AI Datacentres
A note to John Gowings · Gowings Bros · Compute, from the customer's side

Demand

What the customer is buying when they buy compute. The middle of the data flow is the only thing billed. What goes in and what comes back is the customer's own.

IWhat gets billed
Input
a prompt
a query
a document
a question
raw data
Operator inputs
power · space · capital
AUD 45M / MW deployed at construction
Process · Compute
GPU-hours
of work done
the only thing billed to the customer
Operator process
22% opex of revenue
power dominates · ~30–35 FTE · ramp four years
Output
a token
an answer
an image
a prediction
a model
Operator bill
AUD 5.3M / MW / year
anchor rate · +18% over hyperscaler-Sydney baseline
USD 2–4per million output tokens at frontier; USD 1.50–3per H100 GPU-hour at retail.
What gets billed
GPU-hours
The middle of the data flow is the only thing billed. What goes in (a prompt, a query, raw data) and what comes back (a token, an answer, an image) is the customer's own.
INPUT a prompt, a query, raw data the customer's PROCESS · COMPUTE GPU-hours the only thing billed 22% opex of revenue OUTPUT a token, an answer, an image the customer's What the operator sells = the middle
  • Operator inputs · power · space · capital AUD 45M / MW deployed
  • Operator process · GPU-hours of work done 22% opex · ~30–35 FTE
  • Operator bill · anchor rate AUD 5.3M / MW / year
  • Customer pays · token + GPU-hour pricing USD 2–4 per Mtok · 1.50–3 / GPU-hr
What the operator sells The middle of the flow. The customer brings the question and keeps the answer; the GPU-hours between are the billable line — the +18% premium over hyperscaler-Sydney lives there.
IIThe shape of the customer's demand
One customer three days, hour by hour
DAY 1 DAY 2 DAY 3 PEAK IDLE
A typical customer's compute demand: morning ramp, midday peak, evening peak, overnight floor. Bursty.
If the customer owned the fleet capacity sized to peak
OWNED CAPACITY idle most of the time DAY 1 DAY 2 DAY 3
~30%average utilisation. The other 70 per cent is paid for, and idle.
Pooled across many peaks and troughs cancel
POOLED AGGREGATE six customers, each with their own peaks DAY 1 DAY 2 DAY 3
~70%operator utilisation. Your peak is everyone else's trough.
IIIWhen ownership pays, when renting does
$0 $2 $6 $9 $12 USD / GPU-HOUR DELIVERED 0% 25% 50% 75% 100% SUSTAINED UTILISATION OF THE FLEET WHERE MOST CUSTOMERS LIVE BUILD IT YOURSELF cost per useful GPU-hour RENT IT flat ~$3 per GPU-hour, you only pay for what you use crossover at ~67% SUSTAINED ~67%
Below ~67% utilisation

The chip is idle too often to amortise. The all-in cost of build divided by hours actually used exceeds the rental rate. Renting wins. Almost every customer lives here.

Above ~67% utilisation

You can amortise the chip; build undercuts rent. Constant workload (Meta serving global ad inference) or a strategic mandate (sovereign defence) is required. A handful of customers live here.

When ownership pays
≈ 67%
Sustained utilisation. The threshold above which build cost per useful GPU-hour falls beneath the ≈ $3 rent rate, and the chip is amortised. Below it, the chip is idle too often — almost every customer lives there.
$3 RENT 30% $7 · rent 50% $4 · rent 67% $3 · ≈ 80% $2.50 · own 95% $2 · own $0 $3 $6 $9 $12 USD per useful GPU-hour
  • Below 67% · rent wins most below 30% on their own
  • Threshold · ~67% sustained the chip amortised
  • Above 67% · build wins constant load · strategic mandate
  • Operator's arbitrage · pooling bursty demand lands above 67%
What the boutique tier monetises The gap between the customer's ~30% and the operator's pooled ~70% — the same chip, paid for and idle on its own; profitable when pooled.
The investor read

Your peak is everyone else's trough. That arbitrage is what the data centre operator sells. The chips age in three years; the pooled demand had better not. Whether it does is the investor's question — the only real one once the connection is secured.