Tech
What are the storage requirements for AI training and inference? | Computer Weekly
Despite ongoing speculation around an investment bubble that may be set to burst, artificial intelligence (AI) technology is here to stay. And while an over-inflated market may exist at the level of the suppliers, AI is well-developed and has a firm foothold among organisations of all sizes.
But AI workloads place specific demands on IT infrastructure and on storage in particular. Data volumes can start big and then balloon, in particular during training phases as data is vectorised and checkpoints are created. Meanwhile, data must be curated, gathered and managed throughout its lifecycle.
In this article, we look at the key characteristics of AI workloads, the particular demands of training and inference on storage I/O, throughput and capacity, whether to choose object or file storage, and the storage requirements of agentic AI.
What are the key characteristics of AI workloads?
AI workloads can be broadly categorised into two key stages – training and inference.
During training, processing focuses on what is effectively pattern recognition. Large volumes of data are examined by an algorithm – likely part of a deep learning framework like TensorFlow or PyTorch – that aims to recognise features within the data.
This could be visual elements in an image or particular words or patterns of words within documents. These features, which might fall under the broad categories of “a cat” or “litigation”, for example, are given values and stored in a vector database.
The assigned values provide for further detail. So, for example “a tortoiseshell cat”, would comprise discrete values for “cat” and “tortoiseshell”, that make up the whole concept and allow comparison and calculation between images.
Once the AI system is trained on its data, it can then be used for inference – literally, to infer a result from production data that can be put to use for the organisation.
So, for example, we may have an animal tracking camera and we want it to alert us when a tortoiseshell cat crosses our garden. To do that it would infer the presence or not of a cat and whether it is tortoiseshell by reference to the dataset built during the training described above.
But, while AI processing falls into these two broad categories, it is not necessarily so clear cut in real life. It will always be the case that training will be done on an initial dataset. But after that it is likely that while inference is an ongoing process, training also becomes perpetual as new data is ingested and new inference results from it.
So, to labour the example, our cat-garden-camera system may record new cats of unknown types and begin to categorise their features and add them to the model.
What are the key impacts on data storage of AI processing?
At the heart of AI hardware are specialised chips called graphics processing units (GPUs). These do the grunt processing work of training and are incredibly powerful, costly and often difficult to procure. For these reasons their utilisation rates are a major operational IT consideration – storage must be able to handle their I/O demands so they are optimally used.
Therefore, data storage that feeds GPUs during training must be fast, so it’s almost certainly going to be built with flash storage arrays.
Another key consideration is capacity. That’s because AI datasets can start big and get much bigger. As datasets undergo training, the conversion of raw information into vector data can see data volumes expand by up to 10 times.
Also, during training, checkpointing is carried out at regular intervals, often after every “epoch” or pass through the training data, or after changes are made to parameters.
Checkpoints are similar to snapshots, and allow training to be rolled back to a point in time if something goes wrong so that existing processing does not go to waste. Checkpointing can add significant data volume to storage requirements.
So, sufficient storage capacity must be available, and will often need to scale rapidly.
What are the key impacts of AI processing on I/O and capacity in data storage?
The I/O demands of AI processing on storage are huge. It is often the case that model data in use will just not fit into a single GPU memory and so is parallelised across many of them.
Also, AI workloads and I/O differ significantly between training and inference. As we’ve seen, the massive parallel processing involved in training requires low latency and high throughput.
While low latency is a universal requirement during training, throughput demands may differ depending on the deep learning framework used. PyTorch, for example, stores model data as a large number of small files while TensorFlow uses a smaller number of large model files.
The model used can also impact capacity requirements. TensorFlow checkpointing tends towards larger file sizes, plus dependent data states and metadata, while PyTorch checkpointing can be more lightweight. TensorFlow deployments tend to have a larger storage footprint generally.
If the model is parallelised across numerous GPUs this has an effect on checkpoint writes and restores that mean storage I/O must be up to the job.
Does AI processing prefer file or object storage?
While AI infrastructure isn’t necessarily tied to one or other storage access method, object storage has a lot going for it.
Most enterprise data is unstructured data and exists at scale, and it is often what AI has to work with. Object storage is supremely well suited to unstructured data because of its ability to scale. It also comes with rich metadata capabilities that can help data discovery and classification before AI processing begins in earnest.
File storage stores data in a tree-like hierarchy of files and folders. That can become unwieldy to access at scale. Object storage, by contrast, stores data in a “flat” structure, by unique identifier, with rich metadata. It can mimic file and folder-like structures by addition of metadata labels, which many will be familiar with in cloud-based systems such as Google Drive, Microsoft OneDrive and so on.
Object storage can, however, be slow to access and lacks file-locking capability, though this is likely to be of less concern for AI workloads.
What impact will agentic AI have on storage infrastructure?
Agentic AI uses autonomous AI agents that can carry out specific tasks without human oversight. They are tasked with autonomous decision-making within specific, predetermined boundaries.
Examples would include the use of agents in IT security to scan for threats and take action without human involvement, to spot and initiate actions in a supply chain, or in a call centre to analyse customer sentiment, review order history and respond to customer needs.
Agentic AI is largely an inference phase phenomenon so compute infrastructure will not need to be up to training-type workloads. Having said that, agentic AI agents will potentially access multiple data sources across on-premises systems and the cloud. That will cover the range of potential types of storage in terms of performance.
But, to work at its best, agentic AI will need high-performance, enterprise-class storage that can handle a wide variety of data types with low latency and with the ability to scale rapidly. That’s not to say datasets in less performant storage cannot form part of the agentic infrastructure. But if you want your agents to work at their best you’ll need to provide the best storage you can.
Tech
With a Memory Shortage on the Horizon, Here’s Which MacBook to Buy
All of Apple’s processors are scattered throughout different MacBook models. While Apple only currently sells M4 MacBooks, you can find older models at specific third-party retailers online, either completely new or refurbished. If you do stumble upon its older chips (which came out four years ago), you might be wondering how they compare to other options. We break down the differences between each one.
M5 Series
M5: The rollout of the M5 line of chips has just started. The base M5 still has up to a 10-core CPU and 10-core GPU, although there’s also a lower-tier 9-core CPU that’s available in the iPad Pro—and presumably, that’ll also be offered in the M5 MacBook Air at some point. The M5 is around 10 to 15 percent faster in CPU performance, but also takes a significant step up in GPU, AI workloads, and even storage speed.
M4 Series
M4: The M4 originally launched in 2024. It has a 10-core CPU and a 10-core GPU. Apple claims it delivers 1.8 times faster CPU performance and 2.2 times faster GPU performance than the M1. Meanwhile, the neural engine is over three times faster than the original and twice as fast as the M3. It also starts with 16 GB of unified memory, which will help power Apple Intelligence (the company’s suite of artificial intelligence features) a lot more smoothly. It’s available on the 14-inch MacBook Pro (2024), iMac (2024), and MacBook Air (13-inch and 15-inch, 2025).
M4 Pro: The M4 Pro has a 14-core CPU (which Apple claims is up to 1.9 times faster than the M1 Pro) and up to a 20-core GPU, with up to 64 GB of unified memory. Built on a second-generation 3-nanometer process, it also supports enhanced GPU features like mesh shading and ray tracing—the latter of which is now twice as fast as on M3 chips. You’ll find it on the latest MacBook Pro (14-inch and 16-inch) and Mac Mini (2024).
M4 Max: This chip has a 16-core CPU and up to a 40-core GPU with support for up to 128 GB of unified memory. Apple says the CPU is up to 2.2 times faster than the M1 Max, while the GPU is up to 1.9 times faster. As with the M4 Pro, it packs support for mesh shading and ray tracing. The M4 Max is currently the most powerful chip you can get in a MacBook, and is available on the latest 14-inch and 16-inch MacBook Pro. You can also get it as an option in the current Mac Studio.
M3 Series
M3: The M3 is available on the 14-inch MacBook Pro (late 2023), 13-inch MacBook Air (2024), 15-inch MacBook Air (2024), and 24-inch iMac (2023). It packs an 8-core CPU and up to a 10-core GPU with 24 GB of unified memory. When compared to the M1, Apple claims CPU performance is up to 35 percent faster, and GPU performance is up to 65 percent faster. The company says the CPU and GPU are both 20 percent faster than the M2. As with the M1 and M2, it’s great for basic tasks like word processing, sending emails, using spreadsheets, and light gaming. With the 13-inch and 15-inch MacBook Air, you also have support for two external displays (one display with up to 6K resolution at 60 Hz and another with up to 5K resolution at 60 Hz).
M3 Pro: With a 12-core CPU and an 18-core GPU, Apple claims the M3 Pro’s GPU is only up to 10 percent faster than the M2 Pro—making this a marginal upgrade from its predecessor. Compared to the M1 Pro, however, the M2 Pro is up to 40 percent faster in GPU performance and 20 percent faster in CPU performance. It’s available on the 14-inch and 16-inch MacBook Pro from 2023. It’s the ideal in-between for those who need a chip that’s more powerful than the M3 but won’t utilize the full power of the M3 Max.
M3 Max: This is the next step up from the M2 Max and the most powerful of the three chips (but still not as powerful as the M2 Ultra). It has a 16-core CPU, 40-core GPU, and up to 128 GB of unified memory. According to Apple, the CPU performance is up to 80 percent faster than the M1 Max and up to 50 percent faster than the M2 Max. As for GPU performance, it’s said to be up to 50 percent faster than the M1 Max and 20 percent faster than the M2 Max. The M3 Max is available on the 14-inch and 16-inch MacBook Pro (late 2023).
M3 Ultra: While the M3 lineup was introduced in 2023, Apple announced an M3 Ultra in 2025. Confusingly, it remains the most powerful chip in the M-series lineup—even better than the latest M4 Max and M5. It has an up to 32-core CPU (with 24 performance cores) and a GPU with up to 80 cores. Apple claims it’s up to 2.5 times faster than the M1 Ultra. It also comes with 96 GB of unified memory, with the option to upgrade up to 512 GB, while SSD storage can be increased to 16 GB. This chip is currently only available on the 2025 Mac Studio.
M2 Series
M2: You might think the M2 is better than the M1 Pro or M1 Max, but you’d be wrong. It’s an entry-level chip like the M1, with slightly more processing power. It packs an 8-core CPU and up to a 10-core GPU (two more GPU cores than its predecessor), along with support for up to 24 GB of unified memory. Apple says the second-generation chip has an 18 percent faster CPU and a GPU that’s 35 percent more powerful. The M2 is great for daily tasks like word processing and web browsing, but tasks like editing multiple streams of 4K footage and 3D rendering should be reserved for the M1 Pro or M1 Max (or the next two chips). It’s available in the MacBook Air (13-inch, 2022), MacBook Air (15-inch, 2022), and MacBook Pro (13-inch, 2022).
M2 Pro: The M2 Pro is the next step up from the M2. It has up to 12 cores in the CPU and up to a 19-core GPU, with up to 32 GB of unified memory. Apple claims performance is up to 20 percent faster than the 10-core M1 Pro and graphics are 30 percent faster. We recommend this chip for intermediate video and photo editors. It’s a marginal upgrade compared to the M1 Pro, but it’s the best option for those who want a more future-proof processor. You’ll find it in the MacBook Pro (14-inch and 16-inch) from early 2023 and the Mac Mini (2023).
M2 Max: The M2 Max packs up to a 12-core CPU and up to a 38-core GPU (with support for up to 96 GB of unified memory). According to Apple, graphics are 30 percent faster than the M1 Max. The M2 Max is an excellent choice for those who work with graphics-intensive content, including graphic design, 3D modeling, and heavy-duty video footage. But as with the M2 Pro, it’s an incremental upgrade if you’re coming from an M1 Max. It’s available in the MacBook Pro (14-inch and 16-inch) that came out early in 2023 and the Mac Studio (2023).
M2 Ultra: This is the successor to the M1 Ultra. It’s available on the second-generation Mac Studio and the Mac Pro (2023). Composed of two M2 Max chips, using Apple’s UltraFusion technology, the M2 Ultra has a 24-core CPU and a GPU configurable with 60 or 76 cores. Apple claims the CPU delivers up to 20 percent faster performance and a 30 percent faster GPU than the M1 Ultra. This is the chip to get if you’re working with extremely heavy-duty content that you believe the M1 Ultra, M2 Pro, or M2 Max simply won’t be able to handle. You’ll know if you need a chip this robust.
M1 Series
M1: Shockingly, Apple continues to sell the M1 MacBook Air through Walmart for just $599, which is a killer price for this laptop. This was the first custom silicon Apple debuted for its MacBook Air in 2020. It has an 8-core CPU and up to an 8-core GPU. Originally, there was support for up to 16 GB of unified memory (RAM) at an extra cost, but nowadays you can only purchase the 8-GB model. It’s much faster than any previous Intel-powered MacBook Pro, and it is the practical choice for most people, as it’s inside the most affordable MacBook Air you can buy (from third-party retailers). It packs more than enough processing power to get you through common day-to-day tasks, even light gaming, and it can handle more intense jobs like photo editing.
M1 Pro: From there, the next step up was the M1 Pro. It has up to 10 cores in the CPU and up to a 16-core GPU, with up to 32 GB of unified memory. Apple says performance and graphics are both twice as fast as on the M1. We found it to be considerably more capable than the base chip, ideal for anyone who works heavily on MacBooks for music production or photo and video editing. Only the MacBook Pro (14-inch and 16-inch) from 2021 uses this chip.
M1 Max: Like the M1 Pro, the M1 Max has a 10-core CPU but a heftier 32-core GPU (with support for up to 64 GB of unified memory). Apple says it’s four times faster than the M1 in terms of graphics. As proven in testing, this chip is extremely powerful and handles every heavy-duty task with ease. It was the go-to choice if you needed a computer that could handle multiple streams of 8K or 4K video footage, 3D rendering, or developing apps and running demos. You probably already know whether you need this much power. It’s available in the MacBook Pro (14-inch and 16-inch) from 2021.
M1 Ultra: The M1 Ultra was the most powerful of them all. It’s two M1 Max chips connected with a technology called UltraFusion. It packs a 20-core CPU, 64-core GPU (which can be configured with up to 128 GB of unified memory), and a 32-core neural engine—complete with seven times more transistors than the base M1. Even with the M3 Ultra now available, the M1 Ultra remains powerful and a solid option for anyone who needs a heavy-duty processor for working with intense visuals and graphics. It was available only on the first-generation Mac Studio.
Tech
Americans Are Increasingly Convinced That Aliens Have Visited Earth
Americans are becoming more open to the idea that aliens have visited Earth, according to a series of polls that show belief in alien visitation has been steadily on the rise since 2012.
Almost half—47 percent—of Americans say they think aliens have definitely or probably visited Earth at some point in time, according to a new poll from YouGov conducted in November 2025 that involved 1,114 adult participants. That percentage is up from roughly a third (36 percent) of Americans polled in 2012 by Kelton Research, with the exact same sample size. Gallup published polls on this question in 2019 and 2021 that likewise show an upward trend.
Moreover, people seem to be getting off the fence on this issue, one way or the other. Just 16 percent of Americans said they were unsure if aliens had visited Earth in the new poll, down from 48 percent who were unsure in 2012. Meanwhile, even as belief in alien visitation has risen, so has doubt: The new poll shows that 37 percent of Americans said Earth likely hasn’t been visited by aliens, more than double the 17 percent logged in 2012.
It’s impossible to know exactly why Americans have become more receptive to alien visitation from these polls alone; they only include raw statistics, and lack granular details about the specific motivations for the participants’ responses.
“It’s important to note that this is a poll about belief,” says Susan Lepselter, an author and associate professor of anthropology and American Studies at Indiana University who has written extensively on alien beliefs and UFO experiences. “It’s not a poll about experience, contact, feelings—nothing like that.”
“We don’t know what their engagement is; we don’t know if their belief has been life-changing,” she adds. “We just know one thing, which is that the statistics have moved from one set of beliefs to another.”
Of course, it’s still possible—and let’s be real, fun—to speculate on the drivers of the trend. One obvious culprit is a new posture from institutional news sources, such as the US government and legacy media, which have finally started taking unidentified anomalous phenomena (UAP) seriously.
This shift began with the release of mysterious Pentagon UAP videos by The New York Times in 2017, and has since been accelerated by spate of Congressional hearings, and a NASA independent study on UAP. The newly released documentary The Age of Disclosure, which features claims by former military officials that the US government has covered up evidence of aliens visiting Earth, has supercharged the legitimacy to this once marginalized topic.
Tech
UK government confirms Foreign Office cyber attack | Computer Weekly
The UK government has admitted that IT systems at the Foreign, Commonwealth and Development Office (FCDO) were hacked in October, but insists the attack had a “low risk” of personal data being compromised.
During a round of broadcast interviews today (19 December 2025), trade minister Chris Bryant said it was “not clear” who perpetrated the attack, although the first report on the hack, revealed in The Sun, attributed it to a China-based threat actor known as Storm 1849.
The same group was blamed for targeting vulnerabilities in Cisco equipment that led to a National Cyber Security Centre (NCSC) warning in September for organisations using Cisco’s Adaptive Security Appliance family of unified threat management systems. Users were told to replace any devices reaching end-of-life support, noting the significant risks that ageing or obsolete hardware can pose.
Bryant said some of the reports about the FCDO hack were “speculation”, but that the government had managed to “close the hole” quickly, and that security experts were confident there was a “low risk” of any individual being affected. The Sun report claimed hackers accessed confidential data and documents, possibly including thousands of visa details.
The Storm 1849 attack campaign on Cisco equipment was dubbed ArcaneDoor, and targeted two zero-day vulnerabilities. One was a high-severity denial-of-service vulnerability capable of remote code execution; the other was a high-severity persistent local code execution vulnerability.
While government IT systems always face scrutiny over cyber security, the hack will provide further fuel for critics of plans to introduce a national digital ID scheme, many of whom have already raised concerns about the potential risks of gathering citizen identity data.
The development also comes a day after ITV News broadcast a report on the cyber security issues found in One Login – the government single sign-on system that will be at the heart of the digital ID plan – which were first revealed by Computer Weekly in April.
Damaging year
2025 has been a notably damaging year for cyber attacks, with high-profile ransomware campaigns affecting Jaguar Land Rover (JLR), the Co-op and Marks & Spencer.
The Office for National Statistics attributed a November decline in the UK’s economy partly to the impact of the JLR attack, which stopped car production at the manufacturer and had a knock-on impact across the automotive supply chain.
Last month, four London councils – Kensington and Chelsea; Hackney; Westminster; and Hammersmith and Fulham – suffered cyber attacks, disrupting services and prompting an NCSC investigation. Westminster has since admitted that potentially sensitive data was copied from its systems during the hack. Three of the local authorities operate a shared IT service.
-
Business6 days agoHitting The ‘High Notes’ In Ties: Nepal Set To Lift Ban On Indian Bills Above ₹100
-
Politics1 week agoTrump launches gold card programme for expedited visas with a $1m price tag
-
Business1 week agoRivian turns to AI, autonomy to woo investors as EV sales stall
-
Sports1 week agoPolice detain Michigan head football coach Sherrone Moore after firing, salacious details emerge: report
-
Fashion1 week agoTommy Hilfiger appoints Sergio Pérez as global menswear ambassador
-
Business1 week agoCoca-Cola taps COO Henrique Braun to replace James Quincey as CEO in 2026
-
Sports1 week agoU.S. House passes bill to combat stadium drones
-
Tech1 week agoGoogle DeepMind partners with UK government to deliver AI | Computer Weekly
