Glossary

A model trained on large volumes of text that predicts the next token and, by doing so, writes, summarises, classifies and answers questions. In an enterprise setting the interesting question is rarely which model, but what data it is allowed to see and where it runs.

Training is building the model; inference is using it to answer. Almost every business deployment is inference only, which is why a modest amount of hardware goes much further than the headlines about training clusters suggest.

The memory a model holds for a conversation in progress, so it does not re-read the whole context on every token. It grows with context length and with the number of simultaneous users, and on a busy inference server it usually consumes more GPU memory than the model weights do.

An NVIDIA feature that splits one physical GPU into hardware-isolated instances, each with its own memory and compute. Unlike time-slicing, the partitions do not interfere with each other, which is what makes it usable for separate tenants or production workloads.

NVIDIA’s direct GPU-to-GPU interconnect, much faster than sending the same data across PCIe. It matters when a model is too large for one card and has to be split across several; for single-card inference it changes nothing.

Storing model weights at lower numeric precision so they take less memory and run faster. Each step down roughly halves the memory needed and costs some accuracy – how much depends entirely on the model and the task, so it is measured rather than assumed.

Searching your own documents for relevant passages and giving them to the model as context, so it answers from your material rather than from what it memorised in training. Updating a document updates the answer, with no retraining involved.

The operational discipline around a model in production: versioning, deployment, monitoring for drift, evaluation before a change ships, and rollback when it does not work. It is the difference between a prototype and something a business can rely on.

NVIDIA software that lets several virtual machines share one physical GPU by taking turns. More flexible than MIG in how it divides a card, less isolated, and it needs its own licence per concurrent user.

The memory on the graphics card itself. For local LLM work it is the binding constraint: if the weights plus the KV cache do not fit, the model does not run, regardless of how fast the card is.

The time a virtual machine was ready to run but had no physical core available. It is the number that explains a VM that feels slow while its CPU usage graph looks calm, and it usually points at a machine with more vCPUs than the host can schedule at once.

The VMware feature that moves running virtual machines between hosts to balance load. It works well when the machines are sized sensibly and it cannot rescue a cluster full of oversized ones.

Restarting virtual machines automatically on another host when their host fails. It protects against hardware failure within a site – it is not disaster recovery, because it assumes the site itself is still there.

The layer that runs several virtual machines on one physical server and keeps them apart. Everything else in a virtualised estate – licensing, capacity, availability – is a consequence of decisions made at this layer.

Charging by the physical cores in a host rather than by sockets or virtual machines, typically with a minimum number of cores counted per CPU. It makes the physical layout of an estate a financial question: many small hosts cost more than fewer dense ones for the same workload.

A point-in-time marker that lets a virtual machine be rolled back, after which all changes are written to a separate growing file. It is a safety net for an hour, not a backup – snapshots left in place for months are a common cause of full datastores.

VMware’s software-defined storage, built from the disks inside the hosts rather than from a separate storage array. Under current licensing an amount of vSAN capacity is included per licensed core, and it is measured against raw physical capacity, not usable space.

Moving a running virtual machine from one host to another without stopping it. It is what makes patching and hardware maintenance possible during working hours.

Recovery capacity provided as a service rather than built and owned: somewhere for the workload to run, the replication to get it there, and an agreed recovery target. The value is in the rehearsed procedure as much as in the capacity.

A copy that cannot be altered or deleted until its retention period expires, including by an administrator account. Modern ransomware looks for the backup system first, which is what makes immutability the difference between having a backup and having had one.

Requiring a second proof of identity beyond a password. It is the single control with the best ratio of effort to risk removed, and it is the one most often missing on administrative and remote-access accounts specifically.

An EU directive on cybersecurity that extends obligations – risk management, incident reporting, supply-chain security – to far more organisations than the rules it replaces. It applies to entities in listed sectors above a size threshold, and each member state implements it in its own national law, so the operative text is the national one.

How much data you accept losing, measured in time: an RPO of one hour means up to an hour of work has to be re-entered after an incident. It is decided by how often data is copied, not by how fast you can restore.

How long the business is prepared to be down, measured from the start of the incident to people working again. It includes detection, the decision to declare, the recovery itself and the validation afterwards – which is why estimates that time only the restore are always optimistic.

The contractual statement of what is guaranteed: response times, recovery targets, availability, and what happens when they are missed. A recovery target in an SLA that has never been tested is a number, not a commitment.

An approach that stops treating the internal network as trusted and verifies every request by identity and device instead. In practice it is a sequence of changes – identity, segmentation, least privilege – rather than a product you install.

The requirement that data is stored and processed in a defined jurisdiction. Under the GDPR, personal data leaving the EU or EEA needs a lawful transfer mechanism, which is why residency often decides an architecture before cost is even discussed.

Cloud-style provisioning – self-service, elastic, automated – on infrastructure dedicated to one organisation, whether owned or hosted. It is the usual answer when public cloud is ruled out for residency or contractual reasons but the operating model is still wanted.

A data centre certified as concurrently maintainable: every power and cooling component can be taken out of service for maintenance without stopping the equipment. It is the level most enterprise workloads are hosted at, above Tier II and below the fully fault-tolerant Tier IV.

The two VMware Foundation bundles: Cloud Foundation is the full private-cloud stack, vSphere Foundation the virtualisation platform with operations tooling. Both are licensed per core and both include an amount of vSAN capacity per licensed core – a different amount each.