You can tell a lot about an AI infrastructure deal by who's standing behind it.
Microsoft's offtake anchored IREN's A rating. Berkshire wrote a tenth of Alphabet's $84.75 billion equity raise. NATO and ENISA are running pre-release Mythos. Sam Altman chairs Helion.
Big shiny names carry weight at the best of times.
Doubly so in an industry that runs on attention and where the marginal cost of PR is close to zero.
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.
IREN Closes $3.65B Investment-Grade GPU Debt, the First in the US Private Placement Market
This is what institutional looks like.
IREN closed a $3.65 billion investment-grade financing facility on June 1 to fund its AI cloud contract with Microsoft. The structure: $2.10 billion US private placement at SOFR+2.13% fixed, $1.55 billion delayed draw term loan at SOFR+2.25% floating, rate hedges in place. With Microsoft prepayments, the facility covers 96% of the $5.81 billion total GPU capex at a blended 3.31%. Fitch rated it A. DBRS rated it A(low). Goldman Sachs and JP Morgan ran the books, both ratings anchored on Microsoft's offtake covenant.
Why this matters:
GPU debt just crossed from venture-backed to institutional. Insurance companies, pension funds, and sovereigns can buy A-rated paper. They couldn't buy what neoclouds were issuing 18 months ago.
Microsoft's offtake is the credit anchor. The rating exists because the cashflow underneath is a Microsoft contract, not GPU economics. Get an investment-grade counterparty, and investment-grade financing follows.
3.31% blended is hyperscaler-tier cost of capital. Most neocloud debt has cleared at high single to low double-digit rates over the past 18 months. The financing arbitrage between A-rated neoclouds and the rest of the field just opened wider than it has ever been.
Helion Raises $465M Series G at $15.5B Valuation to Build Fusion Power for AI
The power for AI just became institutional too.
Helion announced a $465 million Series G on June 4, led by Thrive Capital. The round triples Helion's post-money valuation to $15.5 billion and brings total capital raised to $1.5 billion. New investors include Alta Park, Anti Fund, BoxGroup, Lux Capital, Peak XV, and Bill Ford. Existing investors who participated include Lightspeed, Mithril, Capricorn, SoftBank Vision Fund 2, Dustin Moskovitz's Good Ventures, and a university endowment.
Why this matters:
The thing AI infrastructure cannot print is power. Helion is the bet that fusion gets there in time to matter. Thrive leading a $15.5 billion valuation is the capital market saying it might.
Sam Altman is Helion's chairman. The OpenAI–Microsoft–Helion triangle is the longest-dated AI infrastructure trade in the market right now.
The round lands days after Helion's 7th-generation Polaris prototype hit two milestones. Polaris became the first privately funded fusion machine to operate with deuterium-tritium fuel, and broke the company's own plasma temperature record at 150 million degrees Celsius. Capital is earmarked to scale manufacturing for commercial fusion deployment, including the previously announced Microsoft power plant.
CoreWeave Brings Up the First Operational Vera Rubin NVL72, Delivered by Dell
A neocloud got there first. Again.
CoreWeave completed bring-up and validation of NVIDIA's first Vera Rubin NVL72 on June 1. The rack arrived from Dell on May 31, passed NVIDIA's L11 diagnostics and a 147-hour stress suite, and was production-validated. Built on Dell's liquid-cooled PowerEdge XE9812, the NVL72 packs 72 Rubin GPUs and 36 Vera CPUs in a single rack, connected through 260 TB/s of 6th-gen NVLink, delivering roughly 3.6 exaFLOPS for trillion-parameter training.
Why this matters:
For the second generation in a row, a neocloud lit up the newest silicon before the hyperscalers did. CoreWeave was first with B200, first with B300, now first with Vera Rubin. Too consistent to be coincidence.
Dell is the OEM partner. PowerEdge XE9812 is the reference design. All the other OEMs now benchmark against Dell's delivery cadence.
NVIDIA claims Vera Rubin delivers up to 10x better inference per watt, one-fourth the GPU count for the same workload, and one-tenth the cost per million tokens versus Blackwell. If these claims are borne out in production, redraw inference unit economics. The companies on Blackwell today face a depreciation question they didn't have last quarter.
Anthropic Scales Project Glasswing to 150 Critical Infrastructure Operators in 15+ Countries
Power. Water. Healthcare. Telecoms. Anthropic just got security access to all of them, across 15 countries.
Anthropic extended Project Glasswing access to approximately 150 new organisations across more than 15 countries on June 2. The cohort spans power, water, healthcare, communications, and hardware infrastructure operators. Named partners include Okta, Samsung, SK Hynix, SK Telecom, NATO, and ENISA, the EU's cybersecurity agency.
Why this matters:
At the core is Claude Mythos Preview, Anthropic's most capable model for zero-day vulnerability discovery. The initial cohort of around 50 partners surfaced more than 10,000 high or critical-severity flaws using Mythos. Public release is gated until safeguards meet what the company calls "highly robust" criteria.
Frontier model deployment is moving from consumer apps to critical national infrastructure. The organisations on this list collectively run the systems 100 million-plus people depend on for power, water, and communications.
Mythos finding 10,000 critical flaws across 50 codebases means defenders just got the same capability attackers will have when an equivalent open-source model ships. Anthropic's release strategy is now a national security decision.
This is also a distribution moat. Anthropic now has 150 institutional partners running pre-release Claude inside their SOCs. Most of them don't switch vendors.
Microsoft Ships Maia 200 into Production, Launches Cobalt 200 VM Preview at Build 2026
Microsoft just told the world it doesn't have to wait for NVIDIA.
At Build 2026 on June 3, Microsoft confirmed Maia 200 is in production in Azure's US Central (Iowa) and US West 3 (Arizona). Italy, Australia, and South Korea are queued next. The chip is delivering 30% more tokens per dollar than what Microsoft calls "the leading GPU today," and now powers Microsoft 365 Copilot inference. Microsoft also opened the early access preview for Cobalt 200 Arm-based VMs across eight Azure regions, with up to 50% better generational performance and 135% better cloud database performance than Cobalt 100.
Why this matters:
Hyperscalers are now genuinely diversified at the silicon layer. Maia 200 in production means Azure no longer needs NVIDIA for every inference workload. Pricing leverage runs both directions.
Cobalt 200 brings Arm-based VMs into agentic AI workloads. ByteDance and Oracle also signed up for Arm AGI CPU adoption this week. The x86-vs-Arm hyperscaler split is closing fast.
"30% more tokens per dollar" is the metric for inference unit economics. If verified at scale, it changes the calculus for every enterprise buyer locked into Copilot, Bedrock, or Vertex.
Alphabet Upsizes Equity Offering to $84.75B, Backed by Berkshire Hathaway
The largest equity offering in tech history priced this week. Berkshire wrote a tenth of it.
Alphabet's upsized $84.75 billion equity raise is the largest in the history of publicly traded technology. The structure: $10 billion private placement with Berkshire ($5 billion Class A at $351.81, $5 billion Class C at $348.20), $40 billion at-the-market programme in Q3, and the balance in priced public offerings. The first tranche was oversubscribed and upsized to $45 billion on the day, and the total grew from $80 billion announced June 1 to $84.75 billion priced two days later. Alphabet now projects 2026 capex of $180 to $190 billion, citing demand for AI from enterprises and consumers "at levels that are exceeding the company's available supply."
Why this matters:
This is hyperscaler capex translated into public equity at scale. Google needs more capacity than it can build, and it just funded the next round of GPU, custom silicon, and data centre orders in one trade.
$180-190 billion of capex in one year is bigger than the GDP of New Zealand. One company. AWS, Azure, and Meta will all match or exceed in 2026.
Berkshire at $10 billion is the signal. The most famously slow-moving capital allocator on earth just doubled down on AI capex on top of its existing Apple position.
SoftBank Commits €75B to Build 5GW of French AI Infrastructure
Sovereign-adjacent AI just got its biggest single-country cheque.
SoftBank Group announced on May 31 a commitment of up to €75 billion to develop 5GW of AI data centre capacity in France, its largest AI infrastructure investment in Europe to date. The first phase commits €45 billion to deliver 3.1GW across three Hauts-de-France sites by 2031: Dunkirk (Loon-Plage), Bosquel, and Bouchain.
Why this matters:
France is the first European country to lock in committed multi-gigawatt sovereign AI capacity at Stargate or UAE scale. €75 billion is roughly 2x the UK's entire planned AI infrastructure budget.
SoftBank picking Sesterce as operating partner anchors a domestic champion. This is the as-yet-undelivered model the UK tried with Nscale.
The Bosquel site runs through a majority-SoftBank JV with French neocloud Sesterce, building a 1GW AI factory. Schneider Electric joins on the Dunkirk industrial production cluster. State-owned EDF transfers a decommissioned Bouchain power plant to SoftBank for conversion into a data centre.
The Rundown
The connecting thread this week was the same word written seven different ways: institutional.
A-rated GPU debt. The largest equity offering in tech history, anchored by Berkshire. A neocloud booting Vera Rubin ahead of every hyperscaler. A frontier model embedded inside critical national infrastructure. Custom silicon at production scale. Fusion power capitalised at institutional scale. Sovereign-adjacent capital backing a country's AI buildout. None of these would have looked plausible 18 months ago. They all happened inside the same five-day window.
When the buyers go institutional, the sellers follow. Cap tables fill with insurance balance sheets and conglomerates. Debt placements rate A. Neoclouds beat hyperscalers to the newest silicon. The companies that were strategic become structural.
The kicker, however, is that while Institutional infrastructure makes the underlying volatility less visible, it does not make it any less real. The risk moves to parts of the balance sheet that move slowly. Which is exactly where, the last time around, it ended up taking everyone by surprise.
See you next week.