance concerns in Issue #72. The $2 billion raise is independently reported and real. It’s a massive achievement - make no mistake there. Especially in Europe. But the pattern the Guardian identifies, governments claiming AI investment wins without verifying delivery, is a risk across every market where political incentives to announce deals outpace the infrastructure to audit them.
Meta Unveils Next Four Generations of Its MTIA Custom Silicon
Meta just mapped out its custom chip roadmap through 2027.
MTIA 300, 400, 450, and 500, all built on RISC-V architecture, co-designed with Broadcom, and fabricated by TSMC. MTIA 300 is already in production, powering ranking and recommendation systems across Facebook and Instagram. MTIA 400 has completed lab testing and is described as "competitive with leading commercial products." MTIA 450 and 500 target early and late 2027 respectively, each adding more high-bandwidth memory for inference. Starting with the 400, Meta has designed entire rack-scale systems around the chips, including liquid cooling.
Why this matters:
Meta is running three external silicon partners and an in-house programme developed with one of them. NVIDIA Blackwell and Rubin GPUs (millions committed). AMD Instinct at 6GW. Broadcom as both a standalone custom accelerator supplier at "multiple gigawatts" and the co-design partner on the MTIA line.
The training ambition is the new signal. MTIA is moving from recommendation inference into generative inference and eventually training. That's where custom silicon stops being cost optimisation and becomes a structural alternative to buying GPUs.
Meta will keep buying GPUs from everyone, but the fastest-growing share of its compute is silicon it designs itself. Long-term implications for NVIDIA's hyperscaler revenue concentration are obvious, even if short-term demand remains enormous.
OpenAI Acquires Promptfoo Red-Teaming Platform
OpenAI just bought the tooling that tests whether AI models are safe to deploy. The timing is not subtle.
OpenAI announced the acquisition of Promptfoo, the open-source LLM evaluation and red-teaming platform used by over 25% of Fortune 500 companies. Promptfoo's CLI and evaluation library have become standard tooling for testing AI applications in production. The acquisition brings evaluation, security testing, and red-teaming capabilities in-house at a moment when OpenAI's models are shipping into classified military environments, and NATO is in discussions for deployment on its networks.
Why this matters:
OpenAI now controls the evaluation pipeline for its own models at the exact moment external scrutiny is peaking. GPT-5.4 launched last week (Issue #95). The Pentagon deal closed the same week. NATO network deployment is under discussion. In-house red-teaming at this stage is an operational requirement, not a product feature.
Promptfoo was open-source and vendor-neutral. Over a quarter of Fortune 500 companies used it to test models from multiple providers. OpenAI's acquisition of it raises the question of whether that neutrality survives.
Competitors relying on Promptfoo for their own evaluation workflows now depend on OpenAI-owned tooling.
Oracle Stood Up 400MW Last Quarter. It Has 10GW of Power Secured Through Partners.
Oracle is scaling AI data centre capacity faster than it can report earnings. But it's not building most of it.
Oracle's Q3 FY2026 results show 400MW deployed in a single quarter and over 10GW of power secured through partners for the next three years, with over 90% of that capacity already funded. Remaining performance obligations (RPO, contracted future revenue not yet recognised) hit $553 billion, up 325% year-over-year. Most of that increase came from large-scale AI contracts where customers either prepay for Oracle to purchase GPUs or supply their own hardware.
Why this matters:
$553 billion in RPO sounds staggering until you read the fine print. A large share of those contracts are customer-funded: prepayments or bring-your-own-GPU models where Oracle provides the facility and the customer provides the silicon. Oracle is scaling as an infrastructure landlord as much as a cloud provider. That's a different margin profile than running your own GPU fleet.
10GW of secured power through partners, not 10GW of Oracle-owned capacity. Oracle leases most of its data centres. The 90%-funded figure means the capital risk sits largely with those partners, not Oracle's balance sheet. The $50 billion financing raise covers the remainder and Oracle's own capex. This model works until the partners' economics don't.
Oracle's RPO-to-deployment ratio is the number to watch: $553 billion in contracted revenue against 400MW per quarter means the backlog is growing far faster than the physical build-out. The gap between what's signed and what's operational is where the risk lives.
xAI Gets Approval for 41 Gas Turbines
Elon Musk's AI company just got permission to burn a lot of natural gas.
xAI has received approval to install 41 natural gas turbines at a site in Mississippi, generating 1.2GW behind the meter for its Colossus data centres. The turbines bypass the local grid entirely, providing dedicated power for xAI's AI training infrastructure. This is xAI's second major behind-the-meter gas deployment. Its South Memphis facility, which we covered expanding toward 2GW in Issue #80, already became one of Shelby County's largest emitters of smog-producing chemicals through a similar setup that experts say violates the Clean Air Act.
Why this matters:
xAI raised $6 billion to double its Memphis supercomputer (Issue #3), then bought a third building to push the campus toward 2GW (Issue #80), and Musk claimed xAI would outcompute "everyone combined" within five years. Mississippi is the next piece: 1.2GW of dedicated gas power for a second major site. The buildout is tracking exactly as Musk outlined.
Mississippi's approval came the same week Trump signed a voluntary pledge with tech CEOs to bring their own power for data centres.
xAI is the clearest case of what that policy produces in practice: fossil-fuelled, behind-the-meter generation that sidesteps grid constraints and environmental review simultaneously.
The Rundown
NVIDIA spent this week telling the market exactly where it stands. Not with the model labs. With infrastructure.
$2 billion into Nebius. A gigawatt into Thinking Machines. A quiet exit from OpenAI and Anthropic equity. The logic: model labs are politically exposed. Infrastructure operators are not. Chips flow to whoever builds the racks, regardless of which model runs on them.
The rest of the market followed the same signal. Nscale raised $2 billion. Oracle reported $553 billion in RPO with 10GW locked in. Meta mapped four generations of custom silicon. xAI got permission to burn gas at 1.2GW because the grid can't keep up.
Physical infrastructure is the scarce asset. Models are abundant. Silicon is available if you can pay. Power is available if you can permit it. The companies assembling all three at gigawatt scale are the ones NVIDIA is backing, Oracle is financing, and pension funds are underwriting.
The model wars make the headlines. The infrastructure build makes the money.
See you next week.
Everything else this week:
The Rundown
NVIDIA spent this week telling the market exactly where it stands.
Not with the model labs fighting over Pentagon contracts. Not with the closed-API providers getting blacklisted or the open ones getting praised.
With infrastructure.
$2 billion into Nebius. A gigawatt-scale partnership with Thinking Machines. A quiet exit from OpenAI and Anthropic equity.
Jensen Huang is building a portfolio of hardware deployment partners, not model company stakes.
The logic: model labs are politically exposed. Infrastructure operators are not.
Chips flow to whoever builds the racks, regardless of which model runs on them.
The rest of the market is following the same signal. Nscale raised $2 billion in Europe's largest Series C. Oracle reported $553 billion in RPO with 10GW of power locked in. Meta mapped out four generations of custom silicon to reduce its dependency on any single supplier. xAI got permission to burn gas at 1.2GW because the grid can't keep up.
Every story this week points the same direction: physical infrastructure is the scarce asset.
Models are abundant and getting more so. Silicon is available if you can pay. Power is available if you can permit it. The companies assembling all three at gigawatt scale are the ones NVIDIA is backing, Oracle is financing, and pension funds are underwriting.
The model wars make the headlines. The infrastructure build makes the money.
See you next week.