Industry
Product

TensorWave announced yesterday that it has signed a six-year agreement to supply OpenAI with 457 megawatts of dedicated AI capacity across several US data center sites, built on AMD Instinct accelerators. The first tranche comes online in the first half of 2027, with full deployment by early 2028. Sit with that for a second. One of the most capitalized AI companies in the world, with every reason to own its own infrastructure, has just signed a six-year lease on other people's hardware in other people's buildings. That is not a failure of ambition. It is a read on what compute actually is right now, which is a scarce commodity with a long queue, where the binding constraint is rarely money. North American data center vacancy touched a record low of around 1.6 percent earlier this year. Capacity is contracted before it is built. For an enterprise with a real AI workload and a finance team asking sensible questions, the decision breaks into three options. This piece is about which one fits which situation, because the answer is genuinely different depending on the workload.
Renting means GPU as a service. You take capacity from a provider, by the hour, the month or on a reserved commitment, and somebody else owns the hardware, the building, the power and the cooling. The case for it is speed and optionality. You can be running this week rather than next quarter. You pay nothing for idle capacity if you size it honestly. And you are not holding an asset that depreciates against a product roadmap moving faster than almost any hardware cycle in recent memory. For a workload that is spiky, experimental, or genuinely uncertain in size, renting is usually the right answer and the arithmetic is not close. The case against gets stronger the more predictable your usage becomes. A rented GPU running at high utilization around the clock stops being cheap fairly quickly, and the crossover point is often sooner than people expect. There is also the question of where your data sits and under what terms, which matters more in regulated sectors than in a pitch deck. And there is availability, which is the one most teams underrate: a reserved commitment is a contract, but on-demand capacity is a queue, and queues in this market have been closing. The useful question is not whether renting is expensive. It is what your utilization will actually be twelve months from now, and whether you believe your own forecast.
Buying means owning the hardware. For most enterprises this is a Supermicro, Dell, HPE or Lenovo chassis with accelerators in it, configured for the workload and installed somewhere you control. The case for it is cost at steady state and control over everything. A machine running continuously for three years is almost always cheaper owned than rented. Your data stays where you put it. You configure what you like. And you have an asset rather than a line of operating expenditure, which some finance teams prefer and others actively do not. The case against is rarely the purchase price. It is the room. An accelerated rack draws many times what a traditional one does, air cooling stops being sufficient at the densities these machines run, and the power feed is a building problem rather than a procurement one. Plenty of teams have bought hardware and then discovered that the comms room it was going into cannot feed or cool it, at which point the asset sits in a crate while somebody prices an electrical upgrade. The other thing people forget is that buying is not one decision. It is hardware, plus somewhere to put it, plus a network fast enough to be worth having, plus somebody to look after it at three in the morning. Each of those is a separate supplier unless you deliberately arrange otherwise.
The third option sits between the two and gets much less attention than it deserves. You own the hardware, and it lives in somebody else's facility, on somebody else's power and cooling, connected to somebody else's network. That combines the cost profile of ownership with the speed of renting, because a carrier-neutral facility already has the power, the cooling and the connectivity that your own building does not. You skip the electrical upgrade entirely. You keep the asset and the control. What you give up is physical proximity, which matters less than most people assume once the network is good enough. How to choose comes down to three honest answers. First, utilization. Below roughly a third, rent. Above two thirds sustained, own. In between, the arithmetic depends on your cost of capital and how long you think the hardware stays current, which is the number nobody can give you with confidence right now. Second, the room. If the facility you would put it in cannot deliver the power density and the cooling, then buying means colocating, and that should be priced from the start rather than discovered later. Third, time. If you need to be running inside a quarter, renting is the only one of the three that reliably delivers that, because hardware lead times and electrical work both run longer than people plan for. The mistakes are consistent. Buying before anyone has checked what the room can feed. Renting on-demand at high utilization and treating the bill as fixed. Sizing from a pilot that ran on clean data and a motivated team. Comparing a rental quote to a hardware quote without including the building, the network and the support around it. And making the decision once, when it is genuinely worth revisiting every time the workload changes shape. Two things deserve more weight than they usually get. The first is the network between the compute and everything else. A cluster that cannot reach its own training data quickly, or cannot be reached by the application it serves, is an expensive machine with a bottleneck in front of it, and that bottleneck is almost always discovered after the hardware decision rather than before it. Where the compute sits determines how hard that is to fix. The second is what happens at the end. Rented capacity ends when the contract ends, which is tidy. Owned hardware has a residual value, a disposal question and data on it that has to be dealt with properly. Neither is difficult, but a three-year decision that has not considered year four is only two thirds of a decision. It is also worth saying plainly that these are not mutually exclusive. A reasonable pattern for a team with both steady and spiky demand is to own enough for the baseline and rent the peaks, which keeps utilization on the owned hardware high and avoids paying ownership prices for capacity used a few weeks a year. That hybrid is harder to arrange than either pure option, because it needs the two environments to work together, and it is usually cheaper than both.
What makes this difficult is rarely the analysis. It is that the three options usually come from three different suppliers who each have a reason to recommend their own. Vocom works across all three. It supplies enterprise AI server hardware as an authorized Supermicro dealer, deploys it inside its own carrier-neutral data centers when you would rather not touch your own building, and connects it on an owned backbone with direct links into the major clouds, with a 24 hour quote turnaround and a 24/7 network operations center behind it. GPU solutions and the wider carrier and facilities services sit alongside. For what makes these machines different in the first place, our earlier piece covers the engineering. If you have a workload and no settled view on which of the three fits it, that is the conversation worth having before anything is ordered. Talk to us.
Get Vocom Insights in Your Inbox
Fiber optic cable lead times, specification guides and AI infrastructure news for network builders.