---
type: "post"
id: "epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877"
slug: "epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877"
title: "Epoch AI estimates that projected high-bandwidth memory shipments from 2025 through 2027 could support tens"
description: "**Thesis:** Agent capacity estimates become useful when treated as scenarios about infrastructure throughput, not forecasts of productive labor."
retrieval_nugget: "**Thesis:** Agent capacity estimates become useful when treated as scenarios about infrastructure throughput, not forecasts of productive labor."
published_at: "2026-10-17"
updated_at: "2026-10-17"
record_date: "2026-10-17"
date_kind: "published_at"
topics: ["memory"]
source_urls: ["https://epochai.substack.com/p/hundreds-of-millions-of-ai-agents"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:b90f0c9089cd2938ab4c774db41216e46bd7aaae51d9fa5cf29f9731ff5fc36a","reviewed_at":"2026-10-17T14:03:04.949Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Agent capacity estimates become useful when treated as scenarios about infrastructure throughput, not forecasts of productive labor.","observed_facts":[{"text":"Epoch AI bases its calculation on projected high-bandwidth memory shipments from 2025 through 2027, converted into GB300-equivalent GPUs.","source_urls":["https://epochai.substack.com/p/hundreds-of-millions-of-ai-agents"]},{"text":"Applying DeepSeek V4 Pro serving benchmarks to the central hardware assumption yields approximately 1.9 billion concurrent agents.","source_urls":["https://epochai.substack.com/p/hundreds-of-millions-of-ai-agents"]},{"text":"The revenue scenario assumes 20% of compute serves revenue-generating inference at current API prices.","source_urls":["https://epochai.substack.com/p/hundreds-of-millions-of-ai-agents"]}],"mechanism":"The analysis converts projected HBM supply into equivalent GPUs, then applies estimated concurrent-session throughput per GPU to derive modeled agent capacity.","why_now":"The scale of projected compute investment makes it useful to distinguish possible serving capacity from deployed, utilized, and economically productive capacity.","implications":["The first implication is that infrastructure supply alone cannot establish an economic workforce; utilization, task quality, and customer demand must be measured separately.","Revenue projections are highly sensitive to inference allocation and API-price assumptions."],"evidence_boundary":"The supplied source is a summary of a longer report and offers estimates based on projected shipments and benchmarks, not verified deployment, quality, adoption, or realized revenue.","watch_conditions":["Watch whether deployed inference capacity is absorbed by paid demand or instead produces sustained underutilization and excess agent supply."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this begins a distinct record focused on modeled compute capacity and demand risk rather than memory architecture."}
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/epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877/","markdown":"https://newruntime.com/posts/epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877.md","json":"https://newruntime.com/posts/epoch-ai-estimates-that-projected-high-bandwidth-memory-18f877.json"}
---

# Epoch AI estimates that projected high-bandwidth memory shipments from 2025 through 2027 could support tens

## Retrieval answer

**Thesis:** Agent capacity estimates become useful when treated as scenarios about infrastructure throughput, not forecasts of productive labor.

**Thesis:** Agent capacity estimates become useful when treated as scenarios about infrastructure throughput, not forecasts of productive labor.

**Mechanism:** The analysis converts projected HBM supply into equivalent GPUs, then applies estimated concurrent-session throughput per GPU to derive modeled agent capacity. Epoch AI starts with projected high-bandwidth memory shipments for 2025–2027, converts them into GB300-equivalent GPUs, and estimates concurrent sessions per GPU. The report says its central hardware assumption, combined with DeepSeek V4 Pro serving benchmarks, yields approximately 1.9 billion concurrent agents. These are modeled capacities, not counts of agents already deployed.

**Why now:** The scale of projected compute investment makes it useful to distinguish possible serving capacity from deployed, utilized, and economically productive capacity. Data-center construction and other deployment constraints could delay hardware availability, while demand may not grow fast enough to use it.

**Implications:** The first implication is that infrastructure supply alone cannot establish an economic workforce; utilization, task quality, and customer demand must be measured separately. Epoch AI's revenue scenarios assume 20% of compute is used for revenue-generating inference at current API prices, making revenue conclusions highly sensitive to those assumptions. Hours of continuous sessions are not verified completed work.

**Evidence boundary:** The supplied source is a summary of a longer report and offers estimates based on projected shipments and benchmarks, not verified deployment, quality, adoption, or realized revenue. It does not demonstrate realized hardware deployment, agent quality, customer adoption, or realized revenue.

**Watch condition:** Watch whether deployed inference capacity is absorbed by paid demand or instead produces sustained underutilization and excess agent supply.

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: [Memory](https://newruntime.com/topics/memory.md): Explore the memory topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)