Predictive intelligence for AI infrastructure
Physical systems. Commercial consequences.
We predict what will delay or interrupt AI compute, so data centre operators and neoclouds get capacity live and keep it dependable.
Predictive models forecast what will slip, and AI agents investigate the consequence and coordinate the response before capacity is lost.
Explore what we're building ↓01 / One site
Finding the one thing that is actually holding up go-live.
Power, electrical, cooling and compute all have to be ready before a customer can run anything, and each one is tracked by a different team in a different system. Veltora looks through all four and finds the item that is setting the date.
02 / The problem
Installed capacity is not usable capacity.
Between the capacity on the plan and the capacity earning revenue sit several distinct states. Capacity leaks out at every one of them, and each leak is owned by a different team, tracked in a different system, and caused by something different.
Installed
Hardware on site and racked
Energised
Grid milestones, transformers, switchgear, electrical commissioning
Healthy
Cooling capability, node validation, burn-in, firmware, network and topology
Customer-ready
Acceptance testing, RMAs, spares, restrictions, approvals
Earning
Allocated to a paying commitment and running workloads
Illustrative. Colocation operators and neoclouds lose capacity at different stages, for different reasons.
01
Infrastructure blocker
Equipment, power, cooling or commissioning slips
02
Capacity cannot go live
Committed MW or GPUs are not customer-ready
03
Financial exposure
Delayed revenue, penalties, idle capital and customer risk
04
Veltora’s economic value
Find the interventions that recover customer-ready capacity sooner
$28–47M
A 16,384-GPU cluster, one month short of customer-ready
~$39M
A 40 MW energisation tranche, five months late
~$2.4M
A 12 MW hall handed to a tenant a month late
SemiAnalysis GPU contract rates and CBRE wholesale asking rates. Revenue alone, before penalties, idle capital and the customer who goes elsewhere.
03 / The platform
One model of your site, built from every system you already run.
Every system an operator runs sees one slice. The ERP knows the order. The schedule knows the plan. The utility knows the grid milestone. The BMS and DCIM know the facility. GPU telemetry knows the nodes. Sales knows the promise. Veltora reads all of it and answers the one question none of them can.
Reads from, replaces none of them
What it does, continuously
01 / READ
Reads reality
Live signals from suppliers, equipment, facilities, the grid and GPUs, not just what someone typed into a plan. A supplier email becomes a structured fact with a source, a timestamp and a confidence.
02 / FORECAST
Forecasts usable capacity
Megawatts and GPUs available on each date, with probabilities, from models trained on what actually happened on real projects.
03 / CALCULATE
Calculates consequence
Which customers are exposed, how much capacity, how many days, how much revenue, and which few issues actually move the date. This arithmetic is deterministic and checkable.
04 / ACT
Drives the fix
Compares the interventions available, recommends the one that recovers the most, and has agents do the chasing. Anything consequential waits for a human.
The loop closes. Every outcome is recorded against the prediction that preceded it, which is what makes the next forecast better.
One dependency model underneath
A live, versioned graph of how everything depends on everything else: which transformer feeds which hall, which hall hosts which cluster, which cluster serves which customer commitment. When one fact changes, Veltora already knows what it touches.
Every edge carries its evidence, its source system, and the date it was last confirmed. Change one fact and Veltora already knows what it touches.
04 / Products
Built for the physical layer of AI.
Each product answers a different question about the same thing: how much capacity you can actually use, and when. They run on one model of your site, so what one product learns, the others already know.
GET IT LIVE / Supply Chain Intelligence
Spot the equipment that will arrive late, weeks before the date officially moves.
Veltora tracks supplier commitments and how their dates have moved across transformers, switchgear, chillers, PDUs and GPUs, and connects every item to the infrastructure waiting on it.
- 01Delivery risk scoring on long-lead equipment
- 02Supplier reliability: who slips, by how much, and how early they say so
- 03Approved alternatives and reallocation options when an order will not make it
- 04Every item linked to the halls, clusters and commitments downstream
A transformer order has quietly moved twice in six weeks. Veltora scores it as likely to move again and shows that Hall 3 and a 12 MW handover sit behind it, eleven weeks before anyone would have noticed.
GET IT LIVE / Power Readiness
Forecast which megawatts the grid and electrical path will actually energise, and when.
Power arrives in tranches gated by utility approvals, grid connections, substations, transformers and electrical commissioning, each tracked by someone different. Veltora maps that path to the halls and commitments it serves.
- 01Deliverable MW per site and hall, with confidence
- 02Every utility, grid and electrical milestone in one place
- 03Which commitments sit on which tranche, and what a slip costs
- 04Recovery scenarios priced side by side
A utility moves a 40 MW tranche from March to August. Halls 3 and 4 move with it and a 12 MW tenant commitment is exposed, around $2.4M of lease revenue per month.
GET IT LIVE / Go-Live Intelligence
Predict when new MW and GPU capacity will actually become usable.
A live model of everything between equipment arriving and a customer running workloads: installation, power, cooling, network, commissioning, validation and acceptance.
- 01Readiness forecast per customer date, with confidence
- 02Which few issues actually control the date, and which do not
- 03Recovery options scored by days recovered, cost and decision deadline
- 04GPU readiness tied to the customer date
- 05A one-page brief per threatened commitment
- 06Alerts when sources disagree or evidence goes stale
A neocloud promised 1,024 GPUs by 15 November. The plan says on track. The supplier moves the transformer to 18 November. Veltora forecasts 21 days late. Of 18 open issues, only 2 matter. Borrowing the Phase 2 transformer recovers 14 days, and the decision window closes in 6 days.
KEEP IT LIVE / Compute Assurance
Predict how much GPU compute will stay available, and what threatens it.
Going live is the first time capacity has to be dependable, not the last. Veltora reconciles installed, energised, healthy, customer-ready and reserved capacity continuously, and forecasts what is about to drop out of service.
- 01Dependable capacity per site, hall and cluster
- 02Every gap explained by cause, in one place
- 03What is likely to fail, and the capacity it would take with it
- 04Maintenance impact checked before the window opens
- 05Whether spares and repair times can keep commitments whole
A cooling loop shows falling pressure and maintenance on the redundant loop is booked for Thursday. Together they expose 128 GPUs on a contracted customer. Move the window a week and the exposure disappears.
KEEP IT LIVE / Operations & Economics
AI agents simulate interventions, put a dollar value on each, and coordinate the best one.
When something material changes, a case opens on its own. Agents trace the dependencies, simulate every realistic way out of it, price the capacity each one saves, and carry through the option you approve.
- 01A case opens the moment something material changes
- 02Every realistic intervention simulated against your live site model
- 03A dollar value on each option, from the capacity it protects
- 04Work ranked by what is at stake, not by alert severity
- 05Tickets, RMAs, supplier confirmations and spares prepared for you
- 06Anything consequential waits for your approval, then is tracked until the capacity is back
Two incidents open the same morning, each holding 128 GPUs. Veltora simulates the options for both and values the premium-contract cluster at roughly three times the older one, ranks it first, prepares the RMA and a maintenance re-sequence, and waits for the fleet lead to approve.
SELL IT SAFELY / Capacity Promising
Test a deal against the live forecast before you sign it, so the date holds.
A proposed deal is run against the readiness forecast, the power tranches, existing allocations and fleet health, and comes back with a date you can stand behind.
- 01The earliest achievable date, and the safe date to promise
- 02Partial delivery options when the full amount will not make it
- 03Conflict checks, so capacity is never promised twice
- 04The dependency the promise rests on, named
- 05One version of reality shared by sales and delivery
Sales is asked for 512 GPUs by 1 March. Full delivery lands 14 March, with a 22% chance of 1 March. 256 GPUs can go by 1 March from Hall 5, so sales offers a phased start instead of missing a promise.
Get AI compute online faster. Keep it dependable.
05 / A decision
One signal. Twenty-one days. Six days to act.
A supplier email arrives on 10 October and moves a transformer by three weeks. Nothing in the plan changes. Veltora reprices the customer date immediately, finds the two issues that control it, and puts a dated decision in front of the person who owns the commitment.
CMT-0142 / 1,024 GPUs / Hall 3
PROMISED
15 Nov
NOW FORECAST
6 Dec
21 days late
PROBABILITY OF 15 NOV: UNDER 2%
$0.8–1.4M
Brought forward by recovering 14 days on a single 1,024-GPU commitment
4–7×
What that one decision is worth against a year of Veltora at target pricing
Capacity × days
How we measure everything. Ten megawatts recovered by fourteen days is 140 MW-days brought forward, converted to dollars only where the contract supports it
06 / Technology
A prediction stack built for physical infrastructure.
Each technology is used only for what it is genuinely best at, and anything that has to be exactly right is calculated rather than generated.
01
Live signals
Veltora reads the systems you already run, and the evidence that never reaches a system, like the supplier email that quietly moves a date.
02
The infrastructure model
One live model of how your site actually hangs together, from equipment through to the customer commitment it serves. When something changes, Veltora already knows what it touches.
03
Prediction
Models trained on how infrastructure gets delivered in practice rather than on what the schedule said. They are calibrated, so a 30% risk happens about three times in ten.
04
The decision engine
Every date, megawatt and dollar is calculated, never generated. You can check any number Veltora puts in front of you.
05
Agents
Agents do the chasing: gathering evidence, reconciling what disagrees, preparing the options and following through once you approve. Anything that leaves your building waits for a person.
07 / Who we serve
Capacity committed to a customer, against a date.
That one condition defines our customer. If capacity is promised and real money is lost when it slips, Veltora has something to protect. Both of these buyers ask the same underlying question, which is why one platform serves both.
01
Data centre operators
Commercial owner-operators undertaking a funded expansion, fit-out or upgrade, with customer capacity committed against a delivery date.
- Promises
- Megawatts handed over to a tenant by a date
- Leaks at
- Equipment, grid power, electrical and cooling commissioning
- Their question
- How many megawatts can we hand over, and when?
- Champion
- Head of delivery or programme director
- Also uses it
- Procurement, commissioning, power and energy, commercial
02
Neoclouds
GPU clouds that directly build or manage physical GPU clusters and own customer readiness commitments.
- Promises
- GPUs customer-ready by a date
- Leaks at
- Node validation, burn-in, firmware, network, RMAs, and the facility underneath
- Their question
- How many GPUs can we sell, and when?
- Champion
- Head of infrastructure or fleet
- Also uses it
- Fleet engineering, procurement, commercial
80%
Utilisation below roughly this level flatlines neocloud returns, so every idle GPU-day comes straight out of the return
McKinsey
Every 3 hours
Unexpected interruption on a 16,384-GPU training cluster, around 78% of them hardware
Meta, Llama 3 infrastructure report
Delivered / Healthy / Reserved / Active
NVIDIA requires its cloud partners to track every GPU in these states. The industry already separates hardware existing from capacity being usable
NVIDIA cloud partner requirements
08 / Build with us
Bring us the problem worth solving.
We work with a small number of operators and neoclouds bringing capacity online right now. A paid pilot runs six to eight weeks on one capacity block and one customer date, credited against year one.
You bring a real commitment with a real date. We replay a completed project of yours event by event to show what Veltora would have said at the time, then run live alongside your own process so you can judge it against something.
sales@veltora.net
Veltora / Working with data centre operators and neoclouds worldwide
Turn physical infrastructure into dependable compute.
Physical infrastructure.
Operational intelligence.
Real-world outcomes.