Good morning 👋

AI infrastructure had one of those days where the headlines look unrelated until you pull back. The pattern is clear: developers are trying to lock down customers, capital, cooling, optics, chips and power years before much of the promised compute actually turns on.

The most important AI-infrastructure disclosure Monday was a glimpse into what developers may be willing to pay to make enormous future campuses financeable.

Draft IPO documents reviewed by The Wall Street Journal show SoftBank-backed SB Energy granted OpenAI warrants valued at an estimated $5.5 billion to secure the AI company as a major data-center tenant. Reuters reported the disclosure but said it had not independently verified the draft documents. The Journal also says SB Energy has nearly 9 GW of contracted computing capacity, mostly tied to southern Ohio, despite having no operating data centers yet; about 800 MW is under construction.

The underlying project is officially confirmed. OpenAI has agreed to secure roughly 8 IT-GW at SB Energy’s PORTS-Pike campus. Nvidia will be the exclusive AI-compute infrastructure provider and has announced a $1.5 billion investment in SB Energy. Capacity is scheduled to begin coming online in phases in 2028 under a 20-year lease.

The scale becomes clearer when you look beyond the servers. The plan calls for at least 10 GW of new generation and at least $4.2 billion of regional grid investment. DOE has described the generation plan as including 9.2 GW of natural gas.

Why it matters: this is no longer a simple “build a data center, then lease it” model. OpenAI supplies long-duration demand; SB Energy develops the site and power; Nvidia supplies compute, invests in the developer and provides credit support; utilities and government entities enable the grid and land.

The warrant package adds another layer: OpenAI is not just the customer, but economically tied to the developer it helps make financeable. That may unlock projects, but it also makes the ecosystem more interdependent.

A signed lease is not a substation. Eight contracted gigawatts are not eight energized gigawatts. The next bottleneck is proving that finance, generation, grid work and construction can all arrive on the same timeline.

SLB agreed to acquire heat-exchanger and thermal-management specialist Kelvion for about $4.1 billion, including assumed debt. The deal is signed and expected to close in the first half of 2027.

Kelvion expects $1.2–1.3 billion of its 2026 revenue to come from data centers, already its largest and fastest-growing end market. That is the number that matters more than the acquisition multiple.

Why it matters: as rack power density rises, heat rejection moves from background facility plumbing into the critical path. SLB is betting that expertise in fluids, thermal systems and modular industrial construction transfers directly into AI campuses.

The bigger signal is strategic. Cooling has become important enough for a major industrial company to spend billions buying scale rather than building it slowly in-house. In the AI buildout, the hottest component may increasingly be the one that keeps everything else from overheating.

Saudi Arabia’s AI buildout crossed status lines Monday, and keeping them separate matters.

HUMAIN, AMD and Cisco said their first AMD Instinct MI355X systems are live and serving customers. They did not disclose the size of that deployment. The next phase is planned for up to 250 MW, with capacity expected in H2 2027.

Separately, HUMAIN and DataVolt said construction is underway on 100 MW at Oxagon in NEOM, inside a 360 MW first phase. Reuters says the broader campus vision is 1.5 GW.

Why it matters: Saudi Arabia is moving from partnership announcements toward infrastructure, but not all megawatts are equal. Some capacity is live, some is under development, and much more remains planned.

Do not simply add the headline MW figures; the disclosures do not establish that every project is non-overlapping. The real signal is execution—and a buildout designed for multiple vendors.

Nvidia invested $3.5 billion in convertible bonds issued by MediaTek, while MediaTek agreed to adopt NVLink Fusion for custom AI accelerators.

The strategic point is bigger than the financing. NVLink Fusion gives hyperscalers, cloud providers and frontier-model developers a way to connect custom chips into Nvidia’s rack-scale interconnect architecture. Nvidia made a similar move with Marvell earlier this year.

Why it matters: custom silicon is one of the clearest ways large AI buyers can reduce dependence on off-the-shelf GPUs. Nvidia’s response is not simply to fight every custom accelerator. It is trying to make those accelerators plug into Nvidia.

If that works, Nvidia can lose some accelerator share while still owning valuable pieces of the surrounding infrastructure—interconnect, networking and rack architecture. The investment also reinforces a broader theme: in AI infrastructure, technology strategy and balance-sheet strategy are becoming increasingly difficult to separate.

Soitec is moving photonics customers toward multiyear capacity-reservation agreements with fixed pricing and deposits for silicon-on-insulator substrates used in optical connectivity.

CEO Laurent Rémont told Reuters that roughly 80% of agreements involving more than 10 photonics customers were expected to be signed within a week or two. Soitec expects photonics-SOI revenue to more than double this financial year from a little over $100 million.

Why it matters: the important signal is the contract structure. Deposits and reserved capacity appear when buyers start worrying that future availability matters more than squeezing today’s price.

Larger AI clusters create a data-movement problem alongside compute, pushing networking toward optics as electrical links hit power and reach limits. Soitec says it can handle near-term growth without a new fab, so this is not proof of a structural shortage. It is evidence that photonics has moved higher on the bottleneck list.

EuroHPC signed a €387.8 million procurement contract with Bull for LUMI-AI, a new supercomputer in Kajaani, Finland, using AMD Instinct MI430X GPUs and sixth-generation EPYC processors. User availability is expected in 2027.

EuroHPC says the system should deliver about 10 times the AI capacity of the current LUMI machine. Reuters reports that demand for Europe’s existing public AI-compute program already exceeds supply strongly enough that some applications are being rejected.

Why it matters: €388 million is tiny beside an 8-GW private campus, but that misses the point. Europe is increasingly treating advanced compute as strategic public infrastructure.

It is also another visible AMD win in a market where buyers have reasons to diversify suppliers: cost, availability and sovereignty. The immediate bottleneck is simple—Europe already has more qualified demand for shared AI compute than it can serve. Governments are becoming infrastructure buyers too.

In plain English:
A gigawatt (GW) is one billion watts of electrical capacity. It’s the scale used when talking about very large power plants and AI data-center campuses.

Why you’re hearing about it:
OpenAI’s Ohio project is planned at roughly 8 IT-GW and calls for at least 10 GW of new power generation.

Why it matters:
AI campuses are reaching utility-scale power demand. At that size, getting electricity and a grid connection can be harder than getting servers.

🤯 Did You Know? The Ohio project pairs roughly 8 IT-GW of planned compute capacity with at least 10 GW of new power generation.

Till next time,
Grid Boiler Team