AI and ML

And as a bonus, it doesn’t require data retention

Anthropic released its Opus 5 AI model on Friday, claiming that it “comes close to the frontier intelligence of Claude Fable 5 at half the price.” That’s welcome news for enterprises beset by rising token costs.

Fable 5 prices tokens at $10 / MTok input and $50 / MTok output, while Opus 5 prices tokens at $5 / MTok and $25 / MTok. OpenAI’s GPT-5.6 Sol sells for $5 / MTok input and $30 / MTok output.

However, the tokens required to complete a task can vary from model to model, so it’s also useful to assess how much each model costs to complete the same task.

Per Artificial Analysis, the weighted average cost (USD) per Intelligence Index task is $2.75 for Fable, $2.03 for Opus 5 (max), $1.04 for GPT-5.6 Sol (max), and $0.95 for Kimi K3. 

Opus 5 leads the Artificial Analysis Intelligence Index at 61, one point ahead of Fable.

“Claude Opus 5 provides greatly improved performance for the same cost as its predecessor, Opus 4.8,” Anthropic claims.

Opus 5 should offer customers more context to fill because Anthropic has put the model’s system prompt on a diet. According to Claude Code engineer Thariq Shihipar, Anthropic removed 80 percent of the Claude Code system prompt for its latest models.

As a result, the company’s guidance for crafting prompts, skills, and CLAUDE.md files has changed. 

“Across your system prompt, skills, and CLAUDE.md files, you may need to simplify just like we did,” said Shihipar, adding that a command called `claude doctor` can help automatically optimize some prompts and skills.

Users of Opus 5 however may find the model’s responses wordier – at their expense. Anthropic cautions, “Claude Opus 5’s default user-facing responses run longer than prior Opus models’.” Customers who don’t want the new verbosity are advised, “To control response length, prompt for it explicitly.”

Opus 5 scores close to Mythos on finding vulnerabilities in open source code (80 percent OSS-Fuzz benchmark score compared to 79.4 percent), but is significantly weaker in weaponizing those findings – it succeeded in 4/14 exploitation attempts compared to 13/14 for Mythos.

The model’s system card states, “Claude Opus 5 is substantially stronger than Claude Opus 4.8 across the board, with the largest gains in agentic coding, computer use, and long-horizon knowledge work.” It is said to be comparable to, or in some cases ahead of, Claude Fable 5 and Claude Mythos 5.

The model is likely to be significantly more useful for cybersecurity and biology tasks because it should refuse to cooperate less often. According to Anthropic, “Opus 5’s cyber classifiers are proportionally less restrictive than those on Fable 5. They allow Opus 5 to find vulnerabilities in source code, but block ‘binary-based’ vulnerability scanning (a method more likely to be associated with malicious actors), penetration testing, and exploit generation.”

Anthropic is also making life easier for those frustrated by model refusals. It’s rolling out an automated fallback function on its API so requests deemed too dangerous to entrust to Opus 5 or Fable 5 can be routed to a less capable model instead of being blocked outright.

Opus 5 is said to be Anthropic’s “most aligned model to date,” meaning it’s the least likely to go off its guardrails.

A more compelling selling point for businesses is likely to be the lack of a data retention requirement. ®

from the how-dare-anyone-expect-the-doj-to-follow-the-law dept

Just last week we had the story of the Trump DOJ issuing very questionable subpoenas of NY Times journalists while trying to track down who leaked information to those reporters regarding the potentially catastrophic security flaws of the “gift” 747 plane he received from Qatar. As we noted in that original post, this appeared to be a wholly abusive use of the government’s subpoena powers, and well outside the norm.

On Thursday, the DOJ agreed to withdraw those subpoenas, but only after a long court hearing in which the DOJ thoroughly embarrassed itself in front of the judge, Arun Subramanian, who noted many, many problems with the subpoenas, which the DOJ tried to tiptoe around, calling them “inadvertent errors.” Most of the media coverage of this is pretty weak, but Matthew Russell Lee of the Inner City Press did a wonderful liveposting of the hearing that suggests just how badly the DOJ fucked this up.

It started out with the DOJ saying they weren’t going to withdraw the subpoenas, and claiming that they believed the subpoenas were “properly” issued. But the court quickly pointed out that there is precedent in the Second Circuit regarding when and how you can subpoena journalists, and the DOJ basically ignored all of that. The DOJ’s Sean Buckley argued that following those rules would amount to conceding the rules applied — something the DOJ apparently didn’t want to admit, leading the judge to say that following the rules wouldn’t be seen as any such admission.

Judge Subramanian kept pressing Buckley on why the DOJ rushed to issue these incredibly broad subpoenas when there appeared to be much more straightforward ways to obtain the information they were seeking. Indeed, another part of what was discussed is that the DOJ’s subpoenas were so broad that they included phone records of reporters’ relatives who had nothing whatsoever to do with the reporting:

A Justice Department lawyer, Sean Buckley, cast the government’s missteps as inadvertent errors and said: “No one was trying to pull a fast one.” Buckley apologized for other subpoenas that sought records for phone numbers belonging to one reporter’s mother and two of the journalists’ spouses.

“That was an error, judge, which we own,” Buckley said. “It was a consequence of trying to move quickly.”

“These things are starting to pile up,” Subramanian said, becoming increasingly testy.

The judge also explored whether or not the DOJ misled the judge who signed off on the subpoenas, by not letting them know that the subpoenas were for information associated with reporting. He even noted that the Assistant US Attorney who got the subpoenas, Kevin Sullivan, was in the room, but not at the table, asking him to come out of the galley and join the DOJ table (this is not something that usually happens).

Following that was an incredible exchange wherein Judge Subramanian asked Sullivan about whether he told the original subpoena-issuing judge that the subpoenas were for reporters, leading Sullivan to say it “was an oversight” and that later on they “did legal research.”

Around that point, a clearly fed up Subramanian said that if this were a normal case, this would be the point where he would issue an order to show cause why the DOJ shouldn’t face sanctions for abusing the subpoena process. There was some more back and forth scolding, including Subramanian pointing out that the “errors” for the DOJ seemed to be “piling up” and asking the DOJ if he should expect to see more mistakes like this moving forward.

Around this point, the DOJ regrouped and changed their stance from earlier in the hearing, saying they were now willing to drop the subpoenas. After the hearing was over, the ever petulant Trump Justice Department quickly whined to the media how unfair it was that federal judges expected them to actually follow the rules and stuff:

After the hearing, the Justice Department lashed out at Subramanian in a statement, saying he “threatened our attorneys with sanctions unless subpoenas were withdrawn, and blocked us from presenting the meticulous process of this investigation.”

“The grand jury has a right to hear testimony from all material witnesses in a federal criminal investigation. This judge’s conduct overrides clear longstanding principles and common sense — blocking the grand jury from receiving core evidence in a national security investigation,” the statement said.

“Make no mistake,” it added, “this investigation remains ongoing, and we will pursue justice against those threatening national security by leaking classified information, a serious federal crime.”

Advertisement

Once again, we have an overly aggressive, understaffed, and generally incompetent DOJ that seems not to realize that there are significant and important Constitutional limits on what it can do. And when a judge calls them out on it, the fact that their immediate response is to start whining about it like they were the victims here suggests a good reason that the entire DOJ will need a massive overhaul post-Trump.

Filed Under: 1st amendment, air force once, arun subramanian, doj, free speech, kevin sullivan, qatar, sean buckley, subpoenas

Companies: ny times

Indeed’s report found that three out of four Ireland-based employees are open to moving into a new career, but are being held back by a number of factors.

Job recruitment platform Indeed has published new research indicating that there are a multitude of factors preventing Ireland’s employees from making a significant career move into another field, despite a significant portion being open to making the change. 

On behalf of Indeed, Censuswide collected data from 1,000 jobseekers and employees based in Ireland in June of this year. A main finding of the research was that, of those who contributed their information, three-quarters of workers are open to a career change, however, there is a lack of understanding into how this may be facilitated. 

31pc of participants explained that they are held back by financial uncertainty, 23pc are concerned about having to start as a beginner all over again, 23pc fear that they don’t have the necessary qualifications and a further 23pc are uncertain about which jobs would suit them. 

Concerns around skill level stood out, as more than 60pc were of the opinion that while they do have transferable skills, they aren’t quite sure where those skills might be most relevant. Half of respondents said they wish they knew what jobs they qualify for based on their skillset and experience.

Commenting on the report, Conor McCarthy, the vice-president of sales for the UK and Ireland at Indeed, said: “Changing careers can feel daunting, especially in a rapidly changing work environment. People know they have valuable skills but cannot see how they translate into new opportunities in today’s workplace.” 

Calm, cool and confident

Indeed’s research found that increased confidence and guidance could potentially encourage people to explore new opportunities. 71pc said they wish they could ask questions about a company or role before applying for it, with an additional 54pc agreeing that they have struggled to explain why they are a good fit for a role during an application or interview.

The report suggests that ambition is not lacking among Ireland’s employees, as 61pc said that they have a dream career they would love to make a reality, while 88pc could imagine moving into a field “where workers will be needed in the long term”.

With longevity in mind, employees are navigating growing uncertainty about the future, with this being a leading concern for of workers that are thinking about their professional future (34pc), followed closely by the impact of AI and automation (32pc).

Career coaching also remains largely untapped, claimed the report, which found that only 21pc of participants have ever used a career coach, with the biggest barriers including never having considered using one (39pc), concerns about cost (28pc), uncertainty about what a career coach does (25pc) and not knowing where to find one (24pc).

McCarthy said: “For many people, career support has either felt out of reach or too generic to be useful.”

Updated, 4.42pm, 24 July 2026: This article has been amended to fix some incorrect statistics.

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.

Quordle was one of the original Wordle alternatives and is still going strong now more than 1,500 games later. It offers a genuine challenge, though, so read on if you need some Quordle hints today — or scroll down further for the answers.

Enjoy playing word games? You can also check out my NYT Connections today and NYT Strands today pages for hints and answers for those puzzles, while Marc’s Wordle today column covers the original viral word game.

The company has comitted to 10 natural gas plants since last year.

The standard computer mouse is a perfectly useful peripheral if your hands work. If you’ve got some trouble in that area, you might appreciate an alternative input solution. To that end, [Varun Adinath Patil] created a neat hands-free solution for moving a cursor around a screen.

The build is based on the Neuro PlayGround Lite, a board built for physiological signal acquisition in the Feather form factor. It’s hooked up to an IMU sensor—both a MPU6050 or BMI270 work—which tracks head movements to allow the cursor to be panned around the screen. Other biological signals are then used to activate other standard mouse functions. Clenching the jaw fires off a left click, while a triple blink fires a right click. Clicking and dragging is achieved by a double-blink. The jaw muscles are sensed via EMG signals picked up with gel electrodes on the skin, while the blinks are detected via EOG signals via the same contact points.

Commercial solutions in this realm exist, but it’s great to see how such a device can be built from the ground up. We’ve looked at other neat applications of head-tracking before, too. If you’re working on your own innovative accessibility tools, don’t hesitate to let us know via the tipsline.

The big picture: McAfee was a trusted name in the antivirus software market during the MS-DOS days. Many computers came with a copy of the company’s VirusScan freeware tool to fend off trojan horses and other virus threats. Today, McAfee is a name that’s almost exclusively trusted in the advertising and B2B markets. End users would very much prefer to avoid McAfee products at this point, which is why they tend to get angry when someone tries to push the McAfee brand through bundled advertising pop-ups.

The controversial advertising partnership between LG and McAfee has apparently come to an end after Microsoft got involved in the matter.

Pavan Davuluri, executive vice president of Windows and Devices, recently confirmed that Redmond got in touch with LG, the company that started it all. LG has “agreed” to remove the McAfee pop-up from its Windows Store app, Davuluri said, which likely means that Microsoft forced the South Korean manufacturer to stop pushing McAfee-related advertising to customers who purchased a new gaming monitor.

The controversy began earlier this month, when users discovered that the “LG Monitor App Installer” app was displaying McAfee promotions. Windows installed the app after affected users connected their brand-new LG monitors, with no easy way to uninstall the tool without resorting to third-party applications or disabling the Microsoft Store itself.

– Pavan Davuluri (@pavandavuluri) July 22, 2026

According to Davuluri’s post, Microsoft “connected” with LG before the latter agreed to remove the McAfee adware message from the LG Monitor App Installer app. Both companies apparently share a common goal of improving the Windows ecosystem and providing customers with the best possible software experience.

Per user complaints, the list of LG monitors reportedly found to automatically push the LG Monitor App Installer app onto users’ PCs includes the following models: 34GX900A-B, 45GX950-B.AEU, 32GS95UE-B, 39GX950B, 27GP83B-B, 27GN800, 32GS95UE-B, and 27GN850-B.

Apps developed through Microsoft’s Universal Windows Platform can be designed to install automatically when a user connects a specific device to a Windows machine. Put simply, Microsoft approved this practice for UWP app developers, and the LG monitor tool was simply working as designed.

The official UWP documentation even states that automatic installation can be a source of confusion because users do not receive any notification about the installation process.

Other companies known to take advantage of this auto-installation feature include Razer, Logitech, Asus, and Gigabyte.

A federal judge has ordered a public iPhone exploit taken offline after Magnet Forensics argued it wasn’t independent security research at all, but instead a stolen trade secret.

U.S. District Judge Victoria Marie Calvert partially approved Magnet’s request for a preliminary injunction. She directed Paradigm Shift and former Magnet exploit engineer Mario Del Gaudio to delete the usbliter8 article, code, technical details, and related materials in their possession by 11:59 p.m. Eastern on July 23.

By July 23, Paradigm Shift had replaced the original article with a page indicating the blog post was unavailable. The preliminary injunction will continue throughout the litigation unless the court removes it in a separate order.

Magnet’s July 7 complaint asserts that usbliter8 originated from a confidential A12 and A13 SecureROM access capability integrated into a commercial forensic product. The company alleges Del Gaudio acquired the technique while employed by Magnet and later shared it through Paradigm Shift.

Paradigm Shift originally presented usbliter8 as newly published security research before releasing it on June 18.

We reported at the time that the exploit affects devices including the iPhone XS, iPhone XR, iPhone 11 lineup, and second-generation iPhone SE. The court hasn’t made a final ruling on liability.

Calvert found that Magnet had established a likelihood of success on its trade-secret and contract claims for purposes of the preliminary injunction, based on evidence the defendants didn’t contest at the July 16 hearing.

The iPhone exploit requires physical access

Usbliter8 targets SecureROM, the immutable code that starts Apple’s secure boot process. It combines a flaw in a USB controller with security settings used on A12 and A13 devices to execute code while a device is in Device Firmware Update mode.

Because SecureROM is built into the processor during manufacturing, Apple can’t replace the vulnerable code through an ordinary software update. It may still be able to develop mitigations that interfere with exploitation or reduce its usefulness.

The flaw doesn’t create a remote attack or automatically expose everything stored on an iPhone. Using usbliter8 requires physical access to the device, a USB connection, DFU mode, and programmable hardware capable of sending specially constructed USB traffic.

The exploit can run unsigned code before the operating system starts, but it doesn’t directly compromise the Secure Enclave or automatically reveal a user’s passcode and encrypted data. Additional vulnerabilities or forensic techniques would be needed to cross those protections.

Those requirements make usbliter8 especially relevant to forensic investigations involving seized devices. Magnet sells investigation products to law enforcement agencies, intelligence services, government bodies, and private organizations.

Magnet says usbliter8 came from a secret capability

Del Gaudio worked as an exploit engineer placed with Magnet from November 2023 through November 2024. He signed an agreement covering confidential information, intellectual property, and continuing restrictions that survived the end of his placement.

Magnet says Del Gaudio had access to a zero-day capability internally called “MSG,” which targeted the same A12 and A13 SecureROM vulnerability later described in the usbliter8 publication.

According to the complaint, Magnet engineers discussed the vulnerability in meetings attended by Del Gaudio by April 2024. The company says it integrated MSG into one of its products in May 2024 and that Del Gaudio used the capability dozens of times while testing another tool.

Magnet says Del Gaudio had access to a zero-day capability internally called “MSG.”

Magnet supplied additional details in a July 17 declaration addressing questions Calvert raised at the July 16 hearing. The company’s director of iOS research said Del Gaudio attended restricted sessions during a company gathering in Denver from March 11 through March 15, 2024.

Fewer than 20 people attended the smaller iOS sessions, according to the declaration. Magnet said the group discussed the SecureROM vulnerability, MSG’s technical architecture, and its development into an access capability for the company’s products.

Magnet links Del Gaudio to the publication

After his placement ended in November 2024, Del Gaudio became affiliated with Paradigm Shift, according to the complaint. The Spanish security company published “Introducing usbliter8: An A12/A13 SecureROM Exploit” on June 18.

A preserved screenshot connected Del Gaudio’s name and photograph to the @NotHdesk account associated with the research, according to Magnet. The company also says the account was linked to an email address known to belong to him.

Those details form part of Magnet’s case that Del Gaudio had access to MSG and was connected to the usbliter8 publication. The public record doesn’t include source-code comparisons, file-transfer records, or a detailed technical analysis showing exactly how MSG and usbliter8 match.

The original usbliter8 article was deliberately omitted from the complaint because Magnet argued that attaching it would further distribute the information it sought to protect. The company offered to provide the material privately for the court to review.

On June 18, Magnet sent Del Gaudio a cease-and-desist demand and contacted Paradigm Shift the next day. The demand sought removal of the article and code, identification of anyone who received the information, preservation of evidence, and return or destruction of Magnet material.

In letters dated June 22 and June 28, Paradigm Shift’s attorneys disputed Magnet’s claims and pressed the company to identify the information it considered a trade secret. Magnet filed the lawsuit on July 7 after the parties failed to reach an agreement.

Neither Del Gaudio nor Paradigm Shift appeared at the July 16 injunction hearing, despite receiving electronic notice. Calvert therefore considered an uncontested record when deciding whether temporary relief was warranted.

The court questioned whether the flaw should remain secret

Magnet argues that publication let competitors study the technique without making the same investment. The company also says Apple could reduce the exploit’s value through mitigations, while the disclosure may weaken customer trust in Magnet’s ability to protect sensitive capabilities.

The dispute raises a security question over whether companies should keep zero-days secret for forensic use or disclose them so manufacturers and device owners can respond.

Magnet argues that publication let competitors study the technique without making the same investment.

Calvert described the public-interest issue as the most difficult part of the case. The judge addressed concerns about companies and government actors stockpiling zero-day vulnerabilities rather than reporting them to affected manufacturers.

The court also questioned whether consumers were better protected by knowing about the vulnerability once its existence had become public. Calvert concluded the court couldn’t resolve that policy debate through an unopposed preliminary injunction motion.

Paradigm Shift, as reported by MacRumors, said it informed Apple Product Security before publishing on June 18. The order doesn’t prevent Apple from using information it already has to mitigate the vulnerability.

Since Apple already has the disclosure, the injunction can’t fully restore the secrecy Magnet claims gave the capability commercial value. However, Calvert found that removing the material could still limit further harm and prevent Paradigm Shift from using the research for promotion.

The central trade-secret question remains unresolved

Magnet also sought extensive forensic access to the defendants’ computers, accounts, and storage. Calvert declined to grant that relief outside the normal discovery process.

The court observed that Magnet accepted Del Gaudio might not have required company hardware or files to replicate the capability. In Magnet’s view, familiarity with the research could have been enough.

A key question remains for future steps, such as whether Del Gaudio copied protected Magnet data or drew on technical understanding and experience kept after departing the firm.

The injunction covers material held by the defendants but can’t remove copies already downloaded or shared elsewhere. The case now turns on whether usbliter8 represents independent research or the disclosure of Magnet’s confidential forensic capability.

The big picture: Nvidia lifted the embargo on its Vera CPU deep dive this week, and the disclosure amounts to a direct challenge to two decades of x86 datacenter design philosophy. This is the most detail the company has shared on the chip since it first appeared on the Rubin roadmap, and it confirms something I have suspected for a while: Nvidia is not treating the CPU as an attach story anymore. It is treating it as a battleground with a lot of potential dollars at play.

Vera is built around the Olympus core, the first custom CPU core Nvidia has ever brought to the datacenter and the first custom core the company has designed anywhere since the Denver and Carmel efforts of the Tegra era nearly a decade ago.

Ryan Shrout is a longtime technology analyst and industry veteran who has spent over two decades covering PC hardware, graphics, and semiconductors. He previously led technical marketing at Intel and was the founding editor of PC Perspective. He is currently President and GM at Signal65. You can follow him on X @ryanshrout.

Grace used licensed off-the-shelf Arm Neoverse V2 cores. Olympus is an Nvidia design from the ground up, a wide, high-IPC core with a 10-wide decode front end that reorders aggressively and prefetches based on patterns like graph structures in memory.

88 of those cores sit on a monolithic compute die, running 176 threads through a partitioned scheme Nvidia calls Spatial Multithreading, a deliberate departure from the opportunistic resource sharing of traditional SMT.

That monolithic choice matters. Nvidia still uses chiplets for the memory controllers and I/O, but the compute die is one piece of silicon connected by a second-generation scalable coherency fabric. The company measures bisection bandwidth across the die at roughly 3.4 terabytes per second.

The memory subsystem is LPDDR5X hardened for the datacenter with ECC and full telemetry, delivering up to 1.2 TB/s of bandwidth, roughly 3x the memory bandwidth per core and about 5x the bandwidth per watt of conventional DDR-based server designs (all based on Nvidia claims).

The headline claims stack up as roughly 2x faster performance from the Olympus core, 3x the core-to-core bandwidth of chiplet-based competition, and 40% lower memory latency under load through the LPDDR5X subsystem.

Vera ships in two forms: a dense liquid-cooled rack packing 256 CPUs and more than 22,000 cores, and a conventional air-cooled 2U with two sockets. Dell has committed to multiple PowerEdge systems built on it. Nvidia sizes the opportunity as a $200 billion expansion of the CPU market, which explains a lot of the recent market dynamics.

The argument behind the architecture

In 2014, a top Xeon carried 14 to 18 cores. Today an Epyc Turin part carries 128. Core counts grew roughly 9x over that stretch because cloud economics rewarded rentable vCPUs, while per-core performance only about doubled. Chiplets kept costs down but taxed memory bandwidth, data movement, and latency along the way.

Nvidia argues that agentic AI breaks this trade. An agent reasons on the GPU, then drops to the CPU for tool calls, SQL queries, API work, and scripting, then goes back to the GPU, sometimes hundreds of times per task. That loop is sequential.

You cannot throw more cores at a sequential loop and make it shorter. Only a faster core, fed with data faster, compresses it. The loaded-latency data Nvidia showed makes the point visually, with chiplet designs hitting a saturation wall just shy of 400 GB/s of memory traffic while Vera keeps scaling.

The company has landed on “max single-threaded CPU at scale” as the category name. It is a mouthful, and I will get to that.

The proof points

The customer data is early but notable. Perplexity ran coding sandboxes on Vera and completed jobs 1.5x faster than the production Xeon fleet it runs today, with concurrent sandbox startup 1.9x faster.

The New York Stock Exchange, which processes 1.1 trillion records a day, tested Vera with the Redpanda streaming engine on HPE systems and measured 6x lower p99 latency versus Epyc Turin, and is now evaluating it as a replacement. Los Alamos National Laboratory saw 7x on an agentic workload and 3x on radiation transport and multigrid simulation codes.

These are real workloads rather than synthetic benchmarks, which I give Nvidia credit for, and the supporting documentation puts names and configurations on most baselines. They remain vendor-supplied and worth reading closely.

The SPEC CPU 2026 numbers carry an estimated label from a pre-production reference system, and every comparison lands on Zen 5 Turin or current and older Intel silicon. The Los Alamos runs were measured against a Sapphire Rapids based supercomputer that launched in early 2023. Beating shipping parts is the right first test, but Venice and Diamond Rapids arrive within the year, and that is the fight that will settle this.

Where there is more detail needed

I asked the Nvidia team directly during the analyst briefing last week what actually separates an agentic CPU from a plain, very good datacenter CPU. We have had big, fast processors running back-to-back loops of VMs and containers for years. Is this genuinely a new workload class, or a fast CPU wearing new marketing?

To their credit, the team acknowledged that “agentic CPU” is the wrong label and would pigeonhole the part. The honest answer is that the fundamentals have not changed, but the rates have.

Agent pipelines hydrate and tear down environments constantly rather than occasionally. Memory pressure is continuous. Latency under load becomes the whole game, because every millisecond the CPU stalls is a millisecond a very expensive GPU sits idle. That is a real architectural argument, and the decision to spend die area on per-core speed instead of core count is a genuine philosophical break from where x86 roadmaps have been heading.

What the market needs now is independent, rigorous CPU measurement built around these agentic pipelines, run across current and next generation parts from every vendor. That is exactly the kind of work we are looking forward to diving into at Signal65.

The ecosystem arrived on day one, to no surprise

The partner roster attached to this launch is unusually deep for a CPU announcement. OpenAI says it will deploy Vera at scale beginning in Q3, and the early adopter list also includes Anthropic, SpaceX, and Perplexity, with Los Alamos, NERSC, and TACC representing the supercomputing side.

Dell, HPE, Lenovo, Supermicro, and Bull all have Vera systems coming, backed by the full ODM bench.

The rack-scale platform is ramping just as visibly. CoreWeave was the first cloud to bring up and validate Vera Rubin NVL72 and published the first measured numbers from live hardware, a 10x gain in tokens per second per megawatt over Grace Blackwell NVL72 on DeepSeek-R1.

Google Cloud stood up the first A5X instance on Vera Rubin for the reinforcement learning startup Ineffable Intelligence, Azure and OCI have racks running, and a newly expanded Microsoft and Mistral agreement puts Vera Rubin at the center of a multibillion-dollar European buildout.

On the CPU specifically, DeepInfra, which serves nearly five trillion tokens a week, measured support for 1.6x more concurrent agents and 2.2x faster orchestration versus Granite Rapids. Nvidia counts 300 partners and more than 350 factory sites in 30 countries behind the ramp.

What this means for AMD, Intel, and the hyperscalers

The timing is not subtle. This disclosure lands the day before Advancing AI opens in San Francisco, the flagship AMD event where the 256-core Zen 6 Venice generation of Epyc and the Instinct MI450 family are expected to headline the keynote from Lisa Su on Thursday. Nvidia just set the terms of the datacenter CPU conversation roughly 24 hours before its biggest rival takes the stage.

If a faster CPU returns GPUs to work sooner, the CPU price becomes a rounding error in rack TCO, and the fight shifts from dollars per core to tokens per rack. That framing is the one AMD and Intel now have to answer, and it is a very different conversation than the one that produced 128-core roadmaps.

AMD is not conceding the frame. It has already argued that rack-level performance per watt favors high-core-count Epyc, and the 256-core Venice generation will sharpen that response. Intel has Diamond Rapids coming. The hyperscalers have Graviton, Axion, and Cobalt, all designed around the scale-out economics Vera explicitly rejects.

But Nvidia is not selling a merchant CPU into a commodity socket. It is selling the CPU as the utilization lever for the most expensive assets in the AI factory. If a faster CPU returns GPUs to work sooner, the CPU price becomes a rounding error in rack TCO, and the fight shifts from dollars per core to tokens per rack.

That framing is the one AMD and Intel now have to answer, and it is a very different conversation than the one that produced 128-core roadmaps.

Nvidia has promised deeper head-to-head benchmark data against both x86 and Arm competition in the coming weeks. That data, and the independent validation that should follow it, will tell us whether Vera resets the datacenter CPU conversation or simply carves out a well-defended niche inside the Nvidia rack.