• Palantir moves AI computing from distant clouds directly into combat zones
• Containerized data centers process battlefield intelligence without permanent cloud connectivity
• Nvidia B300 accelerators power deployable AI systems for frontline operations

Palantir Technologies and its infrastructure partner Armada have begun deploying mobile data centers built inside standard shipping containers for combat zones.

Each unit carries its own computing hardware, storage arrays, networking gear, and cooling systems needed to run advanced artificial intelligence models.

The servers run on Nvidia's B300 AI accelerators, while Palantir's modular AIP platform links the models directly into operational workflows and other company tools.

Processing power without a constant cloud link

Chad Wahlquist, a Forward Deployed Architect at Palantir Technologies, appeared inside one containerised data center during the July 2026 announcement introducing the mobile computing platform.

Unlike conventional cloud-based systems, these containerized units can process information without maintaining any constant connection to public cloud infrastructure.

They can also be fully isolated from external networks whenever operational security requirements demand a stricter level of separation at all times.

The outcome is a self-contained computing node able to analyze intelligence and sensor data right at the point where it is collected.

Distributed computing also offers stronger resilience than depending entirely on one single, centralized data center to run every operation across a wide area.

Losing contact with the cloud, or even losing one mobile node entirely, would not necessarily shut down the whole system.

Other units could keep processing information locally, since the entire network does not depend on one central point of failure.

Bringing AI closer to the battlefield

For military users, the concept brings advanced artificial intelligence much closer to sensors, field commanders, and units already deployed on the ground.

This reduces dependence on communications links that remain vulnerable to jamming, interception, or outright physical destruction during active combat operations.

Instead of sending every piece of battlefield data back to a distant cloud, analysis can occur inside a deployable container placed close to the action.

Alex Karp, the billionaire co-founder and chief executive of Palantir, has praised Ukraine's wartime innovation cycle as instructive for allies.

He has argued that the war is exposing a widening gap between military theory and the reality of modern combat.

Karp credited Ukrainian engineers and soldiers with adapting battlefield technology at a pace unmatched by most peacetime institutions worldwide.

He suggested that this rapid development cycle is teaching the United States and Europe lessons unavailable in traditional laboratories or think tanks.

Whether these containerized AI systems truly perform as promised under sustained combat stress remains an open question.

As of the time of writing, there is no public independent verification of the technology's actual battlefield performance by Palantir, Armada, or any military partner.

Via United24Media

• Tiny desktop system claims massive language model performance at dramatically lower costs
• Proprietary G≡LIX 1 chip delivers 700 TOPS for local AI workloads
• Agent Box runs advanced AI models entirely without relying on cloud services

Singapore-based startup Acrab has unveiled the Agent Box, a mini PC built to run 100 billion parameter language models entirely on local hardware.

The device is powered by a proprietary chip called G≡LIX 1, manufactured using a 5nm process for efficient edge computing.

Acrab claims the system delivers performance close to Nvidia's $5,000 DGX Spark while cutting costs to roughly one-fifth and halving power consumption.

A custom chip built for local AI workloads

The G≡LIX 1 processor combines a 20-core Arm CPU, a GPU rated at 3 TFLOPS, and a multicore neural processing unit tuned specifically for large language model inference.

It draws on 273GB/s of high-bandwidth unified memory alongside an 8MB L1 cache and a 768GB/s L2 cache.

Acrab says the chip reaches 700 TOPS of processing power while keeping the entire system compact enough to sit on a desk.

The NPU is balanced against the general-purpose compute unit, and VAE vector acceleration reportedly speeds up attention processing by up to three times.

Beyond the core processor, the SoC integrates a video DSP cluster, multiple camera input subsystems, and a high-resolution display engine within a single unified architecture.

In internal testing, the company recorded a prefill rate of 1,416.8 tokens per second under a Gemma 26B A4B configuration with a 40K KV cache and a 10K token input.

That compares with 188.9 tokens per second on an Apple Mac Mini M4 Pro, representing a claimed 7.5 times improvement in prefill speed over conventional consumer hardware.

"Generative AI helped people find answers. Agentic AI will help them get things done,” said Dr. Ken Phua, CEO of Acrab.

Software ecosystem expands beyond the hardware

Beyond the hardware specifications, Acrab says Agent Box includes a complete software platform featuring AI runtimes, developer toolchains, operating system capabilities, reference designs, and orchestration software for autonomous agents.

Demonstrations released by the company showed voice commands automatically creating and printing 3D models, operating robotic vacuum cleaners, and interacting with connected devices including smart lights, smart air conditioners, and smart locks.

Acrab says keeping inference on the device can reduce token costs, protect sensitive information locally, maintain operation during limited internet connectivity, and provide persistent memory for personalised AI assistants.

The company also plans to extend its computing platform into AI NAS systems, network-attached storage systems, industrial robots, service robots, and smart vehicles through partnerships with hardware manufacturers.

Although Acrab has disclosed limited corporate information, copyright notices accompanying its announcement reference CATL, the world's largest electric vehicle battery manufacturer.

If that relationship accurately reflects financial backing, the announcement could attract additional attention because of CATL's scale and resources.

“Our goal is to give device makers and developers the foundation to bring agentic intelligence into many different products and environments,” Dr. Phua said.

“Agent Box demonstrates what the technology can do today, while GΞLIX 1 and our full-stack platform are designed to support a much broader ecosystem of devices and applications.”

Via PC Watch

• AI growth keeps enterprise hard drive demand climbing across global cloud infrastructure
• Cloud providers continue expanding storage despite massive investment in AI accelerators
• Enterprise storage remains essential as artificial intelligence generates unprecedented data volumes

Enterprise demand for high-capacity hard drives continues accelerating as artificial intelligence generates expanding datasets requiring reliable, affordable storage throughout increasingly complex cloud environments.

That growing appetite for storage has intensified competition among cloud providers building artificial intelligence infrastructure, keeping enterprise hard drives firmly within long-term investment plans.

Seagate's recent fiscal fourth-quarter and full-year 2026 financial results indicate continued demand from cloud data center customers despite no disclosure of future HDD allocation commitments.

Enterprise customers continue driving demand for mass-capacity HDDs

AI relies heavily on enormous storage capacity because valuable data must remain available long after computations are done.

As organisations deploy larger AI models, the amount of stored information grows alongside computing requirements.

This significantly increases the dependence on enterprise hard drives for economical long-term retention.

Unfortunately, HDDs are becoming harder to secure, with manufacturers already booking long-term supply commitments years in advance.

Many HDD manufacturers have locked in contracts extending into 2028 and 2029, meaning customers wanting drives that far out may already be shut out of available slots

Seagate said robust cloud data center demand kept driving business growth, tying stronger storage purchases to AI infrastructure expanding across global markets.

"As AI accelerates data generation and its value, we see durable long-term demand for mass capacity storage," said Dave Mosley, Seagate's chair and chief executive officer.

Seagate expects strengthening exabyte demand through its Mozaic platform and differentiated HAMR technology roadmap as enterprise deployments continue expanding worldwide.

While Mosley did not explicitly say customers are reserving Seagate's hard drive production through 2029, his remarks make clear that enterprise demand now sits at the core of the company's strategy.

Financial performance supports the broader storage story

Seagate's latest financial results provide additional evidence supporting enterprise storage demand.

Its annual revenue increased during fiscal 2026, while Q4 performance exceeded company expectations for revenue.

Operating cash flow reached $3.7 billion during fiscal 2026, while free cash flow totalled $3.1 billion after fourth-quarter operating cash flow reached $1.3 billion.

Seagate also reduced total debt by $1.4 billion during fiscal 2026, including $302 million during the fourth quarter, ending with outstanding debt of $3.6 billion.

Cash and cash equivalents reached $1.7 billion, while ordinary shares outstanding totalled 227 million and directors approved a quarterly dividend of $0.74 per share.

Seagate expects first quarter fiscal 2027 revenue near $4.1 billion alongside non-GAAP diluted earnings per share of approximately $7.30, subject to stated variations.

"Seagate's strong fourth quarter exceeded our expectations for revenue and non-GAAP EPS, capping a fiscal 2026 in which we grew,” Mosley said.

“Our performance is being driven by robust cloud data center demand and disciplined execution, and we see the momentum continuing in 2027."

Seagate’s rival, Micron, already made the move that Seagate is currently hinting at. It exited its Crucial consumer RAM businesslin 2025 to redirect supply toward enterprise customers.

Micron said the shift was meant to "improve supply and support for our larger, strategic customers in faster-growing segments," cutting consumer buyers off from a manufacturer they had relied on for years.

Seagate has not gone that far with hard drives, but Mosley's insistence that cloud demand is driving the business suggests it could prioritise enterprise demand if current trends continue.

Via TechPowerUp

• Cyberattack hits operational technology at more than 30 Minnesota community water systems, briefly taking treatment plant offline and forcing several towns onto manual operations
• A leaked WaterISAC memo obtained by WIRED relays a state fusion center assessment tying the intrusions to Iranian-affiliated hackers, but no US agency has formally attributed the attack yet
• The CISA had warned four days earlier that Iranian-affiliated actors were compromising internet-facing PLCs, repeating guidance that has mostly fallen on deaf ears

More than 30 community water systems across Minnesota were hit by a coordinated cyberattack, targeting the operational technology running pumps, wells, water towers, and wastewater lift stations rather than the office networks behind them.

Minnesota IT Services disclosed the attack on July 28 2026 and activated a statewide incident response, stating that it wasn't aware of any Minnesota city asking residents to change their drinking water use just yet.

A memo obtained by Wired, circulated by the water sector information-sharing group WaterISAC and carrying a restricted TLP: AMBER handling marking, relays a Minnesota state fusion center assessment that ties the intrusions to Iranian-affiliated hackers.

A warning that was more or less ignored

Four days before the attacks, CISA updated an advisory, AA26-097A, warning that Iranian-affiliated actors were compromising internet-facing programmable logic controllers made by Rockwell Automation, Schneider Electric, Siemens, and possibly others across the US water, energy, and government sectors. It documented confirmed disruptions and financial losses.

This drew no immediate response from state functionaries, and it is easy to see why: the US has tens of thousands of community water systems that serve small populations and lack budgets for dedicated cybersecurity staff.

This is despite there being a precedent: Iranian-linked actors hit the Municipal Water Authority of Aliquippa in Pennsylvania in November 2023, defacing a Unitronics controller with an anti-Israel message. CISA said at the time those devices were exposed to the internet with default passwords still in place. The advice issued then is the same advice being issued this week, which is telling about how little has changed since.

The WaterISAC memo, dated the day the attacks concluded, passes on a state fusion center report linking the activity to Iranian-affiliated actors. That memo was not meant to be public, but it has been corroborated by The New York Times and The Washington Post, which reported that they spoke to federal officials and US intelligence agencies, respectively, lending credence to the claims.

CISA later warned that attackers are changing PLC passwords to lock operators out of their own equipment, and repeated the advice that has been unchanged for years: get PLCs off the public internet and put remote access behind a VPN or gateway.

Who was affected?

The worst of the cyberattack hit Braham, a town of roughly 1,700 people, where the city said attackers shut down the operating controls, taking the water treatment plant and the well offline. Crews restored it manually within a couple of hours, and residents were told to minimize water use in the meantime.

Plymouth, a Minneapolis suburb of about 80,000, found the problem confined to equipment connected over cellular links at two water towers and several wastewater lift stations. Its IT division disconnected the affected kit from the network entirely to stop the attack and prevent retargeting during reconfiguration.

Maple Plain declared a local state of emergency to manage its response, and South St. Paul reported that some automated controls were compromised. In most cases, contingency procedures held and water and wastewater operations continued.

None of this is happening in a vacuum, however. The June memorandum that paused fighting between the US and Iran has broken down, and both are trading strikes, including recent US attacks near the Strait of Hormuz that destroyed a water facility and cut supply to more than 20,000 people. Whoever carried out the Minnesota attacks, the symbolism of targeting municipal water is unlikely to be accidental.

The damage that these intrusions caused was limited; they produced brief outages that trained staff fixed by hand. That is an outcome with a silver lining, and it depended on utilities having people who knew how to run a treatment plant without its automation, even if its digital defenses collapsed outright.

• Robot boat launches drone without crew during challenging ocean demonstrations
• Pioneer vessel survives rough seas while supporting airborne surveillance operations
• Tethered SkyRanger drone provides continuous monitoring from unmanned ships

A robotic boat has launched and recovered its own drone off the deck, without a single crew member aboard either vehicle.

British firm ACUA Ocean and American sensor maker Teledyne FLIR Defense completed the demonstration for the UK Ministry of Defence.

The trial paired an uncrewed surface vessel named Pioneer with a small tethered drone known as the SkyRanger R70 aircraft.

Rough-water stability makes flight operations possible

The entire demonstration depended heavily on the Pioneer's ability to hold steady in genuinely rough conditions each time at sea.

In January 2026, ACUA Ocean said the vessel held steady in waves reaching 4.25 meters, or roughly 13.9 feet high.

The company reported a 60% cut in roll motion compared with traditional single-hull monohull vessels used across the industry.

That stability functions as the mechanical foundation allowing a tethered drone to fly safely off a small robotic vessel at all.

A violently rocking platform would make sustained flight operations for the SkyRanger R70 nearly impossible under open ocean conditions overall.

ACUA Ocean said Pioneer has already completed prior Royal Navy contracts covering surveillance, reconnaissance and anti-submarine warfare tasks before this trial.

Tethered drone extends persistent ocean surveillance

The system pairs a tethered uncrewed aerial system, or T-UAS, with an uncrewed surface vessel known as a robotic boat.

Teledyne FLIR's SkyCarrier is a ruggedized launch platform housed in a container that unfolds into an articulating landing pad system.

During the trial, SkyCarrier successfully launched a SkyRanger R70 drone autonomously from the deck of the moving Pioneer vessel itself.

Once airborne, the SkyRanger R70 stayed connected to Pioneer through a power and data tether, avoiding any need to land and recharge.

"Integrating the SkyRanger R70 with the Pioneer creates something genuinely new: a persistent, deployable capability," said Neil Tinmouth, chief executive of ACUA Ocean.

Richard Cunha of Teledyne FLIR Defense also said pairing the drone with Pioneer opens up an entirely new class of missions.

The SkyRanger R70 runs multiple embedded NVIDIA processors that perform real-time object detection and classification without ground input.

The combined sensor package delivers Group 2 level ISR performance while flying on a much smaller Group 1 airframe type.

UKDI, the Ministry of Defence's early-stage funding arm, backed the demonstration well before it reached any full-scale procurement decision process.

"Securing UKDI funding for this demonstrator is a significant milestone for ACUA Ocean and UK maritime autonomy broadly," Tinmouth said.

This technology will support the Royal Navy priorities including anti-submarine warfare, mine countermeasures and undersea infrastructure protection efforts overall.

That said, a successful demonstration under UKDI funding does not guarantee the system will reach large-scale deployment across naval operations soon either.

Cost, reliability at sea over long periods, and integration with existing fleets remain open questions that testing alone cannot resolve.

Via Defence Blog

• Over two-thirds (69%) of people believe AI generated material can be presented as their own work if they provided the prompt and approved the final version
• Meanwhile, 54% believe AI-based work should be clearly marked
• 7,200 adults were surveyed, located across the US, the European Union, the UK, and Latin America

A growing disparity in how AI generated content, tasks, and tools are assigned ownership and expectations on accurately marking their origins is emerging, according to a new survey by Use.AI.

Furthermore, the findings indicate an emerging dichotomy between human-claimed authorship of an AI work and clear labeling of the work as AI.

Different attitudes based on settings for the AI work have also been identified, such as private emails and school or professional reports.

The chicken or the prompt?

The developing adoption of AI appears to be introducing new attitudes to acceptable creativity, with the emerging model of what is described as “authorship by direction,” where the AI is trusted to create a first draft, which is then revised and claimed by a human.

The statistics shared by Use.AI – a tool that aggregates AI models into a single interface – highlight how AI is being used more extensively, with a human-originating prompt leading to drafts and versions from the machine, followed by editing from the human.

While the AI has contributed, only a human name appears on the finished product, which raises questions of ownership and authorship.

“AI is separating the production of work from the ownership of it. People are increasingly prepared to direct the process, approve the result and take the credit, but there is still no shared standard for when AI’s role should be visible to everyone else,” says Use.AI’s Managing Director, Ihor Herasymov.

With 69% of those surveyed considering AI-generated work their own if they produced the prompt, 46% claiming they would keep AI involvement with their work under wraps, and 54% feeling AI use should be disclosed, we’re some way from agreement on AI use across creative tasks.

Authorship vs. editorship

Herasymov’s observations point to a constantly evolving set of standards around the creation and use of AI-generated materials. Interestingly, this is happening against a backdrop of growing unease over the use of AI in creative tasks.

The implications of claiming ownership of automated work generated upon instructions to an LLM trained on as much human writing, art, music, and cinema as possible are considerable, and go way beyond questions of ethics. Copyright is just one element, and the most obvious, but we’re also moving into a landscape where inputting a prompt on employer-owned hardware takes authorship away from the person at the keyboard.

Ultimately, the task of prompting and revising means anyone using AI in this way stops being the author, and becomes the editor – and that is before the true ownership of the output is determined.

Apple Watch Ultra 3 owners tend to know the headline features by now: satellite messaging, the brighter always-on display, hypertension alerts.

What most people haven't noticed is a much quieter feature that only shows up if you also own a pair of the well-received AirPods Pro 3 — the same earbuds you're probably already using on all those Zoom calls at work every day.

For anyone else trying to squeeze a workout into a packed workday (or if you're trying to travel light on a business trip), I can tell you it's become the one I actually rely on most.

No more packing the chest-strap

AirPods Pro 3 have their own in-ear heart rate sensors, enabled by default, which is easy to miss since Apple doesn't make a big deal about it.

On their own, they're a genuinely solid heart rate source, useful for anyone who wants to track a run or a lift without a watch at all. But when you wear both an Apple Watch and AirPods Pro 3 during the same workout, the two devices don't just duplicate the same job.

They combine into multiple simultaneous streams of heart rate data, and the system automatically leans on whichever source has the highest confidence at any given moment.

That matters more than it sounds like. Wrist-based optical sensors, including the Ultra 3's, can struggle during movements that shift the watch on your wrist: heavy grip work, rowing, anything involving a lot of wrist flexion.

In-ear sensors don't have that problem, but they can be thrown off by their fit or by their own motion. Fusing both sources means the watch is quietly falling back on whichever sensor is actually reading you correctly, workout to workout and rep to rep, instead of you having to guess which device to trust.

Honestly, it's replaced the chest strap by now. I used to strap one on for anything where zone accuracy actually mattered, interval sessions, heavy lifting days, anything where the watch slides around enough on my wrist to throw the optical sensor off.

Is it clinically equivalent to a chest strap? Probably not. But after a few months of comparing the two side by side, it's close enough that I just don't bother anymore.

As a professional, that feels like a win to me. It's one less piece of gear I have to cart around with me on a business trip or to a conference with the rest of my work gear.

Setting it up takes no extra effort. Heart rate monitoring is enabled by default on AirPods Pro 3, so if you already own both devices, this will likely happen in the background the next time you start a workout with both in use.

You can check or turn it off entirely from the Settings app on iPhone, under your AirPods' name, if you'd rather it not run.

It's a small thing, and Apple has barely marketed it. But it's the kind of feature that only reveals itself once you own the specific pair of devices it's built for.

In my experience, it's an ideal setup for busy professionals who still want accurate health data without adding extra setups and gear.

If you've already got an Ultra 3 and a pair of AirPods Pro 3, there's nothing to set up and nothing to buy. You're most likely already using it without realizing it.

• Startup replaces Nvidia GPUs with custom Arm-powered AI processing hardware
• Prometheus packs up to 128TB of unified LPDDR6 memory onboard
• New server promises 1,000× more memory available per processor

Majestic Labs, a startup founded in 2023 by former Google and Meta engineers, has unveiled a server built to rival Nvidia's GPU and HBM combination.

The Tel Aviv-based company argues that pairing costly graphics processors with high-bandwidth memory has become a fundamentally memory-bound and dead-end approach for AI inference.

Its answer is the Prometheus server, which swaps GPUs for Ignite AI Processing Units combining Arm cores with RISC-V vector and tensor engines.

A different way to scale memory

Each Prometheus server can house up to 12 AIUs, sharing between 8 TB and 128 TB of LPDDR6 memory across one contiguous, coherent pool.

That memory pool is accessed through custom memory aggregation chiplets linked by copper cables up to one metre long instead of memory attached directly to GPU packages.

A standard 40U rack can hold four such servers, drawing 120 kW total and cooled through cold-plate liquid systems rather than air.

By comparison, an Nvidia DGX B300 system with eight Blackwell GPUs offers 2.3 TB of HBM3e plus up to 4 TB of DDR5 system memory.

Majestic claims its architecture therefore delivers over 50 times more fast memory than that rival configuration, alongside 1.7 times its interconnect bandwidth.

“One Majestic rack holds the fast memory capacity of 25 Nvidia NVL72 Vera Rubin racks at a fraction of the power,” Majestic Lab said.

“Organizations that could never justify hyperscaler infrastructure can now run any workload. In fact, there can be up to “1000× more memory per processor.”

Performance claims await broader testing

Majestic Labs says the Prometheus server could cost between 10 and 50 times less than a GPU system of equivalent performance once it ships next year, while consuming less electricity per rack.

The server is designed to be OCP-compliant and will support PyTorch, vLLM and OpenAI's Triton frameworks, letting existing AI models run without modification.

Founded by CEO Ofer Shacham, President Sha Rabii and COO Masumi Reynders, the company employs around 40 people across Tel Aviv and Los Angeles and raised $100 million in an A-round late in 2025.

It claims to have already received significant orders from large enterprises, neoclouds and hyperscalers.

Yet several details remain unclear, including how many memory aggregation chiplets a single server actually requires.

If a 128 TB configuration relies on widely available 2 GB LPDDR6 dies, it would need roughly 64,000 of them, implying well over a hundred aggregation chiplets per server.

The numbers Majestic Labs presents are striking, but they remain the startup's own projections ahead of any independent benchmarking or shipped hardware.

Enterprise buyers considering a shift away from established GPU vendors may have to wait for independent validation to verify Majestic's claims.

Via Blocks and Files

• China unveils rugged AS2-W robotic platform with advanced onboard artificial intelligence
• Robot wolf climbs steep slopes while carrying heavy operational payloads
• AS2-W combines wheeled mobility with advanced terrain handling capabilities

Chinese robotics company Unitree has unveiled its new AS2-W, an uncrewed ground vehicle on its official website.

The four-wheeled vehicle weighs 25 kg and can carry a stationary payload of up to 150 kg while sitting still.

It travels roughly 30 km on a single charge without cargo, or about 25 km carrying up to 16 kg.

Sensors and mobility built for rough terrain

According to Unitree, the AS2-W can climb obstacles reaching 80 cm high while traversing slopes of up to 45° without manual intervention.

The robot reaches speeds exceeding 6 m/s using industrial-grade joint modules delivering approximately 95 N·m of peak joint torque.

Unitree also says the platform uses a 648 Wh, 15,000 mAh battery capable of providing more than three hours of unloaded operation.

Its equipment includes industrial-grade 64- to 128-line LiDAR, an HD camera, Wi-Fi 6, Bluetooth 5.2, GPS, 4G connectivity, microphones, speakers, and front lighting.

The AS2-W supports an IP54 dust and water resistance rating and is designed to operate in temperatures ranging from -20°C to 55°C.

Unitree also provides developer software tools, application programming interfaces, and support for artificial intelligence models enabling autonomous interaction and decision-making.

A demonstration video released by the company shows the robot navigating rocky slopes without losing its footing throughout the test.

The same footage shows the machine maintaining full mobility even after falling from elevated platforms during separate testing runs later.

Unitree describes the AS2-W as running a high-performance AI module capable of 150 TOPS, or trillion operations per second, onboard.

The company has not specified whether that processor is actually manufactured by Nvidia, despite widespread reports describing it that way.

Its ISS 3.0 autonomous navigation system is designed to provide centimetre-level positioning accuracy during movement through complex operating environments.

Military interest follows the armed version sighting

The People's Liberation Army has already shown interest in the AS2-W following its appearance during the Chinese-Mongolian Steppe Partner 2026 joint military exercise.

Chinese military reports indicated similar robotic platforms performed reconnaissance and fire support tasks during the exercise alongside conventional military units.

The public unveiling followed those exercises, although Unitree's commercial announcement focused primarily on the platform's technical characteristics rather than operational deployment.

The vehicle's onboard artificial intelligence enables object processing directly on the platform, potentially reducing delays associated with transmitting information to remote operators.

Its combination of autonomous navigation, onboard computing, and off-road mobility reflects broader interest in unmanned ground vehicles across several defence programmes worldwide.

However, the company has not disclosed the full operational limitations, endurance under sustained military use, or performance while carrying heavier payloads.

Unitree's YouTube video of this robot comes with the disclaimer, "Please, everyone be sure to use the robot in a friendly and safe manner."

• Etched reached a $10.3 billion valuation after securing fresh SK Hynix backing
• The startup built custom processors specifically for demanding AI inference workloads
• Etched says conventional GPUs waste power during many inference tasks today

Etched, a US-based AI chip startup, has secured fresh backing from SK Hynix as it hits a $10.3 billion valuation while pursuing processors built specifically for inference workloads.

The company argues that conventional GPUs deliver more computational capability than many inference tasks require, yet offer insufficient memory for increasingly demanding AI models.

Its latest funding will support the production of rack-scale inference systems designed around custom chips, shared memory, and lower power consumption, rather than general-purpose graphics processors.

Etched bets custom inference chips can outperform traditional GPUs

Founded by former Harvard students Gavin Uberti, Robert Wachen, and Chris Zhu, Etched focuses exclusively on AI inference rather than training large language models.

The startup says its systems combine Low Voltage Inference (LVI) technology with Cluster Scale Memory (CSM), allowing processors to access substantially larger shared memory pools than conventional GPUs.

“Today, AI chips can't scale FLOPs without thermal throttling. As FLOPs utilization increases, AI chips draw more power and downregulate clock speed…,” said Chris Zhu

We’ve designed a new architecture to run our chip’s math blocks at under half the voltage of most AI chips. This enables multiple times the FLOPs density of AI chips today.”

The researchers claim that existing processors struggle to increase floating-point performance because higher utilisation raises power consumption and eventually reduces operating clock speeds through thermal limits.

Processors using High Bandwidth Memory (HBM) cannot achieve SRAM-level decoding speeds because memory subsystems and interconnects introduce additional latency during inference workloads.

They therefore created a shared low-latency memory pool connected through what it describes as a proprietary ultra-low-latency, high-bandwidth interconnect to enable faster memory access

“Our HBM/SRAM hybrid design solves both memory capacity and mem2mem latency, enabling high throughput and interactivity simultaneously.

“CSM improves latency and avoids today's cost, reliability, yield, thermal, and compute tradeoffs of SRAM-only chips, 3D DRAM chips, or optics.”

The company also claims many AI models waste time moving data between chips, memory, and networking hardware before processing can continue.

According to Etched, its CSM architecture reduces those delays by minimising additional memory layers during data transfers.

Funding surge accelerates production and global expansion

Etched has attracted funding at a rapid pace, raising $5.4 million in its 2023 seed round, $120 million in 2024, $500 million in 2025, and $300 million in 2026.

Those investments bring total funding to approximately $925.4 million, while the latest financing doubled the company's valuation from $5 billion to $10.3 billion.

The C-round included Sequoia Capital, Andreessen Horowitz, Jane Street, Diffusion, Argo, and SK Hynix, signalling continued investor confidence despite increasing competition within AI hardware markets.

Robert Wachen said, "This round accelerates production of our inference clusters," while confirming an 80,000 sq ft, 10 MW facility opened near Milpitas.

The company now employs more than 400 people, reports customer demand exceeding $1 billion, and has established manufacturing operations in Taiwan.

Etched systems will support conventional large language models, mixture-of-experts architectures, and alternatives including Mamba.

Via Blocksandfiles