HomeSupply and SpendingCompute Supply (Epoch)

Compute Map and Economics

8 charts · Updated 2026-08-31

Who is locking up compute

Compute by User: Who's Locking It Up (Millions of H100e, Including Scheduled Completions)#

By site allocation, OpenAI has the most compute locked up through 2028 with Google close behind; the slope shows how hard each is grabbing capacity.Draft, pending review

How to read this chart

Allocated by each data center's user field, split evenly where there are several users; unlabeled is mostly sites whose customers are undisclosed. The path combines capacity already online and scheduled completions; coverage thins after 2028.

Source: Epoch AI (CC BY 4.0); FinSight compilation and estimates · Updated 2026-08-31

Compute by Owner: Cumulative H100e Stock (Millions)#

Cumulative compute stock by owner: Google, counting TPUs, is largest, followed by Microsoft and Amazon.Draft, pending review

How to read this chart

Stacked area, 1Q22 to 1Q26; in-house chips (Google TPU, Amazon Trainium) are counted under their owners.

Source: FinSight compilation and estimates · Updated 2026-08-31

Latest Share: Compute Owners (Cumulative Stock, Q1 2026)#

Compute owners' share as of the first quarter of 2026, including in-house chips.Draft, pending review

Latest Share: Compute Owners (Cumulative Stock, Q1 2026)

Static image for now; interactive version coming.

Source: FinSight compilation and estimates · Updated 2026-08-31

Latest Share: Compute Users (Mid-2026, Online)#

Users' share of capacity online as of mid-2026; sites with undisclosed customers are shown as unlabeled.Draft, pending review

Latest Share: Compute Users (Mid-2026, Online)

Static image for now; interactive version coming.

Source: FinSight compilation and estimates · Updated 2026-08-31

Upstream capacity: CoWoS and HBM

CoWoS Wafers Consumed (Wafers/Quarter, by Chip Designer)#

NVIDIA consumes most of TSMC's advanced packaging capacity, and its share is a gauge of pricing power.Draft, pending review

Source: FinSight compilation and estimates · Updated 2026-08-31

HBM Purchases ($B/Quarter, by Chip Designer)#

HBM purchases mirror demand at the three memory makers, with NVIDIA still the biggest buyer.Draft, pending review

Source: FinSight compilation and estimates · Updated 2026-08-31

Economics

Compute Cost as a Share of Revenue (%)#

Compute cost falling below 100% of revenue is the business-model turning point: Anthropic is down to 21% in 2026 and OpenAI reached 76% in 2025.Draft, pending review

How to read this chart

Above 100% means compute spending exceeds total revenue.

Source: FinSight compilation and estimates · Updated 2026-08-31

Tokens Processed per Day (Trillions, log)#

Daily tokens processed still rise in a straight line on a log axis; volume is exploding far faster than revenue, so revenue per token is deflating.Draft, pending review

How to read this chart

Different scopes (all products, API only, single company) are drawn separately; absolute values are not comparable across lines.

Source: Epoch AI (CC BY 4.0); FinSight compilation and estimates · Updated 2026-08-31

More pages in this topic