ASUS has expanded its desktop AiO lineup in India with the launch of the ASUS VM240 AiO, a new all-in-one PC designed for everyday work, learning, and entertainment. Priced at ₹66,990, the VM240 combines a 24-inch display, AMD Ryzen processors, and a space-saving design into a single desktop setup.

Alongside the VM240 AiO, ASUS has also introduced four new accessories in India: the Jelly75 mechanical keyboard, Wireless Mouse MD105, Marshmallow keyboard and mouse combo, and the ROG HDMI Cable LHR60.

ASUS VM240 AiO Specs

The ASUS VM240 is aimed at users who want a clutter-free desktop without sacrificing everyday performance. It features a 24-inch Full HD display with an adjustable tilt stand and a slim 22mm profile, making it suitable for home offices, classrooms, and entertainment setups.

Powered by up to an AMD Ryzen 5 processor, the desktop also includes up to 512GB PCIe Gen 4 SSD storage for faster app launches and multitasking. ASUS says the Ryzen 5 variant will become available in mid-September, while the VM240 officially goes on sale from August 20. For video calls, the AiO includes a 1080p webcam with a physical privacy shutter, dual front-facing 2W speakers tuned with Dirac audio, and AI-powered noise cancellation for clearer meetings and online classes.

Four New Accessories

The VM240 launch is accompanied by a refreshed accessories lineup covering productivity, portability, and gaming. The standout product is the ASUS Jelly75 KD201, a gasket-mounted mechanical keyboard with hot-swappable switches, USB-C charging, and support for Bluetooth, 2.4GHz wireless, and wired connectivity across up to five devices.

The Wireless Mouse MD105 supports Bluetooth 5.0 and 2.4GHz wireless, offers adjustable DPI settings up to 2400, and can pair with three devices simultaneously. For users looking for a lightweight setup, the Marshmallow CW200 bundles a compact keyboard and mouse with multi-device support and a dedicated Copilot key. Meanwhile, the ROG HDMI Cable LHR60 supports 48Gbps bandwidth, enabling 8K at 60Hz and 4K at 144Hz output for gaming PCs and current-generation consoles.

The ASUS VM240 AiO and accessories will be available through ASUS eShop, ASUS Exclusive Stores, Flipkart, Amazon, and other retail partners.

EdSurge announced six new members to its Advisory Board, among them a mix of experienced journalists and education leaders.

Todd Frankel, Patrick Harris II, Nneka McGee, Ramona Schindelheim, Avery Thrush and Manoush Zomorodi have joined the board, each bringing a wealth of experience and expertise. The board will serve as a vital resource for the EdSurge newsroom, offering strategic insights and industry perspectives across all areas of coverage.

“Partnering with such a talented, versatile group of journalists and educators is a huge win for EdSurge and our readers,” says Sarah McKibben, EdSurge Editor-in-Chief. “With their unmatched expertise, this board brings both national perspective and deep classroom understanding to our newsroom. They will help us tackle pressing challenges like AI policy while ensuring our early childhood and K-12 coverage remains relevant, cutting-edge, and grounded in real-world impact.”

Designed to maximize impact while respecting members’ time, the board acts as a strategic, high-level advisory body. In accordance with the Institute for Nonprofit News standards, board members provide strategic guidance while holding no editorial influence or veto power on published content.

Meet the Advisory Board

Todd Frankel is a Pulitzer Prize-winning reporter at The Washington Post. The D.C. native has previously worked at newspapers in Missouri, Washington, West Virginia, Kentucky.

Frankel brings award-winning journalism experience to the board, particularly in the role of business and economics in all areas of life.  He anchored the lead story in a series that won the 2024 Pulitzer Prize for National Reporting for coverage of the AR-15 rifle’s role in mass shootings and American culture.

Patrick Harris II is a storyteller and educator on a mission to ensure humanity is centered in the future of education. A former middle school teacher and dean of students in metro Detroit, Harris was named a 30 under 30 Honoree from the International Literacy Association in 2021 and an EdSurge Voices for Change Fellow in 2022-23. His book focusing on new educators, “The First Five: A Love Letter to Teachers,” was named a finalist for Best Professional Audiobook by the Audies. Harris is currently a doctoral student studying popular culture at Michigan State University. 

Having previously written for EdSurge, Harris hopes to use his new involvement as a board member to support EdSurge’s efforts to highlight student and teacher innovation.

“I deeply respect the work of EdSurge and its commitment to editorial independence,” Harris says. “I deeply appreciate the care EdSurge has for teachers and educators and how it comes through the compelling stories they report on and uplift.”

Nneka McGee is an educator and researcher who taught middle school mathematics and served as a district administrator. As part of her doctoral studies, McGee completed research on the impact of artificial intelligence in K-12 learning environments. 

Most recently, McGee served as project lead and principal co-author of “Responsible Assessment in the AI Era” and  “AI + Learning Differences: Designing for a Future with No Boundaries,” published by Stanford Accelerator for Learning. Her latest work, “AI for Young Learners: Balancing Tech Use with Authentic, Off-Screen Experiences,” was released as a jump start guide by the International Society for Transforming Education in early 2026.

“I understand the importance of offering informative content that helps create a better understanding of an ever-changing landscape. The EdSurge Advisory Board provides an opportunity to be part of the conversation,” McGee says.

Ramona Schindelheim is an award-winning journalist and storyteller covering the future of work and pathways to economic mobility. She is a Top Voice on LinkedIn and her monthly newsletter, The Future of Work(ers), features reporting and interviews on the programs, initiatives and people helping American workers prepare for and find quality jobs and economic mobility. Ramona has worked in national and local news, including at ABC, CBS, NBC, CNBC and The Wall Street Journal.  

Schindelheim brings unique perspective and expertise on work readiness and the need for educators to meet the moment. She decided to join the board “to help shape the narrative around the future of work and the role education must play in preparing learners and workers for the changing economy.”

Avery Thrush is an education policy analyst and writer, currently working as a LEE Fellow at the Massachusetts Department of Elementary and Secondary Education. After spending her early career as a sixth-grade science teacher, Thrush has led community organizing efforts and contributed to state-level policy initiatives related to artificial intelligence in education, school funding and strengthening career-connected education. She was selected as a 2022-23 EdSurge Voices of Change Fellow.

Thrush brings versatile expertise in the education field, from rural and Appalachian education to public schools as engines of democracy. She joins the board with the hope of elevating the voices and expertise of students and teachers, since they are the closest to the challenges facing education today.

Manoush Zomorodi is an award-winning journalist, author, and host of NPR’s “TED Radio Hour,” whose work explores how technology is changing the way we think, work, learn and live. Her latest book, “Body Electric: The Hidden Health Costs of the Digital Age and New Science to Reclaim Your Well-Being,” grew out of a groundbreaking NPR–Columbia University public health project involving more than 20,000 participants. 

She is also the author of “Bored and Brilliant: How Spacing Out Can Unlock Your Most Productive and Creative Self,” and her TED talks on the attention economy and the physical effects of technology on our bodies have been viewed more than 7 million times. Zomorodi hosts NPR’s “Body Electric” podcast, a six-part series on the physical toll of digital life and how to reclaim our bodies. A 2025 Rockefeller Foundation Bellagio Center resident, Zomorodi has won two Gracie Awards and a Webby Award and was named one of Fast Company’s “100 Most Creative People in Business.”

Zomorodi brings her expertise in the effect of media and tech on behavior and health to EdSurge’s edtech coverage. 

“I respect the work of the group and its readers immensely and want to support them [as they] figure out this next chapter for education and tech,” Zomorodi says.

Systems

Nv’s new chips need new datacenters, but you can’t have bit barns without power

Nvidia has made its fortune on the AI boom, but the company has a problem. It faces a growing number of roadblocks which, if left unchecked, could pop the bubble. 

Thus, the AI arms dealer has been forced to play a game of Whac-a-Mole, shoring up the foundations of its empire by investing its spoils into ensuring its customers don’t stop buying.

The latest example of this came Friday, when Nvidia announced it’d become a minority investor in Cloverleaf. 

Founded in 2024, Cloverleaf specializes in laying the literal foundations – by which we mean the land and power – necessary to build bit barns so they can be leased by the cloud providers and model devs that actually pay the bills for Nvidia’s hardware.

Nvidia can only sell as many GPUs as there are datacenters to put them, and those datacenters can only be deployed where there’s adequate power to support them. And since its latest round of GPUs can’t just be dropped into any old bit barn, Nvidia is going to need a lot more datacenter capacity, and by extension a lot more power.

A new breed of datacenter

Up until 2024, the majority of Nvidia’s systems were air-cooled. Supercomputers, like those built by Cray and Eviden, were the exception to the rule. Unless you were deploying tens of thousands of GPUs, its accelerators could be deployed just about anywhere with minimal modification to the facility.

That began to change with Nvidia’s Blackwell generation. You could still get air-cooled systems from Nvidia and its partners, and it certainly sold a lot of eight-way HGX B200 boxes, but deploying its most powerful NVL72 systems required liquid cooling that not every facility was capable of supplying.

Liquid-cooled facilities are harder to build than their air-cooled counterparts. Liquid-cooled racks tend to be more compute dense, and therefore heavier and power hungry to operate. That means in addition to larger UPS backups and beefier PDUs, these facilities now need a bunch of coolant distribution units (CDUs) to pump coolant around — about one for every one to three megawatts of compute capacity. Datacenters must be designed to support all this extra weight, and the problem is only going to get thornier as Nvidia transitions from roughly 250kW to 600kW racks starting next year.

With Nvidia’s Rubin generation of GPUs, air-cooled GPUs may as well be a thing of the past, at least in the HGX and NVL-style form factors that power the majority of AI training and inference deployments today. For the first time, Nvidia is offering the chips only in liquid-cooled varieties.

So, to sell its next-generation GPUs, Nvidia not only needs its customers to have adequate capacity, but also needs those facilities to offer the right mix of power and cooling.

Right sizing for deployment

It’s well established that Nvidia prioritizes customers with datacenters ready to fill. However, many of the GPU-slinger’s biggest customers don’t actually own datacenters, but instead lease capacity from others. For example, Crusoe built OpenAI’s Stargate facility in Abilene, Texas, and Oracle runs it. Altman and crew are simply supplying the demand.

So, Nvidia has invested in model devs and neocloud partners by helping them get financing or even investing directly in their businesses. Nvidia’s investments in OpenAI, CoreWeave, and Nebius are just a few examples.

That works so long as the bit barns available for rent are the right size and shape for the hardware. But, as we noted earlier, many existing facilities aren’t.

To further reduce friction and make it easier for its customers to find space to deploy its kit, Nvidia introduced its DSX platform last fall and officially launched it this spring.

In a nutshell, DSX is a set of blueprints outlining exactly how much power, cooling infrastructure, and space are needed to support a given amount of compute. It also provides tools to optimize that compute to minimize wasted capacity. 

The power problem

But even designing to a spec won’t do bit barn builders any good if they can’t get their hands on enough power.

The proliferation of AI datacenters has begun to strain grid capacity, forcing many new ones to employ behind-the-meter power generation. For example, Nebius has begun using Bloom Energy’s natural gas fuel cells to power its datacenters, rather than relying entirely on existing grid infrastructure.

Nvidia’s investment in Cloverleaf is the latest example of how the GPU slinger is using the circular AI economy to its advantage. The terms of the deal haven’t been disclosed, but the benefits are obvious: Cloverleaf gets an influx of cash, while Nvidia gets greater control over the build process.

As part of the deal, Cloverleaf is also embracing Nvidia’s DSX spec. In theory, this should enable bit barn builders to drop a DSX-compliant datacenter on a Cloverleaf-supplied site and be serving customers without delay.

For Cloverleaf, Nvidia’s spec includes a wealth of power-related capabilities, including DSX Flex, which allows datacenters to scale back their workloads in response to grid conditions. So if grid demand spikes during a heat wave because everyone’s AC just kicked on, datacenters know it’s probably not a good idea to kick off another training run. The tech can also be used to offload demand onto onsite power generation or storage systems, like batteries — all things a company tasked with supplying powered sites for datacenters might be interested in.

The deal is the latest example of how Nvidia is using its war chest to shore up cracks in the broader AI ecosystem before they become a problem. We wrote a few weeks ago about how the company was helping AI infrastructure startups secure financing in exchange for a share of the revenues, but it’s now clear Nvidia is attacking potential roadblocks from every angle. ®

At the local level, the robotaxi operator has outspent both Uber and Zoox in lobbying legislators in the District of Columbia. It has pressed local regulators to allow robotaxi services in the nation’s capital, including by inviting residents earlier this year to write in support of its rollout.

In New York state, Waymo has spent more than half a million dollars so far this year—more than double Uber’s lobbying spend—as it seeks a foothold in one of the largest taxi markets.

Its lobbying in the state intensified after Governor Kathy Hochul earlier this year walked back proposals that would have allowed autonomous vehicle companies to run services across most of the state.

Waymo offered to set up a $20 million fund to support drivers affected by the technology’s deployment, according to people familiar with the matter.

In New Jersey, it has engaged lawmakers who are exploring a three-year pilot program to test deployments.

Waymo said in a statement that it was advocating for the rollout of autonomous vehicles and was not pursuing measures that would limit competitors or “prescribe” how the technology was deployed.

Uber, meanwhile, has been racing to make up lost ground in AVs after abandoning its in-house self-driving effort in 2020, committing more than $10 billion over the past year through equity stakes and robotaxi fleet agreements.

The ride-hailing group has warned several Democratic-led states against a carte blanche rollout and has instead lobbied for so-called hybrid networks that route rides to both human drivers and autonomous vehicles.

In New Jersey, Uber lobbyists have proposed that any platform offering robotaxi services also use human drivers for at least 85 percent of all rides during a three-year pilot program.

Critics have accused Uber of seeking regulatory capture by attempting to keep rides flowing through its platform as it grapples with rising competition.

Uber has said it is false that they were “anti-AV or seeking to slow AV deployment.” It said hybrid networks “get the technology to consumers sooner while giving policymakers a practical framework to manage the transition.”

© 2025 The Financial Times Ltd. All rights reserved. Not to be redistributed, copied, or modified in any way.

More than 9,300 Amazon Web Services (AWS) access keys publicly exposed between August 2022 and August 2026 are still active and valid.

Truffle Security has been tracking this exposure for the past four years and says that 817 of the exposed keys were linked to companies, 526 of them being AWS root keys.

According to the researchers, 242 of the keys are associated with Identity and Access Management (IAM) users with the AdministratorAccess policy. This role has full permissions to create, modify, delete, and view virtually all AWS services and resources within an account.

They note that each key of the 768 live keys in the two sets “full control of a company’s AWS account.”

The company found 431,875 AWS secrets across code repositories, Git history, datasets, Docker images, registries, and CI logs and extracted 64,024 unique AWS keys that corresponded to 50,654 AWS accounts after removing duplicates.

However, the subset for which the researchers had complete credentials that could be used for re-verification was 10,616 keys, and 88% of them continued to authenticate as of August 10.

Amazon Web Services (AWS) is Amazon’s cloud-computing platform used by companies to host websites and applications, store data, run databases and servers, manage domains, and operate their online infrastructure.

Full control of a company’s AWS account could allow an attacker to access, exfiltrate, or wipe cloud-hosted data, take control of servers and applications, and create rogue admin accounts for persistent access

Threat actors could also use their access to deploy cryptominers, generating substantial charges for the company. Truffle Security says that only 262 of 2,754 readable accounts had a budget alert set up.

Hugging Face, a popular online platform where developers share AI models, datasets, and applications, was the largest single source of leaked AWS keys, accounting for 8,482 unique key exposures.

Also, 17.9% of those keys were root, meaning the highest-privileged identity, which isn’t restricted by IAM permissions.

Truffle Security found that, for the 2,903 keys with available creation dates, the median age was 1,831 days (about five years), while the oldest had existed for 17.4 years.

Only 398 (13.7%) of those entries had a newer access key associated with the same user, suggesting most had never been rotated.

To defend against potential abuse, the researchers recommend deleting all root access keys, reviewing IAM credentials by age, rotating or revoking exposed keys, and configuring budget alerts.

Also, any credential committed to a public source should be treated as compromised.

Truffle Security said its testing was limited to read-only metadata, and that it has notified all identifiable owners of the exposed credentials.

Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.

The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.

Get the report

When an agentic AI system hands a task from a small model to a larger one — or back down again — it pays a steep tax: the receiving model has to recompute the entire conversation from scratch, driving up compute costs and latency. This is a major bottleneck for enterprises building long-horizon, multi-LLM workflows.

To solve this challenge, researchers at Nvidia have introduced a cross-model KV cache transfer technique that directly maps the prefilled KV cache from a source model into the target model. This technique aligns with real-world agentic applications where large contexts accumulate across many turns. 

For real-world AI applications, cross-model KV cache transfer can reduce compute costs and latency on long-running, multi-LLM workflows — and it does so with simple linear math, not an expensive deep learning model.

Experiments show that, on compatible model pairs, this linear mapping process runs 2.7 to 25 times faster than recomputing the conversation while retaining up to 98% of the target model’s standalone accuracy. 

Why swapping models mid-session is so expensive

Examining how LLMs handle memory helps understand why multi-model workflows hit a performance wall in production. When an LLM receives a prompt, it must first execute the “prefill” stage, which is the initial forward pass that computes the keys and values for all input tokens and populates the Key-Value (KV) cache. 

After that, it enters the “decode” phase, where it computes and generates the next tokens in the sequence. During this phase, the model reads from this KV cache to predict new tokens one by one, bypassing the need to re-evaluate the entire history of the conversation for each new token.

In multi-turn conversations or long-horizon agentic sessions, the context gradually becomes longer. Because the computational cost of the prefill stage scales directly with both model size and input length, processing these long sessions becomes increasingly expensive and introduces significant latency if the KV cache is invalidated.

This invalidation happens whenever the AI system tries to swap models mid-session, such as routing a complex reasoning step to a larger model or dropping to a smaller model to save costs. Because different LLMs have different architectures, they expect their cache inputs in different formats. 

As a result, any model switch forces the receiving model to repay the entire prefill cost from scratch to recompute the KV cache for the accumulated context. 

Mapping memory between models without starting over

The Nvidia researchers studied cross-model KV cache transfer to see how developers can transform the KV cache of one model into the expected format of another without running the prefill phase again. 

If solved, cross-model KV cache transfer has benefits in both directions. Small-to-large model transfer upgrades the quality of the output. For example, a cheap, small model handles the routine parts of an agentic workflow but struggles with a complex reasoning problem, and you map the KV cache to a larger model and continue the process seamlessly.

On the other hand, large-to-small model transfer reduces compute costs. A highly capable, large model might be used to unpack a massive, complex system prompt or synthesize a dense PDF at the start of a session. Once the heavy lifting is done, the session’s KV cache is mapped down to a smaller, more economical model to handle the rapid-fire, conversational turns that follow.

There have been previous efforts to solve the KV cache transfer problem, but they suffer from a few key limitations. These include the need for expensive gradient-based training or very strict architectural constraints.

For this initial study, the authors restricted their focus to within-family transfers, such as transitioning between different-sized models in the Qwen, Llama, or Ministral families. These models share tokenizers, training data DNA, and core architectural styles but differ in size and depth. However, this framework leaves plenty of room for future experiments. The researchers note the technique could eventually be expanded to cross-family transfers, mismatched KV head counts, or hybrid architectures that blend standard attention with other memory mechanisms.

The key finding of the Nvidia study is that cross-model KV cache is a significantly linear structure. This means you can do the mapping with simple algebra tricks and without the need for heavy neural network training. For example, when experimenting on KV cache transfer from a 14-billion parameter Qwen3 model to a 32-billion parameter version, the authors discovered that a simple linear regression mapping from one source layer to a target layer can recover 56% of the variance in the target’s keys and 32% of the variance in its values. When combining multiple source layers, those numbers climbed to 79% and 65% respectively.

To translate this linear relationship into a practical system, the researchers designed a closed-form per-head ridge mapper with three key components:

• Per-head ridge regression: Instead of using complex deep learning to train the system, they fit a simple linear regression using a tiny calibration set of a few hundred text sequences. This technique solves a classic line-of-best-fit problem independently for every attention head.

• Cross-layer source selection: Because the source and target models have different numbers of layers, the mapper evaluates and selects the most predictive source layers to feed into each specific target layer. This way, the system picks only the most helpful pieces of memory from the old model to construct the new model’s memory.

• Content-space mapping: Before translating the data, the mapper strips away the RoPE encodings. RoPE, or Rotary Position Embedding, is a standard mechanism that applies a mathematical, position-dependent rotation to the data so the model understands the order of the tokens in a sequence. Stripping the RoPE values makes it possible for the mapper to generalize to sequences of lengths larger than its training data.

Putting the linear mapper to the test

To test whether the technique works, the researchers evaluated the transfer pipeline across six “matched-KV” model families. Matched-KV means the source and target models share the same KV head count and per-head dimensions, which is typical for different-sized models within the same family.

The model families included Qwen3, Llama 3.1, and Ministral 3, with tests for KV cache transfer across different sizes ranging from 3 billion to 70 billion parameters. Their experiments included a massive 8.8x parameter leap from Llama 3.1 8B to 70B.

To cover a wide range of tasks, they evaluated the models on five core accuracy benchmarks (ARC-Challenge, HellaSwag, WinoGrande, MMLU, and GSM8K) as well as language modeling perplexity on WikiText-2 and a multi-turn conversation task called CoQA. To fit the linear translation mapper, they used a tiny calibration dataset of just 500 text sequences of 1,024 tokens each.

The researchers compared the framework against the baseline ceiling accuracy where the target model does a full, traditional prefill. They also compared their full system against ablated configurations, such as reducing the number of selected layers or deactivating different components. Additionally, they compared their simple method against a deep neural network trained with backpropagation to see if heavier deep learning could recover accuracy on pairs where the linear method struggled.

For four of the six tested pairs, the fast, closed-form linear ridge mapper retained 73% to 98% of the target’s standalone prefill accuracy — including the massive leap from Llama 3.1 8B to 70B, which retained 72.8% of target accuracy.

The mapper also runs between 2.7 and 25 times faster than re-prefilling. For example, when translating a 32,768-token KV cache from a Qwen3 14B to a 32B model, the transfer took just 278 milliseconds, compared to nearly 7 seconds for a standard re-prefill.

The system also demonstrated high stability on tasks that run across many steps. When tested on multi-turn conversations, the drift, or accuracy loss, between the target baseline and the transferred cache remained incredibly small across 10 turns, proving it will not cascade into failure during long agentic sessions.

However, the straightforward linear approach did run into limitations on specific model pairs. For two of the Ministral configurations, the linear mapper degraded sharply because the simple linear fit failed to extrapolate outside calibration data. To fix this, the researchers swapped the linear mapper for a nonlinear multi-layer perceptron (MLP) with two 1,024-unit hidden layers trained on the same data. This added a complexity and training tax to the setup, but it recovered their accuracy to above 90%.

A bigger industry problem than one paper can solve

The introduction of cross-model transfer is part of a broader, industry-wide push to solve the KV cache bottleneck, which has emerged as one of the key hurdles for scaling enterprise AI. As developers push LLMs to process massive documents or code bases and execute long-running reasoning tasks, managing this memory layer is becoming as important as the models themselves.

Over the past year, researchers have attacked this compute and memory problem from multiple angles. For instance, Nvidia recently introduced dynamic memory sparsification (DMS), a technique that intelligently evicts less important tokens from the KV cache to cut reasoning costs by up to 8x. 

Other approaches focus on aggressive data compression. MIT researchers developed an algebraic compaction technique called Attention Matching that compresses the KV cache by 50x without degrading quality. Similarly, Nvidia introduced KV Cache Transform Coding (KVTC), which borrows media compression concepts to shrink memory by 20x without altering the underlying model weights.

Beyond compression, researchers are also attacking the computational overhead of memory retrieval. Optimizers like IndexCache strip away redundant layer calculations to deliver significantly faster time-to-first-token in long-context applications. And models like DeepSeek and the GLM series are optimizing the KV cache through architecture innovations.

As AI systems take on longer-horizon tasks and more complex architectures, the underlying memory infrastructure is becoming as important as the models themselves. Cross-model KV cache transfer gives developers one more tool for keeping inference costs down as they scale multi-model agentic systems.

£44.99 is Prime Day pricing, and Ring’s newest doorbell is down to that price right now.

The Ring Battery Doorbell (2nd Gen) has dropped to £44.99, down 44% from its £79.99 RRP, a £35 saving that puts it firmly inside Prime Day territory months before that event actually rolls around again.