TPU shipments double as datacenter bets pile up
Broadcom's TPU shipments doubled to nearly 2 million H100-equivalents in a quarter, Lambda announced a $33.1 billion Alabama campus, Edera put TPUs inside isolated zones, and Blackfuel emerged with $250 million in contracted revenue.
- Epoch AI reports Broadcom TPU compute shipments doubled from fiscal Q2 to Q3 2026, reaching nearly 2 million Nvidia H100-equivalents. m:…4112
- Epoch AI reports TPU unit volumes rose about 65% quarter over quarter to roughly 800,000 chips in Q3 2026, with the mix shifting toward more powerful chips. m:…4112
- Lambda says its Red Clay Center Campus in Lowndes County, Alabama, is a $33.1 billion investment, with construction starting in early 2027 and taking up to five years. Self-reported m:…3842
- Edera says Google Cloud TPU workloads can now run inside isolated Edera Protect zones, with multiple zones able to share one TPU host. Self-reported m:…9431
- Blackfuel says it has secured more than $250 million in contracted revenue under multi-year agreements, with its first inference-only deployment in Barcelona on AMD Instinct MI355X GPUs. Self-reported m:…9911
A step change in TPU volume
Broadcom's TPU compute shipments doubled from fiscal Q2 to Q3 2026, according to Epoch AI, reaching the equivalent of nearly 2 million Nvidia H100 GPUs. The precise estimate is 1,965,979 H100-equivalents for the quarter. The previous three quarters had grown between 2% and 22% per quarter, so a doubling is a break in the trend, not a continuation of it. m:…4112
Unit volumes rose about 65% quarter over quarter to roughly 800,000 chips, and the production mix shifted toward more powerful chips. TPU v8i shipments alone accounted for 1,752,773 H100-equivalents. The estimates cover shipments to Anthropic and Google, and Broadcom currently oversees virtually all TPU production, with MediaTek expected to ramp production by the end of 2026. A single supplier doubling output while a second supplier prepares to enter is the shape of a market about to get more competitive. m:…4112
The concentration risk is obvious: virtually all production flows through one company today. But the volume itself is the news. Two million H100-equivalents in a quarter is capacity that has to live somewhere, and the datacenter announcements arriving in the same week suggest builders are already planning where it goes. m:…4112
Campuses, zones, and token grids
Lambda plans a large-scale data center campus in Lowndes County, Alabama, a $33.1 billion investment the company calls the Red Clay Center Campus. Construction is expected to begin in the first quarter of 2027 and take up to five years, creating more than 1,000 construction jobs and more than 225 full-time roles. Lambda says it will pay the full cost of power and grid infrastructure to protect Alabama Power ratepayers, and the facility will use closed-loop cooling to reduce water usage. m:…3842
The local economics are part of the pitch: projected local tax revenue above $2.8 billion through 2049, with more than $1.6 billion for Lowndes County Schools, plus a $10 million Community Benefits Agreement and a partnership with N-Drip to support local farmers in water conservation. Whether those projections hold over a 23-year horizon is unknowable today, but the structure is familiar: a very large campus, negotiated in public, with community payments attached. m:…3842
Edera's announcement is smaller but technically sharper. The company enabled Google Cloud TPU workloads to run inside isolated Edera Protect zones, using Xen paravirtualization to run TPU workloads in unprivileged guests, because Google TPU virtual machines do not expose hardware virtualization extensions. Multiple zones can run on the same TPU host, each with its own Linux kernel and IOMMU context for its assigned chips. m:…9431
The current implementation assigns whole TPU chips to zones rather than splitting chips between tenants, so multi-tenancy here means sharing a host, not sharing a chip. Edera says it has used the same paravirtualization and PCI passthrough work to run NVIDIA GPUs inside Xen zones as well. For operators, the point is isolation without dedicated hardware per workload, which is the prerequisite for renting TPU capacity the way GPU capacity is rented today. m:…9431
A new inference cloud, pre-sold
Blackfuel emerged from stealth as an inference-native neocloud provider, saying it has secured more than $250 million in contracted revenue under multi-year customer agreements before its first deployment is widely available. Its first inference-only deployment is in Barcelona, hosted by Digital Realty, built on liquid-cooled Dell PowerRack systems with Dell PowerEdge XE9785L servers, and running AMD Instinct MI355X GPUs with ROCm software. m:…9911
Blackfuel describes its platform as an AI Token Grid, a network of accelerator clusters connected through a common software and commercial layer, and is introducing Logical GPUs that let customers reserve a share of a model's serving capacity for a fixed monthly fee. Its infrastructure uses open serving technologies including vLLM and SGLang, and its token platform plans to support models from Meta, Mistral, DeepSeek, Qwen, GLM, Moonshot AI, OpenAI, MiniMax, Cohere, NVIDIA, Liquid AI, and Google at launch. m:…9911
The company plans to offer dedicated GPU clusters, managed inference, and reserved serving capacity, and is developing an agent-driven process to build, benchmark, and optimize inference runtimes for each model and accelerator architecture. Contracted revenue before launch is a strong signal, but it is still a vendor claim; the test will be whether the Barcelona deployment converts reservations into sustained utilization. m:…9911
What is still unresolved
- Whether Broadcom's TPU volumes stay elevated or revert toward the earlier 2% to 22% quarterly growth range.
- Whether MediaTek's expected TPU ramp by end of 2026 materializes and diversifies supply.
- Whether Blackfuel's contracted revenue converts into operating deployments at scale.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.