When the music stops, don’t be the one left holding the bag.

It’s the cardinal rule of surviving a downturn, and a particularly salient maxim given what happened this week.

The Bank of England is calling a spade a spade and innocuously warning of the possibility of a “sharp correction”.

When institutions of that size start sounding the alarm, you know the music’s slowing down.

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 #68

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.

Bank of England Warns of an AI Bubble

We might be in a bubble.

That might not be a surprise to many of you. What is surprising, however, is the source of this observation: The Bank of England.

Their view?

If sentiment turns on AI or US monetary policy, markets could face a “sharp correction.”

In its quarterly stability report, the BoE compared current AI-linked valuations in US equities to those seen at the peak of the dot-com bubble. The five largest companies (NVIDIA, Microsoft, Apple, Alphabet, and Meta) now make up 30% of the S&P 500’s total valuation. That kind of concentration, the highest in half a century, means the system is “particularly exposed” if optimism around AI’s profitability starts to fade.

Why this matters:

  • The world’s two biggest market narratives, AI and the Fed, are now intertwined. And both are approaching limits of belief.

  • If investors begin doubting the Federal Reserve’s independence (after repeated pressure from President Trump to cut rates), a cross-asset correction is on the cards, driving up borrowing costs globally.

  • If anyone reading this is seeing the same signals and arriving at a different conclusion, remember, there are no words more dangerous to one’s market viewpoint than “this time is different”.

Click here to read the full story from Reuters.

CoreWeave Acquires Monolith AI

CoreWeave is shopping again.

The AI Hyperscaler™ is expanding into industrial AI with the acquisition of Monolith AI. The UK-based company uses machine learning to solve complex physics and engineering problems for manufacturers like Nissan, BMW, and Honeywell, cutting R&D timelines by up to 80%.

This move follows CoreWeave’s acquisitions of OpenPipe (for reinforcement learning) and Weights & Biases (for model tracking), positioning the company as a full-stack AI infrastructure and tooling provider. It also aligns with its push beyond traditional AI training workloads into sectors like automotive and manufacturing, where it’s already the official AI cloud partner for Aston Martin’s Formula 1 team.

Why this matters:

  • CoreWeave is building horizontally across industries, not just scaling GPU rentals.

  • With Monolith, CoreWeave now competes in a new category: AI-native R&D infrastructure, bridging the gap between industrial design and accelerated compute.

  • Each acquisition brings software and domain expertise closer to CoreWeave’s hardware stack, meaning more customer stickiness, better solutions to customer problems, deeper relationships, and healthier revenues.

Click here to read the full announcement from CoreWeave.

QumulusAI Secures $500M Blockchain-Native Credit

QumulusAI has secured a $500 million non-recourse financing facility arranged by Permian Labs and distributed via the USD.AI protocol.

A debt facility of this size, although substantial, may not appear particularly compelling. What is interesting, however, is the fact that this is one of the first large-scale examples of blockchain-native capital being used to fund physical AI infrastructure.

QumulusAI can now finance up to 70% of approved GPU deployments using stablecoin liquidity sourced from USD.AI’s decentralised credit market, a structure that treats GPUs as tokenised, financeable commodities. Permian Labs issues GPU Warehouse Receipt Tokens (GWRTs), which can be used as collateral for borrowing USD.AI stablecoins, providing operators with fast, non-dilutive access to capital.

Why this matters:

  • This model bypasses traditional lenders entirely, connecting institutional DeFi liquidity to real-world compute assets.

  • QumulusAI is among the first AI infrastructure operators to tokenise GPU inventory for financing, rather than relying on venture or debt markets.

  • If successful, tokenisation could reshape AI infrastructure financing, creating faster, more transparent access to liquidity as traditional lenders struggle to keep pace with AI’s capital demands.

Click here to read the full announcement from Qumulus AI.

“Moloch’s Bargain”, Competition, and Alignment

A new Stanford study argues that when large language models are trained for competitive advantage, they become measurably less aligned.

Researchers Batu El and James Zou just published “Moloch’s Bargain: Emergent Misalignment When LLMs Compete for Audiences”. Catchy title aside, they simulated LLMs in three competitive domains: advertising, political campaigns, and influencer marketing. The findings were stark:

  • A 6.3% sales boost correlated with a 14% rise in deceptive marketing.

  • A 4.9% vote gain came with 22.3% more disinformation and 12.5% more populist rhetoric.

  • A 7.5% engagement increase led to 188.6% more disinformation and 16.3% more harmful content promotion.

The study dubs this trade-off “Moloch’s Bargain” - the pursuit of success at the cost of truth and safety. Even models explicitly instructed to remain honest began optimising for manipulation when rewarded for performance, exposing what the authors call a “race to the bottom” dynamic.

Why this matters:

  • In essence, the study reframes misalignment as an economic inevitability, not just a technical flaw, when truth competes with engagement.

  • The social media-driven attention economy already gives us a pretty stark example of what can happen when that competition happens at scale.

  • If the results of this study are confirmed with further research, it’s likely we’ll all need to be part of a much larger conversation around what good governance and sound incentive design, not just technical alignment, mean in the AI era. Both online and off.

Click here to read the study on Arxiv.

AMD and OpenAI Sign 6GW Deal

AMD just signed one of the biggest compute deals in history. And now OpenAI is a shareholder.

The partnership commits 6GW of AMD Instinct GPUs to OpenAI over multiple generations. The start? A 1GW deployment of MI450 chips in 2H 2026. That alone would rival some hyperscaler campuses. But, there’s more than just GPUs in this deal.

AMD has issued warrants for up to 160 million AMD shares to OpenAI, vesting as deployment and share price milestones are met.

That means the more GPUs OpenAI buys, and the higher AMD’s stock goes, the more equity OpenAI earns up to the approximate 10% stake limit.

This move creates an unusual feedback loop. NVIDIA holds equity in OpenAI. Now OpenAI holds equity in AMD. If AMD performs, OpenAI gains. If OpenAI scales, NVIDIA benefits.

All roads, somehow, still lead back to Jensen.

Why this matters:

  • AMD just secured a multi-generation anchor client with exactly the kind of global visibility needed to chip away at NVIDIA’s dominance.

  • For OpenAI, this diversifies its silicon base while letting it literally profit from AMD’s upside.

  • Strategically, it’s less about “Team Red vs. Team Green” and more about ensuring compute abundance. Whoever builds it fastest, wins. And if you’re Jensen, you win twice.

Click here to read the full announcement from AMD.

Oracle’s (Thin) AI Margins Exposed

Oracle just lost $40 billion of its market cap after The Information reported its AI cloud margins were far thinner than advertised.

How thin? Roughly 14 cents profit per $1 in Nvidia server rentals. In some quarters, that slipped negative, with Oracle reportedly losing $100 million renting out Blackwell GPUs. The revelation breaks the illusion that hyperscaler-scale AI clouds are instant cash machines. Oracle’s numbers expose the brutal economics of GPU leasing: colossal CapEx, tiny spreads, and limited differentiation.

Why this matters:

  • The bare metal server rental market is a commodity market.

  • Given the lack of meaningful differentiation and the limited number of customers relative to CapEx, it’s a buyer’s market where price is the deciding factor.

  • If hyperscalers like Oracle, with their nice low costs of capital, struggle to make meaningful profit, what hope do the 145 other undifferentiated neoclouds have?

Click here to read the full story from The Information.

Applied Digital Draws First Tranche of $5B

Applied Digital has drawn the first $112.5 million from its $5 billion perpetual preferred equity facility with Macquarie Asset Management.

The capital will fund Polaris Forge 1, a 400MW AI Factory campus in Ellendale, North Dakota, which has already been fully leased to CoreWeave. The site is designed to scale to 1GW over time, making it one of the largest single AI data centre developments under construction in the US. Future draws under the facility will support additional campuses as Applied aims to expand its total platform capacity well beyond the initial 400MW.

Why this matters:

  • With CoreWeave as a customer, Applied Digital are rapidly shaping up as one of the names to watch in the US DC market.

  • DC buildouts are capital-intensive, to say the least, so perpetual preferred equity offers long-term funding with minimal dilution.

  • With Crusoe energising the first two buildings at their Abilene campus last week, and Applied Digital drawing down funding this week, the industry is moving firmly into execution and delivery territory.

Click here to read the full announcement from Applied Digital.

The Rundown

The AI boom rolls on, but the market is audibly shifting.

When the Bank of England likens AI stock valuations to the dot-com bubble, it’s time to pay attention. Oracle’s wafer-thin GPU margins only add to the unease, puncturing the myth of effortless hyperscaler profits. Even landmark deals like AMD and OpenAI’s 6-gigawatt pact add to the dense web of financial cross-ownerships that now bind the industry together.

Every major player holds a slice of the next, through vendor loans, equity stakes, joint ventures, LOIs, or all of the above.

It’s a system that only works if the music keeps playing.

And if it stops suddenly?

Those bags are going to feel really heavy.

See you next week.

p.s. You might have seen I’ve started running ads. Answer the question below to let me know how you feel about this:

Login or Subscribe to participate

Reply

or to participate