First Move  ·  AI Hardware & Infra  · 
The build-out — Tuesday morning, 07 July

Google Unveils Split TPU 8t and 8i Chips for AI Training and Inference

AI's infrastructure bottleneck shifts from chips to power and land, with global data center demand projected to exceed supply by 500% by 2030, driving hyperscaler capital expenditure to $375 billion this year.

ShareWhatsAppXLinkedIn

The story

Google has revealed its eighth-generation Tensor Processing Units (TPUs), introducing a bifurcated strategy with dedicated chips for AI training and inference. The new lineup includes the TPU 8t for large-scale model training and the TPU 8i designed for rapid, real-time inference workloads.

The TPU 8t superpod is engineered to connect 9,600 chips, offering over 2 petabytes of shared memory and delivering 121 exaflops of compute power. This advanced system incorporates FP4 support for efficient math, Vargo for faster chip-to-chip communication, and Google's own ARM-based Axion CPUs as hosts, signaling a comprehensive AI infrastructure stack.

This development arrives as the industry grapples with increasing infrastructure complexity; for instance, Nvidia's Kyber NVL144 rack architecture has reportedly been delayed by over 12 months, pushing its availability towards 2028. The shift highlights a broader industry trend where the physical demands of scaling AI compute, including power, land, and system integration, are becoming as critical as raw chip performance.

Silicon

TPU 8t

Maker: Google

What: Designed for large-scale AI model training, connecting 9,600 chips with 2PB shared memory and 121 exaflops of compute.

For Whom: Google Cloud and internal AI operations

TPU 8i

Maker: Google

What: Optimized for fast AI inference, AI agents, and real-time responses with features like FP4 support.

For Whom: Google Cloud and real-time AI service deployment

Ascend 950PR

Maker: Huawei

What: Inference-focused neural network processing unit, reportedly delivering 2.87 times the inference performance of Nvidia's H20.

For Whom: South Korean AI accelerator market, as part of Atlas 950 SuperPod deployments

Ascend 950DT

Maker: Huawei

What: Designed for AI training workloads, scheduled for release in Q4 2026.

For Whom: South Korean AI accelerator market, as part of Atlas 950 SuperPod deployments

The build-out

ProjectWhoScaleWhere
Global Data Center CapacityIron Mountain and Structure Research forecastNearly 90GW by 2030Globally
Northern Virginia Data Center HubIndustry projections8.5GW by 2030Northern Virginia, USA
Dallas Data Center HubIndustry projections2.8GW by 2030Dallas, USA
London Data Center HubIndustry projections2.7GW by 2030London, UK
QTS Digital Gateway Data CenterBlackstone-owned QTSProject cancelledVirginia, USA

Supply & policy signals

Global data center demand to exceed supply by 500% by 2030

Implication: Intensified competition for suitable sites, utility access, and pressure to secure capacity ahead of deployment needs.

Hyperscaler capital expenditure projected to hit $375 billion in 2026

Implication: About half of this spending is directed towards servers and GPUs, driving demand for AI infrastructure.

Goldman Sachs forecasts data center power demands to double from 2025 to 2027, reaching 66 GW

Implication: Significant increase in infrastructure investments and enhanced market competitiveness for power solutions.

Server DRAM prices surged nearly 95% this year, with memory accounting for over 50% of new server costs

Implication: Enterprises face rising costs and longer procurement cycles, pushing towards software-defined approaches to optimize existing infrastructure.

Micron Technology's High Bandwidth Memory (HBM) supply chain is contracted and sold out through the end of 2026

Implication: HBM production, consuming three times the wafer supply of standard memory, remains a critical bottleneck for advanced AI accelerators.

Nvidia's Kyber NVL144 rack architecture reportedly delayed by over 12 months, pushing towards 2028

Implication: Provides an opening for rivals like AMD and Google to gain market share in system-level AI infrastructure.

What we'll be watching

Reporting + analyst voices: grounded via Google Search at publish time.