Italian loudspeakers tend to follow their own playbook, and the Opera Callas Diva Special Edition distributed in the U.S. by Fidelity Imports, leans into that identity without apology. Priced at $13,999, this is a reflex, floor-standing design with a rear-firing radiation system (dipole), built around the kind of materials and construction choices that set Italian brands apart: hand-crafted wood cabinetry, leather-clad baffles, and tank-like assembly that feels more atelier than assembly line.

Whether the leather actually changes the sound is still a matter of debate, but as with most things Italian, it’s as much about feel and intent as measurable outcome.

There’s also a clear voicing philosophy here. Like most offerings from Sonus faber and Opera, the goal isn’t clinical neutrality; it’s a more romantic, expressive presentation that leans into tone and texture. That doesn’t mean these speakers lack drama; if anything, they just deliver it with better timing and less shouting over Sunday gravy at Nonna’s house. Think Sophia Loren, not a reality TV meltdown—controlled, confident, and fully aware of the effect… the kind of presence that makes a room go quiet when she crosses her legs, looks your way, and lets you wonder if you’re worth the match.

Fidelity Imports is pushing Opera hard in the U.S. right now, and it’s not difficult to understand why. Paired with electronics from Unison Research, the system synergy is obvious—cohesive, deliberate, and unmistakably Italian. Bellissima, but not in a way that begs for attention. It just assumes you’re paying attention already.

Italian Engineering in a Tailored Suit, Not a Tracksuit

The Opera Callas Diva Special Edition is a reflex loaded, floor standing loudspeaker that combines a traditional forward firing driver array with a rear firing dipole tweeter system. It’s a hybrid approach that aims to balance direct sound with controlled rear radiation, adding spatial cues without turning the room into an echo chamber.

Up front, the speaker uses a single 8-inch long throw woofer paired with a 7-inch midrange driver featuring a re cooked polypropylene cone and phase plug. High frequencies are handled by a 1-inch Scan Speak 9700 tweeter, notably run without ferrofluid and incorporating a double decompression chamber, choices that typically favor openness and low mechanical damping over sheer robustness.

Around back, Opera adds two 1-inch tweeters in what it describes as a “natural dipole” configuration. This rear array expands the soundstage by introducing ambient high frequency energy, effectively making the system a 3-way plus rear dipole design rather than a conventional forward only speaker.

The crossover network is relatively straightforward, using 12 dB per octave slopes across all drivers, woofer, midrange, front tweeter, and rear tweeters, with crossover points centered approximately at 200 Hz and 2,000 Hz. This suggests a focus on phase coherence and smoother driver integration rather than aggressive filtering.

Frequency response is rated at 30 Hz to 25 kHz, covering full range playback without immediate reliance on a subwoofer. Sensitivity is specified at 90 dB (2.83V at 1 meter), making the speaker reasonably amplifier friendly, though the 4 ohm nominal impedance with a minimum above 3.2 ohms means it will benefit from stable current delivery.

Power handling is listed at 240 watts without clipping, and placement guidelines recommend at least 10 cm, about 4 inches, from the rear wall, which is modest considering the inclusion of rear firing drivers.

Physically, the Callas Diva Special Edition is substantial: 116 x 37 x 53.5 cm (H x W x D), approximately 45.7 x 14.6 x 21.1 inches, and each speaker weighs 65 kg, about 143 pounds, including its metal base. This is not a lightweight cabinet, so think carefully about which relative still has the energy to help you move it after sausage and peppers. And don’t forget the cannoli. Marone!

Italian Soul, British Precision, No Passport Required

Fidelity Imports had a lot of rooms at AXPONA. Enough that you start making choices. I only had time for a few. This one, and the Ruark Audio room were the ones that actually made me stop, close my eyes and listen, and silently wish that I didn’t have 30 more rooms to cover on the next two floors.

Part of it was the system; Opera speakers, Unison Research electronics, and the new Michell Gyro Turntable spinning records like it knew that a certain American competitor was MIA and that this was its moment to make everyone take notice.

But it was also the reaction. People didn’t just walk in and walk out. They slowed down. Took a step closer. Leaned in to look at the front baffle, then drifted over to the turntable like it might tell them something if they got close enough. Weird that. Especially because it happened more than a few times.

Nobody rushed. Nobody talked too loud. That’s usually a sign. People stood along the back of the room and listened.

I wasn’t the only one who noticed. And in a show full of rooms fighting for attention, this one didn’t have to. Steve Jain needs to make this set-up a permanent hi-fi show experience.

The room was driven by the Unison Research Unico PRE V2 and Unico DM V2 power amplifier. Together, they retail for $18,498 USD. That’s not inexpensive, but in the context of AXPONA, it sits well below many of the larger systems on display.

The Unico DM V2 is a high power, dual mono hybrid design using Unison Research’s A.S.H.A. Class A-AB output stage. The emphasis is on current delivery and stability into more demanding loudspeaker loads rather than chasing extreme specifications.

The Unico PRE V2 is a fully balanced preamplifier with a tube based input stage. It includes a well equipped MM/MC phono stage with selectable gain and loading, making it a viable option for vinyl playback without requiring an external phono stage.

There is no built in streaming platform or Bluetooth support. That appears to be a deliberate choice, leaving digital source selection to external components.

The PRE V2 does include an internal DAC based on the Sabre ES9018K2M converter. It uses a balanced output stage designed to integrate with the tube input section, with the goal of maintaining consistent tonal balance between digital and analog inputs.

Digital connectivity includes USB-B, two S/PDIF, and two optical inputs. USB supports PCM up to 384 kHz and native DSD up to 256, along with DoP up to 128. S/PDIF and optical inputs support resolutions up to 192 kHz.

The Unico DM V2 is rated at 220 watts into 8 ohms and 340 watts into 4 ohms in stereo operation, with stability down to 2 ohms. In bridged mono configuration, it delivers 650 watts into both 8 ohm and 4 ohm loads.

My biggest takeaway from this room? Synergy matters. A lot.

Having spent time with and reviewed some of Unison Research’s tube amplifiers, the new pairing has a lot more palle, but it doesn’t trade away the qualities that made those designs stand out. The tonal balance, clarity, and sense of flow are still intact. It just brings more control and authority when the music asks for it.

Unison deserves your attention. So do these Opera loudspeakers. They’re expressive without being aggressive. They don’t grab your Members Only jacket and threaten you with brute force. They take a different approach and pull you in, keep you there, and let the music do the work.

There’s something to that. Not everything needs to hit you over the head to make its point.

More info at: operaloudspeakers.com | unisonresearch.com | michellaudio.com

Related Reading:

Marill previously led financial services and automotive industry at Facebook and Instagram.

OpenAI has hired Airbnb’s former European, Middle East and Africa (EMEA) lead Emmanuel Marill as its first EMEA managing director, as the company continues to expand globally with an initial public offering in sight.

Marill, who led Airbnb’s EMEA, Australia and New Zealand operations for a number of years, also previously worked with Meta as their financial services and automotive industry lead for Facebook and Instagram.

Marill’s appointment reflects “strong momentum in EMEA”, OpenAI said. He will be based in Paris and report to chief strategy officer Jason Kwon. He will also collaborate with OpenAI’s EU headquarters in Dublin, which currently has around 80 employees.

“As demand for ChatGPT and Codex continues to grow rapidly all over the world, we are investing significantly in our international leadership and operations”, Kwon said.

Marill added: “There’s real momentum across EMEA, with many countries here leading globally in adoption of AI.”

OpenAI has had a tougher time in Europe than in the US. The company faces increasing regulatory pressures from EU officials, as well as resistance from businesses over digital sovereignty.

However, the company said that weekly active ChatGPT users in the EMEA region have grown by 70pc since last year, with Germany, France, the UK and Spain among its top markets for ChatGPT and Codex.

On its home turf in the US, OpenAI faces an even bigger challenge from the likes of Anthropic, which is fast encroaching on the company’s clientele.

Anthropic appointed long-time technology executive Pip White as the head of its UK, Ireland, Northern Europe and Israeli operations last November.

OpenAI reportedly plans to double its headcount by the end of the year. Late last year, Bloomberg reported that OpenAI and its rival Anthropic were both looking to expand their office footprints in their Dublin headquarters.

Meanwhile, OpenAI also announced its first permanent London office for next year with a capacity of more than 500. The announcement came despite an indefinite hold on the company’s plans for a Stargate UK over energy costs and regulatory burden.

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

Makers are constantly coming up with innovative methods to transform everyday objects into automated displays, and Terence Grover has just added another to the list. His robot flips a rare €1000 Monaco commemorative coin repeatedly, with a live audience cheering and calling the shots the entire time. People watching his YouTube livestream give commands to choose when to flip the coin next, and the robot just does its job without troubling its inventor.

Grover began with a relatively bare-bones setup, utilizing only a bit of cardboard, a basic solenoid, and a 9v battery to get the ball going. He had created a rudimentary prototype that showed promise, but the coin kept landing off-center and refusing to play ball more than once or twice. He then upgraded to a proper design, producing a strong tray on a 3D printer and adding a set of blades placed similarly to a camera’s aperture. After each landing, the new servo motor closes those blades, nudging the coin back to the exact same area above the solenoid so the next flip may take place smoothly right away.

Sale

ELEGOO UNO R3 Smart Robot Car Kit V4 for Arduino Robotics for Kids Ages 8-12 12-16 STEM Science Kits…

• ELEGOO Smart Robot Car: An educational STEM kit beginners (kids) to get hands-on experience about programming, electronics assembling and robotics…
• Complete Package: Contains 24 kinds of module parts including obstacle avoidance, line tracing module, infrared remote control and also you can…
• Easy to Assemble: All the module interface has been modified with XH2. 54 ports as to make it much easier and convenient to assemble the car and…

The power comes from a 12V source, which delivers a fast burst to the solenoid, propelling the coin skyward in a crisp arc. Once it’s returned to the tray, a small security camera above captures a clear 2,000-pixel image. Grover chose this cheap Amazon camera since the original Raspberry Pi camera was too weak for long sessions in changing lighting conditions.

Inside the Raspberry Pi, a Python script is simply sitting there, waiting for that picture to appear. OpenCV converts the image to a grayscale snapshot to look for the coin’s contour, and then a machine learning model trained on over 400 photographs of the coin that Grover had personally labeled determines which side of the coin is visible, heads or tails. The model is performed locally using Tensorflow Lite, so the decision is returned to the stream in a flash, and the results are immediately visible to everyone on the livestream.

Grover set up a simple web server on the same Raspberry Pi so viewers could trigger flips through a neat and clear interface or even just talk to them on YouTube, as the relay board keeps the high-voltage solenoid safe by flipping it on command, preventing damage to the low-power electronics. Grover adjusted every timing variable, down to the fraction of a second, until everything felt absolutely smooth.

Grover was most concerned with getting the device to be reliable over time, since they’d experienced a couple jams early on when the coin landed on its edge or slid too far to one side. Every tiny adjustment to the tray and blade tension reduced the amount of times it would freeze up, eventually allowing it to continue for hours without requiring a human intervention. During a particularly long livestream session in July 2024, the machine managed to flip the coin 50,000 times in a row without requiring any user assistance.
[Source]

Summary: VAST Data raised $1 billion in a Series F at a $30 billion valuation, more than tripling from $9.1 billion, with Drive Capital and Access Industries co-leading and Nvidia, Fidelity, and NEA participating. More than $500 million is secondary capital. The company reports $4 billion in cumulative bookings, $500 million-plus in committed ARR, and is free cash flow positive with revenue roughly tripling year over year. Key customers include xAI’s 200,000-GPU Colossus cluster and CoreWeave’s $1.17 billion agreement.

VAST Data raised $1 billion in a Series F round at a $30 billion valuation, more than tripling the $9.1 billion it was valued at in its Series E in late 2023. Drive Capital and Access Industries co-led the round, with Nvidia, Fidelity Management and Research Company, and NEA participating. More than $500 million of the total is secondary capital, meaning it goes to early investors and employees selling shares rather than into the company’s treasury, a structure that relieves liquidity pressure on long-tenured shareholders and reduces the urgency of an IPO. The round makes VAST Data the most valuable private technology company founded in Israel, following Google’s $32 billion acquisition of Wiz in March.

The valuation is striking not because a company raised a billion dollars in 2026, a year in which record AI funding rounds have reshaped expectations of what venture-scale capital looks like, but because VAST Data sells data infrastructure, the layer of the AI stack that sits between the GPUs and the models. It is not a foundation model company. It is not a cloud provider. It is the company that ensures the data reaches the processors fast enough to keep them busy. Jensen Huang, Nvidia’s chief executive, recorded a personal endorsement at VAST’s Forward 2026 conference, stating that “with VAST Data, we’re transforming the storage of AI infrastructure” and explaining that without VAST’s technology, even the fastest AI processors face severe data bottlenecks. When the company that makes the GPUs tells you the GPUs are useless without a particular data platform, investors listen.

What VAST Data actually does

VAST Data provides what it calls an AI operating system that unifies storage, database, and compute into a single platform. The core architecture, called DASE (Disaggregated and Shared Everything), was announced when the company emerged from stealth in February 2019. It is flash-first and single-tier, eliminating the traditional storage hierarchy in which data moves between fast, expensive tiers and slow, cheap ones. For AI workloads, where training runs consume petabytes of data at sustained high throughput, the elimination of tiering removes a bottleneck that legacy storage systems were never designed to handle.

The platform has expanded well beyond storage. VAST DataSpace provides a globally distributed namespace across on-premises, cloud, and edge locations, scaling to exabytes and trillions of files. VAST InsightEngine automates real-time AI pipelines, handling chunking, embedding, vectorisation, and retrieval for retrieval-augmented generation, semantic search, and classification. VAST DataBase includes an integrated vector store that the company claims supports trillion-vector scale with constant-time search. VAST CNode-X, an Nvidia-certified system, makes GPU servers first-class infrastructure components inside the platform, with a fully CUDA-accelerated version of the operating system designed to run directly on Nvidia-powered servers. The pitch is that VAST is not a storage company that added AI features. It is a data platform that was built for AI from the beginning, and the storage is just the foundation.

The numbers

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

VAST Data has accumulated more than $4 billion in cumulative bookings and reports more than $500 million in committed annual recurring revenue as of the end of fiscal year 2026. CTech, the technology publication of the Israeli financial newspaper Calcalist, reports that total ARR including non-committed revenue has reached $2 billion. Revenue has been roughly tripling year over year. The company is generating more than $100 million in cash per quarter and is free cash flow positive with a positive operating margin, unusual for a company at this growth rate. The customer base has quadrupled among Fortune 1000 companies, with the top 100 new customers spending more than $1.2 million on average. Contracts typically run five to seven years.

The marquee customer relationships illustrate the scale. VAST Data powers the data platform behind xAI’s Colossus supercomputing cluster, a facility with more than 200,000 Nvidia GPUs where VAST says it reduced total cost of ownership by 50%. CoreWeave signed a $1.17 billion commercial agreement in November 2025, using VAST as the primary data foundation for its Nvidia-accelerated computing cloud. Other customers include Pixar, which uses the platform for petabytes of rendered assets as AI training data, NASA, the US Department of Energy, Boston Children’s Hospital, Booking Holdings, and several of the world’s largest banks. Renen Hallak, VAST’s founder and chief executive, said the company is “already supporting AI environments spanning millions of GPUs globally, operating across every layer of the AI stack.”

The data layer thesis

The investment thesis behind a $30 billion valuation for a data infrastructure company rests on a structural argument about how the AI stack works. The industry has spent three years and hundreds of billions of dollars on GPUs. Surging global AI investment, which the Stanford AI Index pegged at $285.9 billion in US private AI capital in 2025 alone, has been concentrated overwhelmingly on compute. But a GPU that is waiting for data is a GPU that is not training. The data layer, the infrastructure that stores, indexes, moves, and transforms the data that feeds the models, is increasingly recognised as the binding constraint on AI performance.

This is why Nvidia is not just investing in VAST Data but actively integrating its technology. The CUDA-accelerated operating system and CNode-X certification mean that VAST’s platform is designed to run on the same Nvidia hardware that runs the models, eliminating the traditional separation between storage infrastructure and compute infrastructure. Nvidia-backed AI infrastructure companies now span the entire stack, from GPU cloud providers to chip fabrication to data platforms, and VAST’s role is to ensure that the data moves as fast as the silicon can process it.

AI infrastructure startup valuations have been climbing sharply across the sector. FluidStack is in talks to raise $1 billion at an $18 billion valuation. CoreWeave, VAST’s largest customer, was valued at $35 billion earlier this year. Enterprise AI infrastructure deals like Jane Street’s $6 billion cloud commitment to CoreWeave, with a $1 billion equity investment attached, illustrate that demand for AI infrastructure is broadening beyond the hyperscalers into financial services, healthcare, and government. VAST’s position at the data layer of these environments, not the compute layer and not the model layer, is what makes the valuation argument distinct from the GPU cloud companies. If the compute layer is the engine, VAST is the fuel line. A $30 billion fuel line is expensive. The argument is that without it, the engine does not run.

The competitive landscape

VAST Data is not the only company building AI-native data infrastructure. DDN and WEKA are the two most frequently cited competitors, both offering high-performance storage platforms optimised for machine learning workloads. Hammerspace provides a global data orchestration layer. The incumbents, Dell, HPE, Hitachi Vantara, IBM, NetApp, and Pure Storage (recently rebranded as Everpure), are all deepening their Nvidia integrations and repositioning their storage portfolios for AI. Pure Storage’s FlashBlade products compete directly with VAST on performance. NetApp has expanded its AI storage services. All of them have larger installed bases and longer customer relationships than VAST.

VAST’s argument is that legacy storage architectures, designed for databases and file servers and retrofitted for AI, cannot deliver the sustained throughput that training runs at the scale of Colossus require. The single-tier, flash-first architecture eliminates the data movement that tiered systems impose, and the integrated database and compute capabilities mean that data transformation, the chunking, embedding, and vectorisation that AI pipelines require, happens within the platform rather than in a separate processing layer. Whether that architectural advantage is durable or whether the incumbents can close the gap will determine whether a $30 billion valuation looks prescient or excessive in three years.

Hallak has told employees and bankers that the company has considered an IPO in the second half of 2026 or later, according to The Information. The secondary-heavy structure of the Series F suggests that timeline is not imminent. VAST Data can afford to wait. It is cash-flow positive, tripling revenue, and sitting at the centre of the most capital-intensive technology buildout since the internet. The question is not whether the data layer matters. It is whether $30 billion is the right price for the company that is building it.

While LG doesn’t seem to be giving up on OLED TV technology any time soon, the company is also hopping on the RGB bandwagon with their own flagship Micro RGB evo series of TVs for 2026. RGB-backlit LCD TVs offer several advantages over traditional MiniLED/LCD TVs including extended color gamut, higher efficiency and higher color brightness. Instead of relying on a white or blue LED backlight, and a subtractive color filter system that diminishes color brightness, LG’s Micro RGB evo TVs use individual red, green and blue backlighting elements, combined with an LCD panel to create the on-screen image.

Because Micro RGB evo TVs use traditional LCD “mother glass,” they can reach larger screen sizes than OLED TVs, and can typically reach higher overall brightness levels than OLED, making them well-suited for bright room viewing.

LG’s flagship MRGB95 Series of Micro RGB evo TVs will be available in screen sizes from 75 inches to 100 inches and will be priced as follows:

• LG 75-inch Micro RGB evo TV (75MRGB95BUA): $4,999.99
• LG 86-inch Micro RGB evo TV (86MRGB95BUA): $6,999.99
• LG 100-inch Micro RGB evo TV (100MRGB95BUA): $7,999.99

Key features of the LG Micro RGB evo TV technology include:

• “Triple 100% Color Coverage:“ LG’s MRGB95 Series has earned certification from Intertek for “full gamut coverage in BT.2020, DCI-P3, and Adobe RGB.” This means LG’s Micro RGB evo TVs can reproduce colors not previously possible on consumer displays.
• “Unrivaled Contrast at Any Size:”  While OLED is still the king of black levels and contrast, LG’s 13 years of OLED development has led to the company’s latest video processor, the Alpha 11 AI Processor Gen 3 which powers the TV’s advanced color, motion and up-conversion processing. Meanwhile the company’s “Micro Dimming Ultra” processing coordinates over a thousand backlight dimming zones for precise control of contrast and detail as well as reduced color blooming and haloing.
• Personalized Viewing Experience: featuring the latest version of LG’s webOS platform (WebOS 26), the Micro RGB evo features Voice ID, AI Picture/Sound Wizard, AI Concierge, and an enhanced customizable “My Page” home screen for a smarter, more intuitive experience. Integration with Google Gemini helps expand the TV’s search capabilities beyond simple content recommendations.

While the company did showcase this Micro RGB evo technology, both at CES and at a press-only TV workshop event at the U.S. headquarters earlier this year, the company also highlighted some of the drawbacks of RGB backlighting, including color bleed from color crosstalk and color blooming on bright objects on dark backgrounds. But LG also stressed that these issues can be greatly reduced via proper RGB backlight control and appropriate use of color filtering. When we spent time with the new MRGB95 at these events, they did not exhibit significant issues with color crosstalk, haloing or blooming, thanks to LG’s superior video processing.

The Bottom Line

While some manufacturers have tried to position their RGB-backlit TVs as “OLED Killers,” LG instead positions their Micro RGB evo TVs as an alternative to OLED technology when customers require bright room viewing or larger screen sizes. But is the enhanced color gamut coverage alone (compared to OLED) worth the premium in pricing over LG’s latest and greatest OLED TVs like the G6 and W6 Wallpaper TV? This remains to be seen.

Related Reading:

LG Enters the Micro RGB LED/LCD TV Race with the MRGB95 Series at CES 2026

LG Reveals Pricing/Availability Details of 2026 OLED TV Line

Apple has released out-of-band security updates for iPhone and iPad devices to fix a Notification Services flaw that could allow notifications marked for deletion to remain stored on the device.

The bug, tracked as CVE-2026-28950, was fixed on April 22, 2026, in iOS 26.4.2 and iPadOS 26.4.2 and in iOS 18.7.8 and iPadOS 18.7.8.

“Notifications marked for deletion could be unexpectedly retained on the device,” reads the Apple security bulletin.

Apple says the flaw was fixed through improved data redaction but provided no additional information.

However, the company has not said whether the flaw was exploited in attacks or why it was addressed outside the normal security update cycle. Apple also did not share technical details about how long notification data remained on the device or how it could potentially be recovered.

While Apple has not explained why it released this emergency update, recent reporting by 404 Media described how the FBI recovered copies of Signal messages from a suspect’s iPhone, even after they had been deleted in the app.

According to trial notes published by supporters of the defendants, the recovered data did not come from Signal’s encrypted message store, but instead from iPhone’s notification storage.

“Messages were recovered from Sharp’s phone through Apple’s internal notification storage — Signal had been removed, but incoming notifications were preserved in internal memory,” the notes state.

404 also reported the notification data was retained even after Signal was deleted from the device.

Apple’s advisory does not reference the case, but its description of notifications being retained on the device closely aligns with the type of data persistence described in that report.

Users are advised to install the latest updates as soon as possible to prevent deleted notification data from being unexpectedly retained on their devices.

Furthermore, it is possible to prevent Signal message content from being retained in the iOS notification data storage by going to Signal Settings > Notifications> Notification content and setting Show to “Name Only” or “No Name or Content”.

BleepingComputer contacted Apple with questions about these updates, but has not yet received a response.

AI chained four zero-days into one exploit that bypassed both renderer and OS sandboxes. A wave of new exploits is coming.

At the Autonomous Validation Summit (May 12 & 14), see how autonomous, context-rich validation finds what’s exploitable, proves controls hold, and closes the remediation loop.

Claim Your Spot

Summary: SpaceX’s confidential S-1 pre-IPO filing warns that its orbital AI data centre plans “involve significant technical complexity and unproven technologies, and may not achieve commercial viability,” contradicting Elon Musk’s January claim at Davos that space-based AI was a “no-brainer” achievable within two to three years. The filing comes as SpaceX targets a $1.75 trillion IPO valuation and has applied to the FCC for one million data centre satellites, while competitors Starcloud, Google (Project Suncatcher), and Blue Origin pursue their own orbital compute programmes.

SpaceX told prospective investors in its confidential S-1 pre-IPO filing that its plans for orbital AI data centres “involve significant technical complexity and unproven technologies, and may not achieve commercial viability.” The company warned that any future space-based compute infrastructure will operate “in the harsh and unpredictable environment of space, exposing them to a wide and unique range of space-related risks that could cause them to malfunction or fail.” The disclosure, first reported by Reuters on Monday, is legally standard for a company approaching what could be the largest initial public offering in history. It is also a remarkable piece of bureaucratic candour from the same organisation whose chief executive described data centres in orbit as a “no-brainer” three months ago.

At the World Economic Forum in Davos in January, Elon Musk said the lowest-cost place to put AI would be in space “within two years, maybe three at the latest.” He called space-based solar “10 times cheaper than terrestrial solar” because “you don’t need any batteries,” described the cooling problem as solved by simply pointing a radiator away from the sun at three degrees Kelvin, and predicted that more AI capacity would sit in orbit than on Earth within five years. In February, SpaceX filed with the Federal Communications Commission to launch and operate up to one million satellites as the “SpaceX Orbital Data Center system” at altitudes between 500 and 2,000 kilometres. The filing described satellites that would “directly harness near-constant solar power with little operating or maintenance cost.” The S-1, filed confidentially with the Securities and Exchange Commission ahead of a targeted June listing at a $1.75 trillion valuation and a $75 billion raise, says something different.

The physics of the problem

The contradiction between Musk’s public statements and SpaceX’s legal disclosures maps onto a set of engineering constraints that have not changed since Davos. In vacuum, all heat dissipation happens through radiation. There is no convection, no liquid cooling, no fans. To radiate just one megawatt of heat at 20 degrees Celsius, an orbital data centre would need roughly 1,200 square metres of radiator surface, the area of four tennis courts. The International Space Station’s entire electrical system produces only 0.2 megawatts; ground-based hyperscale data centres are racing toward gigawatt scale. The three-degree background temperature of space is irrelevant if the radiators needed to exploit it weigh more than the servers they are cooling.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Power is equally constrained. Solar panels in orbit receive roughly five times more energy than on the ground, with no atmosphere, weather, or nighttime in certain orbits. But it would take approximately one square mile of solar array in Earth orbit to produce one gigawatt at 30% cell efficiency. The ISS produces 0.2 megawatts from arrays that span the length of a football field. Scaling to the gigawatts that a single hyperscale data centre consumes on Earth would require deploying and maintaining solar infrastructure orders of magnitude larger than anything humans have built in space.

Hardware obsolescence may be the most underappreciated constraint. GPUs depreciate as new architectures emerge every two to three years. On Earth, racks are swapped continuously. In orbit, every hardware replacement requires a launch, docking, or robotic servicing mission. Radiation exposure causes bit flips and permanent circuit damage. Radiation-hardened chips lag multiple generations behind commercial processors. Triple modular redundancy, running three parallel systems and taking the majority vote, would triple the hardware requirements. The AI’s soaring energy demands, which the IEA projects will push data centre electricity consumption past 1,000 terawatt-hours by the end of 2026, are real. The question is whether solving them in orbit creates more problems than it solves.

The competitive landscape in orbit

SpaceX is not the only company pursuing orbital compute, which makes the S-1 disclaimer more strategically significant than a standard risk factor. Starcloud, formerly Lumen Orbit, launched the first high-powered GPU into orbit in November 2025, an Nvidia H100 that represented 100 times more compute than had ever operated in space. In December, Starcloud became the first company to run a large language model, Google’s Gemma, and the first to perform in-orbit LLM training. By March 2026 it had raised $170 million at a $1.1 billion valuation, the fastest unicorn in Y Combinator’s history. Its next satellite, targeting 200 kilowatts and a cost of roughly $0.05 per kilowatt-hour, is planned for October.

Google’s Project Suncatcher, a partnership with Planet Labs, plans to launch two test satellites carrying Google TPUs by early 2027 and envisions one-kilometre arrays of 81-satellite compute clusters in dawn-dusk sun-synchronous orbit. Google’s analysis suggests launch costs may fall below $200 per kilogram by the mid-2030s, making space data centres cost-comparable to terrestrial energy costs at that point. Nvidia announced Vera Rubin Space-1, a chip system designed specifically for orbital data centres. Blue Origin filed its own FCC application for 51,600 data centre satellites. The a16z-funded startup Orbital is building an AI satellite constellation. The idea is not fringe. It is attracting serious capital and serious engineering talent. SpaceX’s S-1 is notable precisely because the company that controls the launch vehicles and the satellite internet constellation, the company best positioned to make orbital compute work, is the one telling investors it might not.

The terrestrial alternatives

The S-1 disclosure arrives in a week when the terrestrial alternatives are absorbing enormous investment. Massive AI infrastructure deals like Meta’s $27 billion commitment to Nebius illustrate the scale of spending on ground-based compute. Nuclear-powered AI data centres are attracting dedicated funding, with Valar Atomics raising $450 million at a $2 billion valuation to build small modular reactors purpose-built for AI workloads. The US Department of Energy has identified 16 federal sites for data centre construction adjacent to existing nuclear facilities. By 2026, 18 nuclear-powered AI facilities with a combined capacity of 31.2 gigawatts are tracked globally. Microsoft’s Project Natick deployed an undersea data centre capsule designed for AI workloads in February 2025. The tech industry spent roughly $580 billion in 2025 turning deserts and abandoned factories into GPU-packed facilities.

The pattern is consistent: every approach to the AI power problem that keeps the servers on Earth, or at most underwater, is attracting more capital and progressing faster than the orbital alternatives. Nuclear reactors are a proven technology being adapted to a new use case. Orbital data centres are an unproven technology being proposed for a use case that may not require them. The S-1 language suggests SpaceX’s own engineers and lawyers recognise the distinction, even if the company’s public messaging has not caught up.

The IPO context

The S-1 filing serves two masters. SpaceX needs to present orbital data centres as a credible growth story to justify a $1.75 trillion valuation, the highest ever for a pre-IPO company. It also needs to disclose the risks clearly enough to protect itself from securities litigation if the plans do not materialise. The result is a document that simultaneously promotes and disclaims the same initiative. This is not unusual in IPO filings. It is unusual when the chief executive has spent the preceding three months describing the initiative as inevitable, obvious, and cheaper than the alternatives.

The SpaceX-xAI merger in February, an all-stock transaction valuing the combined entity at $1.25 trillion, was explicitly motivated by orbital data centres. Musk said integrating Starlink’s global satellite mesh with xAI’s large language models was a primary rationale. Musk’s AI chip ambitions through the Terafab project with Intel include dedicated processors for orbital deployments. The one million satellites in the FCC filing would represent a hundred-fold increase over the current population of low Earth orbit. Ars Technica estimated the barebones deployment cost at “at least $1 trillion.” The vast majority of more than 1,000 public comments to the FCC opposed the plan, citing debris, light pollution, and the risk of Kessler syndrome, a cascading chain of collisions that could render entire orbital altitudes unusable.

SpaceX may eventually prove that orbital compute works. The company has a record of achieving what others said was impossible, most notably reusable rockets. But the S-1 filing is not the language of a company that has solved the problem. It is the language of a company that wants credit for trying and protection if it fails. The gap between Davos in January and the SEC in April is the gap between a pitch and a prospectus. Both are real. Only one carries legal liability.