Provider profiles lean on network signals, operator classifications and compliance labels that deserve a straight explanation. Hover any dotted term across the directory for the short version, or read the full definitions here.
Every provider in the directory carries one of these classifications, derived from network ownership and infrastructure checks rather than marketing copy.
A direct operator owns and runs the GPU infrastructure it sells, usually on its own network in facilities it controls. Renting from a direct operator means fewer middlemen and clearer accountability for the performance you pay for. viabandwidth verifies this classification from network ownership signals rather than taking the provider's word for it.
A marketplace aggregates GPU supply from many underlying operators and sells it through one platform. Breadth and price competition are the draw, while the tradeoff is that the underlying hardware, location and operator change from listing to listing. Teams with strict compliance or performance needs should verify the source behind a specific offer before committing.
A reseller buys capacity from a larger cloud or infrastructure operator and sells it on with a managed layer. That can simplify procurement, though a margin is built in and transparency over where the hardware actually runs is limited. Comparing against the underlying provider is worth the ten minutes before a large commit.
A carrier's core business is moving traffic: transit, wavelengths, IP connectivity. Some carriers list GPU or compute products as a sideline, which is why they occasionally appear in compute searches. Judge them first on the network, then on the compute offer.
These signals come from observing a provider's own network footprint. They tell you what the company actually runs, which often differs from what its website claims.
An eyeball ISP is industry shorthand for a network whose customers are people watching, browsing and downloading, in other words consumer and small-business internet access. Traffic flows toward those eyeballs rather than out of server racks. When a GPU provider's network reads as an eyeball ISP, its core business is likely internet access rather than data-center compute, which is useful context when weighing its compute claims.
A transit carrier sells passage across its backbone: other networks pay it to carry their traffic to the rest of the internet. The signature is lots of infrastructure in exchange points and long-haul routes with few end-user addresses. For a compute buyer this signals serious network engineering, though it says nothing by itself about GPU capacity.
This traffic profile matches a content-delivery network or public cloud: many points of presence, heavy outbound serving, infrastructure named for caching and edge delivery. Providers with this profile typically run real server fleets, which is a good baseline signal for a compute business.
A hosting profile means the network's address space resolves to servers rather than to consumer connections: dedicated machines, virtual servers, racks of customer gear. This is the most natural network shape for a GPU provider, since selling compute is a hosting business at heart.
Mixed means the network shows more than one dominant role: some address space looks like hosting, some like consumer access, some like backbone. Larger telecoms and national operators commonly read as mixed because they genuinely do several things. It is neither good nor bad on its own; it simply means the network signal cannot pin the company to a single business.
Active means viabandwidth observed live, named infrastructure across the provider's announced address space. Machines answer, names are maintained, the network is clearly in production. This is the strongest baseline signal that an operator runs real infrastructure rather than a brochure.
Sparse means the provider announces address space but only a small share of it shows live, named infrastructure. That can be legitimate, for instance a young operator or one holding room to grow, though it earns a closer look when the marketing claims a large fleet.
Dark means the network announces address space yet none of it showed observable activity when checked. Some genuine operators keep infrastructure invisible by design, so dark is a flag to investigate rather than a verdict. Paired with big capacity claims it deserves hard questions.
Reverse DNS is the practice of giving network addresses human-readable names, something operators only bother with on infrastructure they actually run and maintain. A high coverage percentage means most of the network carries maintained names, which correlates strongly with real production infrastructure. Low coverage on a network claiming a large fleet is worth questioning.
Own DC means the observed infrastructure sits in facilities the operator appears to run itself. Owning the building and power chain gives an operator the most control over density, cooling and expansion, which matters for large GPU deployments.
Leased colo means the operator's hardware lives in colocation facilities run by someone else. This is the normal model for most GPU clouds and carries no stigma; the operator still owns and runs the machines. The practical difference from Own DC is that expansion and power ceilings depend on the landlord.
Cloud reseller means network signals place the service on a hyperscaler such as AWS, Google Cloud or Azure rather than on hardware the operator controls. You may still get a good product, though you are paying two margins and the provider cannot make hardware-level guarantees. Worth knowing before you negotiate.
Certifications shown on a profile were found on the operator's own materials. They tell you what a provider can contractually commit to, from uptime design to handling regulated data.
Tier III is an Uptime Institute data-center rating meaning the facility is concurrently maintainable: every component that matters has a planned path to be serviced without taking customer load offline. In practice buyers read it as roughly 99.98% design availability. It describes the building, not the provider's software or support.
Tier IV is the highest Uptime Institute rating, meaning fault tolerance: the facility rides through a single unplanned equipment failure without dropping load, not just planned maintenance. Design availability is roughly 99.995%. Few workloads truly need it, though for always-on inference it is a meaningful differentiator.
ISO 9001 certifies a quality-management system: the operator documents its processes, follows them and passes independent audits on both. It says nothing specific about security or uptime. What it tells a buyer is that the company is process-driven rather than improvised, which tends to show up in support and change management.
ISO 27001 certifies an information-security management system: access control, risk assessment, incident handling and the rest, verified by an external auditor on a recurring cycle. For buyers with security review processes this is usually the first checkbox. Ask for the certificate scope, since it can cover one facility or the whole company.
ISO 14001 certifies an environmental-management system. For data-center operators it typically covers energy sourcing, waste handling and emissions reporting. Relevant if your procurement carries sustainability requirements.
SOC 2 is an American auditing standard where an independent CPA firm examines a provider's controls for security, availability, processing integrity, confidentiality and privacy. A Type II report covers how the controls performed over months, not just how they look on paper. US enterprise buyers ask for it almost by reflex.
HIPAA is the US law governing protected health information. A provider advertising HIPAA compliance is signalling it can sign a Business Associate Agreement and run workloads that touch patient data. There is no official HIPAA certificate, so ask what an auditor actually attested and whether the BAA covers the specific services you will use.
PCI DSS is the payment-card industry's security standard. Infrastructure that stores, processes or transmits card data must meet it, and a provider's attestation means its platform can sit inside your cardholder-data environment. Mostly relevant to fintech and commerce workloads rather than model training.
GDPR is the EU's data-protection regulation. Every company serving EU users must comply, so a provider citing it is really signalling two things: contractual readiness (data-processing agreements) and often EU data residency options. If residency is the requirement, confirm the actual facility locations rather than the badge.
Terms you will meet while comparing offers and reading pricing tables.
Indicative pricing means the rates were gathered from the provider's published price lists at a point in time. Real quotes move with commit length, capacity, storage and networking, so treat list rates as a comparison baseline rather than the number you will sign. Profiles show when the pricing was last reviewed.
On-demand means renting by the hour or second with no commitment. It is the right mode for experiments, bursts and evaluation, while sustained workloads almost always price better on a reserve or commit. The spread between on-demand and committed rates varies widely between providers, which makes it worth comparing both numbers.
Reserved capacity is a commitment to a block of GPUs for a fixed term, weeks to years, in exchange for a lower rate and guaranteed availability. For training runs that must not be preempted, the guarantee often matters more than the discount.