Downloads have fallen from 20m in January to 8.3m in April, paid conversion is a fifth of ChatGPT’s, the $0.42-per-agency GSA deal is now stalled, and SpaceX has rented out the Memphis Colossus 1 cluster to Anthropic for $1.25bn a month.

SpaceX’s S-1 filed on Tuesday rests on an AI-revenue line Grok is no longer obviously delivering.

Grok is not selling in Washington, and on Thursday it became Wall Street’s problem. Reuters reported that Elon Musk’s xAI chatbot has failed to convert its September 2025 GSA OneGov agreement into the kind of federal-agency adoption that competitors OpenAI and Anthropic are now operationalising.

Only three days after SpaceX filed an S-1 prospectus in which the company’s AI-revenue line is positioned as the growth engine underwriting what would be the largest IPO in history.

The consumer-side numbers are sharper still. Grok downloads falling to about 8.3 million in April from a January high above 20 million.

Paid conversion, on the Reuters reporting, sits at roughly 0.174% of surveyed US consumers and workers in Q2 2026, against more than 6% who pay for ChatGPT.

The growth curve that powered Grok’s 2025 IPO-narrative contribution has reversed across the past four months.

The GSA OneGov agreement Musk signed in September is the part Washington-watchers have been tracking most closely. The $0.42-per-organisation 18-month deal, announced by the GSA in late September 2025, was designed to deliver Grok 4 and Grok 4 Fast to every federal agency at a token price.

Public Citizen has petitioned the OMB twice to suspend federal use of Grok over accuracy and bias concerns, citing prior outputs the group describes as racist, antisemitic and factually incorrect.

Senator Elizabeth Warren has separately pressed Defense Secretary Pete Hegseth over the Department of Defense granting Grok classified-system access despite NSA and GSA concerns.

The compute-side trade is the bit that gives the story its commercial sharpness. SpaceX has rented out the Memphis Colossus 1 data centre, the 220,000-Nvidia-GPU, 300-megawatt facility that was Grok’s primary training environment, to Anthropic for $1.25bn per month through May 2029.

The implication is mechanically straightforward: with Grok’s consumer demand falling, xAI has more compute than it needs, and selling that capacity to Anthropic, the lab whose Mythos model has been displacing Grok on the federal-agency procurement list, is the cleanest way to monetise the shortfall before the SpaceX IPO prices.

The financial picture inside the SpaceX S-1 makes the trade necessary. xAI losing $6.4bn from operations on $3.2bn of revenue in 2025, with revenue growth of about 22% well below the published rates at OpenAI, Anthropic and Google DeepMind.

The structural complication is that the Anthropic deal is xAI selling its own competitor the compute Grok was originally trained on.

Musk’s AI portfolio is ‘falling apart’ in part because the compute-monetisation trade signals to public-market buyers that the underlying product cannot generate enough demand to absorb the capacity Musk built for it.

SpaceX’s roadshow, which begins inside the next two weeks, will be the first formal test of whether institutional buyers are willing to underwrite the AI-revenue-line projection against the federal-stall and consumer-decline data Reuters has now laid out.

Musk’s broader portfolio fortnight has compounded the timing. xAI’s $420 tax-return commitment to employees has slipped past the promised payment window, and the Delaware court’s procedural rulings against him in the OpenAI litigation have continued to compound.

SpaceX itself has not, on the available reporting, addressed the Reuters article publicly. The prospectus does not break out Grok-specific revenue from the broader xAI line, leaving institutional buyers to interpret the federal-adoption stall against the headline AI-line growth figure.

The next visible proof point will be the S-1 amendment expected ahead of the roadshow launch, where any updated Grok-adoption disclosure would be the first formal signal of whether xAI is willing to put numbers behind the consumer and federal demand picture Reuters has now made the public story.

Mercedes-AMG just revealed its most powerful vehicle ever, and this time the GT 4-Door Coupe runs purely on electricity. Available in two versions, the GT 55 and the flagship GT 63, the new model swaps out any combustion engine for three axial flux motors that sit low and deliver instant force without any lag. The top version hits 1,153 horsepower when conditions line up, enough to push it from zero to 60 miles per hour in about two seconds. Even the milder GT 55 produces 805 horsepower and covers the same sprint in roughly 2.4 seconds. Both models share the same long, low body that stretches just over 200 inches from nose to tail, yet they weigh around 5,423 pounds thanks to a mix of aluminum, steel, and carbon-fiber pieces that keep everything stiff.

Power flows through an all-wheel-drive system that can instantly shift torque between the front and back axles. Up front, you get a single motor, which helps with launches and provides more grip when needed, while the back wheels are handled by two larger ones. These axial flux motors are extremely thin, allowing engineers to essentially position them right in the center of the vehicle for better balance. Plus, they produce far more constant power than earlier motors, so even sustained high-speed runs seem effortless. When the optional performance package is added, both versions of the car will reach 186 mph, but the most exciting aspect is how the car handles on a twisty road.

Sale

LEGO Speed Champions Mercedes-AMG G 63 & SL 63 Building Toys – Kids Pretend Play Toy Cars for Boys…

• 2 Mercedes toy cars for kids – LEGO Speed Champions Mercedes-AMG G 63 & Mercedes-AMG SL 63 vehicle playset for boys and girls ages 10 and up and…
• 2 driver minifigures – Each buildable vehicle toy comes with a driver minifigure wearing a Mercedes outfit so kids can role-play fast-paced races
• Authentic Mercedes design – Each collectible car model features design details from the real life versions, including front grilles, hoods, wheels…

There’s a 106-kwh battery pack sitting down on the floor, with 2,660 individual cells cooled by oil that runs directly around them. This ensures that temperatures remain extremely steady even after repeatedly driving the car. The 800-volt system allows you to take up to 600 kilowatts from a fast charger. This implies you can fill up to 80% in around 11 minutes. Official estimates range from 370 to 435 miles on the milder European test cycle, but since the American market will be slightly different, we’ll have to wait and see what it turns out to be.

Mercedes-AMG concentrated on getting the driving experience just perfect, so you won’t miss the old V8. They included a special mode called AMGFORCE Sport+, which, believe it or not, plays over 1,600 recorded sound effects from speakers both inside and outside the car to mimic the AMG GT R’s snarl and burble. They even incorporate seat vibrations to correspond with each simulated gear shift and throttle lift. You also get paddle shifters on the steering wheel, which allow you to manage a simulated nine-speed transmission that provides the appropriate amount of hesitation and power loss, much like a real automatic. The whole thing is so convincing that many people will swear up and down that there is an engine under the hood, but if you don’t care about the sound effects, you can just turn it off in Race mode and enjoy the pure electric rush.

On the center console, you’ll find three large rotary dials that allow you to choose how the automobile behaves. One allows you to adjust throttle sharpness, another controls how violently torque vectoring yanks or pushes the car through corners, and the third allows you to set traction slip on a scale of one to nine. The air suspension automatically adjusts the ride height based on speed, while active roll control maintains the body flat without the use of traditional sway bars. Then there are all those sophisticated aerodynamic gadgets under the car and on the rear spoiler that may add downforce or reduce drag as needed. Finally, all-carbon-ceramic brakes up front and super-sized steel discs in the rear prevent overheating after repeated stops.

Inside the cabin, you’ll find a spacious interior with four strongly cushioned seats and ample storage for golf bags or skis owing to the full width pass-through in the back. A large front trunk provides extra storage, which is always welcome. The center of the dash is dominated by a bank of screens, which includes a digital gauge cluster up front for the driver to monitor, a central display screen, and a separate screen for the passenger to keep them out of the loop. You’ll also discover some proper knobs for the AMG race mode controls that are conveniently situated just where you need them, allowing you to keep your eyes on the road – no fiddling while driving here.

Production of the GT 55 is slated to begin soon at the Sindelfingen facility in Germany, with the GT 55 arriving at US dealers later this year and the GT 63 following early in 2027. Prices will have to be confirmed by Mercedes-AMG, but we anticipate a range of $140k to $210k depending on the features you choose.

It is indeed a weird time to be an automaker, as US federal incentives disappear and support dwindles for newer electric-powered cars. “Manufacturers would really like to know what the future will be and what are the rules,” says Mike Finnern, the senior vice president and zero-emission fleet lead at WSP, a consulting firm. Guarantees of large, future orders from fleet managers like city governments, but also private businesses, “will help them be stable for a while.”

EVs are a nice fit for government fleets, Finnern says. Surveys suggest that regular car buyers are still plenty apprehensive about shifting to a plug-in from gas cars they’re used to, and they want cars with even longer ranges, even if they seldom use the whole battery. But governments know exactly how their vehicles are used, can more precisely control charging, and are able to see that today’s ranges of 250 to 400 miles per charge fit their needs fine. Plus, EVs might help governments save money on fueling and maintenance. Private operators like Amazon aren’t stopping their forays into EVs, and “they wouldn’t do it if it didn’t pencil out,” he says.

“I regret every electric and hybrid vehicle we haven’t bought yet,” says Kerman. “It would’ve shielded us from the doubling of fuel costs that we’re now enduring.” By partnering with the US Department of Transportation, his agency has found that switching to battery electrics improves New York City’s vehicle energy economy by 6 percent.

Still, both governments say they have plenty to learn about how and where EVs fit best and that the partnership will help them share and create best practices so that other cities might eventually follow.

One big takeaway from the government’s experience so far is that officials need to be proactive and mindful about getting city workers on board. There are technical challenges—maintenance workers need to be retrained to maintain EVs instead of gas-powered vehicles, and everyone needs to remember to plug them in—and trickier morale ones, too.

Workers don’t always appreciate sudden changes. And while New York’s data suggests that the intelligent speed assistance built into many of its new EVs reduces speeding and possibly crash severity in city vehicles, employees have lingering worries about workplace surveillance. (In March, the city workers’ union reached an agreement outlining how data collected from city vehicles might be used in disciplinary actions.)

A workforce that’s enthusiastic about EVs can make all the difference. “We’ve seen some deployments be really successful and some, not so much. They have the exact same problems, but some were able to overcome them because their people were excited about it and trained,” Finnern says.

Courtesy of California Internal Services Department

Haynes, who used to work with Kerman in New York before moving to Los Angeles, recalls that he was once an EV skeptic but changed his mind once Kerman coaxed him into trying out a Tesla. It was, above all, fun.

“I will tell you, no one goes into these electric cars, walks out and says, ‘I hate this car,’” Kerman says. “They all say, ‘I love the car.”

Chinese researchers claim a new solid-state battery can survive ultra-fast charging while delivering dramatically higher energy density, potentially reshaping the future of electric vehicles. Researchers at the Chinese Academy of Sciences claim they have developed a new solid-state lithium-metal battery capable of delivering extremely high energy density while surviving ultra-fast charging conditions – a combination the global EV industry has been chasing for years.

According to the research paper published in the Journal of the American Chemical Society, the prototype battery achieved an energy density of 451.5 Wh/kg while maintaining stable cycling performance for 700 charge cycles under a 20C charging rate. In practical terms, that theoretically translates to charging and discharging in roughly three minutes.

If commercialized successfully, the technology could represent a major leap over today’s EV batteries. Most current mass-market electric vehicles from US and European automakers still operate within relatively conservative fast-charging limits. Brands like Tesla, Ford Motor Company, Volkswagen, and Mercedes-Benz Group generally peak between 150kW and 350kW charging speeds under ideal conditions, with many vehicles still requiring 20 to 40 minutes for meaningful charging sessions.

Meanwhile, Chinese automakers and battery firms are rapidly accelerating the development of ultra-fast charging technologies. Companies like BYD, CATL, Ganfeng Lithium, and multiple startups are aggressively pursuing solid-state battery architectures capable of much higher charging speeds and energy density.

China’s battery push is reshaping the industry

The latest breakthrough also arrives as Western automakers increasingly deepen partnerships with Chinese companies to remain competitive in EV technology. Earlier this month, Stellantis expanded its collaboration with Chinese automaker Dongfeng Motor Corporation through a €1.17 billion agreement covering vehicle production, exports, and engineering cooperation. The company has also strengthened ties with Leapmotor to jointly develop electric vehicles for European markets.

Other global manufacturers are making similar moves. Volkswagen has partnered with Chinese EV startups, including Xpeng, while several Japanese and European brands are exploring shared manufacturing and battery development projects with Chinese suppliers.

As Chinese firms continue achieving breakthroughs in battery chemistry and manufacturing scale, these partnerships may allow Western companies to indirectly benefit from China’s rapid technological progress.

High energy density still comes with risks

Despite the excitement, ultra-dense battery chemistries also raise safety concerns. Higher energy density often means greater thermal risk if a battery enters thermal runaway. The industry has already witnessed several high-profile EV fire incidents involving lithium battery systems, including scrutiny surrounding some earlier-generation BYD battery discussions and broader concerns over EV heat management globally.

The Chinese researchers claim their pouch cell passed nail-penetration safety testing, which is often used to evaluate internal short-circuit resistance. However, laboratory results do not automatically guarantee real-world automotive reliability.

That remains one of the biggest caveats surrounding solid-state batteries. While breakthroughs are announced frequently, commercialization can take years due to manufacturing complexity, durability validation, safety certification, and government regulatory testing.

Many battery companies are currently targeting commercialization windows between 2026 and 2028. Until then, traditional lithium iron phosphate (LFP) batteries are likely to remain dominant due to their lower cost, established supply chains, and proven reliability.

Still, the pace of development suggests the EV battery race is entering a far more aggressive phase – and China currently appears to be leading it.

The competition was judged by UCD’s Margaret Kelleher, UCD’s Karlin Lillington and Anne Mulvihill, the sister of Mary Mulvihill.

Cian Morgan, a medical student studying at Trinity College Dublin (TCD) is the 2026 winner of the Mary Mulvihill Award, the science media competition for third-level students that commemorates the late science journalist and author Mary Mulvihill. This year’s theme was on the subject of time and how it is an aspect of our existence that, while difficult to define, deeply pervades our lives and experiences. 

Morgan received the award and a cash prize of €2,000 at a ceremony held at the Dublin Institute for Advanced Studies, while TCD physics student Aoibheann Kearins and Ciaran Lynch, who is studying for a BA in Music and Film at University College Dublin, were highly commended and each received a cash prize of €500.

Morgan’s entry, ‘The Cows of Carlow: A Conversation with My Grandad’, is an essay inspired by his own and his grandfather’s personal and historical reflections on the topic. 

He wrote about Dublin Mean Time, which is Ireland’s national standard time, established in 1880 and was 25 minutes and 21 seconds behind Greenwich Mean Time (GMT). He also wrote of the Time Ball on the roof of the Ballast Office at Aston Quay, Dublin, which was dropped down a pole every day at precisely 1.00pm, to allow sailors on the Liffey to calibrate their marine chronometers. 

Morgan said, “Meanwhile in Tullow, my great-great-grandfather’s hometown in Co Carlow, there was no such sophisticated community timepiece. And so possession of a personal timepiece conferred considerable social status. Yet most people ordered their day around a much looser conception of time, far removed from our current anxious preoccupation with minutes and seconds and even the cows seemed to know what was the ‘right time’.”

Commenting on the essay, judge and UCD professor of Anglo-Irish literature and drama, Margaret Kelleher said, “I really liked it and found it really informative. Cian’s entry has many of the fine qualities of Mary’s work: it conveys substantial information in a way that is very accessible and engaging and is very well researched.”

Kearins is the second person in her family to feature among the prize winners, as her sister Aoife, a TCD graduate also received the highly commended award in 2020 and is now pursuing a PhD on the history of mathematics at the University of Oxford. 

Aoibheann’s piece called ‘Time for you, Time for me’, explores her personal experiences of time over the course of her life, as well as the scientific and philosophical conceptions of time, covering Aristotle’s idea of potentiality and Einstein’s Special Theory of Relativity, which “shows that time is not universal”. 

Lynch’s entry, ‘Timeless’, is an original musical composition, divided into three parts, with iterations and motifs to represent the past, the present and the future. The main melody is played on a grand piano, but Lynch also employs a wide range of percussion instruments to mark time and to introduce dramatic new possibilities to the piece. 

“The theme for this year’s award was ‘Time’, an appropriate topic given that the award is marking ten years and the Award’s committee is wondering where time goes,” said Anne Mulvihill, Mary’s sister and a member of the judging panel.

She added, “It was also appropriate given that in many ways Mary was ahead of her time, pioneering science communication. Once again the judges were impressed and delighted with the wide range of entries on the subject and the winning entries strongly indicate that her legacy has lasted over time.”

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.

For months, scammers have been taking advantage of a loophole that allows them to send spammy emails from an internal Microsoft email address typically used for sending legitimate account alerts.

It’s not clear how the scammers are abusing the system, but they have been able to set up new Microsoft accounts as if they are new customers, and use that access to send out emails purportedly from the tech giant itself, potentially tricking people into thinking that these emails may be genuine.

Microsoft doesn’t yet appear to have gotten a handle on the issue.

Last week, I received several, similarly structured emails containing subject lines and web links to scammy sites from Microsoft across different email accounts. These crudely made emails were sent from msonlineservicesteam@microsoftonline.com, an email account that Microsoft uses to send important notifications to users, such as two-factor authentication codes and other critical alerts about their online account.

Some of these emails’ subject lines resembled official emails that would alert users to fraudulent transactions, while other emails claimed to have a private messaging waiting for the recipient at a web address mentioned in the email body.

In a social post on Tuesday, anti-spam non-profit, The Spamhaus Project, said it had also seen Microsoft’s account notification email address being abused to send spam, and that the activity dated back “several months.”

“Automated notification systems should not allow this level of customization,” wrote Spamhaus. The non-profit added that it has notified Microsoft of the issue.

When contacted by TechCrunch earlier this week, a Microsoft spokesperson acknowledged our inquiry, but has not yet commented or said if the company has stopped the abuse of its account notification email.

This is the latest in a rash of incidents in which hackers or scammers have abused company systems to trick unsuspecting customers in recent months. Earlier this year, hackers broke into a platform used by fintech firm Betterment to send out fraudulent notifications that purported to triple the value of any crypto users send in — a widely known scam used to steal people’s cryptocurrency.

Back in 2023, hackers similarly abused access to an email account run by Namecheap to send out phishing emails aimed at stealing people’s credentials.

Other users commenting on social media say that other companies’ email addresses are also being used to send out spam, suggesting the issue is not limited to Microsoft.

Flipper Devices, the maker of the Flipper Zero pentesting tool, is asking the community to help build Flipper One, an open Linux platform for connected devices.

Unlike Flipper Zero, which focuses on offline access control and radio technologies such as NFC, RFID, infrared, and sub-GHz communications, the Flipper One project is designed as a high-performance, Linux-based platform for networking and hardware experimentation, with sufficient processing power to support SDR (software-defined radio) analysis and local LLMs.

The company underlines that One, which is a portable ARM Linux computer, shouldn’t be seen as an upgrade to Flipper Zero, but rather “a completely different project with its own goals.”

Hardware-wise, Flipper One is built around the Rockchip RK3576 ARM SoC with 8 GB RAM, paired with a Raspberry Pi RP2350 microcontroller in a dual-processor architecture.

The main CPU handles Linux workloads while the MCU independently manages the display, power subsystem, buttons, and boot process. This also practically means that the device remains operational even when the OS is powered off.

Flipper One is also intended to be modular, with support for M.2 and GPIO interfaces, as well as for PCIe, USB 3.1, SATA, UART, I2C, and SIM. This allows adding SDRs, SSDs, Wi-Fi cards, AI accelerators, and 5G or NTN satellite modems.

“You can use Flipper One as a router, a VPN gateway, or a bridge between wired and wireless networks,” Flipper Device says in the announcement today. However, the device could also work as a portable Linux workstation (“survival desktop”), TV media box, and HDMI support.

Call for help and participation

Flipper One has been under development for years, but the project turned out to be a much more difficult challenge than the vendor expected, with several teams working on various aspects of the project: hardware, mechanics, software development for the RK3576 processor, MCU firmware, user interface, documentation, and testing.

“It’s an incredibly hard project, both economically and technically,” Flipper Devices says, adding that anyone can pitch in. “Whether you’re an engineer, software developer, designer, or simply an enthusiastic user with ideas to share, you’re welcome to participate in development and help shape Flipper One.”

The major hurdles the team faces right now are:

• Achieving full mainline Linux support for the RK3576 SoC and removing remaining proprietary components and vendor dependencies.
• Developing and upstreaming the custom dual-processor CPU/MCU architecture and its interconnect drivers.
• Building Flipper OS and the FlipCTL framework to create a new small-screen Linux user experience.
• Resolving hardware compatibility issues involving USB-C DisplayPort Alt Mode, H.264/HEVC hardware encoding, and Wi-Fi analysis features.
• Supporting advanced capabilities like satellite connectivity and offline AI through unfinished software support and external partnerships.

“The current state of ARM Linux is depressing. Every vendor bolts on their own custom mess: closed boot blobs, vendor-specific patches, “board support packages” that nobody outside the chip maker can really understand,” explains Flipper Devices.

Currently, Collabora is helping the project to add full support for the Rockchip RK3576 SoC into the mainline Linux kernel, which Flipper Devices says is progressing well.

Flipper One is an active development project, far from a finished or shipping product. The prototypes Flipper Devices is working on have unfinished parts, some core software support is missing, and architectural decisions remain unresolved.

“There’s a lot of uncertainty in this project, along with technical challenges and financial risks (like the current RAM chip crisis),” stated Pavel Zhovner, Flipper Device founder, adding that the company will try its best to deliver the product.

Updates on the project’s progress will be periodically shared via Flipper’s R&D profile on social media.

Automated pentesting tools deliver real value, but they were built to answer one question: can an attacker move through the network? They were not built to test whether your controls block threats, your detection rules fire, or your cloud configs hold.

This guide covers the 6 surfaces you actually need to validate.

Advertisement

Download Now

Less than a week after completing the largest tech IPO of 2026, Cerebras Systems is making its most aggressive play yet to dominate the fast-growing AI inference market. On Monday, the Sunnyvale-based chipmaker announced that it is now running Kimi K2.6 — a trillion-parameter open-weight model developed by Beijing-based Moonshot AI — for enterprise customers at nearly 1,000 tokens per second, a speed no GPU-based provider has come close to matching.

The result, independently verified by benchmarking firm Artificial Analysis, clocked in at 981 output tokens per second, making Cerebras 6.7 times faster than the next-fastest GPU-based cloud provider and 23 times faster than the median. For a standard agentic coding request involving 10,000 input tokens, Cerebras delivered the full response — including prompt processing, reasoning, and 500 output tokens — in 5.6 seconds, compared to 163.7 seconds on the official Kimi endpoint. That’s a 29-fold improvement in time to final answer.

“We’re really wanting to be very clear and show that we can do the largest models,” James Wang, Cerebras’ director of product marketing, told VentureBeat in an exclusive interview ahead of the announcement. “In this case, Kimi K2.6 — a trillion-parameter MoE model on the wafer-scale architecture — and it runs also at this same incredible speed that we’re famous for.”

The announcement marks a critical inflection point for Cerebras, which has long battled a perception that its unorthodox wafer-scale chips, while blindingly fast, could only handle small and mid-sized models. Kimi K2.6 is the first trillion-parameter open-weight model the company has ever served in production. And with a freshly minted $95 billion market cap and $5.55 billion in IPO proceeds burning a hole in its balance sheet, Cerebras is signaling to Wall Street that it intends to compete not just at the frontier of speed, but at the frontier of model scale.

Why Cerebras chose a Chinese-built model as its trillion-parameter flagship

The choice of Kimi K2.6 reflects both a technical milestone and a commercial calculus. Released on April 20 by Moonshot AI — a Beijing-based company founded in 2023 by Tsinghua University alumni and dubbed one of China’s “AI Tiger” companies — K2.6 is a trillion-parameter Mixture-of-Experts model that has rapidly established itself as the most capable open-weight model available for coding and agentic tasks. The model tops SWE-Bench Pro at 58.6, outperforming Claude Opus 4.6 and matching GPT-5.4, while posting leading scores on agentic benchmarks like Humanity’s Last Exam and DeepSearchQA. Its architecture uses 32 billion activated parameters per token out of a total of 1 trillion, with 384 experts, of which 8 are selected plus 1 shared per forward pass, operating over a 256,000-token context window.

In practical terms, K2.6 is one of the first open-weight models that enterprises can plausibly use as a drop-in replacement for expensive, capacity-constrained closed-source APIs from Anthropic and OpenAI — particularly for the coding and agentic workloads that have become the highest-value application of large language models. The version 2.6 release extends K2.6’s capabilities from front-end design into full-stack workflows, including authentication, database operations, and long-horizon agent execution.

Wang was blunt about what is driving enterprise interest. “They’re very motivated, first of all, to have an alternative to Anthropic,” he told VentureBeat. “Anthropic’s models are fantastic. I use them. I’m sure you probably use them. But they’re quite expensive, and they’re constantly running out of capacity.” He described a personal experience in which an application running on Anthropic’s API failed over a weekend because it ran out of capacity — an anecdote that, he said, resonates deeply with enterprise buyers.

The geopolitical dimension of this arrangement is worth noting, however. Kimi K2.6 is a Chinese-developed model being served by an American chipmaker to American enterprise customers. Moonshot AI operates out of Beijing, and K2.6’s adoption in the West arrives during a period of heightened scrutiny of Chinese AI companies in the U.S. market. Enterprise buyers with strict compliance requirements — particularly those in financial services, healthcare, and defense — will need to evaluate this dimension alongside the model’s technical capabilities.

How wafer-scale chips solve the trillion-parameter speed problem that GPUs cannot

Understanding why Cerebras can achieve these speeds requires understanding what makes its hardware fundamentally different from anything else on the market. Most AI inference today runs on clusters of Nvidia GPUs — typically organized in racks of 72 GPUs, what Nvidia markets as the NVL72 configuration. In these setups, the model’s parameters are distributed across many discrete chips connected by high-speed networking fabric. Data must constantly shuttle between chips, and the interconnect bandwidth between GPUs becomes a bottleneck, particularly for large models with hundreds of billions or trillions of parameters.

Cerebras takes a radically different approach. Its Wafer-Scale Engine 3 is a single chip the size of an entire silicon wafer — roughly the size of a dinner plate — containing 44 gigabytes of on-chip SRAM. Unlike the high-bandwidth memory used in GPUs, SRAM sits directly on the processor die, offering dramatically lower latency and higher bandwidth for data access. For Kimi K2.6, Cerebras stores the model’s weights in their original 4-bit precision while performing computation at 16-bit floating point. The weights are distributed across multiple wafers in a cluster of approximately 20 CS-3 systems, with activations streamed between them. Critically, all the experts for a given MoE layer are placed on the same wafer, meaning the all-to-all communication required for expert routing happens at SRAM speeds. According to Cerebras’ technical description, the on-wafer network fabric delivers over 200 times the bandwidth of NVLink on NVL72.

Wang explained the architecture using an analogy. “Our single units are much larger and much higher capacity — they’re on the order of 20 racks, as opposed to 72 GPUs,” he said. Each layer in the transformer can, in effect, serve a separate user simultaneously. “They’re just like a queue, like you’re queuing for bagels or something — they’re all occupying a different part of the hardware. But because they move across so fast, the actual experience, tokens per second, single user, on your end is still what you’re used to.” Combined with custom kernels and speculative decoding, this allows Cerebras to serve the trillion-parameter MoE model at close to 1,000 tokens per second — a speed the company calls a world record achievable only with wafer-scale hardware.

Fortune 500 companies are already testing Cerebras’ trillion-parameter inference in production

Cerebras is not opening K2.6 to the general public. Instead, the company is positioning this as an enterprise-first offering, with Fortune 500 companies in software, financial services, and healthcare currently running cloud trials of their production workloads on the platform. “These are logos that you’ve definitely heard of,” Wang said, though he declined to identify specific customers due to confidentiality agreements.

The enterprise-first approach is deliberate. Cerebras has historically prioritized its largest customers over its consumer-facing API, in part because of hardware capacity constraints. “Everyone is in a capacity crunch. We prioritize our enterprise customers, so we don’t show it in the consumer-facing gateway or the API, where you get very unpredictable traffic, where a single user can, in effect, take over your whole cluster,” Wang explained. Serving K2.6 also limits the company’s ability to simultaneously offer other large models. “We can’t simultaneously, you know, have six other models,” he acknowledged. “It’s just kind of a mutual constraint of reality.”

On pricing, Wang said that while the enterprise deployment does not carry public pricing, the company’s costs are broadly competitive with GPU-based providers. “On all the models we have served with pricing, the pricing is very comparable — maybe in the middle, kind of middle-upper range of GPU pricing,” he said. “It’s not like, because we run fast, it costs many, many fold more.” He drew a line, however, at the lowest end of the market: if you are willing to run K2.6 at 20 tokens per second on bargain GPU infrastructure, Cerebras will not try to compete on price. “We’re an automaker in the pickup truck market. We don’t do that market,” Wang said. For speed-sensitive workloads — particularly agentic coding, where developers wait in real time for the model to generate and iterate on code — the value proposition is straightforward: comparable per-token cost, but an order of magnitude faster delivery.

The competitive threat from Nvidia’s $20 billion Groq acquisition looms large

Cerebras’ announcement arrives at a pivotal moment in the AI chip industry, one in which the inference market is rapidly overtaking training as the most commercially important compute workload. As AI agents proliferate in enterprise software, the speed of inference directly determines how useful those agents are in practice — and the competitive pressures are intensifying accordingly.

The most significant competitive development in recent months was Nvidia’s acquisition of Groq for $20 billion, a deal that gave the GPU giant access to proprietary inference technology built around specialized Language Processing Units. Wang referenced the deal directly. “I think Nvidia is now sensing fast inference is an extremely important market,” he told VentureBeat. “That’s why they’re willing to spend $20 billion on acquiring a company like that.”

But Wang expressed confidence that Cerebras’ architectural advantages are durable. Both Nvidia and Cerebras operate on roughly annual hardware refresh cycles. “We refresh our hardware on a periodic cycle. You will hear some news about that from us soon,” Wang said, hinting at a forthcoming hardware announcement without providing details. On the software side, Wang pointed to the company’s track record of rapidly adapting to the fast-evolving open-weight model ecosystem. “We started with Llama, we supported all the Qwen models, and then when developers told us they wanted GLM, we brought GLM online. And now they’re telling us Kimi is the best — so we’re giving them Kimi,” he said. “At the same time, we’ve also supported the best companies in running their closed models — OpenAI, Cognition, Mistral.”

The mention of OpenAI underscores one of the most unusual business relationships in the AI industry. OpenAI and Cerebras struck a deal in early 2026 reportedly worth more than $20 billion for computing capacity and related services. Wang confirmed that Cerebras serves OpenAI’s “internal coding models forthcoming” but declined to disclose specifics, as neither party has publicly detailed the technical arrangement.

Inside Cerebras’ plan to serve the smartest AI models faster than anyone else

Wang framed the K2.6 deployment as a stepping stone, not a destination. Cerebras started serving inference in late 2024 with relatively small models and has spent over a year scaling from 70 billion parameters to 1 trillion-plus. “We couldn’t have launched that in November 2024,” he said. “But we’re there now.”

The company’s next challenge is to move from serving the best open-weight frontier model to serving the best frontier models, period — including closed-source models from the likes of Anthropic and OpenAI that sit at the absolute top of the intelligence leaderboards. “This is the first open-weight frontier one that we now have clear demonstrated evidence for,” Wang said. “I think over the course of the year, you will see us serving true frontier, frontier at the speed that we’re famous for. And you should hold us up for that.”

When asked whether the current rollout would be overtaken by the pace of hardware improvement at Nvidia and others, Wang was unfazed. “Nvidia has a very clear roadmap. They publish every year at GTC. They’re roughly on a yearly product cycle, and so are we. You will hear some news about that from us soon,” he said, hinting at new hardware without offering details.

He also addressed the question of vendor lock-in — a concern that any CTO evaluating a single-vendor inference provider would raise. “These enterprises rarely commit fully to one vendor,” Wang said. “They have strategies to make sure that some traffic can go to us, some traffic can go to someone else, and there’s load balancing between the two. This is not a new problem. This is just generally how you manage cloud resources.”

The pitch, ultimately, is about more than speeds and feeds. Wang sees the AI industry converging on a world in which autonomous agents — not human developers — are the primary consumers of inference compute, and in which the speed of those agents determines competitive outcomes for the companies that deploy them. “The world economy is kind of getting rebuilt on agents,” Wang said. “Speed will determine who wins or loses.”

It is a bold claim from a company that, until last week, had never traded on a public exchange. But for Cerebras, the logic is straightforward: if the future of enterprise software is built by AI agents that think at the speed of their hardware, then the company that provides the fastest hardware provides the fastest thinking. And in a market where enterprises are spending billions to shave seconds off their AI response times, a company that can serve a trillion-parameter model in the time it takes to pour a cup of coffee might just have the most compelling pitch in Silicon Valley.

Looking for the most recent regular Connections answers? Click here for today’s Connections hints, as well as our daily answers and hints for The New York Times Mini Crossword, Wordle and Strands puzzles.

Today’s Connections: Sports Edition is pretty tricky, with some words that could fit in a couple of categories. If you’re struggling with the puzzle but still want to solve it, read on for hints and the answers.

Connections: Sports Edition is published by The Athletic, the subscription-based sports journalism site owned by The Times. It doesn’t appear in the NYT Games app, but it does in The Athletic’s own app. Or you can play it for free online.

Read more: NYT Connections: Sports Edition Puzzle Comes Out of Beta

Hints for today’s Connections: Sports Edition groups

Here are four hints for the groupings in today’s Connections: Sports Edition puzzle, ranked from the easiest yellow group to the tough (and sometimes bizarre) purple group.

Yellow group hint: Spotted at a game.

Green group hint: Charlie Brown’s baseball home.

Blue group hint: Snow sport.

Purple group hint: Queen City teams.

Answers for today’s Connections: Sports Edition groups

Yellow group: Seen at a college sporting event.

Green group: Pitching mound.

Blue group: Alpine skiing disciplines.

Purple group: Charlotte ____.

Read more: Wordle Cheat Sheet: Here Are the Most Popular Letters Used in English Words

What are today’s Connections: Sports Edition answers?

The yellow words in today’s Connections

The theme is seen at a college sporting event. The four answers are band, cheerleaders, dance team and student section.

The green words in today’s Connections

The theme is pitching mound. The four answers are bump, hill, mount and rubber.

The blue words in today’s Connections

The theme is Alpine skiing disciplines. The four answers are combined, downhill, slalom and Super-G.

The purple words in today’s Connections

The theme is Charlotte ____. The four answers are 49ers, FC, Hornets and North.

Toughest Connections: Sports Edition categories

The Connections: Sports Edition puzzle can be tough, but it really depends on which sports you know the most about. My husband aces anything having to do with Formula 1, my best friend is a hockey buff, and I can answer any question about Minnesota teams.

That said, it’s hard to pick the toughest Connections categories, but here are some I found exceptionally mind-blowing.

#1: Serie A Clubs. Answers: Atalanta, Juventus, Lazio, Roma.

#2: WNBA MVPs. Answers: Catchings, Delle Donne, Fowles and Stewart.

#3: Premier League team nicknames. Answers: Bees, Cherries, Foxes and Hammers.

#4: Homophones of NBA player names. Answers: Barns, Connect, Heart and Hero.

The classroom laptop fight just got a real-world stress test. Kansas City Public Schools has already bought more than 4,500 MacBook Neo units for students in 8th grade and up, putting Apple’s new low-cost Mac into schools at a scale that goes well beyond a pilot program.

The district plans to retire more than 30,000 existing devices over time. That gives Apple a visible education-sector win as cheaper classroom laptops become more competitive, and it gives school IT teams another reason to rethink the old Windows, Chromebook, and Mac divide.

Why is KCPS going all Apple

KCPS says the move is meant to simplify how students and teachers work across devices. Instead of supporting several platforms at once, the district is moving toward one Apple-based setup across classrooms.

The first wave gives the plan real scale. Older students are getting MacBook Neo laptops, while iPads and existing MacBook Airs are expected to cover other grade levels as the transition continues. That creates a cleaner device lineup for the district, though KCPS still has to show how well the approach works once more schools are folded into the rollout.

How Windows stays in the fight

The MacBook Neo gives Apple a lower-cost Mac for an education market where Chromebooks and budget Windows laptops have built a strong position. For school IT teams, that puts macOS into a price range where it can be compared more directly with cheaper classroom machines.

Intel is also trying to keep Windows PCs in the conversation. Recent reporting says its Project Firefly push is aimed at sub-$600 Windows laptops built around more standardized designs, with cheaper Macs, Chromebooks, and Arm-based machines all adding pressure. Schools also have to weigh repairability, ports, battery life, software support, and fleet management before committing to one platform.

KCPS gives Apple a live classroom test instead of a spec-sheet argument. Thousands of students will use these machines for daily assignments, and that is where battery claims, durability, app access, and support costs start to count.

What should schools watch next

A first shipment is easier than a multi-year fleet change. KCPS still has to manage the long replacement cycle, support teachers through the transition, and keep costs predictable as older machines leave service.

The strongest signal will be whether KCPS can control repair, training, and management costs as the replacement cycle expands. If it can, other districts may look at the MacBook Neo as a more realistic Chromebook and Windows alternative. If it can’t, cheaper classroom laptops will still have a simple argument to make.