• AMD guided Q4 2026 revenue to about $17 billion, above Intel's $15 to $16 billion forecast, which would make it the larger company in terms of raw revenue for the first time
• Data center sales nearly doubled to $8.5 billion in Q3 as Instinct MI455X GPUs and Helios racks started shipping
• The lead, interestingly, is in revenue, not volume: Intel still ships about two-thirds of x86 CPUs, and AI accelerators are set to outsell Epyc inside AMD this quarter

AMD has told investors to expect fourth-quarter 2026 revenues of around $17 billion, plus or minus 3%.

If that lofty forecast holds, AMD would bring in more money than Intel in a single quarter for the first time in the two companies' decades-long rivalry.

For context, AMD's outlook arrived with a record third quarter in which revenue rose 45% year over year to $13.4 billion, even as Intel, which reported a week earlier, has guided to between $15 billion and $16 billion for the same three months.

A numbers game in revenue, not volume as data centers do the heavy lifting

Using the midpoint of both companies' estimates and assuming both predictions come true, AMD is essentially set to generate $1.5 billion more in revenue than its traditional chipmaking rival.

Much of this comes from AMD's data center gains: Data center revenue climbed 97% to a record $8.5 billion, or 64% of total sales, up from 47% a year earlier, with the segment's operating margin widening to 34%.

AMD CEO Lisa Su said Q3 was the first quarter AMD shipped Instinct MI455X accelerators and Helios racks for revenue, with Anthropic, Meta, Microsoft, OpenAI and Oracle bringing their first racks online.

Revenue shipments of 2nm Epyc "Venice" processors also began, although, as TechRadar Pro reported at launch, the full Venice stack will not ship until 2027, so there is potentially more uplift in the pipeline assuming demand remains on track.

The picture elsewhere was mixed at best: Client revenue rose 7% to $2.9 billion, gaming fell 40% to $773 million on weaker console chip sales, and embedded grew an impressive 29% to $1.1 billion.

For Q4, CFO Jean Hu guided data center revenue up about 50% sequentially, which Tom's Hardware puts at roughly $12.8 billion, with client and gaming down more than 20% and embedded up by double digits. She added that Q4 will be the first quarter in which AMD's data center AI business, meaning Instinct GPUs and racks, outsells its server CPUs, and by a wide margin.

Intel's numbers, on the other hand, are not essentially a clean baseline, even if AMD comes out ahead regardless: Its outlook accounts for lost output from the October 14 explosion at Fab 28 in Kiryat Gat, Israel, which killed two workers and shut the plant for around two weeks. "Our guidance includes an impact of roughly $300 million to revenue," CFO Dave Zinsner told analysts, according to Calcalist, almost all of it in 10nm server products. Add that back to Intel's high end, and you get $16.3 billion, still below AMD's estimated floor.

The crossover does not seem to have been on the cards so soon, however: Wall Street was looking for about $17.05 billion from Intel, according to Futurum, and around $16.2 billion from AMD, per Opening Bell Daily. As recently as April, consensus estimates had AMD surpassing Intel in revenue in 2027, and AMD's moves seem to have been considerably ahead of analyst estimates, resulting in the stock surging more than 100% YTD despite a steep AI selloff lately.

The move is a revenue milestone, not a volume one; Intel still ships most x86 processors, holding 65.9% of the market in Q2 2026, according to Mercury Research figures reported by HotHardware. AMD's revenue also includes businesses Intel does not run at comparable scale, from Xilinx-derived embedded chips and console silicon to, above all, Instinct accelerators, making for a skewed direct comparison between the two, especially as the latter is reaping AI bonuses at a scale the former cannot currently hope to match.

In fairness, AMD's data center segment has outearned Intel's equivalent group before: TechRadar Pro reported the first such quarter in Q3 2024, when AMD's $3.5 billion edged past Intel's $3.3 billion, but its recent revenue explosion has carried AMD into what seems to be an insurmountable short-term lead. In Q3 2026, it was $8.5 billion, compared with $7.3 billion for Intel's Data Center and AI group. But with GPUs set to overtake Epyc inside AMD this quarter, the overall crossover owes more to a market Intel has barely entered than to the CPU fight that defined the rivalry. Intel's own data center accelerator, the GPU D-series formerly codenamed Crescent Island, is still upcoming.

Other factors

It would also be a mistake to read Intel's forecast as an exodus of customers: Intel's Q3 revenue rose 21%, and its data center group grew 78%, and CEO Lip-Bu Tan said the immediate priority is to "unlock additional supply to meet extraordinary demand for our Xeon family."

Part of Intel's problem is capacity, layered on top of a PC slump that is hitting both firms: AMD is planning for PC unit shipments to fall by a double-digit percentage in 2027.

AMD CEO Lisa Su was candid about the fact that AMD cannot build everything it could sell. "It is true that demand is higher than supply," she told analysts.

The Q4 forecast rests on a roughly 50% sequential jump in data center sales, much of it from a rack platform that shipped for revenue for the first time last quarter.

That ramp has costs: Inventory rose about $3 billion to $11.5 billion, and free cash flow was a relatively measly $554 million. AMD's biggest AI deals also come with customer warrants, which produced a $1.1 billion non-cash charge in Q3 that the company expects to grow. Likewise, revenue is concentrated in a handful of frontier labs and hyperscalers even as circular financing concerns continue to grow.

AMD now expects data center sales to make up more than 80% of revenue in 2027 and server CPU revenue to double. Delivering that depends on partners such as HPE, which became Helios' first major OEM backer and has since said Vultr will deploy the first Helios systems it builds. It also depends on the extra capacity Su has promised for 2027, which TechRadar noted is aimed at AI rather than consumer chips.

If both companies land inside their ranges, AMD will have outearned Intel for the first time. The more telling story is where that money comes from: not a knockout in the old CPU war, but a shift in the industry's revenue toward accelerators and racks that now fill data centers.

• A Hong Kong-led venture plans a 168m, 20,000-seat LED sphere in Hangzhou for ¥8 billion (~$1.2 billion)
• Its 72,000m² exterior screen would be a third larger than the Las Vegas Exosphere, the biggest of its kind, at least on paper
• The project is still at the planning stage, with no construction dates, screen vendor or display specs announced

A developer in eastern China has unveiled plans for a spherical entertainment venue that looks to beat the Las Vegas Sphere on every headline measurement, for less than half the price.

The catch is that, so far, the project amounts to a signing ceremony, a rendering and a set of very ambitious forecasts.

The project, dubbed the Eye of the World in English, was reported by AV Magazine after its public debut on October 30 2026 at a services-sector conference in Hangzhou's Qiantang District.

A bigger sphere on paper

According to City Express, the plan calls for a sphere roughly 168 meters across and 122 meters tall, skinned in about 72,000 square meters of exterior LED, with a 270-degree wraparound screen and 20,000 seats inside.

It would occupy a 56,000-square-meter plot on the Qiantang River in the Jinshazui New Town area, with around 160,000 square meters of neighboring land earmarked for supporting development.

On the spec sheet, the comparison is flattering when one compares it to the Las Vegas Sphere. AV Magazine puts the Las Vegas venue at 157 meters in diameter and 112 meters tall, with a 54,000-square-meter Exosphere and 17,600 seats, which makes the Hangzhou design about 7% wider and 9% taller, with a third, more exterior screen. The seating gap, however, is narrower than it looks, as Las Vegas reaches around 20,000 when its floor is set up for standing crowds.

The headline budget is ¥8 billion (~$1.2 billion), against the $2.3 billion Las Vegas eventually cost, and the project's forecast annual revenue is ¥2.6 billion, or $400 million. At the (relatively harsher) exchange rate implied elsewhere in Chinese coverage, the two figures come closer to $1.1 billion and $360 million.

The venture pairs Hong Kong-based Xinshi International Investment Co. with Vision Eye (Hangzhou) Digital Entertainment Technology Co., and Invidis reports that it is led by Huang Junhao, investment president of the former and general manager of the latter. He says the domestically designed project will show that Chinese applied technology and cultural innovation have reached "world-leading levels," and at the launch he described it as "a spectacle of unimaginable size and unimaginable form."

Interestingly, for an LED project, the most telling detail is what's missing: The invidis report also notes that no vendor or installer has been appointed for either the exterior screen or the interior wraparound display, that design, zoning and infrastructure arrangements are still in progress, and that no construction start or completion date has been given.

Nobody has published a resolution, pixel pitch or playback architecture either, and that matters, because the systems behind a screen like this are as demanding as the panels themselves.

In Las Vegas, TechRadar Pro reported that 150 Nvidia RTX A6000 GPUs drive nearly 750,000 square feet of LED with three layers of 16K video at 60 frames per second, fed by a Hitachi Vantara system of 27 nodes and 4PB of flash storage delivering 400GB/s.

The structure won't be simple either: One analysis notes that Las Vegas used parametric modeling to optimize its geodesic lattice shell and cast-steel nodes, cutting weight by around 40%, and warns that a larger sphere demands even tighter tolerances, where millimeter-level errors could show up as visible seams or patches of uneven brightness. Any claim to be the world's largest, it concludes, depends on the engineering meeting the paper specifications.

An analysis on a Chinese publication, The Paper's government affairs channel, argues the cost gap with Las Vegas rests on China's domestic LED sphere supply chain, which now has a track record of delivery, and that the real difficulty is keeping brightness, color, and seams consistent across tens of thousands of square meters of curved outdoor LED through direct sunlight and temperature swings.

The launch coverage on Sohu flags acoustics as another headache: a concave sphere focuses sound unevenly, and any absorptive treatment has to sit inside the lattice shell without breaking up the screen.

A bloated forecast with questionable demand in tow?

The commercial forecasts look even bolder than the project at hand: The developer expects 9 million to 13 million visitors a year. Spread evenly, that is roughly 24,700 to 35,600 people every day, so even the low end exceeds a sellout of the 20,000-seat bowl 365 days a year.

The Paper's analysis supplies useful context. In 2025, Hangzhou's entire performing arts sector drew about 10.25 million attendees across 19,194 shows and took ¥3.12 billion at the box office.

One venue is forecasting roughly the whole city's 2025 audience and revenue, totaling about 83% of that box office total, though the projected figure presumably also includes advertising, sponsorship, and food and drink. The same analysis also describes The Eye of the World as "still just a drawing, with construction yet to begin."

The owner of the original isn't standing still. Sphere Abu Dhabi, on Yas Island, carries a $1.7 billion construction budget and will match Las Vegas with a capacity of up to 20,000, depending on configuration, a 16K-by-16K interior screen, and 167,000 speakers, with completion due by the end of 2029. A 6,000-seat venue at National Harbor in Maryland is targeting 2030.

Sphere is also lining up the technology behind its screens. Announcing a multi-year partnership with media server specialist Disguise in September, Sphere executive vice president Ryan Dolan tied the deal to "offering these experiences to audiences in new Sphere venues around the world."

TechRadar Pro's June 2026 coverage of China's first Sphere-style domes analyzed a completely different story about going smaller and cheaper: Qianjiang's Tiangong Nova, with 7,500 square meters of exterior LED, and Shanghai's West Bund Starry Sky, budgeted at ¥1.2 billion for around 3,000 visitors at a time.

Hangzhou keeps prices low and scales up, but not everyone thinks the world needs it. Discussing China's smaller domes on an AVNation panel, AVIXA's Jennifer Weaver asked: "Massive replicas of the Vegas sphere all over the place? Probably not." Hangzhou is betting they are wrong; for now, though, it has a rendering, a riverside plot, and a set of forecasts, while the original, by all accounts, remains busy every day of the year as a revenue-generating juggernaut, bolstered, at least in part by a much more affluent audience to boot.

• Rmatamini's P99S packs a 16GB Tesla V100, Core i7-10850H, 16GB RAM and 512GB SSD into a 5.6L box for $699 on Newegg
• The same 16GB V100 PCIe card costs a whopping $584 new on Amazon, and renewed units can be had for $410
• Nvidia's R580 driver branch is Volta's last; CUDA 13 no longer targets the card, and potential buyers need to contend with a relatively new seller in tow

A new Newegg listing squeezes an Nvidia Tesla V100 with 16GB of HBM2 into a 5.6-liter box - and asks just $699 for the lot.

The Rmatamini P99S "Mini AI Workstation" pairs Nvidia's 2017 data center accelerator with a mobile-centric Intel Core i7-10850H, 16GB of DDR4, a 512GB NVMe SSD, a 700W power supply, and Windows 11 Pro, and ships free from China.

The number that makes this interesting sits on another site: HP's version of the same 16GB V100 PCIe card is selling new on Amazon for $584 through a third-party seller with only two units left in stock; Take that at face value, and the rest of the computer costs as little as $115.

A $584 card in a $699 box?

Amazon's pricing is where one would currently find a brand-new V100 GPU, which has seen a surge in demand due to both the relevance of older silicon and the fact that it uses HBM2, which holds its own in bandwidth compared to newer consumer-grade GPUs.

One could argue that a fairer comparison would involve a potentially refurbished unit that can be found for as low as $410 on Amazon, visible from the same listing as the new SKU. This means the mini PC itself costs either ~$115 or ~$290, depending on how one values the Tesla V100 GPU in a 16GB trim.

The V100, which once commanded a staggering $10,000 or higher price tag, is arguably dated silicon before fading away, but it is drawing attention again, and as much to do with today's market as with the card. Fast memory has become the most expensive part of any AI build.

TechRadar's analysis tied runaway graphics card pricing to AI demand for RAM in January, and a TechSpot analysis, published in February, found 16GB RTX 5060 Ti cards up 22% globally in roughly three months, a trend that has yet to break. Nvidia itself raised the price of the DGX Spark from $3,999 to $4,699 in February, citing memory supply constraints, for a machine whose early reviews had already flagged memory bandwidth as a limitation. Compact Ryzen AI Max+ 395 systems such as the MINIX ER939-AI start at $3,150.

Against that backdrop, the V100's spec sheet looks oddly current. Nvidia's datasheet rates it at 900GB/s of memory bandwidth, about twice the 448GB/s of an RTX 5060 Ti 16GB. Token generation in large language models leans heavily on bandwidth, which is why hobbyists such as Oscar Molnar spent 2026 bolting retired V100 GPUs into desktops.

The best public numbers come from YouTube channel Hardware Haven, whose tests Wccftech reported in May 2026. A modded V100 16GB produced about 130 tokens per second on gpt-oss-20b, compared to roughly 90 for an RX 7800 XT, and 108 tokens per second on Gemma 4 e4b, compared to 76 for an RTX 3060. Capped at 100W, it still managed 95 tokens per second while the whole system drew 170W at the wall.

Plenty of caveats in tow

In practice, 16GB holds mixture-of-experts models such as Gemma 4 26B-A4B entirely on the card, according to Andrew Leech's V100 LLM kit, but denser 30B-class models will not fit, and 16GB of system RAM leaves little room to offload the overflow to the CPU. RmataMini PC also sells a 32GB SKU of the same, with a bump to 32GB of DDR4 RAM, for an eyewatering $1,500, marked down from a sticker price of $2,000.

The biggest long-term caveat is software: Nvidia's driver deprecation schedule names the R580 branch as the last to support Volta, alongside Maxwell and Pascal. ComputingForGeeks notes that CUDA 13 can no longer build code for the V100, and that vLLM now requires a newer compute capability than Volta's 7.0. Ollama y llama.cpp still works on CUDA 12 builds, and community forks such as the 1Cat vLLM build keep the card useful, but the window is closing rather than opening. As Leech's kit notes, Volta's tensor cores support only FP16, so BF16 and newer low-precision formats are off the table.

The host platform is modest but adequate. The Core i7-10850H is a 45W laptop chip from 2020 with six cores, 12 threads, and 16 PCIe 3.0 lanes, matching the V100's PCIe 3.0 interface. Because the V100 has no outputs, displays are driven by the CPU's iGPU.

The seller is also relatively new, and RmataMini PC has a single seller rating; the product has none; warranty questions are referred back to the seller; and returns are for refund only, within 30 days.

For someone who wants a small, ready-made CUDA box with 16GB of 900GB/s memory and has no interest in sourcing adapters, fans, and ducts, $699 is an excellent price in a market where memory keeps getting more expensive.

But a finite software runway from Nvidia and a considerably pricier, but available 32GB SKU mean that users have much to consider before they take the plunge here for anything that could require more memory or a more current CUDA version in the near future.

• Google, Microsoft, OpenAI, and Anthropic say they will honor one robots.txt group as a cross-company opt-out from AI training
• It governs how content is used rather than whether bots can crawl it, defaults to allowing training, and does nothing for content already scraped
• Anthropic has started implementing it, OpenAI promises to do so in the coming months, Microsoft starts February 1, 2027, and Google targets 2027, while Meta is supportive but uncommitted

Google and Microsoft have joined OpenAI and Anthropic in promising to honor a single line in robots.txt that tells AI developers whether a website's content can be used to train their models.

Until now, a site that wanted to say no had to write a rule for each company's crawler, and sometimes for several crawlers run by the same company. The shared signal is meant to replace that sprawl with a single entry that every participating provider reads the same way.

That is genuine progress for overworked webmasters, but the promise is narrower than much of the coverage suggests: The new line states a preference about how content may be used rather than stopping bots from fetching it; it is not an approved standard, does nothing about content already collected, and none of the four major players fully honor it yet.

Doing the bare minimum with differing definitions of what the commitment entails

OpenAI's announcement drew much of the attention, but the company casts itself as one participant among several in an Internet Engineering Task Force (IETF) effort, and the mechanism isn't its own. It comes from an Internet-Draft published the same day by Kevin Kelley of Anthropic, which names a policy token, "__usage_ai-training", in the User-Agent line of an otherwise ordinary robots.txt group.

Training is defined by the IETF's AI Preferences (AIPREF) working group as the process of changing a generative model's learned parameters, excluding models used purely for tasks such as classification or ranking. OpenAI describes what it will honor as "foundational generative AI training," which sounds narrower than the draft's definition.

Calling it a standard may be premature: The IETF's own tracker labels the document an individual draft that the IETF has not endorsed, and that has no formal standing, and it lapses on May 3, 2027 unless revised.

Anthropic is the only one of the four major AI players to say it has started. OpenAI aims to implement the signal over the next few months but has given no date. Microsoft will begin honoring the shared preference as it crawls from February 1, 2027, while Google is aiming for 2027.

Each company is also wiring the line into its own controls differently. Google says that when a site uses the shared token to opt out without separately specifying Google-Extended, it plans to treat that "as if the publisher had disallowed Google-Extended." Because Google-Extended also governs grounding in Gemini Apps, a site that only wants to stop training could lose its citations and clicks in Gemini answers too, and Search Engine Journal notes that Google's documentation describes no way to separate the two.

Interestingly, Microsoft goes the other way. It adopts the shared line for training but adds a separate, page-level AI Grounding preference, including a DisallowExtended option that keeps content available to its own products such as Copilot while excluding the grounding services it sells to other businesses.

Meta, not one of the four, has been warmer than it has been about its commitment. Its VP of communications, Andy Stone, called the common signal "an important step," but SiliconANGLE reports the company has not committed to honoring it. The timing is awkward: Meta has pushed back to November 20 revised commercial terms that would let it crawl websites linked to its business accounts, after a backlash forced it to strip training from the listed uses.

With training removed but not expressly banned, Sam Quigley of the News/Media Alliance told the Mercury News last month, "It's still potentially on the table."

Opting out of AI data scraping: wishful thinking or a step in the right direction?

Much of the coverage has billed this as a way to opt out of being scraped outright, but it really isn't.

The draft is explicit: "This policy token is not a directive for crawling behavior." Bots can keep fetching every page, and the draft merely permits a crawler that gathers content only for training to skip pages it can't use. Stopping the crawlers themselves still takes the per-bot rules sites rely on today, which is why OpenAI, Google and Anthropic all stress that their existing controls aren't going anywhere. Microsoft puts it most plainly: "These settings govern use of content, not crawl access."

These further caveats matter at a time when the commitment itself is both voluntary and interpretations and pledges vary across the board even as certain players have higher exposure than others.

The New York Times sued OpenAI and Microsoft in December 2023, and both companies are invoking fair use in their defense. Anthropic, whose engineer wrote the draft, agreed to pay $1.5 billion to settle authors' claims over pirated books used in training, and Penske Media, owner of Rolling Stone, sued Google over AI Overviews in 2025.

It also relies on crawlers choosing to comply; the draft lets private agreements override it, and in 2025 Cloudflare accused Perplexity of disguising its crawlers to scrape sites that had blocked it, a charge the latter denied.

Microsoft's place in the coalition is arguably the most awkward: In 2024, its AI chief Mustafa Suleyman described open-web content as "freeware" for training, carving out only sites that explicitly object, which is roughly the exception this signal now formalizes. Unsealed filings in the Times case also show that a Microsoft director, Brent Hecht, described AI scraping in a 2023 internal memo as possibly "the largest theft of labor in human history."

The wider market, meanwhile, has been moving toward enforcement and payment rather than polite requests. Cloudflare has blocked AI crawlers by default for new customers since July 2025 and lets participating sites charge them per crawl. Google also says the shared token aligns with the EU's code of practice for general-purpose AI, which treats robots.txt as publishers' primary machine-readable opt-out, a sign that regulators, not just goodwill, are pushing the big labs to listen. For bots that ignore robots.txt entirely, we have rounded up heavier options, from login walls to IP rate limiting.

For now, one line that the biggest AI companies agree to read is a real step forward for honest crawlers and the people who run websites. But it is a request rather than a lock; it only looks forward, and it comes with none of the payment publishers have spent years fighting for.

• Google global head of data center energy says space data centers could cost up to 100 times more than terrestrial ones
• Amanda Peterson Corio's remarks came weeks after Google launched its first Suncatcher prototype, a roughly 1kW satellite that has since run Gemma 4 on an orbiting TPU, one chip at a time
• Corio says the real bottleneck on Earth is permitting rather than power, the same politics that backers of orbital compute claim they can escape

Google has spent the past month showing off AI chips working in orbit, but the executive responsible for powering its data centers back on Earth has just put a sobering price on the idea.

Speaking at The Information's recent WTF Summit in Napa Valley, Amanda Peterson Corio, Google's global head of data center energy, said data centers in space could cost up to 100 times as much as those built on the ground.

Her advice to the rest of the industry was simple and (pun intended) grounded: "Let's explore all the terrestrial options before we go extraterrestrial."

A very expensive proposition?

The remarks land at an awkward moment, because Google is testing exactly this kind of hardware right now. Its first Project Suncatcher prototype, built with satellite operator Planet, launched on SpaceX's Transporter-18 rideshare mission on October 1.

Google recently also announced its Hello World milestone: an onboard Gemma 4 model, running on a TPU, wrote a short haiku about being in space, which was downlinked as the satellite passed over Svalbard, Norway.

The company bills this as the first AI response from a data center-class TPU in orbit. That is a carefully measured claim, given that Starcloud put an Nvidia H100 into orbit a year ago and was running AI models on it within weeks.

Scale is the bigger caveat and one that can be easily overlooked amidst the hype: The satellite carries four TPUs but generates only about 1kW of solar power, so it can run just one chip at a time for roughly 20 minutes per session, according to The Next Web. A single 1 GW AI campus, the kind now under construction on the ground, draws about a million times that much power.

Google has described Suncatcher as a "moonshot" since it was unveiled a year ago, and project lead Travis Beals told NPR on launch day that he did not expect orbital compute to become cheaper than the terrestrial kind within the next five years.

Not the first or last skeptic of orbital datacenters

This did start a somewhat interesting conversation on Wall Street; Short seller Jim Chanos read Corio's statement as $5 trillion per gigawatt of compute, and posted on X that the gap between her view and the estimates of SpaceX management and bulls "should give pause to anyone trying to model this business out."

TIKR pushed the arithmetic further by examining recent claims: SpaceX's CFO told investors last month that new compute pays for itself in under a year at current rental rates. Built in orbit at up to 100 times the cost and rented at the same price, TIKR calculated that payback stretches to as long as a century, or about a decade even at a tenth of Corio's stated ceiling.

That matters because orbital compute underpins part of SpaceX's roughly $1.8 trillion market value at the time of writing. The company has asked the FCC for permission to launch up to a million data-center satellites, and Musk has posted that the orbital compute SpaceX deploys each month will probably overtake terrestrial compute around the end of 2028. Yet SpaceX's own IPO paperwork warned investors of "significant technical complexity and unproven technologies," and said its orbital data centers might never become commercially viable.

Corio is hardly the first skeptic. OpenAI's Sam Altman called the idea "ridiculous" under current conditions earlier this year, even as Nvidia lined up a Space-1 Vera Rubin module aimed squarely at orbital data centers.

This is an extremely interesting take from Google's end. Alphabet is one of SpaceX's largest outside shareholders, with a stake disclosed at 6.11% at the end of 2025, according to Bloomberg, and Google has reportedly been in talks with SpaceX about launching future Suncatcher hardware. Its energy chief has questioned and potentially poured cold water on a thesis that props up part of an asset Alphabet owns, while Google's own moonshot depends on cheap SpaceX launches arriving on schedule.

Corio's broader argument, somewhat ironically, was about the ground rather than the sky. According to The Information, she said the US has enough power and that the real bottlenecks are permitting and grid interconnection studies.

She blamed much of the local backlash against data centers on newer, less experienced developers, without naming any, and said Google prepays for grid upgrades over 15 to 20 years in more than 9 states, posts restoration bonds, and has stopped using nondisclosure agreements with local governments.

There is, however, a problem with this argument: Established firms such as Oracle have also faced local resistance, which complicates the idea that inexperienced newcomers are the problem. The backlash is anything but marginal: opposition has spread across the US, and Data Center Watch counted 75 projects worth about $130 billion delayed or blocked in the first quarter of 2026 alone.

There is a political angle built around these arguments that keeps orbital data centers in play, and SpaceX is essentially building on that promise. After this week's midterm elections, in which data center construction was a campaign issue in several races, including the contest for Texas governor, Musk posted, "Data centers in space looking pretty appealing right now."

Leaving the planet does not, however, mean leaving the permit queue; SpaceX's million-satellite request and Blue Origin's 51,600-satellite Project Sunrise both need FCC approval, and debris and spectrum concerns give regulators and rivals plenty to object to, just at a higher altitude. If power really is plentiful and permission is the bottleneck, as Corio argues, then orbit's main selling point is jumping that queue, and by her numbers it is a very expensive way to do it.

For now, until someone flies more than a kilowatt-scale test bed and publishes what it actually costs, the multiple between orbit and the ground, whether it turns out to be 3x, 10x, or as much as 100x, remains an educated guess. Corio's stated ceiling is arguably as unproven as Musk's timetable. Only one of those numbers, however, is already priced in considerably into a stock's current valuation, and SpaceX investors will be waiting for financials to back its CEO's claims with bated breath to say the least.

• Raspberry Pi has raised the 16GB Pi 5's price by $80 to $385, more than triple its $120 launch price
• Company CEO blames lack of alternative 16GB DRAM supply, compounded by wider component shortages
• Raspberry Pi called the hikes temporary, but memory makers expect tight supply to last until 2028 to 2030

Raspberry Pi has raised the price of its 16GB Raspberry Pi 5 yet again, and the flagship board now sits far from the brand's budget roots.

In a blog post, company CEO Eben Upton confirmed an $80 increase that takes the board to $385, which is more than triple the $120 it cost at launch in January 2025.

The 16GB variants of Compute Module 5 increase by $40 in the same round, while the Raspberry Pi 500+, the company's keyboard computer that also offers 16GB of RAM, remains at its existing price.

A far-reaching memory crisis dictates prices

The threefold figure is cumulative rather than a single jump: The November 2026 increase on its own amounts to roughly 26%, but the rest built up over a run of rises that began in December 2025, when the 16GB Pi 5 went from $120 to $145. February added $60, leaving the board 70% more expensive than at launch, and April added a further $100.

That April 2026 round also brought a new 3GB Raspberry Pi 4, introduced to give buyers a cheaper middle option, with Upton saying at the time that the LPDDR4 memory used in the Pi 4 and Pi 5 had risen sevenfold in price over the previous year.

November's change also comes barely a month after Raspberry Pi raised prices on its 2GB boards by $12.50 each, having decided that shielding those models from earlier price increases was no longer sustainable.

Upton's explanation is more specific than a general complaint about RAM prices. He blamed "a lack of alternative supply for 16GB DRAM, compounded by other ongoing component shortages," which suggests the bottleneck is sourcing at that particular density rather than cost alone.

Raspberry Pi, for its part, has been qualifying additional memory suppliers through 2026 to widen its options, and the wording implies those alternatives remain thinnest at the top of the range.

Interestingly, despite the price increases, demand hasn't fallen but rather grown considerably. Direct sales of the Pi 5 rose 47% in the half, and Raspberry Pi says demand for the 8GB variants continued to grow despite higher prices.

For buyers, the new pricing produces some odd comparisons. The 8GB Raspberry Pi 5 lists at $175, which means the extra 8GB on the 16GB model now costs $210, more than an entire 8GB board.

The spared-but-similarly-specced Pi 500+ sharpens the point. For $25 more than the bare 16GB board, it pairs the same processor and memory with a 256GB NVMe SSD and a mechanical keyboard. It's no substitute for embedded projects that need the bare board's footprint, but for desktop use, the standalone 16GB Pi 5 is now difficult to justify once one considers the cost of a standalone SSD alone, even if the keyboard is considered a non-factor.

Buying a cheaper board and upgrading it isn't a workaround either, since Raspberry Pi firmware locks boards to the memory capacity they shipped with. For those who don't need 16GB at all, we have found that an 8GB starter kit can often be a better buy than the board on its own.

Upton's own advice is to avoid paying for memory you don't need by choosing lower-density variants or older models, or by optimizing applications. He also warned that some sellers are exploiting the shortage by inflating prices or misrepresenting products, and urged buyers to stick to approved resellers.

A "temporary" crisis might be an understatement

Raspberry Pi continues to describe the memory squeeze as temporary and says it intends to roll back its increases once costs fall. Its own definition of temporary is fairly elastic, though.

In the company's September results, Upton described the current cycle as "a large, but ultimately typical, instance of a phenomenon that periodically affects the industry," with meaningful new supply not expected until 2028.

Memory makers are gloomier still. SK Hynix titled a section of its third-quarter results "Memory supply shortage to persist through 2030," and Micron has warned that shortages could run into 2028 or beyond.

TrendForce, whose latest outlook Upton linked in his post, expects DRAM suppliers to keep prioritizing server products through the fourth quarter, leaving the wider market undersupplied even as contract prices rise slowly.

The entry-level boards, however, remain a bright spot, with the 1GB Pi 4 and Pi 5 still at $35 and $45, respectively. Rising prices risk putting the Pi out of reach of schools and kids, and for now the bottom of the range is still holding the line. At the top, though, the 16GB Pi 5 has become a premium product by necessity rather than design, and the CEO's optimism might not be shared by many consumers in the market for one in the near future.

• Sam Altman met about 50 Samsung executives, including its chip chief, then lunched with SK's Chey
• Samsung billed the visit as an executive AI seminar, and neither company announced a new memory deal
• Samsung's 2027 HBM is sold out, and SK hynix is booked through 2029, while Microsoft and OpenAI both report demand outrunning capacity

OpenAI CEO Sam Altman recently spent time in a room with around 50 of Samsung's most senior executives, including the man who runs its chip business.

Officially, the session at Samsung's Human Resources Development Center in Yongin, just south of Seoul, was a seminar on how AI will reshape management.

Investors on Wall Street and Bloomberg saw it differently, however, with the latter describing it as an effort to deepen cooperation on memory chips and AI infrastructure amid intensifying global shortages.

A training day with a chip chief in the room?

While both narratives can be true at the same time, the timing is arguably intriguing, coming days after Altman complained that the world's memory supply is badly allocated, and at a moment when nearly every big AI player, OpenAI included, is running into hard limits on how much memory, compute, and money it can get hold of.

The Korea Herald claims the meeting was held under the theme "The Next 10 Years Shaped by AI: The Enterprise's Choice" and was the second intensive AI training program Samsung has held for its top leadership in 2026. Executives from its electronics, finance, and biotechnology affiliates went through hands-on generative AI training before Altman joined them.

The guest list, though, was interestingly not purely management; SBS reported that attendees included Vice Chairman Jun Young-hyun, who heads Samsung's Device Solutions chip division, and Song Yong-ho, head of that division's AI Center. Altman brought four OpenAI executives with him, among them Richard Ho, the company's vice president of hardware.

Samsung's statement only stated that the two sides "discussed the direction of AI technology development and its potential applications," and announced no new memory supply agreement.

Altman's own message to the room was about using AI rather than buying chips. "It is difficult to secure a sustainable competitive advantage simply by adopting AI," he said, according to a Seoul Economic Daily report.

He also handed out OpenAI's $230 Codex Micro, a 13-key keypad built with Work Louder to control the company's Codex coding agent, with Song among the recipients, marking an unusual gift for a company that makes some of the most advanced memory on the planet.

Samsung was also not Altman's only stopover: He also had lunch in Seoul with SK Group Chairman Chey Tae-won and senior executives from SK hynix and SK Telecom before leaving the country that afternoon, according to The Herald Business, which noted that the details of their discussion were not disclosed.

In a nutshell, Sam Altman met both of Korea's big memory makers within hours of each other, with his hardware chief in tow, and left without announcing anything. A Korea Herald analysis notes that neither side confirmed any discussion of chip pricing or investment commitments. A push for more RAM is a reasonable reading of the weekend but not a forgone conclusion just yet.

What makes that reading plausible is that capacity, rather than demand, has become the binding constraint on AI. Altman said as much on Stripe co-founder John Collison's Cheeky Pint podcast, released two days before the trip. Alongside his now-notorious claim that it is "quite terrible" for so much of the world's DRAM to sit in personal computers, he described a "very supply-constrained environment" and predicted that memory makers were each "going to make like $100 billion of free cash flow in a year."

The suppliers' own numbers back him up: Samsung said last week that its 2027 HBM supply is already sold out, while SK hynix says its DRAM and HBM output is effectively spoken for through 2029. "Customers are requesting more than twice the volume we can supply," is what SK hynix's head of HBM sales and marketing, Kim Ki-tae, told analysts.

The two firms also control most of that supply. TrendForce data gives Samsung 39.4% and SK hynix 24.9% of second-quarter DRAM revenue, roughly 64%, even as Counterpoint Research figures cited by Tom's Hardware put their share of AI-specific HBM revenue at an even higher 83%.

The squeeze is visible further down the chain, too. Microsoft said last week that demand for its AI infrastructure is outpacing available capacity, and that spending will pick up as delayed third-party capacity comes online. Roughly three-quarters of its $45.2 billion in quarterly capex went to short-lived assets such as GPUs and CPUs, a bill swollen considerably by rising memory and storage costs.

Capacity alone does not fix AI's problems

AI providers are rationing what they already have: In September, OpenAI paused new sign-ups for its most expensive ChatGPT Pro tier after demand for its Astra model outstripped capacity, Fortune reported, with OpenAI's Thibault Sottiaux saying the company was "pulling all the levers possible to sustain the demand."

More recently, Elon Musk said xAI had cut Grok Bot's usage limits because "the cost of serving Grok Bot was unsustainably high relative to revenue." That is a cost problem rather than a capacity one on its face, but the two are linked when the memory inside every accelerator is being sold at sellers' prices.

All of this combined turns securing more memory into as much a financing exercise as a diplomatic one. As we reported in August 2026, Samsung is moving 60 to 70% of its capacity onto five-year supply agreements, with price floors on mainstream products and substantial advance payments required as a contractual term.

OpenAI's existing paperwork with Korea is also softer than it looked a year ago. In October 2025, Samsung's letter of intent projected that OpenAI's memory demand could reach 900,000 DRAM wafers per month, a figure the Korea Herald described as a production target rather than a binding order.

When OpenAI Korea's chief hinted at deeper chip cooperation in September, an OpenAI spokesperson framed any expansion as happening "in the context of our existing relationship," pointing back to that same letter.

There is also the question of where all the extra hardware will be housed; Stargate is roughly halfway to its goal of adding 10 GW of US AI compute by 2029, and OpenAI is reportedly considering rebranding its largest data centers as "AI compute campuses" to soften a growing local backlash, according to Axios.

Altman is hardly the only customer knocking. AMD CEO Lisa Su made her own trip to Seoul in early October 2026 for back-to-back HBM talks with Samsung and SK hynix leadership, stressing partnerships of three to five years or more, while SK's Chey has warned that "the current shortage could continue until 2030."

The real test for Sam Altman's perceived diplomacy is whether OpenAI can turn its Korean partnerships into firm supply contracts and funded projects. With Samsung's 2027 HBM already allocated and SK hynix booked through 2029, any new memory OpenAI secures is likely to arrive late, cost more, and come on the suppliers' terms, a relatively easy conclusion to come to given the lack of leverage in play here.

• Around four dozen US recon and intelligence Marines are joining NATO's Baltic Sentry alongside Finland's Coastal Brigade
• They were re-tasked from training rotations already in Europe and will feed drone and sensor data into NATO's shared picture
• It is the Corps' second Baltic Sentry stint, after roughly 40 Marines deployed to Finland in February 2025

A small team of US Marines is being redirected to the Baltic Sea to help NATO keep watch over the cables and pipelines on its seabed, some of the most exposed infrastructure in Europe.

Around four dozen Marines from 2nd Reconnaissance Battalion, 2nd Marine Division, and the II Marine Expeditionary Force Information Group will operate alongside Finland's Coastal Brigade in support of NATO's Enhanced Vigilance Activity Baltic Sentry.

Both units are based at Marine Corps Base Camp Lejeune in North Carolina, and Military Times reported a headcount of 48, in line with the recent release announcing the move at the start of October 2026.

Eyes on the water not on the helm this time

What the team brings is surveillance. The rotation will use small uncrewed aerial systems and littoral sensors to feed data directly into NATO's shared common operational picture.

Washington is also keen to stress who is in charge to avoid overstepping: Baltic Sentry is led by European allies under Allied Joint Force Command Brunssum and Allied Maritime Command, with the Marine Corps serving in a "supportive, enabling capacity."

Maj. Gen. Daniel Shipley, who commands US Marine Corps Forces, Europe and Africa, said working side by side with Finnish allies would build "a cohesive, lethal team to deter hybrid threats in the Baltic littoral."

The Finnish Navy's chief of operations, Capt. (N) Johan Tillander, said the Marines' experience, skills and systems were "perfectly suited for supplementing our situation picture."

This is not the first time recon Marines have taken on the task. In February 2025, roughly 40 Marines from the same battalion deployed to Finland with RQ-20 Puma drones, in what the Corps described as the first activity between Finland's armed forces and the Marine Corps since a US-Finland defense cooperation agreement took effect in September 2024. At the time, the Corps said its expertise in coastline environments made it "a logical partner" for the mission.

Baltic Sentry itself launched in January 2025, after a run of incidents put the region on edge. Two fiber-optic cables more than 100 nautical miles apart were severed on November 17 and 18, 2024, and on Christmas Day the Estlink 2 power cable between Finland and Estonia was damaged along with four telecom lines, according to a Baird Maritime timeline.

Evidence of Russian interest in cables is more than just an assumption: A Telegraph investigation covered by TechRadar Pro logged more than 500 incidents in which 17 Russian trawlers loitered over or traced cables linking Britain to Europe and the US. UK telecoms minister Ian Murray said Russia was "starting to use commercial proxies," and the UK has proposed replacing its 1885 cable law with penalties of up to 14 years' imprisonment.

In the Arctic, British, Norwegian and American forces disrupted a Russian exercise near Svalbard this spring involving vessels tied to GUGI, Russia's deep-sea research directorate, which was reportedly rehearsing with a device designed to cut cables while leaving little evidence behind. Norway's defense minister, Tore Sandvik, said the operation "sent a clear message to Russia that they cannot operate covertly," although reporting so far shows the device was rehearsed rather than used against a live cable.

The deployment lands amid sharpening tension with Moscow. Stars and Stripes reports that allies have accused Russia of repeated airspace violations, that Estonia has arrested suspects over an arson attack it says Russian special services ordered, and that Moscow has warned it could resort to nuclear weapons if NATO tried to isolate Kaliningrad. NATO Secretary General Mark Rutte dismissed the threat, saying "if anything happens, we are ready to counter it."

At the same time, Pentagon officials have reportedly asked Lithuania, Latvia and Estonia to tone down their public rhetoric on Russia, according to four people who spoke to Reuters, with policy chief Elbridge Colby among those delivering the message. One European official summed up the request as "Just don't be loud. Don't make noise." The Pentagon declined to discuss private conversations, and a review of US force posture in Europe could yet bring a sharp drawdown.

Read against that backdrop, a few dozen Marines re-tasked from an existing rotation, wrapped in language about European leadership, looks less like an escalation than a low-cost signal: Washington is still present in the Baltic, but quietly and in a supporting role. Whether that proves enough is only something that time will tell.

• French deputies voted 11 to 10 to tax content providers €0.055 for every gigabyte sent to French internet subscribers
• Backers claim €10 billion a year in taxes, but the government disputes both the figure and whether traffic can be measured fairly
• Five days later, the Assembly rejected the budget's whole revenue section 458 to 0, wiping out the tax with it

France's lower house had voted to charge content companies for every gigabyte they deliver to French internet users - but the measure lasted only five days before it was shot down unanimously.

Deputies rejected the revenue half of the 2027 budget by a vote of 458 to 0, LCP reported.

Every amendment adopted during three weeks of debate fell with it, including the per-gigabyte levy; this means that what now goes to the Senate is the government's original bill, which never contained it to begin with.

How a gigabyte would have become a tax bill

The amendment, tabled by Socialist deputy Laurent Lhardit, would have required content providers to pay €0.055 (roughly $0.06) per gigabyte delivered to customers of French internet providers.

It passed at the opening of the 9 am sitting on Friday, October 30, with just 35 of the National Assembly's 577 deputies taking part: 11 voted for, 10 against and 14 abstained, according to the official session transcript. Both the government and the finance committee's general rapporteur, Philippe Juvin, opposed it. Its authors said it would raise around €10 billion ($11.2 billion) a year.

The mechanics are spelled out in the amendment itself: Internet providers, along with any company selling internet access to households or businesses, would count the data flowing to their customers and attribute it to whoever sent it, under the supervision of telecoms regulator Arcep and France's tax authority.

Senders would be identified by company name and their Internet Assigned Numbers Authority (IANA) registration number, and no record of what the data contained or where it went could be demanded.

The move has not only found obvious opposition within French politicians after its surprise initial approval in a nearly empty chamber, but has found surprisingly vocal critics at home: VideoLAN, the nonprofit behind VLC, warned the levy would cost it more than €1 million a year, MacGeneration reports, though Lhardit replied that the caps would protect it. The founder of a national ISP, Free, Xavier Niel, dismissed the deputies on X as "ignorant incompetents."

OVHcloud founder Octave Klaba, whose company raised some server prices by up to 87% this autumn over memory costs, calculated that on a 100Gbit/s transit port costing around €6,300 a month, the tax on his assumed traffic would reach €1.02 million a month. He put OVHcloud's exposure at €23 million to €70 million a year.

Even the operators who would have done the counting were unenthusiastic. The French Telecoms Federation, which represents Orange, Bouygues Telecom and SFR, does not reject the idea that big platforms should contribute more to networks, according to Freenews, but it is worried about the technical and administrative load of a scheme whose revenue would go to the treasury rather than to networks, and would rather see commercial negotiations with the platforms.

The solemn vote on the revenue section was due on Tuesday, November 3, but it slipped because 758 amendments were still unexamined at midnight on Monday, NosParlementaires noted. It came on Wednesday evening instead, and no group voted for the amended text, essentially killing the entire proposal in its tracks before the French Senate got involved.

To reappear, the tax would need a senator to refile it and a chamber controlled by the center-right Les Républicains to back it, which Next considered unlikely.

The idea itself is not going away: The Socialists put a €0.10-per-gigabyte version in the program they presented this summer, ahead of next year's presidential election. There is, however, international pressure and considerations that would ensure this never actually happens, even as countries vie to attract the most AI infrastructure investments, and this could be the biggest dealbreaker for multiple players.

Beyond France, it runs against the tide. Brussels spent years fielding telecom operators' demands that Big Tech chip in for network upgrades, the so-called "fair share" fight, and in January the European Commission's Digital Networks Act proposed only a voluntary cooperation mechanism rather than mandatory network fees.

Then there is Washington: Donald Trump has long threatened tariffs on countries that tax US tech firms, and in June 2026 he warned of 100% tariffs on any country implementing a digital services tax on American companies. A per-gigabyte levy whose biggest payers would be Google, Meta, Amazon, and Netflix would be hard to present as anything other than that.

For now, the proposal stands killed; nothing has changed on anyone's phone or internet bill, and, seemingly, for the 2027 budget at least, nothing will. But the idea could show up in other forms in the near future.