Nvidia licenses Groq technology as Meta acquires Manus

Groq and Nvidia announced a licensing agreement for Groq’s inference technology on December 24, 2025. Groq founder Jonathan Ross, president Sunny Madra, and other team members will join Nvidia to develop and scale the licensed technology. Groq said it would continue operating independently.¹

The agreement comes as specialized chips and low-latency systems compete with Nvidia’s GPU-based approach to AI inference.² Citing a Groq investor, CNBC reported that Nvidia would pay about $20 billion for Groq assets. Neither company has confirmed that amount or an acquisition of Groq as a company.³¹ The public announcements also leave the extent of any asset transfer beyond the license and staff moves unclear.⁴¹

Meta announced its acquisition of AI agent startup Manus in late December 2025.¹² Based in Singapore, Manus develops agents that can carry out tasks autonomously.¹ Meta plans to bring the technology into its consumer and business products, including Meta AI.¹² It also plans to keep operating and selling the existing Manus service.² Meta did not disclose a purchase price; The Wall Street Journal reported that the deal exceeded $2 billion.²³

The acquisition adds to Meta’s push into general-purpose AI agents as other major technology companies also invest in AI.⁴ For Manus customers, the distinction matters: integrating its technology into Meta products is not, under the announced plans, a shutdown of the existing service.² The companies have not specified which features will become available in Meta products or when.¹²

Federal prosecutors in Texas unsealed documents on December 8, 2025, concerning an alleged smuggling network with links to China. Investigators say the network attempted to move at least $160 million worth of export-controlled NVIDIA H100 and H200 chips to China between October 2024 and May 2025. The investigation is known as “Operation Gatekeeper,” and one participant has pleaded guilty.¹²

The case illustrates the difficulty of enforcing chip export controls across hardware supply chains. Investigators allege that the network used shell companies, false shipping documents, and relabeled equipment to conceal the shipments. Allegations against the other defendants have not been established by a final conviction. The $160 million figure describes attempted exports, not a confirmed total of chips delivered to China.¹²

OpenAI combined its audio research, engineering, and product teams over the past two months, according to TechCrunch.¹ The company is reportedly building a model designed to sound more natural, handle interruptions better, and respond while a user is still speaking.¹ The report expects the model in early 2026 and a related device roughly a year later. Those timelines have not been officially confirmed.¹

Such a device could make AI conversations less dependent on screens. Its form and intended uses remain unclear, however.¹ OpenAI already documents models for real-time voice conversations, but that documentation does not establish what the proposed device would do or when it might launch.² For businesses, improvements to conversational audio are a more concrete development to watch than the prospect of new hardware.

Europe’s planned AI data center expansion is putting power supply in sharper focus. The International Energy Agency says the EU’s goal of expanding data centers will require solutions to grid strain.¹ Shortly before that warning, Goodman Group and Canada Pension Plan Investment Board agreed to partner on data center projects in Frankfurt, Amsterdam, and Paris.² EdgeMode also reported progress in securing power supply for its planned sites in Spain.³

The consequences extend beyond electricity bills. A French study points to rising power consumption and environmental pressures involving emissions, water, and rare metals.⁴ Operators also need grid connections and competitively priced electricity.⁴ Whether the announced projects can obtain grid access and clean power quickly enough remains unclear. Investment commitments alone do not settle the tension between AI ambitions and climate goals.⁵¹

DeepSeek researchers presented the publicly available paper “mHC: Manifold-Constrained Hyper-Connections” in late December 2025.¹² Their proposal constrains the matrices used to mix residual paths in Hyper-Connections to the Birkhoff polytope. This is intended to preserve the identity mapping that helps stabilize residual networks but can be weakened by earlier Hyper-Connections.¹

Hyper-Connections provide additional routes for information between network layers, but can make training unstable.¹ mHC aims to limit that instability, while engineering optimizations are intended to keep the added memory and compute demands manageable.¹³ The reported experiments point to more stable training. Whether the approach works as well in other model architectures or production settings remains unclear. Companies evaluating it will need reproducible comparisons of memory use and compute costs, not just training curves.¹³

OpenAI was recruiting a Head of Preparedness in late December 2025. The leader would develop and coordinate capability evaluations, threat models, and mitigations for risks from advanced AI systems. The remit includes model misuse, cybersecurity threats, and potential mental health harms. The position sits within the company's safety organization; the announcement was a job posting, not confirmation that someone had been hired.¹²

The role is intended to lead the technical execution of OpenAI's Preparedness Framework. Its scope puts attention on how the company assesses and limits severe risks before deploying new systems. Former Head of Preparedness Aleksander Madry moved to another role in July 2024. Whether a new hire will change specific evaluation or release decisions remains unclear.¹³

Instagram chief Adam Mosseri argued in late December 2025 that real photos should be verifiable from the moment they are captured.¹² Camera makers could cryptographically sign images when they are taken.¹² That would provide a way to trace their origin rather than relying solely on attempts to identify AI-generated images afterward.¹² Mosseri presented this as a possible approach to synthetic media, not a new Instagram feature.¹²

In his view, AI-generated images are becoming harder to recognize by appearance alone.¹² Provenance would shift attention to how an image was created. The approach would require camera makers to support signatures and platforms to make the information accessible to users.¹² Whether Instagram will introduce such verification remains open: Mosseri gave no timeline and announced no new feature.¹²

Microsoft CEO Satya Nadella published his outlook for AI in 2026 at the end of the year.¹² He argued that AI should support people rather than replace them.¹³ Instead of putting individual models at the center, the industry should build coordinated systems. It should also be more deliberate about where AI provides real value and earns public trust.³²

The remarks describe a direction for the industry, not an announced Microsoft program.³ Nadella sees a gap between advancing model capabilities and their usefulness in everyday settings: capable models are not enough without systems that make them dependable in practice.² For businesses, that puts reliable applications and measurable results ahead of impressive demonstrations.³² Whether Microsoft will turn these ideas into specific products remains unclear.³