Google is expanding external adoption of its custom AI chips, the TPUs, through financing and ecosystem-building tactics modeled on Nvidia's own approach, as it challenges Nvidia's grip on the AI accelerator market, according to multiple reports.
AI Accelerators · Google vs Nvidia
Google Borrows Nvidia's Playbook to Sell Its TPU Chips
Alphabet is opening its in-house Tensor Processing Units to outside buyers — backing data centers with billions in financial guarantees and launching direct sales to chip away at Nvidia's 90%+ grip on the AI accelerator market.
Billions pledged to fill TPU data centers
$1.4B
Colorado City, TX
$3.2B
Lake Mariner, NY
$5B
Blackstone AI
$7B
River Bend, LA
Column heights scaled to dollar value · "circular financing" lets customers raise cheap capital to use TPUs.
8th-gen TPU — split into two chips
TPU 8t
Training-focused
12.6 FP4 PFLOPs/chip 216 GB HBM 6,528 GB/s bandwidth 121 FP4 exaflops / Superpod (9,600 chips)
TPU 8i
Inference-focused
10.1 FP4 PFLOPs/chip 288 GB HBM 8,601 GB/s bandwidth 384 MB SRAM — triple prior gen
Platform scales to clusters of up to 1 million TPUs .
Gains vs prior-gen "Ironwood"
2.7×
better performance-per-price on TPU 8t (training)
+80%
improvement on TPU 8i (inference)
90%+
of the accelerator market still held by Nvidia
Customers see upside
Citadel Securities: costs cut 30% , speed up to 4× on some workloads
Anthropic trains large models under multi-gigawatt TPU contracts
Developers cite price-performance & energy efficiency as a realistic alternative
The skeptics
Nvidia's Jensen Huang: ASICs have "limited market reach"
CUDA ecosystem lock-in persists
Operators fear losing Nvidia allocation if they switch
Within Google Cloud, mixing TPUs with Nvidia GPUs stays mainstream — "partner and competitor"
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…