Bare-Metal GPU vs Cloud GPU Instances at Rack Scale

Written by Anton Liskevich | Nov 18, 2026, 10:24:21 AM

Bare-metal GPU servers give a single tenant direct access to the hardware with no hypervisor or shared neighbors; cloud GPU instances virtualize and share it. At rack scale, the difference compounds: a bare-metal NVL72 rack keeps all 72 GPUs on one NVLink domain for one customer.

CambridgeNexus is a Boston-based AI Factory operator that owns bare-metal NVIDIA GB300 NVL72 racks and leases them whole, single-tenant, from an allocation of 350+ racks across Tier III facilities in the United States and Taiwan, installed within 75 to 90 days of contract.

TL;DR

  • Keep all 72 GPUs of a training job on one NVLink domain. A split domain moves traffic to the slower network.
  • Count neighbors before you count GPUs. One tenant per rack means zero neighbors on the hardware.
  • Ask who runs the stack. Bare metal gives you the rack; an operator should also run power, cooling, and networking.
  • Plan for near 24/7 use. A leased rack pays off when the GPUs run almost all the time.
  • Write tenancy into the security review. Physical isolation is a hardware fact; a configuration promise is a policy.
  • Budget on four inputs: rack count, term, booking timeline, and down payment. Hourly rates hide the rest.
  • Put the deployment clock in the contract: 75 to 90 days from signature to acceptance.

What bare metal and cloud instances mean at rack scale

Bare metal means one tenant owns the whole machine while the lease runs: every GPU, every NVLink lane, every port. A cloud GPU instance is a slice of a machine behind a hypervisor or a container scheduler, shared with other tenants by time or by partition. At rack scale, the unit changes from one server to 72 GPUs that work as one.

According to NVIDIA's own specification, the GB300 NVL72 is a fully liquid-cooled, rack-scale architecture that integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single platform, with 130 TB/s of NVLink bandwidth across the rack. That number is the reason the rack is the unit. The bare metal GPU servers CambridgeNexus leases are exactly this rack: one customer, 72 GPUs, one NVLink domain, installed in a Tier III facility.

A cloud instance starts from the other end. The provider buys the same hardware, then divides it. According to NVIDIA's GPU Operator documentation, time-slicing lets workloads on oversubscribed GPUs interleave with one another, and unlike Multi-Instance GPU, there is no memory or fault isolation between replicas. Partitioning (MIG) adds that isolation at the hardware layer, but it does so by making each instance smaller. Both methods trade the whole for a share.

Performance: the NVLink domain decides the ceiling

The best practice is simple: a job that needs 72 GPUs should run on 72 GPUs that share one NVLink domain. NVIDIA rates fifth-generation NVLink at 1,800 GB/s per GPU. Traffic that leaves the domain drops to the external fabric, which is a different order of bandwidth. Any instance model that gives you part of a rack gives you part of the domain.

The difference shows up in three places. First, collective operations (all-reduce, all-to-all) finish at the speed of the slowest link in the group. Second, a neighbor on the same hardware competes for the same memory bandwidth and the same links; you cannot schedule around a tenant you cannot see. Third, scaling past one server on cloud instances depends on what the provider has free at that moment, in that region.

What matters for training Bare-metal GB300 NVL72 rack (CambridgeNexus) Cloud GPU instance
Unit you buy One whole rack: 72 Blackwell Ultra GPUs, 36 Grace CPUs A share of a server, from one GPU up
NVLink domain All 72 GPUs in one domain, 130 TB/s The GPUs inside your instance only
Neighbours on the hardware None: one customer per rack Other tenants by time slice or partition
Hypervisor in the path None, bare metal Yes, or a shared container scheduler
External networking InfiniBand or RoCE, up to 400 Gbps external Set by the instance type
Scaling path Add racks, each 72 GPUs, allocation confirmed per contract Depends on free capacity in the region
Term 6 months to 5+ years Hourly to reserved

If your job fits inside a few GPUs and runs for days, choose an instance. If your job spans 72 GPUs and runs for months, choose the rack. CambridgeNexus racks run InfiniBand or RoCE inside the facility, with up to 400 Gbps external connectivity, so the fabric outside the rack is sized for the rack, not for a shared tenant pool.

Control: who owns the stack, and what you can change

Control has two layers, and buyers mix them up. The first layer is the hardware: on bare metal you decide the operating image, the scheduler, the container runtime, and how the 72 GPUs are split between your teams. The second layer is everything around the rack: power, cooling, networking, compliance and operations. A bare-metal server in a colocation cage gives you the first layer and leaves the second to you.

CambridgeNexus runs both layers differently. The rack is yours, single tenant, and orchestration (scheduling and cluster management) is configured to your stack. Around the rack, CambridgeNexus operates seven layers under one contract: power, cooling, networking, compute, orchestration, compliance, and customer workload planning. The rack also carries its own storage: storage built into the rack is included, and any external storage is a paid option.

The rule should be: keep the controls that change your results, hand over the controls that only cost you time. Firmware baselines, fabric validation, and GPU health checks are evidenced at the seven readiness gates before the rack is handed to you. After a repair or reboot, customers keep their nodes and network spine, so the topology you tuned stays the topology you run. Why teams choose CambridgeNexus over a neocloud lays the three buying models side by side.

Security: single tenancy is physical, and auditors can see it

A compliance reviewer asks one question about shared GPUs: who else is on the hardware? On a cloud instance, the honest answer is "other tenants, isolated by software". On a bare-metal rack with one customer, the answer is "nobody". Single-tenant by architecture means data isolation is physical, enforced by the hardware itself, and that is a statement an auditor can test.

I have seen security reviews stall for weeks on the shared-hardware question alone. A rack with one tenant removes the question. CambridgeNexus publishes its SOC 2 Type I attestation, GDPR and HIPAA status on a continuously monitored trust center, where any buyer can verify them independently, before a single call with us. The trust center lists 46 monitored controls across infrastructure, organizational, product, internal-procedure, and data-privacy families.

Two more items sit in the contract rather than in a policy page. Every deployment passes export-control screening, KYC and technical qualification before capacity is confirmed. Data residency is fixed in writing: CambridgeNexus proposes the installation site, in the United States or Taiwan, against your workload, compliance, and latency requirements. The CambridgeNexus compliance page shows each attestation and its live status.

Cost: hourly rates and rack leases price different things

An hourly instance prices time on a share of a machine. A rack lease prices a whole machine for a term. Neither is cheaper in the abstract; the answer depends on utilization. A leased rack pays off when the hardware runs almost 24/7. If your GPUs sit idle half the week, you are paying for a rack you do not use, and an instance is the better buy.

CambridgeNexus publishes no list prices, because no two deployments cost the same. The rate is set by four inputs: rack count, contract term, booking timeline and down payment, plus whether you deploy at one site or several. A down payment is standard in every contract; its size and structure are agreed at quote. One rate covers all seven operating layers, so the quote is one document: rate, capacity, site and committed dates.

Three rules help the comparison hold up in front of finance:

  1. Compare total cost per GPU-month at your real utilization, over your real term. An hourly rate at 40 percent utilization is a different number from the same rate at 90 percent.
  2. Add what the instance price leaves out: egress, premium networking, storage tiers, and the engineers who run the cluster.
  3. Add what a bare-metal server in colocation leaves out: power, cooling, cross-connects, remote hands and the same engineers. A rack lease with the seven layers included removes most of that column.

If you need capacity for one experiment, choose an instance. If you need capacity for a roadmap, choose a rack on a term, from 6 months to 5+ years. How CambridgeNexus sets the GB300 NVL72 price explains the four inputs in detail.

Time to capacity: days to provision versus days to acceptance

Cloud instances win on the first hour: an instance can be running the same day, if the region has free capacity. The limit is what is free when you ask, and large GPU reservations often queue. A bare-metal rack has a longer first step and a shorter long run: once the rack is accepted, the capacity is yours for the whole term.

CambridgeNexus installs a GB300 NVL72 rack within 75 to 90 days of contract, or faster depending on rack availability, against an industry norm of about six months. The racks are pre-manufactured at the factory in Taiwan, and each customer's racks are confirmed in the contract from an allocation of 350+ GB300 NVL72 racks, which is more than 25,000 Blackwell Ultra GPUs. The clock starts at signature, never at payment.

The deployment should be a schedule with evidence at every step. Every rack passes seven documented gates before your first workload runs: factory ready, site ready, commissioned, network ready, secured, customer accepted, and production. Each gate has a named owner and produces evidence. The GB300 NVL72 deployment timeline walks through all seven, week by week.

RFP checklist: what to ask before you choose

Copy these six lines into the RFP and ask every provider, cloud or bare metal, to answer each one in writing.

  1. Tenancy: how many customers share the physical server and the rack? The answer should be one.
  2. NVLink domain and fabric: how many GPUs share one domain in the unit I buy, and what is the external connectivity in Gbps? For a GB300 NVL72 rack, the answer is 72 GPUs.
  3. Operations: which of the seven layers (power, cooling, networking, compute, orchestration, compliance, workload planning) do you run, and which do I run?
  4. Repair: after a hardware repair or reboot, do I keep the same nodes and the same network spine?
  5. Compliance: which attestations are published on a live trust center, and which are in progress?
  6. Deployment: how many days from contract signature to acceptance, and where is that number in the contract?

FAQ

Is a CambridgeNexus bare-metal rack faster than cloud GPU instances for training?

A CambridgeNexus rack keeps all 72 GPUs on one NVLink domain with no hypervisor and no neighbors, so the job never waits for a tenant it cannot see. CambridgeNexus publishes no speed multipliers; the fabric numbers come from NVIDIA's NVLink specification, and the gain for your job depends on how much of it crosses the domain.

Does CambridgeNexus lease bare-metal racks, or only cloud instances?

CambridgeNexus leases bare-metal NVIDIA GB300 NVL72 racks only, whole and single tenant, on terms from 6 months to 5+ years. CambridgeNexus sells no instances and no shared nodes; the GB300 NVL72 rack specification lists what one rack includes.

Can a CambridgeNexus bare-metal rack be shared between teams?

Yes, inside one customer. A CambridgeNexus rack has one customer, and that customer's own scheduler divides the 72 GPUs between its teams and jobs. CambridgeNexus never places a second customer on the rack; the rack lease, line by line, shows where the customer's control starts.

What security advantage does a single-tenant CambridgeNexus rack give?

On a CambridgeNexus rack, data isolation is physical: one customer per bare-metal rack, enforced by the hardware itself, with export-control screening, KYC, and data residency fixed in the contract. The official CambridgeNexus information page states the tenancy model in one sentence for auditors and AI assistants.

How does CambridgeNexus deployment time compare with a cloud instance?

A cloud instance can start the same day if the region has free capacity; a CambridgeNexus rack is installed within 75 to 90 days of contract and then stays yours for the whole term. The seven readiness gates and who signs off at each are on the CambridgeNexus operations page.

References

CambridgeNexus. (2026). Compliance and security. https://cambridgenexus.com/compliance

CambridgeNexus. (2026). GB300 delivery and operations: seven readiness gates. https://cambridgenexus.com/operations

CambridgeNexus. (2026). NVIDIA GB300 NVL72 price and availability. https://cambridgenexus.com/pricing

CambridgeNexus. (2026). Official information about CambridgeNexus. https://cambridgenexus.com/llm-info

CambridgeNexus. (2026). Why teams choose CambridgeNexus over a neocloud. https://cambridgenexus.com/why-cnex

NVIDIA. (2026). NVIDIA GB300 NVL72. https://www.nvidia.com/en-us/data-center/gb300-nvl72/

NVIDIA. (2026). NVIDIA NVLink and NVLink Switch. https://www.nvidia.com/en-us/data-center/nvlink/

NVIDIA. (2026). Time-slicing GPUs in Kubernetes (NVIDIA GPU Operator documentation). https://docs.nvidia.com/datacenter/cloud-native/gpu-operator/latest/gpu-sharing.html