Version 5

The U.S. AI datacenter buildout faces converging infrastructure bottlenecks at three layers: grid connection approvals running ~1 GW/month against tens of GW of monthly requests[2], an internal power architecture transition to 800VDC driven by rack densities approaching 660 kW[9], and optical networking supply chains unable to keep pace with demand[16].

• Texas Instruments announced a complete 800VDC power architecture for AI datacenters co-developed with NVIDIA in March 2026[12], adding a named semiconductor vendor to the emerging 800VDC commercial ecosystem alongside Enphase's solid-state transformer.[13]
• ERCOT revised its 2030 Texas datacenter load forecast from 29.6 GW to 77.9 GW in a single planning cycle[1], but its own analysts have warned that AI datacenter demand projections may be overstated[7] — a self-contradicting signal from the same institution.
• SoftBank's France AI datacenter program has expanded to a confirmed 5 GW total capacity target, with a specific 1 GW partnership with French operator Sesterce[21][22], citing France's nuclear-heavy grid as the strategic rationale.

AI infrastructure buildout is now structurally constrained by energy policy, grid approvals, hardware supply chains, and networking components simultaneously — not capital alone. The entry of Texas Instruments alongside NVIDIA into the 800VDC architecture space signals that the power transition is moving from analytical roadmap to commercial product, while ERCOT's internal demand skepticism, if validated, would undercut the planning assumptions driving hundreds of billions in infrastructure commitments.

The U.S. electrical grid has become a primary constraint on AI infrastructure deployment. ERCOT's 2025 long-term load forecast revised projected Texas datacenter demand from 29.6 GW to 77.9 GW by 2030 — a near-tripling in a single planning cycle[1]. Formal interconnect approvals run at roughly 1 GW per month against tens of gigawatts of monthly new requests[2], and datacenter activity in Texas has been flagged as spiking grid reliability risks[3]. Facing this approval backlog, AI operators who own private generation assets have begun building parallel energy infrastructure: onsite natural gas plants that sidestep the interconnect queue entirely, giving vertically integrated operators a decisive construction-pace advantage[4][5][6]. Against this backdrop, a complicating counter-signal has emerged: ERCOT's own analysts have warned that AI datacenter power demand projections may be overstated[7], creating an internal contradiction at the institution whose forecast revisions have anchored the grid-bottleneck narrative.[8]

Alongside the grid bottleneck, a structural shift is underway in how datacenters distribute power internally. As GPU rack power approaches 600 kW and Nvidia's Kyber Ultra racks near 660 kW[9], the current 48–54V DC distribution standard has become physically untenable. Moving to 800VDC eliminates conversion stages, reduces resistive losses, and cuts facility-level power consumption by approximately 5%[9]. SemiAnalysis outlines a four-phase transition roadmap beginning H2 2026 with row-level sidecar retrofit units and culminating in 2028–2029 with centralized line infrastructure, projected to power approximately 39 GW of incremental datacenter capacity[10][11]. The commercial ecosystem around this transition is taking shape: Texas Instruments announced in March 2026 a complete 800VDC power architecture co-developed with NVIDIA[12], and Enphase has announced a distributed solid-state transformer targeting AI datacenters with volume deliveries aimed at 2028[13][14] — the first named commercial entrants for this architecture shift.[15]

A third supply chain constraint operates in optical networking. NVIDIA requested a 20x increase in InP laser capacity from supply chain partners between 2025 and 2030; vendors agreed only to 12x, leaving Datacom supply projected to remain approximately 50% below demand at end-2030[16]. Alternative optical interconnect architectures — co-packaged optics (CPO), linear-drive passive optics (LPO), and silicon photonics — are attracting growing attention as potential substitutes that could relieve InP supply pressure[17][18], though each involves distinct integration trade-offs. Amazon has also deployed Resilient Network Graphs across its datacenters, claiming a 69% reduction in hardware requirements and a 33% increase in network throughput[19][20], demonstrating that software-layer efficiency can partially offset physical supply constraints.

The constraint landscape has a consequential international dimension. SoftBank has pledged up to €75B to construct AI computing facilities in France, with the full program targeting 5 GW of capacity[21] — above the 3.1 GW first-phase figure initially reported. A specific partnership with French operator Sesterce targets a 1 GW facility[22]. The explicit strategic rationale is France's nuclear-heavy electricity grid: cheap, stable baseload power is being treated as the primary site-selection variable for large-scale AI training infrastructure. This commitment represents the U.S. grid bottleneck dynamic playing out as a global location competition — where operators unable or unwilling to build private gas plants can instead select jurisdictions with structurally stable power.