Insights
Notes on GPU compute deals
Advisory and build notes for GPU compute operators and the teams financing them: cluster and datacenter architecture, hardware configuration, off-take readiness, and applied AI systems.
Written by Jason Sun, Principal Consultant. Book a call
Latest GPU Compute
Is my building ready to host GPU servers? A readiness checklist for California property owners
Most buildings fail on power or cooling, not on floor space. Seven checks, in the order that kills deals, with California utility tariffs, Title 24, air permit and submetering questions, and the arithmetic for a five-rack room.
The compute series
Contracts, quotes, acceptance, and the arithmetic behind GPU cluster deals.
-
Cost of Capital
Should we buy GPUs or keep paying for Copilot? What private AI costs per seat
A rack of your own GPUs can serve a thousand staff for a few dollars a head each month, if a mid-size open model is good enough for the work. Here is the arithmetic, the assumptions to benchmark, and where the comparison is unfair.
-
GPU Compute
Is my building ready to host GPU servers? A readiness checklist for Ontario property owners
Most buildings fail on power or cooling, not on floor space. Seven checks, in the order that kills deals, with the arithmetic for a five-rack room, so you can say go or no-go before you spend money on an engineer.
-
GPU Compute
Does everyone run the NVIDIA GPU Operator?
The NVIDIA GPU Operator is a useful way to manage an NVIDIA Kubernetes stack, but each accelerator needs its own runtime plan.
-
Compute Financing
Should you sell your idle GPU time?
Owning only beats renting past a utilization line, and that line moves with every assumption behind it. Selling the leftover hours raises a second set of questions, about the counterparty and the payout, that the hourly rate doesn't answer.
-
GPU Compute
How do you burn in a GPU cluster before you accept it?
A cluster can pass every acceptance test on delivery day and still fail in week two. Burn-in is the multi-day soak that finds the marginal parts while fixing them is still the supplier's problem. Put it in the deal before you pay.
-
Compute Financing
Can you resell GPU compute you already bought?
A reservation is a forward contract on compute, not a sunk cost. Your right to get out of it is something you negotiate before you sign, not something you find out about after.
-
Cost of Capital
What does it actually cost to run a GPU cluster?
I priced a 576-GPU cluster four honest ways. Moving the cost of capital from 15 to 7 percent is worth more than most of the price gaps people actually fight over.
-
Cost of Capital
What does a dollar-per-GPU-hour actually include?
Two vendors can quote the same dollar-per-GPU-hour and mean completely different things. The rate is a bundle. Here is what's usually inside it, what bills separately, and how to line up two quotes before you compare them.
-
GPU Compute
How do you tell if a GPU cluster actually works?
The acceptance tests I'd run, in order, before paying for a GPU cluster. Six measured numbers, how to compute each one from public specs, and how to tie the final payment to them.
-
Compute Financing
How do you vet a GPU cloud provider before you wire the deposit?
Six questions about the company on the other side of the contract, not the cluster, with dated public examples of a good answer and a bad one.
-
GPU Compute
What counts as downtime in a GPU cluster contract?
The provider's dashboard says the cluster was up. Your job failed anyway. Whether you get paid for that gap depends on a definition of availability buried in the contract, measured by whoever holds the logs. Read it before the outage, not during the dispute.
-
GPU Compute
What does a colocation quote for AI hardware actually include?
The per-kW-month number on a colo quote is a reservation fee for space and cooling, not your electricity. Power is billed on top, the kW is usually committed rather than metered, and the extras live outside the headline. Here is how to read one before you sign.
-
Cost of Capital
Where is the money in the AI compute stack?
The money in this stack sits at the two ends, in the silicon and in the software that routes work to it. The middle seat, owning depreciating hardware financed by someone else's prepayment, is project finance wearing a GPU costume.
-
Data Center Economics
Are AI data centers getting smaller?
The headline capacity numbers keep getting bigger, but a lot of the sites actually being built are not. Here's my take on why both are true, and what to check on a deal either way.
Earlier work
Case studies, research, and company updates.
- Beyond Extraction: AdaExtract's Confidence Layer AI Research
- Q1 2025 Recap - AdaNomad Consulting Drives Innovation from Core Infrastructure to Client Solutions Company Update
- Building Apps on AceTeam Platform App Development
- AdaExtract - Multimodal Knowledge Graph AI Research
- Automated Sales Outreach Tool AI
- AI for legal Case Study
- Data Warehouse Insights Case Study
No articles match this topic yet.