ExascaleOpen terminal

Industries · demand, by where it comes from

Twelve sectors. Four with public evidence we will stand behind.

Every compute company publishes an industries page and every one of them reads the same, because the numbers are borrowed from vendor marketing. This one says which claims have a citable public source and which do not. The sectors without one are still real demand — they are just not proven by anything you could check, and pretending otherwise is how a page like this becomes worthless.

12sectors
4with citable evidence
8stated as unevidenced
0customer logos claimed

NoteNo company named on this page is an Exascale customer. Names appear only where a company has published a figure itself, or a named outlet has reported one, and they are cited so you can read the original. We would rather show you four checkable facts than twelve unverifiable logos.

Sector by sector

What each one runs, and what it pays for

Financial services

Buyer
New York, London, Singapore, Frankfurt, UAE

Workloads

  • Document and KYC extraction at portfolio scale
  • Risk and scenario simulation (tabular, forecasting)
  • Customer service agents across chat and voice
  • Surveillance, AML and fraud detection on streaming data

How it pays for itself

Replaces per-case human handling time in operations, and shortens risk cycles that currently run overnight to intraday.

Public evidence

Klarna reported in February 2024 that its AI assistant handled two-thirds of customer service chats — 2.3 million conversations in the first month, “doing the equivalent work of 700 full-time agents”, with an estimated $40M profit improvement for 20241. By May 2025 the company had reversed course and resumed hiring humans, with its CEO conceding the result was lower quality2. Both halves matter: the compute demand was real and the durable headcount saving was smaller than announced.

Healthcare & life sciences

Buyer
US, UK, Switzerland, Denmark, Japan

Workloads

  • Ambient clinical documentation (speech → structured note)
  • Imaging triage and segmentation (realtime vision)
  • Molecular property prediction and docking (simulation)
  • Prior-authorisation and claims document processing

How it pays for itself

Returns clinician minutes per encounter and compresses discovery cycles that are measured in months of wall-clock simulation.

No citable figure

We have no audited public figure we are willing to cite for clinical AI ROI at system scale. Vendor-published time savings exist but are self-reported and rarely controlled. Treat this sector as high compute demand with weak public evidence — which is what it is.

Manufacturing & industrial

Both
Germany, Japan, South Korea, US Midwest, China

Workloads

  • Realtime vision on production lines for defect detection
  • Predictive maintenance on sensor streams (tabular)
  • Digital twins and physics simulation
  • Synthetic data generation for rare-defect training

How it pays for itself

Scrap rate and unplanned downtime are the two line items. Both are measured continuously already, which makes attribution unusually clean here.

No citable figure

Firm-level results are almost never published with enough detail to cite. The sector is notable on our side for a different reason: large manufacturers hold on-premise GPU fleets with poor average utilisation, which makes them credible sellers as well as buyers.

Retail & e-commerce

Buyer
US, UK, Germany, China, Brazil, India

Workloads

  • Product imagery and video generation at catalogue scale
  • Demand forecasting and markdown optimisation (tabular)
  • Conversational shopping and post-purchase support agents
  • Search relevance and reranking

How it pays for itself

Two distinct effects: content production cost per SKU falls, and forecast error translates directly into inventory carrying cost.

No citable figure

The Klarna figures above are the closest well-documented analogue for the support workload and are cited there rather than restated here. Catalogue-generation savings are widely claimed and rarely published in auditable form.

Media & entertainment

Buyer
Los Angeles, London, Seoul, Mumbai

Workloads

  • Text-to-video and image-to-video generation
  • Dubbing, voice conversion and subtitling at library scale
  • Restoration and upscaling of archive material
  • Previsualisation and 3D asset generation

How it pays for itself

Substitutes for production and localisation spend that is currently linear in minutes of finished output.

No citable figure

No citable public ROI figure. Worth stating plainly that video is the most compute-hungry modality we sell, and the one where demand is least sensitive to price at the top end.

Energy & utilities

Both
US, Norway, UAE, Australia, Gulf states

Workloads

  • Grid load and generation forecasting
  • Seismic and reservoir simulation
  • Asset inspection from drone and satellite imagery
  • Trading and scheduling optimisation

How it pays for itself

Forecast accuracy is settled financially every day in these markets, so improvements price themselves.

Public evidence

The structurally interesting sector for us. Data centres consumed roughly 415 TWh in 2024 and are projected to exceed 945 TWh by 20306, which puts energy companies on both sides of our book — buying simulation compute, and increasingly selling capacity sited against their own generation.

Public sector & defence

Buyer
US, EU, UK, UAE, Singapore, India

Workloads

  • Document digitisation and records modernisation (OCR)
  • Language services and translation at population scale
  • Geospatial analysis and change detection
  • Sovereign model training on domestic infrastructure

How it pays for itself

Backlog clearance in records and casework, plus a sovereignty requirement that is a placement constraint rather than a cost saving.

Public evidence

Sovereign requirements are why region is part of the instrument definition rather than a delivery detail. Compute is already geographically concentrated — the US at 45% of data-centre electricity consumption, China 25%, Europe 15%3 — and public buyers are frequently unable to transact outside their own jurisdiction.

Telecom

Both
US, EU, Gulf, South Korea, Japan

Workloads

  • Network optimisation and anomaly detection
  • Customer service agents across voice and chat
  • Edge inference for low-latency services

How it pays for itself

Care cost per subscriber, and capex deferral where optimisation delays a build.

No citable figure

No citable figure. Telecoms appear on the sell side more often than expected — edge sites with power and space are a natural fit for latency-bound inference capacity.

Logistics & transport

Buyer
Netherlands, Singapore, UAE, US, China

Workloads

  • Route and network optimisation (tabular, forecasting)
  • Document processing for customs and freight
  • Realtime vision in yards, ports and warehouses

How it pays for itself

Fuel, empty miles and dwell time. All three are already instrumented, so the counterfactual is measurable.

No citable figure

No citable public figure at firm level.

Robotics & autonomy

Buyer
US West Coast, Germany, Japan, China

Workloads

  • World-model training and rollout
  • Simulation and synthetic data at scale
  • On-vehicle and on-robot realtime vision

How it pays for itself

Simulation substitutes for physical trials, where the binding constraint is wall-clock rather than headcount.

No citable figure

The most compute-intensive buyer per dollar of revenue on this page, and the least able to forecast its own demand — which is exactly the profile that needs forwards and options rather than reservations.

Software & AI-native

Both
SF Bay Area, New York, London, Tel Aviv, Beijing

Workloads

  • Model training and post-training
  • Inference serving at product scale
  • Coding agents against large repositories

How it pays for itself

Compute is cost of goods sold. A basis point on the input price is a basis point on gross margin.

Public evidence

The most price-sensitive buyer, and the reason a reference rate matters. The open-weight frontier — Kimi K2.6 shipped as a 1T-parameter model with 32B active parameters and a 256K context window5 — means serving-side demand is no longer gated on access to a closed lab, only on access to capacity.

The shape of the demand

Why this matters to a venue rather than to a vendor

A cloud provider cares which sectors buy. An exchange cares whether their demand is correlated, because correlation is what makes a book thin.

Uncorrelated demand is the asset

Retail peaks at Q4, media peaks around release slates, robotics is steady and enormous, financial services peaks at close. Demand that peaks at different times is what lets one pool of capacity serve all of it — and what makes a spread tradable rather than theoretical.

Some of these are sellers

Manufacturing, energy and telecom all hold underutilised fleets or sites with power. Four of the twelve sectors here appear on both sides of the book, which is unusual and is the reason supply acquisition does not depend solely on neoclouds.

Geography is part of the instrument

Sovereignty rules, latency budgets and power availability mean an H100-hour in Virginia and one in Abu Dhabi are different instruments. That is a constraint for a reseller and a product for an exchange.

The buildout is the backdrop

Roughly $725B of guided 2026 hyperscaler capital expenditure4 sets the supply curve every sector on this page buys against. None of them can forecast their own share of it, which is the market’s founding problem.

Next

Where to go from here

Open a trading accountSee the credit typesRead the fee schedule

Sources

Every figure above is numbered to an entry here. Links last read 27 July 2026.

  1. 1

    Klarna AI assistant handles two-thirds of customer service chats in its first month

    Klarna · 27 February 2024 · Primary

    The AI assistant has had 2.3 million conversations, two-thirds of Klarna's customer service chats … it is doing the equivalent work of 700 full-time agents … it is estimated to drive a $40 million USD in profit improvement to Klarna in 2024.
  2. 2

    Klarna Reverses AI Push, Says Customers Prefer Human Support

    Forbes · 18 May 2025 · Reporting

    ParaphraseSiemiatkowski conceded the company had focused too much on cost: 'the result was lower quality', and Klarna began recruiting human agents again.

    The counterweight to the 2024 figures. Cited alongside them deliberately — the 2024 press release on its own overstates the durable result.

  3. 3

    Energy and AI — Energy demand from AI

    International Energy Agency · April 2025 · Primary

    The United States accounted for the largest share of global data centre electricity consumption in 2024 (45%), followed by China (25%) and Europe (15%).
  4. 4

    Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026, up 77% from last year

    Tom's Hardware · February 2026 · Reporting

    Google, Amazon, Microsoft, and Meta collectively plan to allocate $725 billion to capital expenditures in 2026 — up 77% from last year's $410 billion.

    A sum of separate company guidance ranges, not a reported figure. Individual guidance: Amazon ~$200B, Google $175–185B, Meta $115–135B, Microsoft $110–120B.

  5. 5

    Kimi K2.6 — open-weight frontier model

    Moonshot AI (via Interconnects) · April 2026 · Reporting

    ParaphraseKimi K2.6 is a 1T-parameter mixture-of-experts model with 32B active parameters and a 256K context window, released under a permissive open-weight licence.
  6. 6

    Energy and AI — Executive summary

    International Energy Agency · April 2025 · Primary

    Electricity demand from data centres worldwide is set to more than double by 2030 to around 945 terawatt-hours (TWh) … slightly more than the entire electricity consumption of Japan today.

Where a claim rests on a third-party estimate rather than the party that owns the number, the entry says so. Figures that are Exascale’s own — our rate card, our fee schedule — carry no citation, because they are ours to set rather than facts about the world.