gshc2020.com

The AI infrastructure race is most frequently described in terms of models, chips, power and data center capacity. Those are important measures, but they only tell part of the story. Behind an eerie AI deployment lies a layer of physical infrastructure that must be designed, manufactured, qualified and installed. Investors and policymakers must look beyond these traditional signs of success and shift their focus to the hardware that powers, processes and transmits computational data.​

That infrastructure includes racks, enclosures, cabinets, power distribution systems, cooling equipment, containment structures, cable management and other mechanical assemblies. These components and manufacturing systems rarely receive the attention that chips and AI models do, but they can determine how quickly new capacity actually comes online. AI may feel almost intangible to the people using it, but underneath the model layer and the cloud, it is intensely physical. ​

With hyperscalers projected to pour trillions of dollars into AI infrastructure, the question is no longer simply how much computing power they want to build. It’s whether the manufacturing ecosystem can build fast enough to keep pace. The companies, factories and supply chains behind that effort may attract a fraction of the attention given to chipmakers and AI developers, but they ultimately help determine whether massive capital spending translates into actual operating capacity. ​

​The Physical Infrastructure Behind AI’s Growth​​

The scale of today’s data center buildout is historic. Bank of America Global Research forecasts that AI-related capital spending by hyperscalers could approach $800 billion in 2026 and surpass $1 trillion in 2027. That spending is translating into extraordinary physical demand. Goldman Sachs Research estimates global data center power demand will surge from roughly 62 gigawatts to 92 gigawatts by 2027, with AI consuming an increasing share. ​

The scale of that investment has put chips, power and data center capacity at the center of the conversation. But capital does not become capacity until the equipment is designed, manufactured, qualified and installed. That leaves hyperscalers confronting a fundamental challenge: AI demand can grow at the speed of software, but the physical infrastructure required to support it cannot. ​

​Where Engineering Meets Manufacturing​

AI infrastructure is being called on to scale while the underlying technologies continue to mature. Higher densities, new cooling requirements and increasingly complex power systems are driving more frequent engineering changes across the mechanical components that support them. These supply chains were largely designed around relatively predictable product releases. Now, they must adapt to frequent engineering changes while simultaneously scaling production.

As AI infrastructure evolves, engineering changes ripple through the mechanical supply chain at an increasing rate. Hyperscalers can commit billions to new capacity, but those commitments still have to move through the manufacturing chain, from design and prototype to NPI and eventually production. When demand is accelerating while designs continue to evolve, the ability to move through these stages becomes a constraint in its own right.

The question then becomes not just how much capital is being committed to AI infrastructure but also whether the manufacturing ecosystem can convert that capital into qualified equipment quickly enough. ​

The Capacity To Keep Up​

Large-scale manufacturing works best when designs have stabilized. AI infrastructure has not reached that point, with much of it still in a period of rapid development. This has forced OEMs and infrastructure suppliers to manage more engineering changes while production is already underway. ​

That puts greater value on manufacturing partners that can move between engineering and production without creating delays. Capabilities ranging from design and engineering feedback to machining, inspection and assembly become a part of keeping projects on schedule.​

The same pressure extends to the workforce. The Manufacturing Institute and Deloitte estimate U.S. manufacturing may need as many as 3.8 million additional employees between 2024 and 2033, with as many as 1.9 million positions potentially going unfilled. Meeting the demands of AI infrastructure requires machinists, welders, programmers, engineers and supply chain professionals who can work through changing technical requirements. ​

For investors, this makes the manufacturing companies supporting AI infrastructure worth closer attention. They may sit well below the chip and cloud layers in visibility, but they provide the capacity and expertise required to turn engineered designs into deployable systems. ​

What Investors Should Watch

​For investors evaluating the next phase of AI infrastructure spending, the relevant question is not simply which companies have exposure to data center growth; it’s whether their manufacturing operations can keep pace with the speed and complexity of that growth. ​

That means evaluating how early a manufacturer is brought into the engineering process, how quickly it can respond to design changes and how effectively it can move a part from prototype through qualification into production. ​

Investors should also consider whether the company has the inspection, documentation, supplier relationships and workforce required to maintain quality as volumes increase. These capabilities become particularly important when production cannot wait for every design decision to be finalized. ​

The winners in the coming wave of AI infrastructure will be the companies that can convert innovation into deployable capacity with speed and discipline. As AI moves from a race for computing power to a race to build it, the manufacturing layer may prove just as important as the technologies it supports.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

To be fair, I can't even seem to find evidence that this was an April Fools joke. But Corsair implies as much:

"From April Fools, to Computex, to Gamescom and beyond - here by popular demand: you can now print your own Corsair fan cube! Just a simple cube that holds 6 fans—no fuss, no greater meaning, only fans."

"No greater meaning," that's what I like to hear. In a world that seems to be collapsing around us in every direction, sometimes it's nice to have something pointless to grasp on to.

At any rate, if you fancy making one, you can download the files from Printables. They include three main parts—the base and both sides of the cube—plus a connector, which you print four of. Apart from this, all you need is M4x14 mm screws to connect the base.

Judging from the picture, it looks like the cube supports standard 120 mm fans, but don't quote me on that. So if you have a few spare case fans lying around, as well as a 3D printing machine, why not? Well, I suppose there's the fact that it'll be a waste of energy, but hey, that's RGB anyway, right?

The PC Gamer Circuit Breaker newsletter breaks down the latest PC gaming tech, and finds you the best deals on the best products

Jacob got his hands on a gaming PC for the first time when he was about 12 years old. He swiftly realised the local PC repair store had ripped him off with his build and vowed never to let another soul build his rig again. With this vow, Jacob the hardware junkie was born. Since then, Jacob's led a double-life as part-hardware geek, part-philosophy nerd, first working as a Hardware Writer for PCGamesN in 2020, then working towards a PhD in Philosophy for a few years while freelancing on the side for sites such as TechRadar, Pocket-lint, and yours truly, PC Gamer. Eventually, he gave up the ruthless mercenary life to join the world's #1 PC Gaming site full-time. It's definitely not an ego thing, he assures us. %!s()

This week, both parties filed opening submissions at an employment tribunal in Glasgow, which Rockstar parent company Take-Two Interactive have emailed to RPS and Game Developer, among others. Taken together, the filings amount to over 100 pages of claims and counterclaims, with thousands of pages of evidence to be discussed in the coming weeks. Amongst other things, the docs give us more sense of events leading up to the dismissals, including the revelation that Rockstar had a mole or "whistleblower" passing them information on the people using the Discord server.

According to the court filing from the ousted workers, Rockstar employees affiliated with the IWGB Game Workers union set up what would become known as the Rockstar Game Workers Union Discord Server in 2022. Prior to this, the lawyers claim, there was "no official workplace channel through which concerns could be discussed independently of management." By the time of the October 2025 dismissals, the server had around 340 members.

Rockstar management did not have direct access to the server prior to October 2025, but the former workers and their legal team allege that "from at least February 2024, [Rockstar HR director] Charlie Kinloch had been receiving information from a junior employee on trade union activity inside the Server". They claim that "over 21 months, at least 26 meetings or communications took place between Mr Kinloch and [the junior employee] concerning the Server, roughly half of them initiated by Mr Kinloch himself".

The "mole" is said to have "supplied screenshots and recordings, reported union meetings, training sessions and union membership numbers, and identified individual union activists". On 7th October 2025, they allegedly informed Kinlock that the union were 12 members away from the statutory threshold for recognition - 10% of the workforce. The union eventually reached this milestone on 18th October, with Kinloch learning of this on 20th October, the fired workers claim. At some point in this period, they go on, Rockstar management got hold of logins for the server, with Rockstar's vice president of information security Abid Ansari eventually signing in to download "the contents of the Announcements and General channels". Rockstar began firing members of the Discord at the end of the month.

Rockstar's own court submission chimes with some of this. The company's lawyers mention three "whistleblowers" on the Discord who reported security concerns, including one member who had been leaking info to Rockstar management "for over two and a half years". Rockstar's lawyers say that between 13th and 16th October, a member of the server supplied Rockstar bosses with login details; prior to this, they claim, Rockstar were "largely in the dark" about events on the server, save for "occasional glimpses" care of their primary "whistleblower".

The court filings offer wildly contrasting views about the security of the union Discord server itself. The unionised workers argue that the server was "invite-only, access-restricted and subject to verification through Rockstar’s own internal Slack platform", explaining that people with an invite link were placed in a waiting room till they had been approved by mods, who could verify them using Rockstar's own Slack channels before granting "substantive access". They point out that "no confidential information is known to have leaked from the Server into the public domain".

Rockstar, however, argue that "there were numerous ways in which the information on the Discord was vulnerable, including by a deliberate leak from a discontented Discord user, by loss or theft of the personal devices of the Discord users, by successful phishing or social engineering of one of the Discord users, or by hacking of Discord itself." Amongst other things, they support this by saying that somebody was, in fact, leaking info from the Discord: the aforesaid "whistleblower".

Rockstar offer a layperson breakdown of the Discord server's structure in their filing, portraying it as essentially consisting of three "rooms": one for IWGB union committee members only, or members of specific working groups, which Rockstar couldn't access at all before a court-ordered disclosure on 3rd August; a second for general IWGB members, which Rockstar again couldn't access initially; and a third "room" consisting of almost 350 people, some of whom were anonymous, while others identified themselves as Rockstar employees.

Rockstar claim that "26 on an initial analysis" of the people in the third "room" were not actually employees, while 10 others were former staff, including a few who had "left in poor standing with the business and were therefore well motivated to cause harm to it". The ousted workers acknowledge that "former employees" were members of the Discord, but say they were confined to certain channels, and not present in the General and Announcements channels to which Rockstar initially gained direct access. Rockstar also repeat an earlier accusation that one member of the wider Discord group was a working journalist.

As for the incriminating info being circulated, Rockstar allege that it ranged from discussion of digital and real-world security measures, through internal correspondence such as HR emails, to "closely guarded secrets" about game projects, including an "online 32 player format" for a "highly confidential" project. They also claim that Discord users shared "many, many pieces of information" regarding GTA 6's development, including discussion of delays, overtime and crunch.

The former workers, meanwhile, argue that "no posts were witnessed as disclosing the date on which GTA VI would be released, the date of any trailer, or comparable confidential release information", and that "the posts generally discussed workplace policies which were essential to trade union organising". They comment that many of the posts about office goings-on considered "egregious" by Rockstar were "trivial as matters of discipline but... legitimate as matters of concern to a trade union representing its members".

Both docs have a whole section on the other side's use of bad language, with Rockstar's lawyers accusing the union Discord members of "openly and vituperatively disparaging and defaming" company management, while the former workers point out that "profanity was routinely used by management and staff at [Rockstar] without disciplinary consequence". According to the former workers, one employee was told after joining Rockstar "that the use of the word 'cunt' was an 'apparently acceptable use of language in Scotland' and was used by management regularly".

The lawyers for the fired workers allege that the abrupt dismissal of 34 staff had "an immediate chilling effect" on the Discord server. Membership apparently fell "from around 390 to 261 within days, with 136 people leaving and only nine new joins", while "11 Rockstar employees cancelled their IWGB memberships in the following week." They broadly claim that Rockstar "identified and removed members and activists of a rapidly growing trade union, including many of those central to its organising structures, for the purpose of disabling the wider organisation and of deterring the remainder, and, in particular, preventing the union obtaining the right to negotiate pay, hours and holidays with [the company]."

The warring parties naturally disagree about whether Rockstar are hostile in general towards unions. Rockstar's lawyers point out that Rockstar did not dismiss every last union member on the Discord server in October 2025, and that three of the workers dismissed were not union members (the latter are or were based in Canada and so, outside the scope of any UK union, though the lawyers for the fired workers argue that they "were interested in unionising their own workplace").

Rockstar's lawyers further comment that the company "has never sought to prevent a recognition request and now one has been made is engaging constructively with it". Regarding the timing, they contend that the management didn't realise the extent of the "misconduct that was rife on that platform" until gaining direct access. They argue that if Rockstar had harboured real concerns about the union reaching the statutory requirement for official recognition, they would have acted on them much earlier.

The ousted workers, meanwhile, cite several alleged formal and verbal warnings their members had received from managers about soliciting union organisation, prior to the events of October 2025. They suggest that union advocacy or membership of the union Discord server could be a negative factor in discussion of flexible working, performance and promotions.

They also claim that when the IWGB presented a petition about return-to-office policies in February 2023, Rockstar refused to engage with union members as a collective: "instead, management circulated lists of the signatories to HR managers and contacted them individually, declining to engage with the union".

The court proceedings continue. There is a lot more in the full filings, including a lengthy characterisation from Rockstar's lawyers of the company's "embedded culture of confidentiality" and its "reliance on a culture of respect, non-disparagement, trust and confidence". If you'd like to read more, Game Developer news editor Chris Kerr has been posting large chunks from each filing on Bluesky. %!s()

But AI agents can call APIs, modify systems, access databases, trigger workflows and initiate transactions. A hallucination with a generative AI tool creates bad information, but an agent with authority can turn bad reasoning into a business event.

That distinction is becoming increasingly important as organizations move from AI that advises people to AI that acts on their behalf.

The risk has shifted from the model to execution.

In August 2026, the U.K.’s AI Security Institute disclosed that agents participating in controlled cybersecurity evaluations had taken autonomous, unsanctioned actions on the live internet. Across 122 evaluation runs, AISI identified 19 unsanctioned actions across 10 runs, including attempts involving real people and organizations. The institute found no evidence of resulting real-world harm, and the evaluations were deliberately permissive.​

Once an AI system can interact with external tools and environments, the risk shifts away from the sentences the model produces to the permissions granted to the system as a whole—what it is allowed to access, trigger, modify or execute.

This concern has already appeared in commercial systems.

In 2025, Replit’s AI Agent deleted data from a production database used by SaaStr founder Jason Lemkin despite instructions intended to prevent changes during a code freeze. Replit later acknowledged the incident and strengthened separation between development and production databases, including preventing the agent from changing the production database during development.

The industry will continue to improve its safeguards, but the lesson beyond any particular incident is that telling an agent what not to do is different from designing a system that makes certain actions impossible.​​

Governance has to exist at runtime.

Traditional AI governance often operates before deployment: Policies are written, models are validated, risks are classified and approvals are documented.

Agentic systems challenge that model because they operate continuously. They invoke different tools, interact with multiple systems and can execute chains of actions faster than a human reviewer could examine each step.

That means governance has to participate at the point where an action becomes executable.

A system should not merely record that an agent called an API after the fact. It should be able to determine whether that specific agent, acting for that specific user, in that specific context, is authorized to perform that action—and whether additional approval is required before execution.

While observability remains essential, seeing an unauthorized transaction after it occurs is not the same as preventing it.

The hardest failure may be completely authorized.

There is an even harder problem: An agent may perform a sequence in which every individual action appears permitted, yet the combination creates an outcome nobody intended to authorize.

For instance, an agent might be allowed to read customer records, generate a report and send an email. Each capability may be legitimate on its own. But if the agent combines them in the wrong context, it can create an unauthorized disclosure without ever invoking an individually forbidden tool.

This is an authorized-sequence problem.

Governance, therefore, cannot stop at tool-level permission checks. Organizations need controls that understand delegated intent, context and the cumulative effect of an agent’s actions.

AI agents are becoming privileged digital identities.

Enterprises already know how dangerous privileged identities can be. Administrators, service accounts and automated processes are normally constrained through authentication, least privilege, segregation of duties and audit trails.

AI agents increasingly belong in the same conversation. An agent capable of acting across databases, cloud services, payment systems or internal applications is effectively a privileged digital identity. Giving it broad, persistent access because doing so makes automation easier creates the same structural risk organizations have spent decades trying to reduce elsewhere in technology.

Agent permissions should therefore be narrow, contextual and revocable. High-consequence actions may even require explicit human approval. Credentials should be scoped to the task rather than inherited broadly from the user. And organizations should be able to suspend authority quickly when behavior diverges from expected boundaries.

Regulated industries will feel this first.

Regulated sectors will face these questions earlier because governance obligations already extend beyond model accuracy.

For high-risk AI systems within its scope, the EU AI Act requires capabilities for automatic event logging and effective human oversight. Those requirements do not apply to every enterprise agent, but they illustrate where regulators are heading: Organizations increasingly need evidence of what systems did, who or what authorized the action and where human control existed.

Financial institutions already apply comparable concepts to other critical technologies through permissions, transaction controls, monitoring, separation of duties and auditability. Agentic AI will make those principles more important.

The next AI bottleneck is deploying agents safely.​

Organizations already have increasingly capable models, but that creates risk until leaders have put in place strong governance about how much authority to delegate to agents.

Governance involves the ability and infrastructure to answer in real time: Who delegated this authority? What is the agent allowed to do? What requires approval? What actually happened? And can the system stop the next action before a mistake becomes a business event?

As AI moves from generating answers to executing decisions, authorization becomes part of the AI architecture itself.​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

Using advanced computer simulations, the researchers examined the enormous impact thought to have occurred about 4.5 billion years ago between the young Earth and a Mars-sized object. Their results reveal strikingly different outcomes depending on the temperature and structural properties of the two bodies.

The findings were published in The Astrophysical Journal Letters. For the first time, the simulations account for the material strength of the two ancient worlds. That addition could reshape scientists' understanding of the Moon-forming collision and may also help narrow down when it happened.

"We discovered that the preexisting geology of the Mars-sized proto-moon matters," said Adeene Denton, a former postdoctoral researcher at Lunar and Planetary Laboratory who is now at SwRI. "When you simulate the Earth and the Moon as colliding bodies with geologic properties, it changes how the Moon forms out of that impact - that's something we considered unnecessary before."

Revisiting the Giant Impact Theory

One of the leading explanations for the Moon's origin is the giant impact scenario. According to this idea, a Mars-sized body called Theia collided with the early Earth. The impact destroyed Theia and sent its remains into orbit around Earth, creating a disk of debris that eventually assembled into the Moon.

A foundational 2001 study of this scenario was led by Robin Canup, vice president of SwRI's Solar System Science and Exploration Division in Boulder, Colorado, and Erik Asphaug, a professor at the Lunar and Planetary Laboratory and co-author of the new study.

Those early simulations, along with many later models, did not account for material strength. Scientists assumed that the collision was so energetic that the rocky bodies could effectively be treated like fluids.

Denton and her colleagues decided to test whether that assumption was justified. Using modern computational methods, they incorporated temperature-dependent geologic strength into Moon formation simulations for the first time.

"Because the collision was considered violent enough to melt and vaporize large portions of Earth and Theia, those previous papers assumed that it is okay to approximate them as fluids," Asphaug said. "Based on our new results, however, we think that it is time to reconsider that."

Adding Realistic Geology to the Collision

To revisit the giant impact, Denton used a more advanced form of smoothed particle hydrodynamics simulations, or SPH. These simulations account for the structural strength of the objects involved rather than treating them as entirely fluid.

"It turns out material strength is really important when you're studying collisions between smaller bodies like asteroids, dwarf planets and moons," said Denton, who got the idea of applying the concept to Earth's moon while working on a previous paper about the formation of the Pluto-Charon system. "We weren't sure if it would matter for our moon or not. When we did the simulations, we found it actually matters quite a bit."

The version of SPH used in the study was developed at the U of A and the University of Bern in Switzerland. Its strength model allows simulated planetary material to resist deformation in ways more consistent with real geologic substances, including the rock and metals that likely made up Theia and the proto-Earth, or solid ice.

The Moon Could Form in Very Different Ways

Temperature turned out to be especially important. Hotter planetary bodies are mechanically weaker than colder ones, and the simulations showed that this difference can dramatically affect what happens after impact.

Under some conditions, the collision destroys Theia and creates a broad protolunar disk around Earth. Material in that disk then gradually comes together to form the Moon.

But other simulations produced a much more dramatic result. Instead of forming slowly from debris, a fully intact Moon appeared within only a few hours.

Because young protoplanets generally begin hot and gradually cool as they age, this result creates a potentially important link between the timing of the giant impact and the way the Moon initially formed.

"Depending on how hot the Earth and Moon are prior to the collision, the impact can destroy Theia and produce this massive disk of debris that eventually forms the Moon," Denton said. "But when we used the same parameters as original impact modeling - down to the equal temperature structures inside both bodies - within around five hours, an intact moon emerged."

Previous simulations have also produced intact moons, but the new work is the first to demonstrate that material strength and temperature can be central in determining whether the Moon forms largely intact or gradually assembles from processed debris in a protolunar disk.

New Clues to the Timing of the Moon's Birth

The findings could potentially connect the Moon's present-day properties with the physical state of Earth and Theia billions of years ago.

"These surprising and exciting new results imply a potential connection between the physical properties of the Moon today, including perhaps its volatile content, and the thermal state of the Earth and Theia at the time of the giant impact," said Canup, who was not involved in the study. "This in turn might help scientists better constrain when the moon-forming event occurred."

One major mystery, however, remains unresolved.

Earth and the Moon have remarkably similar compositions, something that giant impact models have long struggled to fully explain. The new simulations do not eliminate that problem.

"Because Earth and Mars formed in the same neighborhood of the solar system, they are like siblings," Denton said. "The Moon and Earth are more like fraternal twins."

One possibility is that Theia and the proto-Earth formed from material in the same region of the early solar system. Mars, whose composition differs from both Earth and the Moon, may have formed farther away.

A New Window Into the Moon's Origin

By showing that the internal condition of Earth and Theia could strongly influence the aftermath of their collision, the study gives researchers another way to investigate one of planetary science's biggest mysteries.

"We now know that the geophysical state of Earth and Theia play a fundamental role in shaping the outcome of the collision," said study co-author Namya Baijal, a doctoral student in Asphaug's group. "This gives us a new way to explore the conditions of the impact and what they might reveal about the Moon's origin." %!s()