---
type: "post"
id: "much-of-broadcom-s-increase-in-shipped-compute-532104"
slug: "much-of-broadcom-s-increase-in-shipped-compute-532104"
title: "Much of Broadcom’s increase in shipped compute came from chip volume: Broadcom shipped 65% more chips, or roughly 800,000 in total."
description: "Epoch AI’s AI chip sales database estimates shipments, compute capacity, power use, and spending across major accelerator designers."
retrieval_nugget: "Epoch AI’s AI chip sales database estimates shipments, compute capacity, power use, and spending across major accelerator designers. Its method highlights why aggregate supply can rise through both greater chip volume and stronger chips; builders should not treat a headline compute figure as a measure of online, usable capacity."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["google"]
entities: ["Epoch AI","Nvidia","Google","Amazon","Broadcom","AMD","Huawei","TPU v7"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Epoch AI estimates Broadcom shipped roughly 800,000 AI chips, a 65% increase, including TPU v7."
prior_refs: []
source_url: "https://epoch.ai/data/ai-chip-sales?cumulative=false&tab=h100_equivalents&view=graph"
source_urls: ["https://epoch.ai/data/ai-chip-sales?cumulative=false&tab=h100_equivalents&view=graph"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:much-of-broadcom-s-increase-in-shipped-compute-532104"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/much-of-broadcom-s-increase-in-shipped-compute-532104/","markdown":"https://newruntime.com/posts/much-of-broadcom-s-increase-in-shipped-compute-532104.md","json":"https://newruntime.com/posts/much-of-broadcom-s-increase-in-shipped-compute-532104.json"}
markdown_url: "https://newruntime.com/posts/much-of-broadcom-s-increase-in-shipped-compute-532104.md"
---

# Much of Broadcom’s increase in shipped compute came from chip volume: Broadcom shipped 65% more chips, or roughly 800,000 in total.

## Retrieval answer

Epoch AI’s AI chip sales database estimates shipments, compute capacity, power use, and spending across major accelerator designers. Its method highlights why aggregate supply can rise through both greater chip volume and stronger chips; builders should not treat a headline compute figure as a measure of online, usable capacity.

## Summary
Epoch AI’s AI chip sales database estimates shipments, compute capacity, power use, and spending across major accelerator designers. Its method highlights why aggregate supply can rise through both greater chip volume and stronger chips; builders should not treat a headline compute figure as a measure of online, usable capacity.

## What happened
Epoch AI describes its AI chip sales explorer as an open database of estimated sales and shipments for dedicated AI accelerators. The page says it tracks Nvidia, Google TPUs, Amazon Trainium and Inferentia, AMD Instinct, and Huawei Ascend, with estimates broken down by manufacturer, chip model, and period. The dataset is based primarily on chip volumes, revenue, and prices drawn from company earnings commentary, company disclosures, analyst estimates, and media reports. Epoch says the goal is to understand AI progress and societal impacts, rather than to provide industry or financial analysis.

The explorer presents several measures: compute capacity in H100-equivalents, chip power, estimated cost, and number of units. Epoch defines an H100-equivalent as the ratio of a chip’s peak dense 8-bit operations per second to the Nvidia H100’s specification, multiplied by the number of chips. The organization gives the example that one TPUv7 can perform about 2.3 times as many operations as an H100, so one million TPUv7s represent 2.3 million H100-equivalents. It chose the H100 as a familiar reference chip and 8-bit formats because Epoch understands them to be widely used in inference and at major labs for training.

Epoch explicitly calls H100-equivalent compute an imperfect proxy. It reflects paper specifications, not necessarily real-world performance: memory capacity and bandwidth, software, bugs, utilization, and networking can change what a chip or cluster delivers. The database does not show how chips are arranged into servers and clusters. Its power estimates multiply each chip’s thermal design power by estimated quantities; Epoch says a data center’s total draw is typically roughly twice the chips’ combined TDP because of infrastructure overhead. Cost estimates multiply approximate average customer purchase prices by chip counts and do not include servers, clusters, or facilities.

The page says the underlying figures are estimates because manufacturers do not consistently disclose exact sales. Epoch models 90% confidence intervals when feasible; most span roughly a factor of two, or about 1.5 times above and below the median. It describes Nvidia estimates as its highest-confidence figures, based on compute revenue, extensive reporting, analyst coverage, and direct shipment disclosures. Amazon Trainium estimates are the least certain, relying on indirect evidence and limited analyst coverage. Epoch aims to track chips delivered and ready for installation or use, not necessarily already online. It says estimates for Nvidia and AMD reflect deliveries recognized as revenue, Google TPU estimates use Broadcom revenue and likely reflect chips produced and delivered to Google, while methods differ for other designers.

The page says the dataset was updated October 21, 2026, and that it dates back to 2022 for Nvidia, 2023 for Google, and 2024 for other designers. It covers what Epoch considers the large majority of global AI compute capacity, but excludes several smaller or less publicly documented suppliers and products such as consumer GPUs. Epoch offers downloadable datasets and says the data may be reused under Creative Commons Attribution with source and authors credited.

## What the archive adds
New Runtime’s earlier coverage introduced Epoch AI’s Inference Performance Index, which compares chips by modeled concurrent inference capacity rather than raw FLOP/s alone ([Epoch AI’s Inference Performance Index](https://newruntime.com/posts/epoch-ai-launched-the-inference-performance-index-scoring-1bb9b1/), November 10, 2026). This is a new data point in a related thread: the chip-sales explorer tracks estimated shipments and capacity, while the earlier index focuses on inference performance.

## Trend signal
This supports the broader infrastructure thread that chip supply and usable compute need more than one measure. Epoch’s database separates unit counts from H100-equivalent peak compute and warns that neither alone captures memory, software, utilization, or cluster networking. It does not directly establish a cost-per-completed-task trend. This is a useful measurement signal, not a direct fit for a tracked pattern.

## Why it matters
Teams planning model training or inference capacity can misread aggregate compute growth if they treat it as a single number. A larger shipment total may reflect more chips, newer generations, or both; the operational result depends on whether the hardware is delivered, installed, networked, and supported by software that keeps it busy. H100-equivalents make unlike accelerators easier to compare on peak 8-bit operations, but they do not settle questions about throughput, latency, cost per request, or effective capacity for a specific workload.

Epoch’s estimates are therefore best used as a starting point for scenario planning, not a procurement forecast or benchmark. Builders should pair shipment and specification data with workload-specific tests, memory and interconnect requirements, availability dates, and facility constraints. The distinctions are especially important when custom accelerators are involved: the buyer, the chip designer, and the party reporting revenue may not be the same organization.

## Limits and uncertainty
Epoch says its figures are estimates, and the confidence and methodology vary by chip designer. Many 90% intervals span about a factor of two; some estimates depend on indirect evidence, fiscal periods, or interpolation into calendar quarters. H100-equivalent performance is based on peak specifications and does not measure realized workload performance or cluster topology. Delivered chips may take time to become operational, and the database excludes some suppliers and chip categories. The estimates do not account explicitly for chip retirements, and cost figures omit the wider capital costs of compute infrastructure.

## Reader takeaway
Epoch AI’s database is useful for tracking the scale and composition of accelerator supply, but its H100-equivalent totals are not a substitute for real-world performance or installed capacity. Read unit counts, chip generation, uncertainty, and delivery status alongside any aggregate compute estimate.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)