News analysis MLPerf Inference 6.1: what the new results mean for buyers
The latest results broaden the view of inference. Compare the complete workload, not just the headline score.
Read the analysisSearch the catalogue
Search by manufacturer, model or catalogue reference.News, analysis & opinion
News analysis, industry perspectives and procurement guides for teams building AI infrastructure.
News analysis The latest results broaden the view of inference. Compare the complete workload, not just the headline score.
Read the analysisPlatforms, capacity deals and the companies building them.
Memory, power, networking and the installation.
System selection, comparable quotes and a clear business case.
Editorial illustrationThe latest results broaden the view of inference. Compare the complete workload, not just the headline score.
Read article
Manufacturer imagePreviously deployed hardware needs a documented condition and test scope. A useful report connects the unit's identity, operating conditions and results to the equipment you are purchasing.
Read article
Editorial illustrationSeptember 2026: the next contest reaches beyond accelerator performance into memory, software, electricity and the ability to deliver usable capacity.
Read article
Event photographNVIDIA's next platform is moving through the supply chain. Production status and a deliverable system are different pieces of information.
Read article
Announcement artworkA major Anthropic commitment puts AMD's rack-scale approach in focus. Software readiness will matter as much as the accelerator specification.
Read article
Manufacturer imageThe September agreement concerns growing CPU workloads. It is a reminder that an AI service contains much more than its model-serving GPUs.
Read article
Architectural visualisationThe planned Finnish system combines hardware, delivery, installation and maintenance. Its timetable is as important as its headline value.
Read article
Editorial illustrationThe July call is a milestone. For startups and smaller businesses, the surrounding support and access arrangements deserve equal attention.
Read article
Manufacturer imageTPU 8t and TPU 8i point to a market where the best architecture depends increasingly on the job being done.
Read article
Editorial illustrationCustom silicon is changing the capacity market. It also asks buyers to compare platforms and commercial commitments, rather than chips alone.
Read article
Editorial illustrationCapacity, bandwidth and the supply chain affect useful compute in different ways. They should not be collapsed into a single specification.
Read article
Announcement artworkThe IEA's 2026 update shows rising datacenter electricity demand alongside efficiency gains. Both developments can be true at once.
Read article
Announcement artworkA new alliance proposes changing how AI facilities use power. The difficult part is deciding which work can move without breaking the service.
Read article
Editorial illustrationInference v6.1 adds RAG and edge-agent workloads. That makes the benchmark suite more relevant, while leaving the usual comparison rules intact.
Read article
Editorial illustrationModels such as Qwen3.8 make self-hosting a practical subject for evaluation. The software licence is only one part of the operating decision.
Read article
Editorial illustrationModel weights are only part of the footprint. The way a service handles context can change the hardware it needs.
Read article
Manufacturer imageOptical networking and rack-scale systems receive attention for a reason. Multi-GPU performance depends on how data moves between the machines.
Read article
Manufacturer imageIntegrated racks can simplify delivery. The facility connections, service responsibilities and acceptance checks still need to be specified.
Read article
Manufacturer imagePerplexity's Windows expansion brings another practical example of local model execution. Hardware sizing and data boundaries still need careful reading.
Read article
Manufacturer imageCoreWeave's quarterly figures show strong demand alongside the cost of building capacity. A backlog is not the same thing as a running datacenter.
Read article
Manufacturer imageOur view: a system's generation matters less than the useful work it can deliver within your project's budget and timetable.
Read article
Editorial illustrationOur view: the address of a datacenter is one part of control. The ability to move the workload is another.
Read article
Manufacturer imageOur view: competition becomes useful when your team can actually deploy the alternative.
Read article
Editorial illustrationOur view: capacity planning should start with completed useful work, not with a utilisation chart or a list of new accelerators.
Read article
Manufacturer imageChoose the right unit of purchase before comparing prices. A PCIe card, an accelerator module and a populated server are very different orders.
Read article
Manufacturer imageCompare the GPU topology, populated system and installation requirements. The headline accelerator name is only the first line of the specification.
Read article
Manufacturer imageA lower headline price can leave out the GPUs, host components or support you expected. Use a common specification to compare the actual order.
Read articleNo matching articles. Try a different term or clear the filters.
We link to the sources behind our reporting and clearly label opinion pieces. We distinguish announced projects and future delivery plans from equipment already operating.
Our editorial team writes about AI infrastructure, equipment procurement and the industry behind it. News analysis distinguishes reported developments from our conclusions; opinion articles are labelled as such.
Technical and industry references are linked within each article. Publication dates describe when an article was written, rather than implying that every specification or market condition remains unchanged.