Signal Briefing: June 10, 2026

Broadcom, Apollo, and Blackstone Launch $35B, 20GW XPU Platform for Anthropic

Broadcom, Apollo Global Management, and Blackstone have jointly launched a 20GW XPU compute platform with an initial $35 billion commitment, structured to serve Anthropic’s infrastructure needs via Fluidstack, according to Data Center Dynamics (source). The platform combines Broadcom’s custom ASIC design capability with Apollo and Blackstone’s balance sheets to fund large-scale inference and training capacity outside the hyperscaler model.

Why this matters. This is the clearest signal yet that frontier AI labs are building parallel capital stacks — purpose-financed infrastructure backed by private equity rather than hyperscaler cloud contracts — which restructures how inference capacity gets funded and who captures margin. At 20GW, the platform would represent a material fraction of current total U.S. data center power capacity; even partial execution changes the supply curve for AI compute.

Confidence: high — reported by Data Center Dynamics with named principals and dollar figure; no confirmed contradicting sources.

---

OpenAI in Talks for a 10GW Ohio Data Center on Federal Land

OpenAI is in active negotiations to lease a 10GW data center from SB Energy in Ohio, sited on federal land, according to Data Center Dynamics (source). No build timeline or lease terms were disclosed. The project is separate from the Stargate consortium campuses already announced.

Why this matters. A 10GW single-site commitment would be roughly ten times the scale of today’s largest operational AI campuses; even at a fraction of nameplate power, it signals that OpenAI is pursuing dedicated sovereign infrastructure at a scale that outpaces what hyperscaler co-location can realistically absorb. The federal land siting also suggests possible regulatory or permitting advantages that private land acquisitions don’t offer.

Confidence: medium — single-outlet report, negotiations ongoing, no signed agreement disclosed.

---

China Drafts $295B National AI Data Center Grid Targeting 80% Domestic Silicon

China is drafting a plan to deploy roughly 2 trillion yuan (~$295B) over five years to build a nationwide AI data center grid, with a target of running on 80% domestically produced chips by 2028, per Tom’s Hardware (source). The plan remains in draft form and the outlet notes the 2028 timeline may strain the current output capacity of China’s domestic chip industry.

Why this matters. The 80% domestic silicon target is the structural bet worth watching: it is simultaneously an industrial policy lever and a constraint that will force Chinese hyperscalers and cloud providers to absorb the cost and performance gap of non-TSMC, non-NVIDIA hardware at national scale. If the domestic chip supply falls short — a real risk given CoWoS and HBM constraints outside the U.S.-allied supply chain — the grid build could proceed at degraded utilization, creating a test case for whether state-directed capex can substitute for chip supply access.

Confidence: medium — reported from a draft policy document; final figures and adoption timeline unconfirmed.

---

SpaceX Reveals AI1 Satellite Architecture and Gigasat Factory, Targets 1GW Orbital Compute by 2027

SpaceX detailed the AI1 satellite design ahead of a Friday IPO: the spacecraft spans wider than a Boeing 747, carries a 120kW compute payload peaking at 150kW, and uses an interchangeable chip architecture, per Data Center Dynamics (source) and Tom’s Hardware (source). Simultaneously, SpaceX unveiled an 11-million-square-foot Gigasat manufacturing facility targeting 1GW/year of space AI compute capacity by late 2027 and 100GW/year by 2030.

Why this matters. The 100GW/year 2030 projection is almost certainly aspirational, but the manufacturing facility announcement and the IPO-eve disclosure suggest SpaceX is treating orbital compute as a capital markets story, not just an R&D project. The interchangeable chip payload is the architecturally interesting detail: it implies a servicing model that lets the compute layer evolve independently of the orbital platform — a potential answer to the obsolescence problem that has historically made space-based compute economically unviable.

Confidence: medium — primary disclosure from SpaceX ahead of IPO, but projections (especially 100GW/year by 2030) are management guidance with no independent verification.

---

Apple Routes Private Cloud Compute to Google Cloud on Nvidia GPUs

Apple’s Private Cloud Compute (PCC) — the server-side inference layer for Apple Intelligence — will run on Google Cloud infrastructure using Nvidia GPUs with confidential computing enabled, per Data Center Dynamics (source). Apple had previously emphasized PCC’s privacy architecture as running on Apple-controlled hardware; the shift to third-party cloud marks a notable change in the infrastructure model.

Why this matters. Apple is the largest consumer hardware platform deploying on-device and server-side inference at scale; routing PCC to Google Cloud rather than building owned capacity means Google captures inference revenue from Apple’s installed base, and Nvidia captures the GPU attach. It also confirms that even a company with Apple’s capital and vertical integration finds it faster to lease inference capacity than build it — reinforcing the pattern of inference workloads flowing toward whoever can provision GPU clusters fastest.

Confidence: high — reported by Data Center Dynamics; the confidential computing and Nvidia GPU details are specific enough to suggest primary sourcing.