Techin' Geek

During his keynote on Monday that launched the company’s annual GTC Conference in San Jose, California, Nvidia CEO Jensen Huang shared a multitude of figures — predominantly technical ones.

However, there was one financial statistic that surely caught the attention of investors: his forecast indicating there would be $1 trillion in orders for Nvidia’s Blackwell and Vera Rubin chips, a financial indicator of a flourishing AI market.

About an hour into his address, Huang pointed out that last year Nvidia experienced around $500 billion in demand for its Blackwell and forthcoming Rubin chips up to 2026.

“Now, I don’t know if you guys feel the same way, but $500 billion is an enormous amount of revenue,” he stated. “Well, I’m here to tell you that from where I stand — just a few months post-GTC DC, one year post the last GTC — right here where I am, I foresee through 2027, at least $1 trillion.”

The Rubin computing chip architecture, first unveiled in 2024, has been characterized by Huang as the pinnacle of AI hardware, surpassing its Blackwell predecessor. The company announced in January, when it formally commenced production of Rubin, that it would operate 3.5 times faster than the Blackwell architecture on model-training tasks and 5 times quicker on inference tasks, achieving up to 50 petaflops.

Nvidia has indicated that it anticipates increasing production in the latter part of the year.

Senator Elizabeth Warren (D-MA) addressed a letter to Defense Secretary Pete Hegseth on Monday voicing her concerns regarding the Pentagon’s choice to grant Elon Musk’s firm xAI access to classified networks.

“Grok, the contentious AI model created by xAI, has produced troubling results for users, including offering ‘advice on committing murders and terrorist acts,’ generating antisemitic material, and creating child sexual abuse content,” the correspondence states. 

Warren noted Grok’s “clear absence of sufficient safeguards” could present “serious threats to the safety of U.S. military personnel and the cybersecurity of classified systems.” She insisted that Hegseth provide details on how the Department of Defense intends to “address these possible national security dangers.”

Warren is not the first to raise concerns about Grok, xAI’s contentious chatbot, obtaining access to classified systems. Last month, a coalition of nonprofits called on the government to promptly halt Grok’s deployment in federal agencies, including the DoD, after X users persistently urged the chatbot to transform actual photos of women, and in some instances children, into sexualized images without their consent. On the same day Warren dispatched her letter, a class action lawsuit was initiated against xAI claiming Grok had produced sexual content from authentic images of the plaintiffs as minors. 

The letter follows the Pentagon’s ruling to classify Anthropic as a supply chain risk after the AI company declined to grant the military unrestricted access to its AI systems. Until recently, Anthropic was the sole AI firm with systems ready for classified use. Amid that dispute, the DoD secured an agreement with OpenAI and xAI to utilize their AI systems within classified networks, as reported by Axios. 

A senior official at the Pentagon verified that Grok was integrated for use in a classified environment but is not currently operational.

“It remains uncertain what guarantees or documentation xAI has presented to the Department of Defense regarding Grok’s security measures, data management practices, or safety protocols, and whether the DoD has assessed those assurances prior to reportedly permitting Grok access to classified systems,” Warren states.

Warren sought a copy of the agreement allegedly made between the DoD and xAI concerning the use of Grok in classified systems and an explanation of how the department plans to prevent Grok from being vulnerable to cyberattacks and ensure it will “not disclose sensitive or classified military information.”

(Last week, a former employee from Musk’s Department of Government Efficiency supposedly stole personal data from Americans at the Social Security Administration and saved it on a thumb drive — marking the latest allegation of DOGE-related data leakage.)

Chief Pentagon spokesperson Sean Parnell stated that the department “anticipates deploying Grok to its official AI platform GenAI.mil in the very near future.”

GenAI.mil is the military’s secure enterprise platform for generative AI, providing DoD personnel access to large language models (LLMs) and other AI resources within government-sanctioned cloud environments. It is primarily designed to assist with non-classified tasks such as research, document drafting, and data analysis. 

Shawn Shen asserts that for AI to thrive in the physical realm, it must have the capability to remember what it observes. Shen’s venture, Memories.ai, employs Nvidia AI technologies to establish the framework for wearables and robotics to retain and retrieve visual recollections.

On Monday, at its GTC conference, Memories.ai revealed a partnership with the semiconductor powerhouse Nvidia. This collaboration allows Memories.ai to leverage Nvidia’s Cosmos-Reason 2, a reasoning vision language model, along with Nvidia Metropolis, a tool designed for video search and summarization, to further enhance its visual memory capabilities.

Shen (shown above left) informed TechCrunch that he and his co-founder and CTO, Ben Zhou (depicted above right), conceived the idea for the company while developing the AI behind Meta’s Ray-Ban sunglasses. The creation of these AI glasses prompted them to consider how users would effectively engage with the technology in real-world scenarios if they could not remember the recorded video data.

They searched for existing solutions focused on that specific visual memory aspect for AI but, finding none, made the decision to depart from Meta and create the solution independently.

“AI is excelling in the digital arena. How about in the physical space?” Shen remarked. “AI wearables and robotics require memories too. Ultimately, visual memories for AI are essential. We envision that future.”

The capacity of AI systems to retain information is a relatively recent development. OpenAI enhanced ChatGPT to start remembering prior conversations in 2024 and refined that capability in 2025. Elon Musk’s xAI and Google Gemini have both introduced their own memory functionalities over the last two years.

However, Shen pointed out that these developments primarily concentrate on text-based memory. Text-based memory is more organized and simpler to index but falls short for physical AI applications that predominantly engage with their environment visually.

Founded in 2024, Memories.ai has secured $16 million to date, garnered from an $8 million seed round in July 2025 complemented by an $8 million extension. The funding round was spearheaded by Susa Ventures and included participants such as Seedcamp, Fusion Fund, and Crane Venture Partners.

Shen indicated that creating this visual memory layer successfully necessitated two elements: establishing the required infrastructure to assimilate and organize videos into a data-friendly format for storage and retrieval, along with gathering the necessary data to train the model to achieve that goal.

The organization unveiled its large visual memory model (LVMM) in July 2025. Shen mentioned it as a smaller counterpart to Gemini Embedding 2, a multimodal indexing and retrieval model released earlier this month.

For data collection, the company developed LUCI, a device worn by its “data collectors” that captures video to train the model. Shen stated they do not intend to evolve into a hardware company or market these devices, but rather created their own as they were dissatisfied with existing high-definition video recorders that consumed too much battery.

The firm has rolled out the second iteration of this LVMM and has partnered with Qualcomm to operate on Qualcomm’s processors starting later this year.

Memories.ai is collaborating with some major wearable companies already, according to Shen, though he refrained from naming them. Despite current demand, he anticipates even greater potential in wearables and robotics in the future.

“In terms of commercialization, we are concentrating more on the model and infrastructure, because we ultimately believe the wearables and robotics market will emerge, but it may not be immediate,” Shen explained.

In the past ten years, the electrical grid has undergone more transformation than in the previous fifty. Solar power, wind energy, and battery technology have shifted electricity generation from centralized producers. However, the grid continues to face the same core issues.

“The issue with the grid is primarily a peak issue. Typically, power levels are sufficient, but during peak periods, there may not be enough available,” stated Michael Phelan, co-founder and CEO of GridBeyond, in a conversation with TechCrunch.

Currently, these power shortages are especially affecting technology firms and data center operators, which require substantial electricity to train and run AI systems. 

“However, if you have adequate energy stored in a battery or can reduce an industrial load, which can reach hundreds of megawatts, you can initiate the establishment of those hyperscalers,” Phelan added.

GridBeyond is developing both hardware and software to integrate various segments of the grid into larger virtual power plants. The startup is already responsible for managing approximately 1 gigawatt of solar, batteries, wind, and hydropower, and has “several gigawatts” on the demand side across commercial and industrial sites, according to Phelan.

To broaden its offerings, GridBeyond recently secured a €12 million ($13.8 million) equity round led by Samsung Ventures, as exclusively reported by TechCrunch. Other investors include ABB, Act Venture Capital, Alantra’s Energy Transition Fund, Constellation Technology Ventures, EDP, Energy Impact Partners, Enterprise Ireland, Klima, Mirova, and the Japanese electronics and software company Yokogawa.

The startup has implemented its hardware controllers in batteries and renewable energy facilities, as well as substantial commercial and industrial operations across Australia, Ireland, Japan, the U.K., and the United States.

Similar to various virtual power plant companies, GridBeyond, based in Dublin, originated on an island. As Ireland began incorporating wind energy, Phelan noted, “they encountered an issue due to their island status, requiring them to balance the grid. Consequently, it was very applicable for them to have flexible loads they could introduce into the market.”

For an extended period, grid operators have requested large consumers to curtail their power consumption during extreme heat events, often providing financial incentives for reducing peak demand. This method is more cost-effective than constructing new transmission infrastructure or energy generation facilities. As renewable energy has proliferated, this strategy has been adopted more widely, enabling industrial and commercial clients to lessen their energy usage at night or during periods of low wind.

Recently, batteries have introduced a new element. GridBeyond oversees several significant energy storage projects, including a 200-megawatt battery in California. This flexible supply source helps address fluctuations in renewable energy output.

Batteries present another benefit: they can respond to demand much more swiftly than conventional peaking power plants, which may take minutes to activate. This capability allows the company to engage in rapid buying and selling of power through arbitrage. 

Additionally, it opens up new opportunities for data centers. Many facilities do not require a constant power supply, instead peaking during AI training sessions. These peaks can disrupt the grid—“akin to the issue that caused the collapse of the Spanish grid, which is not desirable,” Phelan explained. Batteries located at data centers can mitigate much of this load, stabilizing the site’s demand pattern on the grid to prevent unwanted fluctuations. 

By connecting to a nearby virtual power plant or utilizing batteries on-site, “it’s evidently simpler for them to establish a connection,” he mentioned.

Update 5:30 pm ET: The article was revised to specify that the investment originated from Constellation’s venture division, Constellation Technology Ventures.

Apple has purchased MotionVFX, a firm specializing in the creation of plug-ins, templates, and advanced features for Apple’s Final Cut Pro video editing application. The financial details of the transaction remain undisclosed.

“We are incredibly thrilled to announce that MotionVFX is becoming part of the Apple family to further enable creators and editors to excel in their work,” MotionVFX stated in a post on its website. “For more than 15 years, our goal has been to develop top-tier, visually impactful content and effects for video editors. From the outset, our focus has been on quality, user-friendliness, and exceptional design. These principles are also what we respect most in Apple’s offerings, and we are excited to unite our visions.”

Apple, which seldom reveals its acquisitions publicly, did not provide a response to TechCrunch’s inquiry for comment. TechCrunch will revise this article should Apple reply.

Founded in 2009 and based in Warsaw, MotionVFX provides subscription services starting at $29 monthly for access to its professional-level video editing solutions, graphics, and templates. Apple’s acquisition of the company is expected to lead to the integration of MotionVFX’s tools within the tech titan’s offerings.

Moreover, this strategy could enhance Apple’s competitiveness against Adobe Premiere Pro and the entire Adobe Creative Cloud ecosystem.

In January, Apple introduced Creator Studio, a subscription package that gives access to six creative applications along with premium content in iWork applications. The package is priced at $12.99 per month or $129 annually and features access to Final Cut Pro, Logic Pro, and Pixelmator Pro on both Mac and iPad, as well as Motion, Compressor, and MainStage on Mac. It also offers premium content for Keynote, Pages, and Numbers.

Through the acquisition of MotionVFX, Apple aims to draw more subscribers to its Creator Studio package.

In recent years, Apple’s services sector has emerged as one of its most significant growth catalysts. During the last fiscal year, this division represented over 26% of revenue, a remarkable rise compared to the 8.5% reported in 2015.

Three unidentified plaintiffs contended in a lawsuit submitted on Monday in California federal court that Elon Musk’s company xAI must be held liable for permitting its AI systems to create abusive sexual images of recognizable minors.

The trio of plaintiffs seeks to initiate a class action lawsuit on behalf of any individual whose genuine images as minors have been transformed into sexual content by Grok. They claim that xAI failed to implement fundamental safety measures employed by other leading laboratories to stop their image models from generating pornography involving real individuals and minors.

The legal case, Jane Doe 1, Jane Doe 2, a minor, and Jane Doe 3, a minor versus x.AI Corp. and x.AI LLC, has been filed in the United States District Court for the Northern District of California.

Other deep learning image-generating technologies utilize diverse strategies to avert the production of child pornography from ordinary photographs. The lawsuit asserts that xAI did not embrace these standards.

Significantly, if a model is capable of generating nude or erotic content from real pictures, it becomes nearly impossible to prevent it from creating sexual content featuring children. Musk’s public endorsement of Grok’s capability to produce sexual imagery and portray real individuals in revealing outfits is a fundamental aspect of the case.

The company has not commented on a request from TechCrunch.

One plaintiff, Jane Doe 1, had images from her high school homecoming and yearbook altered by Grok to show her naked. An anonymous informant who reached out to her on Instagram alerted her that the images were spread online, providing her with a link to a Discord server containing sexualized pictures of her and other minors she recognized from school.

A second plaintiff, Jane Doe 2, was alerted by law enforcement about altered, sexualized depictions of her generated by a third-party mobile application that utilizes Grok models. A third individual, Jane Doe 3, was also informed by law enforcement who found an altered, pornographic image of her on the device of a suspect they had detained. The plaintiffs’ attorneys assert that since third-party utilization still necessitates xAI code and servers, the company should be deemed responsible.

All three plaintiffs, two of whom are minors, report severe distress related to the dissemination of these images and the implications it may hold for their reputations and social interactions. They are requesting civil penalties under various laws designed to safeguard exploited children and avert corporate negligence.

During the Nvidia GTC keynote on Monday, CEO Jensen Huang unveiled DLSS 5, an advanced iteration of the chipmaker’s AI graphics technology aimed at enhancing video game realism while reducing computational requirements. 

The innovative DLSS 5 framework integrates conventional 3D graphics data with generative AI models capable of predicting and completing sections of images, enabling Nvidia’s GPUs to render intricate scenes and realistic characters without having to recreate every detail from scratch. 

“We merged controllable 3D graphics, the authentic representation of virtual environments, the organized data… with generative AI, probabilistic computing,” Huang stated in his keynote. “One is entirely predictive, while the other is probabilistic but highly realistic.”

Huang remarked that the fusion of these two elements—structured data and generative AI—empowers developers to craft content that is “stunning, incredible, and also manageable.”

“The integration of structured information with generative AI will recur across various sectors,” Huang predicted. “Structured data serves as the backbone of reliable AI.”

Although gaming currently constitutes a smaller segment of Nvidia’s revenue compared to its historical significance, it remains the industry that established Nvidia’s trajectory. Huang presented DLSS 5’s methodology as a representation of a wider computational transition, implying its potential application beyond gaming into enterprise computing. 

The billionaire leader highlighted enterprise data solutions like Snowflake, Databricks, and BigQuery as instances of structured datasets that upcoming AI systems may analyze and derive insights from. 

“Looking ahead, these data structures will be leveraged by AI, which will operate at a significantly higher speed than humans,” Huang declared. “Future agents will utilize both structured databases and unstructured, generative databases. These databases encompass the vast majority of global information.”

Apple’s latest MacBook Neo is not only the least expensive MacBook yet — it also ranks as the most repairable MacBook in “approximately fourteen years,” based on a thorough teardown by the how-to site iFixit.

According to iFixit, the “key highlight” is the battery. While previous MacBook batteries were allegedly glued in, the Neo’s battery is located in a tray fastened with 18 screws. Although this is a considerable number of screws, iFixit stated that “screws always outperform adhesive.” This new setup — which should facilitate battery replacement for users — “elicited cheers throughout the iFixit workspace.”

Additional modifications that supposedly enhance the repairability of the MacBook Neo include a straightforward disassembly tree, the Repair Assistant’s acceptance of replacement components without issues, and a display and keyboard that are simpler to replace.

In the end, iFixit had enough reservations — such as soldered RAM and storage — to assign the Neo a repairability rating of 6 out of 10. They noted that for a MacBook, “this is a solid score.”

In 2012, Apple revealed a redesign of its MacBook line featuring the new Retina Display MacBook Pro. This redesign marked a shift towards handling repairs internally or through Apple-endorsed shops, sacrificing much of the DIY repairability offered by previous versions. Many components in the new MacBooks, such as memory cards and batteries, were glued down; iFixit rated that model a 1 out of 10 for repairability.

Since that time, Apple has been criticized for its opposition to certain right to repair legislation in various states and for creating what are referred to as parts-pairing barriers, which restricted or prevented devices from functioning with non-manufacturer components. Legislative changes that permitted individuals to repair gadgets in states like Oregon prompted the company to ease limitations on repairing iPhones with older components.

The new MacBook Neo still has soldered memory and storage, complicating upgrades. “However, the parts that typically fail first are more accessible than they have been in many years on a MacBook,” iFixit remarked. “The battery is secured with screws rather than glue. The ports are modular. The display is easier to replace. The internal layout is surprisingly logical.”

Nvidia is set to commence its yearly GTC developer conference in San Jose, California, on Monday, with CEO Jensen Huang’s keynote planned for 11 a.m. PT / 2 p.m. ET.

GTC — which signifies GPU Technology Conference — is Nvidia’s premier yearly gathering, taking place from March 16 to March 19. The semiconductor giant generally utilizes this platform to reveal new products, highlight collaborations, and present its vision for the future of computing. Huang’s keynote will emphasize Nvidia’s contribution to the future of computing and AI. Attendees can view the two-hour address live at the SAP Center or stream the presentation on the event’s website. The YouTube livestream is integrated below.

The extensive four-day event concentrates on future AI developments across various sectors, such as healthcare, robotics, and self-driving vehicles.

On the software front, there are speculations that Nvidia may unveil an open-source platform for enterprise AI agents, referred to as NemoClaw, as initially reported by Wired. This platform would provide companies a systematic approach to create and implement AI agents (software capable of performing complex tasks independently) and would position Nvidia to align with similar offerings from firms like OpenAI.

On the hardware front, the company is also believed to be introducing a new chip intended to enhance the AI inference process — the method whereby an AI model utilizes learned information to generate answers or make choices, distinguishing it from the initial training phase that demands significantly more computational resources. Accelerating and reducing the cost of inference is widely recognized as one of the remaining hurdles to broad AI application scalability. This chip would signify Nvidia’s latest effort to lead not only the training sector, where it already holds an estimated 80% market share, but also the inference market, where custom chips from Google, Amazon, and others are rapidly emerging as competitors.

Moreover, a variety of partnership announcements and demonstrations highlighting Nvidia’s AI capabilities across sectors are expected.

Kevin Cook, a senior equity strategist at Zacks Investment Research, informed TechCrunch that attendees should also anticipate insights into what the company intends to do regarding its partnership with Groq, the inference firm Nvidia reportedly invested $20 billion in late last year to acquire its technology. There is considerable interest around this collaboration, particularly since Jonathan Ross, Groq’s founder; Sunny Madra, Groq’s president; and additional members of the Groq team have agreed to join Nvidia to further develop and scale that licensed technology.