2026-05-18 22:00:00
A new study challenges the idea that consciousness is necessary to make sense of language.
Our brains keep on whirling long after we drift off to sleep.
Each night, the hippocampus, a major hub for learning, replays experiences from the previous day and etches them into memory. And even in deep sleep, neurons in sensory regions of the brain spark with activity when they receive new stimuli, like sounds.
This raises a provocative question: How much is consciousness required to make sense of the world around us?
A new study suggests the unconscious brain can handle far more than simple sensory cues. Recording electrical activity from patients under general anesthesia, a team at Baylor College of Medicine and collaborators found the hippocampus continued processing sounds, words, and speech while patients listened to alternating tones and podcast clips.
Groups of neurons shifted their activity depending on the type of word spoken—nouns or verbs, for example—and predicted the next word in sentences.
“Our findings show that the brain is far more active and capable during unconsciousness than previously thought,” study author Sameer Sheth said in a press release. “Even when patients are fully anesthetized, their brains continue to analyze the world around them.”
Scientists have long thought that language processing, a complex computation, relied on awareness. Anesthesia disrupts large-scale communication across the brain, seemingly making complex language processing impossible. But the new findings suggest that even as global brain dynamics break down, some local circuits retain the ability to process sophisticated information—and, at least for storytelling, predict what comes next.
To be clear, it doesn’t mean that participants were secretly awake. Whether the brain retains local processing power during sleep, coma, or other states of unconsciousness is also up for debate.
But “this work pushes us to rethink what it means to be conscious,” said Sheth. “The brain is doing much more behind the scenes than we fully understand.”
We slip into unconsciousness every night. The brain shifts gears.
Compared to when we’re awake and alert, the mind’s activity patterns change dramatically. The hippocampus reactivates neurons involved in recent learning, rapidly replaying their activity patterns to strengthen neural connections. Elsewhere, the brain generates short bursts of electrical activity called sleep spindles, which shut off communication between regions necessary for processing new information from the outside world. These unique electrical signals are crucial for sorting new experiences and integrating them into long-term memory.
The brain is clearly busy during unconsciousness, but it also seems largely sealed off from its surroundings. Over the past two decades, however, scientists have increasingly realized the sleeping brain remains surprisingly alert.
In one study, volunteers repeatedly exposed to unfamiliar sounds during sleep were able to identify them after waking up. In another, participants hearing their own names or angry voices triggered brain activity even in deep sleep, a phenomenon called “sentinel processing.”
Scientists have also recorded directly from the brains of people with epilepsy, who had electrodes implanted to pinpoint the source of seizures. The researchers confirmed that the auditory cortex—the first region involved in processing sound—lit up with activity, but it appeared disconnected with regions responsible for interpreting meaning.
Similar patterns emerged under other states of unconsciousness. After receiving propofol, a common drug used to induce general anesthesia, patients still showed activity in their auditory cortex, but information relay to higher regions involved in cognition seemed to break down.
Or did it?
“The brain has developed such amazing, sophisticated mechanisms for doing all these complex tasks all day long, that it can do some of these things even without us being aware,” Sheth told Nature. They decided to take another look.
The team focused on the hippocampus, best known as the brain’s memory center. Linking it to language processing seems like a stretch. But mounting evidence suggest the hub is responsible for far more than memory. It may also help organize information more broadly, from the mapping of physical spaces to watching other unfolding events like language.
It’s still a niche idea, said Sheth. But the hippocampus could play a much broader role in structuring the world around us—even without awareness. “How is the world organized? The hippocampus may be part of that as well,” he said.
To test the idea, the team recruited seven people undergoing epilepsy surgery. While they were under propofol anesthesia, the team inserted tiny probes into the hippocampus. Called Neuropixels, the implants are thinner than a human hair but packed with over a thousand sensors that eavesdrop on the electrical chatter of hundreds of neurons at once.
The team first played repetitive beeps to three participants, occasionally interrupted by random boops at a different pitch. In the beginning, neurons were indifferent to the oddball sounds. But within 10 minutes, their activity levels showed they were getting better at separating the unexpected tones from the normal ones.
“They learned over time to pay more attention to oddball sounds,” even while the person was fully unconscious, said Sheth.
A second test took things further. The team played 10-minute snippets from The Moth Radio Hour, a storytelling podcast featuring speakers from all walks of life, each with distinct intonations, turns of phrases, and accents.
Across the recordings, specific groups of hippocampal neurons responded to different linguistic features. Some were attuned to uncommon words like “cosmos.” Others tracked grammatical structure, responding differently to nouns, verbs, or adjectives.
The neurons also cared about semantic meaning, or the relationships between words. For example, they seemed to recognize that “cat” is conceptually closer to “dog” than an unrelated word like “pen.” The hippocampus also seemed to anticipate upcoming words based on the context of a sentence, with activity patterns similar to those seen in the awake brain.
“We are always making predictions about what we’re about to hear next,” said Sheth. Even under anesthesia, these neurons appeared to keep track of the narrative, indicating a “very sophisticated form of processing of the natural speech that they’re listening to.”
Despite intense neural activity, patients didn’t remember any of the podcast stories upon waking. Still, traces of the experience may have lingered unconsciously. In future studies, the team plans to test for this by exposing unconscious participants to different podcasts then later asking which ones feel familiar. They also want to explore whether the hippocampus processes stories told in unfamiliar languages.
The findings are preliminary, drawn from a small group of people under one type of anesthetic. The sleeping or comatose brain may work differently. But the work could help scientists decipher brain activity in people with severe traumatic brain injuries in a vegetative state. It could also guide the development of implants to rewire damaged neural circuits to other parts of the brain and reboot communication.
“Maybe the most important thing is what can we do about this,” said Sheth. For someone who’s unconscious, “can we bring them back?”
The post The Fully Anesthetized Brain Can Still Track a Podcast appeared first on SingularityHub.
2026-05-16 22:00:00
Unitree Will Sell You a Massive ‘Transformable Mecha’ for $650,000Jess Weatherbed | The Verge
“Unitree is already one of the most recognizable names in the humanoid robot industry, but now it’s pursuing even nicher sci-fi tech: giant mech suits. The Chinese robotics company has debuted the GD01, which it describes as ‘the world’s first production-ready manned mecha,’ and it can be yours for a paltry $650,000.”
How an ‘Impossible’ Idea Led to a Pancreatic Cancer BreakthroughGina Kolata and Rebecca Robbins | The New York Times ($)
“A drug nearing regulatory approval, daraxonrasib, is the first to substantially extend the lives of patients with pancreatic cancer. It works by targeting a cellular protein that fuels not just nearly all pancreatic tumors, but also many lung and colon cancers. …Now, some scientists predict that the approach could wind up being the most significant advance in cancer treatment in 15 years, since the arrival of immunotherapy.”
Software Developers Say AI Is Rotting Their BrainsEmanuel Maiberg | 404 Media
“Developers talk not just about how the AI output is often flawed, but that using AI to get the job done is often a more time consuming, harder, and more frustrating experience because they have to go through the output and fix its mistakes. More concerning, developers who use AI at work report that they feel like they are de-skilling themselves and losing their ability to do their jobs as well as they used to.”
A Plan to Make Drugs in Orbit Is Going CommercialAntonio Regalado | MIT Technology Review ($)
“Varda Space Industries, a startup that’s been pitching its ability to perform drug experiments in space, says it has signed up the pharmaceutical company United Therapeutics in what may be remembered as a notable step toward in-orbit manufacturing.”
Rebooting Stem Cells Builds Aged Muscles and Assists Injury RecoveryAlice Klein | New Scientist ($)
“Old mice grow bigger muscles and recover from injuries better when stem cells are taken out of their aged muscles, given a reboot, then put back in. A similar approach may allow rejuvenation of aging muscles in people too. ‘In theory, if you took an elderly person’s muscle stem cells out, charged them up and put them back in, they would probably be more functional,’ says James White at Duke University in North Carolina.”
Google Stopped a Zero-Day Hack That It Says Was Developed With AIStevie Bonifield | The Verge
“It’s the first time Google has found evidence that AI was involved in an attack like this, although Google’s researchers note that they ‘do not believe Gemini was used.’ Google says it was able to ‘disrupt’ this particular exploit, but also says hackers are increasingly using AI to find and take advantage of security vulnerabilities.”
Can Some Very Tiny Particles Cool the Planet? One Tech Company Says Yes.Eric Niiler | The New York Times ($)
“Stardust executives said that initial effort to begin atmospheric cooling would cost about $10 billion. …By adding 10 million tons of the reflective particles to the atmosphere over the course of several years, the atmosphere could be cooled by 1.5 degrees Celsius, the company said.”
Anthropic Blames Dystopian Sci-Fi for Training AI Models to Act ‘Evil’Kyle Orland | Ars Technica
“Those with an interest in the concept of AI alignment (i.e., getting AIs to stick to human-authored ethical rules) may remember when Anthropic claimed its Opus 4 model resorted to blackmail to stay online in a theoretical testing scenario last year. Now, Anthropic says it thinks this ‘misalignment’ was primarily the result of training on ‘internet text that portrays AI as evil and interested in self-preservation.'”
Forget Smart Glasses, These Earbuds Can See, Hear, and Remember Everything for YouShimul Sood | Digital Trends
“Smart glasses have always felt a little awkward to me. Sure, they can play music, take calls, snap photos, and even throw notifications in front of your eyes, but at the end of the day, they’re still just tiny screens sitting on your face. Now imagine removing the screen entirely. That’s exactly what this new pair of AI-powered earbuds is trying to do. …And honestly, this might be one of the more interesting directions wearable AI has taken so far.”
A Single Infusion Could Suppress HIV for Years, Study SuggestsApoorva Mandavilli | The New York Times ($)
“For about a decade, scientists have had remarkable success curing some blood cancers by modifying a patient’s own immune cells to recognize and kill the malignant cells. That same approach may help control HIV, among the wiliest of viruses, scientists will report on Tuesday. After a single infusion of immune cells engineered to recognize the virus, two people in a new study have suppressed their HIV to undetectable levels, one of them for nearly two years.”
The Tesla Semi Could Be a Big Deal for Electric TruckingCasey Crownhart | MIT Technology Review ($)
“Globally, trucks and buses represent about 8% of total vehicles on the road, but they create 35% of carbon dioxide emissions from road transport. Tesla’s latest addition to its vehicle lineup, the Class 8 Semi, could be part of the solution to cleaning up this polluting sector.”
World’s First Native Color Lidar Gives Machines Human-Like VisionOmar Kardoudi | New Atlas
“LiDAR sensors—the laser-based eyes of self-driving cars, industrial robots, and inspection drones—build precise 3D maps of their surroundings, but everything is built of monochrome geometric shapes. Ouster’s new Rev8 sensor family aims to change that, not by bolting a camera onto a LiDAR unit, but by fusing color directly into every point of data the sensor captures.”
The Creative Risk of Letting AI Do All the WorkNatalie Nixon | Fast Company
“[MIT’s Sinan Aral] calls this ‘diversity collapse,’ the slow homogenization of output that occurs when AI, trained on the same publicly available internet, starts flattening the edges that make creative work distinctive. The more a team delegated to AI, the more productive they became—and the more vulnerable they were to this collapse.”
The post This Week’s Awesome Tech Stories From Around the Web (Through May 16) appeared first on SingularityHub.
2026-05-15 22:00:00
Self-assembling swarms of microrobots could someday deliver drugs and pull toxins from water.
For most of us, a locust swarm sounds like an utter nightmare. For roboticists, it’s inspiration.
Nature abounds with creatures that cooperate with a “hive mind.” From bees gathering pollen to schools of sardines grouping to avoid predators, individuals seamlessly move together in ever-changing configurations. Roboticists inspired by these dynamics have designed microrobots—often no more than the width of a human hair—to mimic their behavior.
These tiny machines show promise in medicine and environmental cleanup. They easily sail through blood vessels to remove blood clots, deliver chemotherapy to tumors, and bring medicines to the eye and gut. In the wild, they remove plastics and heavy metals from water.
Researchers usually steer microbots with sound, magnets, or light. But few systems are able to assemble into swarms and disassemble on command. A University of San Diego team has now engineered a part-biological microbot swarm controlled by shifting colors of light. The swarm is made of living algae and nanoparticles and can coalesce into various shapes on demand.
In one test, the researchers shaped the living robots to match damaged tissue in a simulated wound. They then assembled the robots on a smart “Band-Aid” and released them into the wound, concentrating treatment exactly where it was needed.
Microbots that deliver drugs, perform surgery, or act as environmental sentinels are no longer science fiction. Swarms of these robots have especially captured the imagination of roboticists. Tweaking a swarm’s shape and size can allow it to tunnel into small spaces and do work that would thwart any single sophisticated robot.
Early versions use a variety of synthetic materials to mimic natural swarms. Some made of tiny iron-based particles shapeshift from chains to vortexes and ribbons after scientists strategically apply magnetic forces. Certain configurations offer strength and stability; others are more steerable, like robotic sentinels from the Matrix movies. Another class of nanomachines respond to light or sound waves for navigation.
Synthetic microbots can mimic swarm behavior, but they’re limited by a material’s physics. So researchers are turning to nature too, building biohybrid bots powered by living cells.
Swimming bacteria are a popular choice. Tethered to nanoparticles carrying drugs, these robots can navigate liquid environments to kill pathogens, trap microplastics, or deliver antibiotics. But their relatively large size makes it hard to access tight or delicate spaces.
Algae could be an alternative. These single-celled organisms swim using long, whip-like arms called flagella that act as microscopic propellers. Roughly 10 micrometers across—about the size of an average skin cell—they’re small enough to thread their way through tiny spaces.
Researchers can coat nanoparticles with drugs or chemical sensors and attach them to the algae. These bots have already been used to deliver antibiotics for bacterial pneumonia in mice. Other designs have been tested as a treatment for inflammatory bowel disease, a chronic disorder that affects millions worldwide. Here, scientists engineered nanoparticles to absorb and neutralize inflammatory chemicals in the gut. Packed into a pill, the algae-powered bots dispersed throughout the treatment area while largely avoiding other organs.
But the microbots are still hard to control. Researchers don’t understand their collective behavior and how they form assemblies, wrote the authors of the new study.
The team picked Chlamydomonas reinhardtii for their robots. Commonly found in freshwater puddles and soil, these single-celled algae are a staple of lab research. They have two powerful arms and are sensitive to various colors of light, making them easy to control.
In a test, the team projected blue or red light onto petri dishes crowded with the algae. They shaped the swarms with masks—basically, stencils—patterned to look like different continents. Blue light caused the algae to cluster in swarms matching the mask . Red light dispersed them. The team shaped the living swarm to resemble the Americas and Afro-Eurasia within minutes.
Using a mask shaped like an arrow, the team moved the swarm several millimeters while maintaining its shape. Other masks transformed the swarm into stars, letters, and triangles. By further tuning the duration and intensity of red and blue light, the researchers coaxed the swarm to double its size while maintaining a circular shape or split into four smaller parts. They used the results to write an algorithm predicting how light alters swarm activity.
The team next attached the algae to nanoparticles to see if they could target a simulated wound on a dummy hand coated in lifelike “phantom skin.” A thin coat of artificial wound fluid, made up of proteins and chemicals usually found after a scrape, made the test more realistic.
They used an AI system to analyze images of the wound, segmenting regions into healthy, inflamed, or potentially infected tissue, and then laser-printed a custom mask matching the infected area. Under blue light, the microbots assembled on a piece of medical tape in the exact geometry of the wound. After applying the custom Band-Aid, a burst of red light released over 90 percent of the bots to the target area in less than two minutes.
The work is still early though. In future studies, researchers will have to load nanoparticles with medication and test how the swarms behave in real wounds and living tissue. And because the system relies on light for control, it’s currently limited to surface-level applications.
That said, because they can now more reliably control the swarms’ shape, size, and position, the technology could prove quite useful in medical applications, wrote the team.
The post New Algae Robots Swarm Like Locusts at the Flick of a Switch appeared first on SingularityHub.
2026-05-15 04:17:56
Photons traveling straight through a cloud of gas appear to exit, on average, before they enter.
As Homer tells us, Odysseus made an epic journey, against the odds, from Troy to his home in Ithaca. He visited many lands, but mostly dwelt with the nymph Calypso on her island.
We can imagine that his wife, Penelope, would have asked him about that particular time. Odysseus might have replied, “It was nothing. In fact, it was less than nothing. Negative five years I dwelt with Calypso. How else could I have arrived home after only ten years? If you don’t believe me, ask her.”
Quantum particles, it turns out, are just as wily as Odysseus, as my colleagues and I have shown in an experiment published in Physical Review Letters. Not only can their arrival time suggest that they dwelt with other particles for a negative amount of time, but if one asks those other particles, they will corroborate the story.
Our experiment used photons—quantum particles of light—and the against-the-odds journey they must undertake to pass straight through a cloud of rubidium atoms.
These atoms have a “resonance” with the photons, meaning the energy of the photon can be transferred temporarily to the atoms as an atomic excitation. This allows the photon to “dwell” in the atomic cloud for a time before being released.
For this resonance to be effective, the photon must have a well-defined energy, matching the amount of energy required to put a rubidium atom into an excited state.
But, by a form of Heisenberg’s famous uncertainty principle, if the energy of the photon is well defined then its timing must be uncertain: The pulse of light the photon occupies must have a long duration. This means we can’t know exactly when the photon enters the cloud, but we can know on average when it enters.
If a photon like this is fired into the cloud, the most likely outcome is that its energy will be transferred to the atoms and then re-emitted as a photon traveling in a random direction. In such cases, the photon is scattered and fails to arrive at its Ithaca.
But if the photon does make it straight through, a strange thing happens. Based on the average time when the photon enters the cloud, one can calculate the expected average time it would arrive at the far side of the cloud, assuming it travels at the speed of light (as photons usually do).
What one finds is that the photon actually arrives far earlier than that. In fact, it arrives so early it appears to have spent a negative amount of time inside the cloud—to exit, on average, before it enters.
This effect has been known for decades and was observed in a 1993 experiment. But physicists had mostly decided not to take this negative time seriously.
That’s because it can be explained by saying that only the very front of the long-duration pulse makes it straight through the atomic cloud, while the rest is scattered. This leads to a successful (non-scattered) photon arriving earlier than would be naively expected.
However, Aephraim Steinberg, one of the authors of that 1993 paper, was not so quick to accept this dismissal of the negative time as an artifact. In his laboratory at the University of Toronto, he wanted to find out what happened if one queried the rubidium atoms in the cloud to find out how long the photon had spent dwelling among them as an excitation. After an initial experiment with inconclusive results, he asked me, as a quantum theorist, for help in working out what to expect.
When we talk of querying the atoms, what this means in practice is continuously making a measurement on the atoms while the photon is passing through the cloud to probe whether the photon’s energy is currently dwelling there. But there is a subtlety here: Measurements in quantum physics inevitably disturb the system being measured.
If we were to make a precise measurement of whether the photon is dwelling in the atoms, at each instant of time, we would prevent the atoms from interacting with the photon. It is as if, merely by watching Calypso closely, we would stop her getting her hands on Odysseus (or vice versa). This is the well-known quantum Zeno effect, which would destroy the very phenomenon we want to study.
The solution is to make, instead, a very imprecise (but still very accurately calibrated) measurement. That is the price paid to keep the disturbance negligible. Specifically, we fired a weak laser beam—unrelated to the single photon pulse—through the cloud of atoms, and measured small changes in the phase of the beam’s light to probe whether the atoms were excited.
Any single run of the experiment gives only a very rough indication of whether the photon dwelt in the atoms, but averaging millions of runs yields an accurate dwell time.
Amazingly, the result of this weak measurement of dwell time, when the photon goes straight through the cloud, exactly equals the negative time suggested by the photons’ average arrival time. Prior to our work, no-one suspected that these two times, measured in entirely different ways, would be equal.
Crucially, the negative value of the weakly measured dwell time cannot be explained by imagining that only the front of the photon’s pulse gets through, unlike the time inferred from the arrival time.
So what does this all mean? Is a time machine just around the corner?
Sadly, no. Our experiment is fully explained by standard physics.
But it does show that negative dwell time is not an artifact. However paradoxical it may seem, it has a directly measurable effect on the atomic cloud that the photon traverses. And it reminds us that there are still lands to discover on the odyssey that is quantum research.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
The post Physicists Have Measured ‘Negative Time’ in the Lab appeared first on SingularityHub.
2026-05-13 02:28:00
The wireless rings read 100 common signs from two sign languages and “autocomplete” sentences.
At the turn of the 20th century, William Hoy transformed Major League Baseball. The most prominent deaf player in history, he taught his team American Sign Language (ASL) to communicate on the field while keeping opponents in the dark. His silent speech, a legacy well over a century old now, also inspired umpires to make calls using hand gestures.
ASL is one of some 300 sign languages used today by roughly 70 million deaf people worldwide. But only a sliver of society understands signs. Everyday tasks, like ordering at a restaurant or meeting people at social events can be difficult. To bridge the gap, a South Korean team developed smart rings to translate finger motions into text.
Older devices usually require a jungle of cables to connect sensors. But the new rings are wireless, freeing people to use natural hand motions. The rings also stretch to fit different finger sizes. These upgrades make them more comfortable and reliable, wrote the team. Each ring is powered by a replaceable 12-hour battery.
Fluent signers can communicate at speeds of around 100 to 150 signs per minute, similar to spoken conversation. Devices need to keep up with that speed to avoid uncomfortable pauses. So the team developed AI-based “autocomplete” for the system that, like typing, guesses the next word based on what’s already been signed to generate phrases and sentences on the fly.
Trained on 100 common words in ASL and International Sign Language (ISL), the wearable was over 88 percent accurate in tests, even for users with no experience.
The rings are a step toward “seamless interaction between signers and non-signers,” wrote the team.
There are a variety of devices that translate sign language into text or speech, some already on the market.
One design is a bit like virtual reality gaming. It uses cameras and computer vision software to recognize hand gestures. The approach is reasonably fast and accurate in the lab, but struggles in simulated real-world scenarios, where changes in lighting or background confuse the system.
Devices worn by users are more reliable. WearSign, for example, uses sensors to capture the electrical activity of muscles during signing and translates it into text. Often, these devices need to be tailored to the user, a hurdle that limits use, as some can’t commit to the training.
Engineers have also tried embedding tracking sensors in a smart glove. The sensors send signals through cables to a shared wireless transmitter. But it’s a bit like using tools wearing a heavy winter glove. The devices limit natural movement and are uncomfortable for daily use.
They also usually come in only one size with fixed sensor placements, wrote the team. So, depending on hand size, the sensors may be out of place, reducing accuracy.
To overcome these problems, the team built AI rings to track the seven most dominant fingers in signing. (The right pinkie, left middle finger, and thumb didn’t make the cut.) The rings are worn right below the second knuckle to allow natural movement.
Each device is made of stretchy material to accommodate different finger sizes and looks more like a translucent Band-Aid than a typical ring. A tiny accelerometer captures movements like bending, curling, and holding still. The sensors are cheap, low-power, and already used in Apple Watches, Fitbits, and other wearables. There are also onboard chips to manage power use, wafer-thin Bluetooth transmitters, and common replaceable batteries that last nearly 12 hours.
The rings broadcast signals to a host device, which processes the data and maintains a timeline of each movement so incoming signs aren’t scrambled in translation.
To identify words, the system matches gestures to a database of 100 ASL and ISL signs. For example, closing both open palms into fists means “want.” The rings can also pick up signs in motion, like “dance” or “fly,” and those with fingers held still, like “I” and “you.” In first-time users, the system was 88 percent accurate for both ASL and ISL.
To make sure that conversations flow naturally, the team added an AI to track conversations and predict what word comes next. In tests, the system autocompleted simple phrases, like “family want beautiful animal.”
While still experimental, the rings could also translate between sign languages. Because the AI learns from gestures alone, with enough training data, it could eventually turn into a kind of Google Translate for signing.
But finger gestures fail to capture the full spectrum of sign language. Facial expressions, mouth movements, shoulder and body posture, speed, and rhythm all carry critical information, including meaning and emotion. Without this context, the system could easily miscommunicate intent. Some efforts are now returning to older video-based systems to better capture the entire signing experience, this time with sleeker hardware and far more processing power.
The team thinks the rings might be useful elsewhere too, like for use in virtual or augmented reality, touchless computer interfaces, and tracking hand movements in rehabilitation.
The post These Seven AI Rings Translate Sign Language in Real Time appeared first on SingularityHub.
2026-05-11 22:00:00
Far from shore, the server farms would be powered by waves, cooled by seawater, and networked by satellite.
As AI demand for computing power surges, companies are searching for new ways to fuel data centers. One startup is now proposing floating data centers powered by ocean waves, and they just raised $140 million to bring the idea to fruition.
Tech companies are planning to spend roughly $750 billion on data centers this year. But the elephant in the room is figuring out how to power these facilities. They’re already straining electrical grids across the world, and the pace of the buildout is far surpassing our ability to bring new power online.
This energy shortfall is leading tech companies to invest in a series of increasingly outlandish fixes from restarting shuttered nuclear reactors to developing novel geothermal energy technology and even launching data centers into space.
Now, several leading Silicon Valley figures, including Palantir’s Peter Thiel and Salesforce’s Marc Benioff are backing Oregon-based startup Panthalassa. The startup is developing floating data centers that generate their own electricity from waves. These investors recently joined a $140 million series B round that will allow the company to complete a pilot manufacturing facility near Portland and begin deploying the latest generation of its devices, or “nodes.”
“There are three sources of energy on the planet with tens of terawatts of new capacity potential: solar, nuclear, and the open ocean,” CEO Garth Sheldon-Coulson said in a press release. “We’ve built a technology platform that operates in the planet’s most energy-dense wave regions, far from shore, and turns that resource into reliable clean power.”
The company’s nodes are nearly 300 feet long. A bulbous sphere at the top floats on the ocean’s surface, and a lengthy tube-like housing beneath holds computer servers. As the node bobs up and down on the waves, the movement forces water up through a tube into a pressurized reservoir where it drives a turbine to generate electricity for the chips.
Besides powering the data center with renewable energy, the nodes also use the surrounding seawater to cool the chips—a much more sustainable solution compared to land-based facilities, which use significant amounts of water and electricity to manage heat.
The data centers transfer information via SpaceX’s Starlink satellite network. This does away with the need for cabling, either for power transmission or networking, and allows the nodes to operate autonomously from anywhere in the ocean. They’re also self-propelling, can navigate to their deployment location, and can stay in position without external help.
The company designed the hardware with minimal moving parts, so it can operate for extended periods without maintenance—a crucial factor for operating far from shore. Panthalassa validated the concept with a three-week trial of their second-generation node Ocean-2 off the coast of Washington state in early 2024.
This isn’t the first attempt to harness the power of waves to generate renewable energy. The company’s main innovation is that it skips the complexities of getting power back to shore. “One of the key insights we had…was that it’s very important to use the electricity in place,” Sheldon-Coulson told the Financial Times. “We will never be transmitting electricity back to shore. That makes us very different from all other ocean energy that’s been tried in the past.”
The latest funding will be used to complete a pilot manufacturing facility near Portland and deploy Panthalassa’s next-generation Ocean-3 nodes, which are scheduled for testing in the northern Pacific later this year. The company says it’s targeting commercial deployment in 2027.
The approach does face some major hurdles though, Benjamin Lee, a computer architect at the University of Pennsylvania, told Ars Technica. While relying on satellite communication does away with power transmission headaches, these links have significantly lower bandwidth compared to the optical fiber normally used to network data centers. Combined with the potential for signal delays, this could limit how useful they are for the heavy AI workloads they’re meant to handle.
However, the approach has clear parallels with another idea that’s seized Silicon Valley—orbital data centers. Rather than using wave energy and ocean water for cooling, these facilities would rely on abundant solar energy and the frigid deep-space vacuum to chill their chips. But going orbital would be far costlier and more complex, suggesting Panthalassa’s approach may be a more viable alternative.
The sea is a cruel mistress though, and deploying and maintaining a fleet of ocean-going data centers won’t be simple. Nonetheless, if they can pull it off, the idea may ease the AI energy crunch.
The post In the Scramble to Power AI, Investors Bet $140 Million on Data Centers at Sea appeared first on SingularityHub.
