AI has reached a critical juncture, becoming more intelligent and useful due to its reasoning ability. This advancement has led to a significant increase in computational requirements, with the industry needing much more computing power than previously anticipated. 

The generation of tokens for reasoning is a key factor in this increased demand, according to NVIDIA CEO Jensen Huang, who recently addressed the future of AI and computing infrastructure at the GTC 2025 summit in San Jose earlier this week. 

His keynote highlighted AI’s rapid evolution and the immense computational power required to support its growth. “Every single data centre in the future will be power-limited. We are now a power-limited industry,” he said.

With AI models growing exponentially in complexity and scale, the race is on to build data centres, or what Huang calls “AI factories”, that are not only massively powerful but also energy-efficient.

The Rise of the AI Factory

Huang introduced the concept of AI factories as the new standard for data centre infrastructure. These centres, which are no longer simply repositories of computation or storage, have a singular focus—to generate the tokens that power AI. 

He described them as “factories because it has one job, and that is to generate these tokens that are then reconstituted into music, words, videos, research, chemicals, or proteins”.

AI factories, according to Huang, are becoming the foundation for future industries. “In the past, we wrote the software, and we ran it on computers. In the future, the computer is going to generate the tokens for the software.”

Huang predicts a shift from traditional computing to machine learning-based systems. This transition, combined with AI’s growing demand for infrastructure, is expected to drive “data centre buildouts to a trillion-dollar mark very soon”, he believes.

Power Problem is Also a Revenue Problem

As data centres expand, they will face significant power limitations. This underscores the need for more efficient technologies, including advanced cooling systems and chip designs, to manage energy consumption effectively.

Huang noted that the computational requirements for modern AI, especially reasoning and agentic AI, are “easily a hundred times more than we thought we needed this time last year”. 

This explosion in demand places enormous strain on data centres’ energy consumption. His keynote made it clear that moving forward, energy efficiency isn’t just a sustainability concern; it will be directly tied to profitability.

“Your revenues are power limited. You could figure out what your revenues will be based on the power you have to work with,” he said. 

This shift will influence everything from how AI models are trained and deployed to how entire industries operate. In this regard, power is the ultimate constraint in AI-dominated computation. This limitation is reshaping both the design and operation of data centres around the world.

“The more you buy, the more you make,” Huang quipped, encouraging businesses to view their investments in NVIDIA’s accelerated computing platforms as the key to unlocking the full potential of AI-driven value creation.

Scaling Up Before Scaling Out

Huang explained NVIDIA’s approach to managing this power limitation, which would be a fundamental rethinking of scale. 

“Before you scale out, you have to scale up,” he stated. NVIDIA’s new Blackwell platform demonstrates this principle with its extreme scale-up architecture, featuring “the most extreme scale-up the world has ever done”. 

A single rack delivers an astonishing one-exaflop performance within a fully liquid-cooled, high-density design.

By scaling up, data centres can dramatically reduce inefficiencies that occur when spreading workloads across less integrated systems. 

Huang explained that if data centres had scaled out instead of scaling up, the cost would have been way too much power and energy. He pointed out that, as a result, deep learning would have never happened.

Blackwell, a Path to 25x Energy Efficiency

With the launch of NVIDIA’s Blackwell architecture, Huang highlighted a leap in performance and efficiency. According to him, the goal is to deliver the most energy-efficient compute architecture you can possibly get.

Huang believes NVIDIA has cracked the code for future-ready AI infrastructure by combining innovations in hardware, such as the Grace Blackwell system and NVLink 72 architecture, with softwares like NVIDIA Dynamo, which he described as “the operating system of an AI factory”.

Explaining the broader significance, he said, “This is ultimate Moore’s Law. There’s only so much energy we can get into a data centre, so within ISO power, Blackwell is 25 times [better].”

AI Factories at Gigawatt Scale

NVIDIA’s ambitions don’t stop with Blackwell. Huang outlined a roadmap extending years into the future, with each generation bringing new leaps in scale and efficiency. 

Upcoming architectures like Vera Rubin and Rubin Ultra promise “900 times scale-up flops” and AI factories at “gigawatt” scales.

As these AI factories become the standard for data centre design, they will rely heavily on advancements in silicon photonics, liquid cooling, and modular architectures. 

Huang likened the current AI revolution to the dawn of the industrial era, naming NVIDIA’s AI factory operating system Dynamo in homage to the first instrument that powered the last industrial revolution. 

“Dynamo was the first instrument that started the last industrial revolution—the industrial revolution of energy. Water comes in, electricity comes out. [It’s] pretty fantastic,” he said. “Now we’re building AI factories, and this is where it all begins.”

Cohesity, a data security platform, has announced a significant expansion of Cohesity Gaia, its enterprise knowledge discovery assistant. This development introduces what is claimed to be one of the industry’s first AI-powered search capabilities for backup data stored on-premises. 

This marks a major leap in the enterprise data management ecosystem. By leveraging NVIDIA’s accelerated computing and enterprise AI software, including NVIDIA NIM microservices and NVIDIA NeMo Retriever, Cohesity Gaia seamlessly integrates generative AI into backup and archival processes. 

This enables enterprises to enhance efficiency, innovation, and overall growth potential through deeper data insights. 

Pat Lee, vice president of strategic enterprise partnerships at NVIDIA, highlighted the benefits of this collaboration, and said, “Enterprises can now harness AI-driven insights directly within their 8 to preserve data accessibility and security while unlocking new levels of intelligence.”

This solution will be compatible with Cisco Unified Computing System (UCS), Hewlett Packard Enterprise (HPE), and Nutanix and offer various deployment options.

Moreover, customers like JSR Corporation, a Japanese research and manufacturing company, are also evaluating the benefits of this solution.

As enterprises adopt hybrid cloud strategies, many retain critical data on-premises to meet security, compliance, and performance requirements. By extending Gaia to these environments, organisations can adopt high-quality data insights while maintaining control over their infrastructure.

Sanjay Poonen, CEO and president of Cohesity, also emphasised the importance of on-premises AI solutions.

Cohesity Gaia now offers enterprises enhanced speed, accuracy, and efficiency in data search and discovery. Its multi-lingual indexing and querying capabilities allow global organisations to analyse data in multiple languages.

The infrastructure is scalable and customisable to meet business requirements, with a reference architecture designed for seamless deployment across hardware platforms. 

Pre-packaged large language models (LLMs) on-premises ensure that backup data remains secure without cloud access. Its optimised architecture allows efficient searches across petabyte-scale datasets, making retrieval fast and reliable.

Cadence Design Systems, a leading computational software company, has announced that it is expanding its multi-year collaboration with NVIDIA, focusing on accelerated computing and agentic AI. 

This partnership addresses global technology challenges by driving innovation across various industries and involves Cadence leveraging NVIDIA’s latest Blackwell architecture to accelerate its engineering and scientific solutions. 

This includes reducing computational fluid dynamics simulation times by up to 80 times, from days to minutes, and accelerating the Cadence Spectre X Simulator by up to 10 times.

Jensen Huang, NVIDIA CEO, noted, “Accelerated computing and agentic AI are setting new standards for innovation across industries.” 

Using its Fidelity CFD Platform, Cadence also successfully ran multi-billion cell simulations on NVIDIA GB200 GPUs in under 24 hours. It would have previously required a top 500 CPU cluster with 100,000 cores and several days to complete. 

The company expressed that it will continue to leverage Blackwell for simulation and help the aerospace industry reduce the amount of wind tunnel tests by reducing cost and expediting time to market.

New Era for Accelerated Computing

Additionally, the partnership involves the companies working together on a full-stack agentic AI solution for electronic and system design, as well as science applications. This will integrate Cadence’s JedAI Platform with NVIDIA’s NeMo generative AI framework and the Llama Nemotron Reasoning Model. 

Anirudh Devgan, president and CEO at Cadence, says, “We’re enabling the delivery of today’s infrastructure AI and agentic AI and transforming the principled simulations that underpin physical AI and sciences AI.”

The collaboration is expected to transform industries by enabling complex simulations that were previously impossible, driving efficiency, and fueling scientific discovery. It will also deliver breakthroughs in simulation, optimisation, and design. 

In his keynote address at the NVIDIA 2025 GTC summit, he mentioned that until now, the giant had been using general-purpose computers running software super slowly to design accelerated computers for everybody else. 

But with the entry of optimised CUDA software, “now our entire industry is going to get supercharged as we move to accelerated computing.”

Cadence Molecular Sciences (OpenEye) is also integrating NVIDIA BioNeMo NIM microservices with its cloud-native molecular design platform, Orion. Cadence has also been one of the first adopters of NVIDIA Omniverse Blueprint for AI factory digital twins. 

Cadence and NVIDIA are leading the way in creating an ecosystem of high-quality models, allowing equipment manufacturers and data centre companies to quickly create digital twins.

At NVIDIA’s GTC 2025 event on Tuesday, the company delivered a variety of new advancements across AI hardware, personal supercomputers, self-driving cars, and humanoid robots. Moreover, the event took an unexpected turn when an unlikely guest made an appearance.

Surely, if Pat Gelsinger was still the CEO of Intel, there’s no way he’d be seen mingling with CEO Jensen Huang at an NVIDIA event. That said, Gelsinger certainly didn’t hold back and offered a few strong takes on the industry. 

He participated in a panel discussion alongside the hosts of the Acquired podcast and several other industry experts. While Gelsinger applauded NVIDIA’s accomplishments in the present era of AI, he disagreed with Huang on certain key issues—specifically, the timeline for the arrival of quantum computing and the use of GPUs for inference. 

‘Data Centres Will Have CPUs, GPUs, and QPUs’

Gelsinger, who is notably bullish on quantum computing, stated that it could be realised within the next few years. 

This stands in contrast to Huang’s comments earlier this year, where he said that bringing “very useful quantum computers” to market could take anywhere from 15 to 30 years. His statements triggered a massive selloff in the quantum computing sector, wiping out approximately $8 billion in market value. 

“I disagree with Jensen,” said Gelsinger, adding that the data centres of the future will have quantum processing units (QPUs) handling workloads, along with GPUs and CPUs. 

Similar to how GPUs are deployed to handle tasks for training AI models in language and human-like behaviour, Gelsinger believes it is only appropriate to have a quantum computing model for the complex parts of humanity. “Most interesting things in humanity are quantum effects,” he said. 

He added that many unsolved problems today run on quantum effects, and quantum computers would help realise many ideas like superconducting, composite materials, cryogenics and medical breakthroughs, among others.

“That’s why this is a thrilling time to be a technologist. I just wish I was 20 years younger to be doing more,” he said. 

While Gelsinger differs from Huang, he shares an optimistic view with Microsoft co-founder Bill Gates and Google. 

“There is a possibility that he (Huang) could be wrong. There is the possibility in the next three to five years that one of these techniques would get enough true logical qubits to solve some very tough problems,” said Gates to Yahoo Finance. 

Besides, even Microsoft and Amazon have already taken major strides in quantum computing within the first three months of the year. On the flipside, Meta CEO Mark Zuckerberg resonated with Huang. “My understanding is that [quantum computing] is still ways off from being a very useful paradigm,” Zuckerberg had said in a podcast episode a few months ago. 

Ironically, NVIDIA does seem to have huge plans for quantum computing. The company announced at the GTC event that it is building a Boston-based research centre to advance quantum computing. 

‘Huang Got Lucky With AI’

Besides, Gelsinger clarified that he isn’t a fan of GPUs for AI model inference—the process in which a pre-trained AI model applies its learnings to generate outputs.

He reflected on the early days when a CPU, or a cluster of them, was the undisputed “king of the hill” for running workloads on computer systems. When Huang decided to use a graphics device (GPU) for the same purpose, Gelsinger said that, in the end, he “got lucky” with AI. 

While he acknowledged that AI and machine learning algorithms demand the GPU architecture, which is where most of the developments are being made today, he also pointed out, “There’s a lot more to be done, and I’m not sure all of those are going to land on GPUs in the future.” 

While GPUs work well for training, Gelsinger added that there needs to be a more optimised solution for inference. “A GPU is way too expensive. I argue it’s 10,000 times too expensive to fully realise what we want to do with the deployment of inference of AI.” 

His sentiments are also reflected by the growing ecosystem of inference-specific hardware that is overcoming the inefficiencies posed by GPUs. Companies like Groq, Cerebras, and SambaNova have achieved tangible and useful real-world results for providing high-speed inference. 

For instance, French AI startup Mistral recently dubbed its app ‘Le Chat’ the fastest AI assistant by deploying inference on Cerebras’ hardware. 

Even Huang has acknowledged this in the past. In a podcast episode last year, he said that one of the company’s challenges is to provide efficient, high-speed inference. Having said that, companies working on AI inference hardware may not compete with NVIDIA after all.  

Jonathan Ross, CEO of Groq, said, “Training should be done on GPUs.” He also suggested that NVIDIA will sell every single GPU they make for training. 

All things considered, Gelsinger’s first outing post-resignation involved several strong statements. However, it remains clear that he’s still a massive fan of Huang and the work NVIDIA has accomplished. 

When DeepSeek made a significant impact on NVIDIA’s stock price, Gelsinger argued that the market reaction was wrong. He also revealed that he is an NVIDIA stock buyer, expressing that he was “happy” to benefit from the lower prices. 

Tech Mahindra and NVIDIA have announced a collaboration to develop an autonomous pharmacovigilance solution for improving drug safety management. According to the company blog, the solution uses Tech Mahindra’s TENO framework alongside NVIDIA AI Enterprise software, including NeMo, NIM microservices, and AI Blueprints.

The solution is being showcased at NVIDIA’s GTC 2025.

“AI is ideal for monitoring medicines throughout their lifecycle to support safety. Integrating AI into the Tech Mahindra TENO framework with NVIDIA AI Enterprise software enhances pharmacovigilance by augmenting human capabilities to help identify potential safety issues more effectively,” said John Fanelli, vice president of enterprise software at NVIDIA.

The goal is to reduce the risk of human error. The AI system automates pharmacovigilance workflows, handling case intake, data transformation, quality control, and compliance management. In this system, AI agents classify, prioritise, and verify pharmacovigilance emails. 

According to the companies, the solution can reduce turnaround times by up to 40%, increase data accuracy by 30%, and lower operational costs by 25%. The system processes adverse drug reaction (ADR) cases and supports regulatory compliance through autonomous decision-making.

Nikhil Malhotra, chief innovation officer at Tech Mahindra, said the collaboration with NVIDIA will help the pharmaceutical industry manage large volumes of data more efficiently. By applying generative AI and multi-agent systems, they will improve drug safety.

“Together, we are revolutionising drug safety management and using the innovative AI-driven framework to develop multiple use cases for our global customers,” Malhotra said 

Alongside Tech Mahindra, Wipro has also launched sovereign AI services with NVIDIA to help governments and businesses develop country-specific AI solutions. The services use Wipro’s WeGA Studio and NVIDIA AI software to support local language models, data privacy, and AI governance. The applications include healthcare, banking, education, and emergency services, focusing on data security and sovereignty.

If last year’s GTC was a pop concert, this year’s event felt like the Super Bowl of AI, said NVIDIA chief Jensen Huang. Only this time, there are no losers, just winners.

“I’m up here without a net—there are no scripts, there’s no teleprompter, and I’ve got a lot of things to cover,” he quipped. The event was attended by over 25,000 in-person attendees, while 3,00,000 joined virtually. This prompted Huang to joke that the only way to accommodate more people at GTC was to physically expand the size of San Jose, where the event was held.

This year, Huang, unfazed by DeepSeek’s impact, only means business. 

“The entire world got it wrong—the computation requirement and the scaling law of AI are more resilient and, in fact, hyper-accelerated. The amount of computation we need at this point, as a result of agentic AI and reasoning, is easily a hundred times more than we thought we needed this time last year,” Huang claimed.

Huang emphasised the shift from “retrieval computing” to “generative computing”,  where AI generates answers based on context. He introduced “agentic AI”, which involves AI perceiving, reasoning, and planning actions. Huang introduced the open Llama Nemotron family of models, which are equipped with reasoning capabilities. 

According to him, this evolution, coupled with “physical AI” for robotics, has significantly increased computation needs.

Building on this, Huang announced that  NVIDIA’s Blackwell architecture is now in full production, delivering 40 times the performance of Hopper. The Blackwell architecture boosts AI model training and inference, improving efficiency and scalability. 

Notably, Blackwell Ultra is set to hit systems later this year, but the real powerhouse is coming soon. Named after an American astronomer known for her research on dark matter, Vera Rubin is NVIDIA’s next-generation GPU and is expected to debut in 2026. Huang added that the Rubin chips will be followed by Feynman chips, which are slated to arrive in 2028.

To accelerate large-scale AI inference, Huang introduced NVIDIA Dynamo, an open-source platform that powers and scales AI reasoning models within AI factories. “It is essentially the operating system of an AI factory,” Huang said

Comparing Blackwell with older Hopper GPUs, Huang said, “The more you buy, the more you save. But now, the more you buy, the more you make.” He even joked that when Blackwell hits the market, customers won’t be able to give away Hopper GPUs and encouraged them to buy Blackwell. 

He further quipped that he is the ‘chief revenue destroyer’ as his statement might alarm his sales team, discouraging customers from purchasing remaining Hopper inventory. Ultimately, he stressed that technology is advancing so rapidly that buyers should invest in the latest and most powerful options available rather than settling for older models.

Moreover, he expects data centre revenue to reach one trillion dollars by the end of 2028.  “I am fairly certain we’re going to reach that very soon.”

NVIDIA is not just about big data centres. The company also unveiled two new supercomputers, DGX Spark and DGX Station, powered by the Grace Blackwell platform, allowing AI developers, researchers, and students to prototype, fine-tune, and run large models on desktops.

Cute Disney Robots on the Way

Huang is all in on robotics and is convinced that physical AI is the next big thing. On stage, he was joined by ‘Blue’, a small AI-powered robot resembling Disney’s Wall-E. He also introduced Newton, an open-source physics engine for robotics simulation developed in collaboration with Google DeepMind and Disney Research.

Not stopping there, NVIDIA unveiled Isaac GROOT N1, the world’s first open Humanoid Robot foundation model. If that wasn’t enough, Huang also announced NVIDIA Isaac GR00T Blueprint, a system that generates massive synthetic datasets to train robots—making it way more affordable to develop advanced robotics.

NVIDIA’s physical AI experience extends to autonomous vehicles. General Motors announced it will work with NVIDIA to optimise factory planning and robotics. Moreover, GM will use NVIDIA DRIVE AGX for in-vehicle systems to support advanced driver-assistance systems and enhanced safety features.

Huang also mentioned the importance of safety in Autonomous Vehicles (AVs) and launched NVIDIA Halos, a comprehensive safety system for AVs. It integrates NVIDIA’s automotive hardware and software safety solutions with AI research to ensure safe AV development from cloud to car. 

While AI remains a priority, NVIDIA is now focusing on quantum computing. It will host its first Quantum Day on March 20—an interesting move considering Huang once claimed that “quantum computers are still 15 to 30 years away”.

NVIDIA has announced two new personal AI supercomputers to handle AI workloads at the GPU Technology Conference (GTC) 2025 event. The company announced the DGX Spark and DGX Station, which are powered by the NVIDIA Grace Blackwell platform. 

These are aimed at helping AI developers, researchers, data scientists, and even students prototype, fine-tune, and infer from large language models on a desktop. Models can be run locally or deployed on any cloud-based platform. 

“DGX Spark and DGX Station bring the power of the Grace Blackwell architecture, previously only available in the data centre, to the desktop,” the company said. Original equipment manufacturer (OEM) partners like ASUS, Dell, HP and Lenovo are set to develop DGX Spark and the DGX Station. 

The DGX Spark, formerly known as ‘Project DIGITS’, is dubbed the world’s smallest AI supercomputer. It can deliver 1,000 trillion operations per second (TOPS) using AI models. It is priced at $3,000. The NVIDIA DGX Spark (5.91″ × 5.91″ × 1.99″) is only slightly larger than the Apple Mac Mini (5.00″ × 5.00″ × 1.96″).

On the other hand, the more powerful NVIDIA DGX Station features a massive 784 GB memory to accelerate AI workloads. It is also the first desktop system built with the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. NVIDIA says the DGX Stations are purpose-built for teams who need the best desktop AI development platform. 

“This is the computer of the age of AI. This is what computers should look like,” said CEO Jensen Huang in the keynote. 

The company has opened reservations for DGX Spark Systems. The DGX Station is expected to be available later this year. 

Besides, NVIDIA had plenty of announcements to make at the GTC event. The company also partnered with General Motors (GM) to develop AI-powered self-driving cars and HALOS, a new AI-enabled automotive safety platform. Furthermore, NVIDIA also announced plenty of updates in the robotics sector. 

Earlier this year, the company unveiled the GeForce RTX 5090 GPU, which is 30% smaller in volume and 30% better at energy dissipation than the RTX 4090.

NVIDIA CEO Jensen Huang announced a collaboration with General Motors (GM) during his keynote speech on Tuesday at the NVIDIA GTC conference in San Jose. The partnership aims to use AI for next-generation vehicle experience and manufacturing.

GM will leverage NVIDIA’s accelerated computing platforms, including NVIDIA Omniverse with NVIDIA Cosmos, to optimise factory planning and robotics. 

Additionally, GM will use NVIDIA DRIVE AGX for in-vehicle systems to support advanced driver-assistance systems and enhanced safety features. This system will be capable of performing up to 1,000 trillion operations per second, facilitating the development of safe AVs at scale. 

The collaboration extends GM’s existing use of NVIDIA’s GPU platforms for AI model training across various operations. 

Mary Barra, chair and CEO at GM, noted, “AI not only optimises manufacturing processes and accelerates virtual testing but also helps us build smarter vehicles while empowering our workforce to focus on craftsmanship.” 

The partnership aims to create digital twins of assembly lines using NVIDIA Omniverse, allowing for virtual testing and production simulations to minimise downtime. GM will also train robotics platforms for material handling and precision welding to enhance manufacturing safety and efficiency. 

“The era of physical AI is here, and with GM, we’re transforming transportation, from vehicles to the factories where they’re made,” Huang said.

Also, a Safety System for AVs

In addition to this announcement, Huang also mentioned the importance of safety in Autonomous Vehicles (AVs) and launched NVIDIA Halos, a comprehensive safety system for AVs. 

Halos integrates NVIDIA’s automotive hardware and software safety solutions with AI research to ensure safe AV development from cloud to car. 

Halos operates on three levels: technology, development, and computation. The technology level includes platform, algorithmic, and ecosystem safety. The development level incorporates guardrails for design time, deployment time, and validation time.

Key elements of Halos focus on platform, algorithmic, and ecosystem safety. Halos complements existing safety practices and can potentially accelerate standardisation and regulatory compliance.

US-based Keysight Technologies has partnered with Samsung and NVIDIA to develop AI models to improve radio access network (RAN) performance for 5G-advanced and 6G technologies. This collaboration enables Samsung to integrate these AI models into its virtual RAN (vRAN) software, which will be showcased at the Mobile World Congress (MWC) 2025 in Barcelona.

Developed under the AI-RAN Alliance, the initiative addresses key challenges in traditional RAN systems, such as limited throughput, high latency, and inefficient resource utilisation. 

This development comes right after the company’s recent collaboration with Hyundai to complete what is claimed to be the industry’s first end-to-end reduced capability (RedCap) trial over a private 5G network. 

Samsung will also unveil its first Android XR headset, Project Moohan, by integrating multimodal AI with advanced XR capabilities. 

Charlie Zhang, senior VP at Samsung Research America, stated, “This collaboration marks a significant milestone as we move toward AI-native and sustainable telecommunication networks.” 

AI-based channel estimation models have demonstrated a 30% improvement in cell edge throughput during lab tests compared to conventional methods. These advancements optimise resource allocation, enhance system capacity, and reduce power consumption.

Samsung’s AI model evaluation was conducted using an end-to-end setup featuring its radio point and distributed unit (DU) on NVIDIA’s AI Aerial platform powered by the GH200 Grace Hopper Superchip. 

Keysight’s Channel Emulation Solutions enabled precise testing under various conditions. Giampaolo Tardioli, GM and VP at Keysight, highlighted the transformative potential of this partnership in advancing energy-efficient networks.

Soma Velayutham, GM for AI, 5G and telecoms at NVIDIA, also emphasised the flexibility and efficiency gains unlocked by AI in signal processing pipelines. The collaboration is expected to drive innovation and accelerate the global adoption of AI-enhanced RAN technologies.

At MWC 2025, Jio Platforms Limited (JPL), along with AMD, Cisco, and Nokia, also announced plans to develop an Open Telecom AI Platform. It offers a ‘multi-domain intelligence’ framework to introduce AI and automation to ‘every layer’ of network operations. 

Even AWS has entered the game with its new cloud solutions to boost 5G networks for telecom operators. 
At the same event, SynaXG (a provider of AI-RAN solutions), in collaboration with Kyocera (Japanese manufacturer), and NVIDIA claimed to have unveiled the world’s first software-defined mmWave 5G vRAN solution.

NVIDIA announced another record-breaking quarter on Wednesday. Revenue surged to $39.3 billion, a 12% increase from the previous quarter and a 78% rise from the previous year. 

“Demand for Blackwell is amazing as reasoning AI adds another scaling law—increasing compute for training makes models smarter and increasing compute for long thinking makes the answer smarter,” said Jensen Huang, founder and CEO of NVIDIA.

He said that the world is at a nascent stage of reasoning AI and inference-time scaling, and multimodal AIs, enterprise AI, sovereign AI, and physical AI are right around the corner. “We will grow strongly in 2025,” said Huang.

He further noted that much has been accomplished with AI in two years, whereas it took decades to develop certain technologies, highlighting a greater potential for the AI ecosystem. 

“No technology has ever had the opportunity to address a larger part of the world’s GDP than AI. No software tool ever has. And so, this is now a software tool that can address a much larger part of the world’s GDP more than any time in history,” Huang said. 

The company’s CFO, Colette M. Kress, stated that it generated $11 billion in Blackwell revenue to meet the increasing demand, marking the fastest product ramp in its history.  

She added that at the upcoming GTC event, which is to be held between March 17 and March 21, the company will discuss Blackwell Ultra, Vera Rubin, and new computing and networking products.

Kress explained that with Blackwell, clusters of 100,000 GPUs or more will become common. “Shipments have already started for multiple infrastructures of this size.”

Last year, Microsoft became the first company to launch the Azure ND GB200 V6 VM series based on the NVIDIA GB200 Grace Blackwell Superchip, which features NVIDIA Grace CPUs and NVIDIA Blackwell GPUs. 

Most recently, Google Cloud announced ​​that it is bringing the Blackwell GPUs to Google Cloud with a preview of A4 VMs powered by NVIDIA HGX B200. Oracle also hosts Blackwell GPUs on its Zettascale cloud computing clusters.

DeepSeek Couldn’t Shake NVIDIA Yet

The launch of DeepSeek’s latest model, R1, which the company claims was trained on a $6 million budget, triggered a sharp market reaction. NVIDIA’s stock tumbled 17%, wiping out nearly $600 billion in value, driven by concerns over the model’s efficiency.

The model was launched in January, and its impact on compute demand may not be evident in the current Q4 results.

However, Huang said the company’s inference demand is accelerating, fuelled by test-time scaling and new reasoning models. “Models like OpenAI’s, Grok 3, and DeepSeek R1 are reasoning models that apply inference-time scaling. Reasoning models can consume 100 times more compute,” he said. 

“DeepSeek-R1 has ignited global enthusiasm. It’s an excellent innovation. But even more importantly, it has open-sourced a world-class reasoning AI model.”

Experts speculate that the Chinese company may not be revealing the whole truth. During an interview, the CEO of Scale AI, Alexandr Wang, said that he believed DeepSeek possessed around 50,000 NVIDIA H100s, but wasn’t permitted to talk about it.

Notably, Elon Musk’s xAI used 200,000 GPUs to train Grok 3. According to reports, tech giant Meta Platforms is discussing constructing a new data centre campus for its AI projects, with potential costs exceeding $200 billion. Also, the U.S. government recently announced Project Stargate, a $500 billion AI infrastructure initiative backed by tech titans like Oracle, Softbank, and OpenAI. 

These developments indicate a growing need for more compute in the future. Apple, too, recently announced a $500 billion investment in the United States over the next four years to build AI infrastructure. Although Apple does not use NVIDIA, its investment still reflects the broader direction of the industry.

Inference is Tough

NVIDIA will face increased competition from inference players like Groq, Cerebras, and SambaNova. Perplexity AI recently announced that its in-house LLM, Sonar, built on Llama 3.3 70B, now runs on Cerebras’ inference infrastructure.

French AI startup Mistral recently launched the Le Chat app for iOS and Android. According to the company, Le Chat is 10 times faster than GPT-4o, Claude Sonnet 3.5, and DeepSeek R1, thanks to Cerebras’ inference technology.

Similarly, in a recent interview, Groq founder Jonathan Ross said that NVIDIA dominates AI model training, and Groq sees no reason to compete in that space. Instead, they focus on faster and cheaper inference.

“They don’t offer fast tokens, and they don’t offer low-cost tokens. It’s a very different product. But what they do very, very well is training. They do it better than anyone else,” said Ross. He added that Grok’s chips cost more than 5x less than NVIDIA’s.

Ross argued that raw specs like teraflops per second are meaningless—what truly matters is tokens per dollar (cost efficiency) and tokens per watt (energy efficiency). Microsoft CEO Satya Nadella recently echoed a similar sentiment. 

Yet, Microsoft has been sending mixed signals. A recent report revealed that the tech giant cancelled leases for significant data centre capacity in the US, raising concerns about the long-term sustainability of AI infrastructure investments.

What About China? 

Following DeepSeek’s success, demand for NVIDIA GPUs has surged nationwide. A recent report states that Chinese companies are ramping up orders for NVIDIA’s H20 chip to support the growing demand for DeepSeek’s low-cost models.

During the earnings call, Huang said that China’s contribution to NVIDIA’s revenue has remained stable as a percentage of overall revenue compared to Q4 and previous quarters. 

However, he acknowledged that China’s share has dropped to half of what it was before US export controls limited NVIDIA’s ability to sell high-end AI chips to Chinese companies. After the US imposed new export restrictions in October 2023, NVIDIA introduced the H20 as its main legally permitted chip for the Chinese market.

SynaXG (a provider of AI-RAN solutions), in collaboration with Kyocera (a Japanese multinational electronics manufacturer) and NVIDIA, has claimed to have unveiled the world’s first software-defined mmWave 5G vRAN solution at MWC Barcelona 2025. 

This technology is expected to transform industries with better connectivity and edge AI processing, targeting widespread deployment in sectors like smart manufacturing and healthcare. 

Trials with select partners are planned for late 2025.

“AI-RAN represents the next frontier for wireless networks, taking 5G, 5G Advanced, and O-RAN to new heights and paving the way for 6G as the future of connectivity,” noted Soma Velayutham, general manager for AI in 5G and telecoms at NVIDIA.

Lab Test Results

In recent lab tests, the solution demonstrated handling five carriers with 400 MHz bandwidth each on a single NVIDIA GH200 superchip. This resulted in a speed of up to 30 Gbps, comparable to custom solutions but with more flexibility and upgradability.

These tests also showed potential for creating highly energy-efficient networks. AI algorithms improved data transfer efficiency and reduced power consumption, making mmWave technology promising for future AI-native wireless communications.

The solution combines advanced radio technology with network control units, enabling both AI-and-RAN and AI-for-RAN capabilities across Layer 1 to Layer 3 functions on the NVIDIA GH200 Grace Hopper Superchip. 

“This achievement lays the foundation for AI to improve and run on super-fast mmWave networks,” said Mantosh Malhotra, chief business officer at SynaXG. 

The collaboration leverages Kyocera’s strengths in mmWave technology and NVIDIA’s AI Aerial platform to deliver high performance, optimised costs, and improved sustainability for next-generation networks. 

In December last year, American telecommunications company Verizon announced a collaboration with NVIDIA to deliver a solution enabling AI applications to operate on Verizon’s 5G private network with mobile edge computing (MEC).

Verizon’s secure 5G infrastructure and private MEC were to be integrated with NVIDIA AI Enterprise software and NIM microservices. According to the company, demonstrations of this solution were to begin early this year.

In November 2024, NVIDIA, along with SoftBank Group’s telecoms division, SoftBank Corp., launched AITRAS, the world’s first AI and 5G telecommunications network.  

Artificial Intelligence Radio Access Network (AI-RAN) came in, which refers to the integration of AI and machine learning (ML) into the radio access network (RAN) of cellular systems like 5G and 6G. This combination enables smarter, more efficient network performance and management.

In the past month, while AI enthusiasts celebrated the Chinese AI startup DeepSeek’s low-cost R-1 model, which was built with minimal GPUs and capital, the market responded rather brutally. NVIDIA lost nearly $600 billion as its stocks tumbled 17% owing to the efficiency with which the model was built. 

More recently, DeepSeek-V3 used just 2,048 NVIDIA H800 GPUs to outperform most open-source models. In contrast, xAI’s Grok-3 was trained on over 100,000 GPUs, yet beats DeepSeek-R1 by just a small margin. 

This made the chip market question whether such large numbers of GPUs were even needed to train these models. However, the story has multiple layers, and DeepSeek’s achievements may not hurt the chip demand after all. 

Experts speculate that the Chinese company may not be revealing the whole truth. During an interview, the CEO of Scale AI, Alexandr Wang, said that he believed DeepSeek possessed around 50,000 NVIDIA H100s, but wasn’t permitted to talk about it.

Jensen Huang Clarifies

NVIDIA chief Jensen Huang recently addressed this market reaction, asserting that investors misinterpreted the implications of the Chinese firm’s developments. 

During a conversation with DDN’s Alex Bouzari, he emphasised that DeepSeek’s R1 model was built using less powerful chips and significantly lower funding than their Western counterparts. This led to the dramatic sell-off of NVIDIA stocks.

“From an investor perspective, there was a mental model that the world was pre-training, and the inference was that you ask an AI a question, which instantly gives you an answer. I don’t know whose fault it is, but obviously, that paradigm is wrong,” he said.

However, Huang believes this reaction stemmed from a misunderstanding of the AI landscape, particularly the ongoing need for high-performance computing in post-training processes, which is essential for AI functionality.

He pointed out that while DeepSeek’s innovations are exciting and energising for the AI sector, they do not diminish the demand for NVIDIA’s chips. Huang explained that post-training methods—where AI models make predictions or draw conclusions after initial training—remain critical and require substantial computing power. 

As DeepSeek continues to generate interest and discussion across various tech sectors, including earnings calls of major companies like Airbnb and Palantir, it is clear that its impact on AI development will be significant.  

Analysts have noted that while DeepSeek’s cost-effective approach may disrupt traditional chip demand, it could also increase competition and innovation within the semiconductor industry. Companies like AMD and Intel may find opportunities to expand their market presence as AI adoption grows, driven by more accessible technologies.

ASML Saw Nearly 9% Rise in Shares

Shortly after the NVIDIA sell-off, Netherlands-based ASML, a leading manufacturer of chip-making equipment, saw shares jump nearly 9% when the company reported a sharp rise in net bookings in the fourth quarter ending December 2024. 

Earlier that week, ASML’s stock had suffered a blow amid the global tech sell-off following DeepSeek’s rollout of its R1 reasoning model, which claims to outperform OpenAI’s o1 in both cost and efficiency. 

The launch had raised concerns that AI firms might reduce spending on advanced chips, potentially impacting ASML’s extreme ultraviolet (EUV) machines, which produce high-end semiconductors. However, CEO Christophe Fouquet dispelled the fears of a slowdown, stating that lower AI costs could drive greater demand for semiconductors.

“A lower cost of AI could mean more applications. More applications mean more demand over time. We see that as an opportunity for more chip demand,” Fouquet told CNBC.  As a result, ASML said net bookings totalled €7.09 billion, marking a 169% increase from the previous quarter. The figure exceeded analyst expectations of €3.99 billion, signalling continued demand for its chipmaking tools.  

The company’s net sales reached €9.26 billion, surpassing the expected €9.07 billion, and its net profit was €2.69 billion, slightly above the €2.64 billion forecast.  
Despite concerns, ASML’s latest results suggested continued strength in semiconductor demand, reinforcing its position as a key supplier in the AI-driven chip industry.

California-based nonprofit Arc Institute and Stanford University, in collaboration with NVIDIA, unveiled Evo 2 on Wednesday as the largest publicly available AI model for genomic data. Evo 2 can predict and design the genetic code—DNA, RNA, and proteins—of all domains of life. 

The model has been trained on nearly 9 trillion nucleotides, the building blocks of DNA and RNA. “We make Evo 2 fully open, including model parameters, training code, inference code, and the OpenGenome2 dataset, to accelerate the exploration and design of biological complexity,” the researchers said in the official paper.

“Deploying a model like Evo 2 is like sending a powerful new telescope out to the farthest reaches of the universe,” said Dave Burke, Arc’s chief technology officer. “We know there’s immense opportunity for exploration, but we don’t yet know what we’re going to discover.”

NVIDIA said the model can be used for biomolecular research applications, including predicting protein structures, identifying novel molecules for healthcare and industrial use, and evaluating how gene mutations affect function.

“Evo 2 represents a major milestone for generative genomics,” said Patrick Hsu, Arc Institute cofounder and core investigator, and an assistant professor of bioengineering at the University of California, Berkeley. “By advancing our understanding of these fundamental building blocks of life, we can pursue solutions in healthcare and environmental science that are unimaginable today.”

The model is available as an NVIDIA NIM microservice, allowing users to generate biological sequences with customisable settings. Researchers can also fine-tune Evo 2 on proprietary datasets through the open-source NVIDIA BioNeMo Framework.

“Designing new biology has traditionally been a laborious, unpredictable and artisanal process,” said Brian Hie, assistant professor of chemical engineering at Stanford University and Arc Institute innovation investigator. “With Evo 2, we make biological design of complex systems more accessible to researchers, enabling the creation of new and beneficial advances in a fraction of the time it would previously have taken.”

Arc Institute, founded in 2021 with $650 million in funding, supports long-term scientific research by providing multiyear funding and dedicated lab space. Scientists at the institute focus on disease areas, including cancer, immune dysfunction, and neurodegeneration.

NVIDIA contributed computing resources by providing access to 2,000 NVIDIA H100 GPUs via NVIDIA DGX Cloud on AWS. The AI platform includes NVIDIA BioNeMo software, featuring optimised microservices and BioNeMo Blueprints. NVIDIA researchers also collaborated on AI scaling and optimisation.

Evo 2 processes genetic sequences up to 1 million tokens in length, enabling a broader analysis of the genome. This capability allows scientists to explore relationships between genetic sequences and cell function, gene expression, and disease.

“A single human gene contains thousands of nucleotides—so for an AI model to analyse how such complex biological systems work, it needs to process the largest possible portion of a genetic sequence at once,” said Hsu.

In healthcare and drug discovery, Evo 2 could help researchers identify gene variants linked to specific diseases and design molecules that precisely target them. In a separate study by Stanford and Arc Institute, researchers found that Evo 2 could predict with 90% accuracy whether previously unrecognised mutations in BRCA1, a gene associated with breast cancer, would affect gene function.

In agriculture, the model could support food security efforts by improving understanding of plant biology, leading to the development of climate-resilient or nutrient-dense crops. Evo 2 could also be used to engineer biofuels or proteins that break down plastic or oil.

Positron, an AI chip startup that aims to go head-to-head with NVIDIA, has raised $23.5 million in funding from investors, including Flume Ventures, Valor Equity Partners, Atreides Management, and Resilience Reserve. The company said it will use the fund to scale the production of its energy-efficient AI chips, offering businesses a more cost-effective alternative to NVIDIA’s hardware.

The company, launched in 2023, is led by Mitesh Agrawal, with co-founders Thomas Sohmers and Edward Kmett.

“With this funding, we’re scaling at a pace that AI hardware has never seen before–from expanding shipments of our first-generation products to bringing our second-generation accelerators to market in 2026,” Agrawal said in a statement.

He added that their solution is growing rapidly because it outperforms conventional GPUs in both cost and energy efficiency while delivering AI hardware that eliminates reliance on foreign supply chains.

More Rivals to NVIDIA AI Chips 

Positron, as a startup, claims to have already shipped products to data centres and neoclouds around the US.

According to the company, their Atlas systems are presently achieving 3.5 times better performance per dollar and 3.5 times greater power efficiency than NVIDIA H100 GPUs for inference. Its servers powered by field-programmable gate array (FPGA) support models with up to a trillion parameters while offering plug-and-play compatibility with Hugging Face and OpenAI APIs.

The system uses a memory-optimised architecture that uses more than 93% of the bandwidth. Traditional GPUs often consume upwards of 10,000 watts per server, creating a major hurdle for data centres with limited infrastructure. However, the chip’s energy-efficient architecture makes it cost-efficient, allowing traditional data centres to harness AI computing without needing to overhaul the infrastructure completely.

Addressing the chip’s unique value proposition in a statement, one of the investors, Rob Reid, the co-founder of Resilience Reserve, said, “What sets Positron apart is not just its cost efficiency, but its ability to bring AI hardware to market at an unprecedented speed and provide high performance per watt. Their innovative approach is enabling businesses to scale AI workloads without the typical barriers of cost and power consumption.”

Positron mentioned that it has built a fully American supply chain, which ensures that its AI hardware is designed, fabricated, and assembled in the US. With developments like this, including OpenAI’s aim to produce AI chips, it should be exciting to see who succeeds in taking off NVIDIA’s AI crown.

Ola chief Bhavish Aggarwal has announced Krutrim AI Lab and the launch of several open source AI models tailored to India’s unique linguistic and cultural landscape. This includes the launch of Krutrim 2, the startup’s second LLM consisting of 8 billion parameters.

“While we’ve been working on AI for a year, today we’re releasing our work to the open source community and also publishing a bunch of technical reports,” announced Aggarwal, founder and CEO of Ola and Krutrim. 

Furthermore, the Krutrim AI Lab includes:

Chitrarth 1: A Vision Language Model capable of interpreting images and documents.

Dhwani 1: A Speech Language Model designed for tasks such as speech translation.

Vyakhyarth 1: An Indic Embedding model optimised for applications like search and retrieval-augmented generation.

Krutrim Translate 1: A text-to-text translation model facilitating seamless translation between languages.

Recognising the absence of a global benchmark for Indic language performance, Krutrim AI Lab has also developed “BharatBench,” a comprehensive evaluation framework. 

The lab has also published several technical reports and papers to further the research community’s understanding of these models.

In partnership with NVIDIA, the lab is set to deploy India’s first GB200 supercomputer by March, with plans to scale it into the nation’s largest supercomputer by the end of the year. This infrastructure will support the training and deployment of AI models, addressing challenges related to data scarcity and cultural context. The lab has committed an investment of ₹2,000 crore into Krutrim, with a pledge to increase this to ₹10,000 crore by next year. 

By open sourcing the models, Aggarwal said that he aims to foster collaboration within India’s AI community, accelerating the development of a world-class AI ecosystem.

Last week, Krutrim also brought China’s DeepSeek models to its cloud infrastructure.

India’s AI Mission Continues

This announcement aligns with India’s broader AI ambitions. The IndiaAI Mission seeks to build a comprehensive ecosystem that fosters AI innovation by democratising computing access, enhancing data quality, and developing indigenous AI capabilities. 

A key component of this mission is the establishment of a common computing facility powered by approximately 18,693 GPUs, including high-end models like the NVIDIA H100 and H200. 

This facility aims to provide accessible computing power to startups, researchers, and academia at a fraction of global cost benchmarks.

Union Minister Ashwini Vaishnaw has emphasised the importance of accessible computing power, stating that it is ‘the most important part of the mission.’ He also announced plans to develop 6-8 large language models by the Indian tech ecosystem, supported by this robust computing infrastructure.

Safety and ethical deployment of AI models remain top priorities for the government. To this end, an AI Safety Institute is being established, adopting a techno-legal approach to ensure responsible AI development.

The DeepSeek effect is real. On Monday, NVIDIA’s stock saw a sharp decline of about 17%, ending the day at roughly $118.58. This plunge wiped out nearly $600 billion in market value, setting a new record for the largest single-day loss in market capitalisation for any company on Wall Street.

The market has attributed the decline to DeepSeek, which recently released its latest model, DeepSeek-R1, trained using NVIDIA’s lower-capability H800 processor chips with a budget of under $6 million.

However, the tech industry may be misinterpreting the situation as DeepSeek-R1 was also trained on NVIDIA GPUs. 

Currently, the only potential concern for NVIDIA is that the compute power and cost required to develop next-generation models might decrease in the near future. However, this would be positive for the industry, as demand for these models increases over time, and more consumers will adopt them.

After its shares plunged and DeepSeek became the talk of the town, NVIDIA released a statement saying that its chips are proving valuable in the Chinese market and more will be needed to support DeepSeek’s growing demand.

“DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely available models and compute that is fully export control compliant,” the company said. 

The company added that inference requires a large number of NVIDIA GPUs and high-performance networking, mentioning that there are now three scaling laws, pre-training, post-training, and the new test-time scaling.

During a recent interview, the CEO of Scale AI, Alexandr Wang, said that he believes DeepSeek possesses around 50,000 NVIDIA H100s, though they are not permitted to talk about it.

Notably, the company recently launched Janus-Pro-7B, an open-source multimodal AI model created to challenge industry leaders like OpenAI’s DALL-E 3 and Stability AI’s Stable Diffusion in text-to-image generation.

Is DeepSeek the Real Cause of the Market Crash?

Before the release of DeepSeek-R1, the AI research lab launched DeepSeek V3, which, according to the company, was trained on a cluster of 2,048 NVIDIA H800 GPUs with a budget of only $5.576 million.

“Wow… NVIDIA dropped by 17% because of DeepSeek. I wonder if the investors realise that NVIDIA and DeepSeek aren’t competition. DeepSeek was trained using NVIDIA GPUs… People installing and running it locally are mostly using NVIDIA GPUs too…,” said Matt Wolfe, founder of Future Tools. 

Similarly, Aashay Sachdeva, an engineer at Sarvam AI, expressed confusion over why NVIDIA is losing money due to DeepSeek. “RL is so behind, they also need even more GPUs to train it with more data for longer periods,” he wrote in a post on X.

He went on to say that DeepSeek’s base model is open-source, which will likely lead to more smaller labs participating, adding that more high-end GPUs will be required for inference. “100k output tokens in open-source models are coming soon,” he added.

The sentiment was echoed by Microsoft CEO Satya Nadella, who remarked, “Jevons paradox is coming again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

Jevons paradox suggests that when technology improves and allows us to use a resource more efficiently (for example, using less coal to produce the same amount of energy), the cost of using that resource decreases. This, in turn, can make the resource more attractive and lead to an increase in demand for its use, even though it’s more efficient.

Likewise, OpenAI chief Sam Altman finally took notice of DeepSeek-R1 and said, “DeepSeek’s R1 is an impressive model, particularly in terms of what they’re able to deliver for the price.”

However, he added that as research progresses, more compute will be required. “We are excited to continue executing on our research roadmap and believe that more compute is now more important than ever to succeed in our mission,” he added.

It makes sense, as the startup is part of the Stargate Project, which will build a $500 billion AI infrastructure in Texas.

“I believe that Jevons Paradox could actually make NVIDIA far bigger than what it currently is, as democratised AI expands the demand base of their GPUs. The current stock dip is a knee-jerk reaction,” said Tech Whisperer founder Jaspreet Bindra. 

However, not everyone believes in Jevons Paradox. With the rise of small language models and reduced reliance on GPUs, users will soon be able to run them on mobile devices and laptops. “Jevon’s Paradox? Phones and low-end laptops will be running powerful 1.5B-parameter models within a year or two,” said KissanAI founder Pratik Desai.

Similarly,  Mansi Gupta, senior analyst at Everest Group, said, “DeepSeek is likely to create ripple effects for chipmakers like NVIDIA, which could potentially witness reduced demand for their higher-end, premium chips, as the model creators optimise their models for better cost-performance ratio.”

However, even though the cost of intelligence will drop to zero, there will still be a demand for more compute, as AI applications continue to grow. Lower costs lead to stronger distribution, which in turn means broader application. And that, in turn, leads to more compute and more users.

NVIDIA revealed Project DIGITS, a new $3,000 small supercomputer, at CES 2025. It targets AI researchers, data scientists, and students. As more efficient models are developed, more customers will purchase the supercomputer and run LLMs locally. 

Meanwhile, Meta CEO Mark Zuckerberg has announced plans to invest $60-65 billion in capital expenditure during 2025 to expand the company’s AI infrastructure and computing capabilities.

NVIDIA Loves China 

At Donald Trump’s inauguration, a notable gathering of tech leaders was observed, including Tesla CEO Elon Musk, Amazon founder Jeff Bezos, Meta CEO Mark Zuckerberg, Alphabet CEO Sundar Pichai, and Apple CEO Tim Cook. 

However, NVIDIA CEO Jensen Huang decided to skip the event and instead visited China, stopping in Beijing, Shenzhen, and Shanghai to celebrate the Lunar New Year with local staff. 

Despite US chip export restrictions, Huang affirmed that NVIDIA remains dedicated to investing in China, where its workforce now totals approximately 4,000 employees, he mentioned at the Beijing office’s annual meeting. 

China is an important market for NVIDIA. The company recently strongly criticised the Biden administration’s new “AI Diffusion” rule, set to impose restrictions on global access to AI chips and technology. 

The US has banned the export of NVIDIA’s H800 to China and prevented the company from selling chips even with a reduced transfer rate. However, the GPUs are still smuggled to China. 

While there is no official disclosure of the number of H800 GPUs exported to China, an investigation suggested that there is an underground network of around 70 sellers who claim to receive dozens of GPUs every month. 

Meanwhile, NVIDIA has developed modified versions of its chips, such as the H20, L20, and L2, which comply with US export regulations. These chips are designed to have reduced capabilities compared to their full versions, allowing them to be legally sold in China. 

Another report revealed that NVIDIA chips are being used in server products from Dell, Supermicro, and others in China. Recently, the US department of commerce asked NVIDIA to investigate how its products reached China.

The GPU giant also faces competition from local competitors like Huawei. With its Ascend series of data centre processors, particularly the Ascend 910B and the upcoming Ascend 910C, Huawei is actively working to challenge NVIDIA’s dominance in AI computing. 

The company has informed potential clients that its upcoming Ascend 910C processor is on par with NVIDIA’s H100. Interestingly, while DeepSeek trained its models on NVIDIA H800 GPUs, it is now running inference on Huawei’s new domestic chip, the 910C.

Known for its powerful GPUs, NVIDIA is now turning its attention to robotics. The company recently launched Cosmos, a platform to accelerate the development of physical AI systems, such as autonomous vehicles (AVs) and robotics. 

“NVIDIA will be a robotics company at the end of the day, not just semiconductor. Few understand the moves they are making at the lowest level from manufacturing up to software,” Dylan Patel, founder of Semianalysis, said.

In a recent interview, NVIDIA chief Jensen Huang said that the world needs an AI that understands the physical world. “It has to understand the dynamics of the physical world, like gravity, inertia, or friction, and it has to understand spatial and geometric relationships,” Huang said.

According to him, the world needs more robots as today there aren’t enough workers. “There’s an ageing population, a changing preference in the type of work people want to do, and the birth rate is declining. The world needs more workers, so the timing is really quite imperative that we have robotic systems.”

NVIDIA is working to integrate robotics primarily into manufacturing processes, autonomous vehicles, and the healthcare sector. For example, humanoid robots in manufacturing can perform repetitive tasks, handle materials, and collaborate with human workers. Huang predicts that the $50 trillion manufacturing industry will become software-defined.

The future Huang envisions is already taking shape. For example, BMW is using the Figure 02 humanoid robot on its production line. The company claims that Figure 02 can now operate as an ‘autonomous fleet’, with a 400% increase in speed and a sevenfold improvement in success rate.

The Foundation of NVIDIA Robotics

Huang referred to Cosmos as the “ChatGPT or Llama of world foundation models”. The platform has been trained on 20 million hours of video, focusing on dynamic interactions like walking humans and hand movements. 

He further said that the real magic happens when Cosmos is integrated with Omniverse. The combination of the two provides “ground truth” for AI, which helps it understand the physical world. Huang compared the connection of Cosmos and Omniverse to the concept of LLMs connected to retrieval-augmented generation (RAG) systems.

Moreover, Huang introduced the concept of three fundamental computers essential for building robotic systems. The first is Deep GPU Xceleration (DGX), which is used to train AI. Once the AI is trained, the next computer, AGX, is employed to deploy AI into real-world applications such as cars or robots. 

The third component is the Digital Twin, a simulated environment where the AI practices, refines its abilities and undergoes further training before deployment. 

Backed by Strong Research

This is not the first time NVIDIA is discussing humanoids and autonomous vehicles. For the past year, the company has been actively researching this field.

“It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. Our children will grow up as ‘robot natives’. They will have humanoids cook Michelin dinners, robot teddy bears tell bedtime stories, and FSD (full self-driving) drive them to school,” Jim Fan, senior research manager and lead of Embodied AI (GEAR Lab) at NVIDIA, said. 

In another interview, Fan said that the company chose humanoid, considering that the world is built around the human form factor. “All our restaurants, factories, hospitals, and all equipment and tools are designed for the human form.” 

Notably, the company recently unveiled project Eureka and demonstrated a demo where they trained a five-finger robot hand to spin a pen.

Besides, NVIDIA recently developed HOVER (humanoid versatile controller), a 1.5 million parameter neural network designed to coordinate the motors of humanoid robots for locomotion and manipulation.  

“Not every foundation model needs to be gigantic. We trained a 1.5 million-parameter neural network to control the body of a humanoid robot,” Fan revealed. 

NVIDIA launched Project GR00T and the Isaac platform last year. GR00T is a framework that allows developers to generate extensive synthetic datasets from a limited number of human actions. 

The company has also developed Jetson Thor, a new generation of compact computers for humanoid robots, which is slated for release in the first half of 2025.

NVIDIA is working with companies such as 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Robotics and XPENG Robotics to build humanoid robots.

World Models x Humanoids 

It seems that NVIDIA is not alone in the robotics race. According to OpenAI’s career page, the startup is hiring for roles in mechanical engineering, robotics systems integration, and program management. The goal is to “integrate cutting-edge hardware and software to explore a broad range of robotic form factors”.

Last year, the company hired Caitlin Kalinowski to lead its robotics and consumer hardware divisions. Previously at Meta, she oversaw the development of Orion augmented reality (AR) glasses. OpenAI has also invested in Figure AI and robotics AI startup Physical Intelligence.

Similarly, Apptronik, one of the leaders in AI-powered humanoid robotics, recently announced an exciting new partnership with Google DeepMind’s robotics team to create truly intelligent and autonomous robots.

Tesla is also upping its game. At the ‘We, Robot’ event at Warner Bros. Studio in California last year, CEO Elon Musk showcased the company’s humanoid robot, Optimus. It can walk dogs, mow the lawn, do household chores, babysit, and even speak like the GenZ while making smooth hand gestures.

Meanwhile, Fan’s mentor and ‘AI godmother’ Fei-Fei Li, recently founded her own company, World Labs, to build large world models (LWMs) that can perceive, generate, and interact with the 3D world.

E2E Networks, an NSE-listed cloud infrastructure company in India, unveiled AI Lab as a Service (AILaaS), designed to help educational institutions. AILaaS offers scalable, customisable, and cost-effective AI infrastructure using a pay-as-you-go model, providing access to NVIDIA GPUs, including the H100 and H200.

This initiative addresses a prominent challenge faced by colleges and universities: the prohibitive costs and complexities of setting up on-premise AI labs. As the demand for AI proficiency increases, access to advanced tools and resources becomes crucial for students’ success.

AILaas for Education

AILaaS bridges this gap, enabling institutions to build and scale AI capabilities without the burden of specialised infrastructure maintenance or obsolescence.

AILaaS comes equipped with pre-configured datasets, models, and tools for hands-on learning in fields such as machine learning, AI agents, and AGI. It also supports platforms for inference, retrieval-augmented generation (RAG), and AI agent pipelines. 

Customisable packages tailored for institutions of varying sizes, from 50 to 200 students, ensure flexibility to meet specific requirements. The pay-as-you-go model ensures cost efficiency, allowing colleges to pay only for the resources they utilise.

E2E’s AI Progress

The cloud infrastructure company, supported by MeitY, recently partnered with People+ai, which is an initiative of Nandan Nilekani-founded EkStep Foundation). The collaboration looks to meet the increasing demand for cloud GPU and computing power in India and bring customisable cloud resources to a broader audience.

The partnership centres around People+ai’s Open Cloud Compute (OCC) project, which aims to create a network of micro data centres across India. The OCC project promotes an open, decentralised cloud infrastructure to enable better collaboration among businesses, developers, and government organisations.

NVIDIA has strongly criticised the Biden administration’s new “AI Diffusion” rule, set to impose restrictions on global access to AI chips and technology. The company argues that the regulation, expected to take effect in 120 days, threatens to undermine U.S. leadership in artificial intelligence and stifle innovation worldwide.

Ned Finkle, vice president of government affairs at NVIDIA, said, “The Biden Administration now seeks to restrict access to mainstream computing applications with its unprecedented and misguided ‘AI Diffusion’ rule, which threatens to derail innovation and economic growth worldwide”.

“For decades, leadership in computing and software ecosystems has been a cornerstone of American strength and influence worldwide,” he added.  

He further noted that under the previous Trump Administration, policies had fostered a competitive environment that allowed U.S. industries to lead in AI innovation without compromising national security.

The Biden administration has introduced new restrictions on the export of US-developed computer chips used in artificial intelligence (AI) systems to prevent rivals like China from accessing advanced technology. The move comes just a week before President-elect Donald Trump’s inauguration.

These measures, part of a long-standing effort to curb China’s progress in military and industrial leadership, are likely to heighten tensions between Washington and Beijing. The 200-page regulation imposes limits on exporting advanced AI chips and technology to countries like China while introducing a Pentagon initiative to use AI models for cyber defence and a pilot program to bolster energy sector cybersecurity.

In a recent appearance on Joe Rogan’s podcast, Meta chief Mark Zuckerberg revealed that the Biden administration had pressured Meta to censor content related to COVID-19 vaccines.

“They pushed us super hard to take down things that were honestly true,” said Zuckerberg. He claimed that administration officials would “scream” and ‘curse” at Meta employees during discussions on content moderation.

Last year, a16z venture capitalist Marc Andreessen also expressed serious concerns about the Biden administration’s approach to AI regulation. After meetings with government officials in May 2024, Andreessen described the experience as ‘absolutely horrifying.’ He claimed that officials discouraged investments in AI startups, saying, “They actually said flat out to us, ‘Don’t do AI startups, like, don’t fund AI startups.’”

A single statement from Nvidia CEO Jensen Huang during an analyst event at CES has triggered a massive selloff in the quantum computing sector, erasing approximately $8 billion in market value, according to reports. 

Huang suggested that bringing “very useful quantum computers” to market could take 15 to 30 years, citing the need for quantum processors, or qubits, to increase by a factor of 1 million.  

Quantum computing stocks crash after AI godfather Jensen Huang exploded the quantum computing bubble w/3 lines: "If you said 15 years for very useful quantum computers, that would probably be on the early side. If you said 30 is probably on the late side. But if you picked 20, I… pic.twitter.com/8XG4xYZEBv

— Holger Zschaepitz (@Schuldensuehner) January 8, 2025

Market Fallout

Huang’s comment had a ripple effect. The quantum computing companies’ stocks have witnessed a sharp decline ever since. For instance, IonQ shares fell over 31.65%, while Rigetti Computing dropped by 37.25%, and D-Wave Quantum saw its stock tumble down by 25.61% after Huang’s statement. 

The remarks undermined the optimism that had been building in the sector, particularly following Google’s announcement of a breakthrough with its Willow quantum chip in December. 

Google revealed progress in creating a 105-qubit chip, part of its roadmap to develop a quantum system with 1 million qubits. This news led to a great exchange between Sundar Pichai and Elon Musk with dreams of building quantum clusters in space. 

Industry Pushes Back

However, countering his claim, Quantum leaders were quick to challenge and form an alternative narrative. Alan Baratz, CEO of D-Wave Quantum, dismissed Huang’s comments as “dead wrong.” 

Baratz told CNBC that the reason was “that we at D-Wave are commercial today.” He pointed to clients like Mastercard and NTT Docomo, which are already leveraging their quantum systems for business operations.  

$NVDA $QBTS

TODAY JENSEN HUANG SAID THAT QUANTUM COMPUTING WAS 20 YEARS AWAY FROM BEING USEFUL.

DWave Quantum $QBTS was down as much as 49% after these comments and ended the day down 36%.

The stock is up 1000% from the September lows.

The CEO of DWave says in this clip… pic.twitter.com/2hGf9ZSS0R

— amit (@amitisinvesting) January 8, 2025

Baratz acknowledged that Huang’s timeline might apply to gate-based quantum computers but argued it was “100% off base” for annealing quantum computers, which D-Wave specialises in. 

He also said that D-Wave quantum computers solve in minutes what supercomputers would take millions of years, challenging Huang’s views on current tech capabilities. He publicly offered to meet with Huang to clarify what he described as “knowledge gaps” in the CEO’s understanding.  

Another user on X took to the platform to address this, saying, “NVIDIA is literally hiring quantum engineers right now.”

So let me get this straight…

Jensen is hiring people right now for a technology that he thinks won't be useful until 30 years from now?$NVDA is literally hiring quantum engineers right now.

The math on this timeline isn't math-ing. $IONQ pic.twitter.com/8UYhznX3J2

— Ashton Cheekly (@elwalvador) January 8, 2025

Similarly, others have also posted images of NVIDIA job postings for a quantum computing director and related positions; the very next day, Huang expressed his views.

The selloff followed a period of intense investor interest in quantum computing. While Huang’s projection has sparked debate, it underscores the technical and commercial challenges facing the quantum computing sector. 

For now, Huang’s remarks have cast a shadow over what was previously seen as a fast-moving and highly promising market.

After debunking myths about Bengaluru’s infamous traffic, Joint Commissioner of Police (Traffic), MN Anucheth, shared his belief that self driving vehicles and Level 5 autonomous cars could navigate the city’s chaotic roads—but he questioned whether there’s a real need for such technology in Namma Bengaluru.

Bengaluru Traffic Commissioner said that while fully autonomous (Level 5) vehicles could theoretically navigate the city’s chaotic roads, their practicality and adoption remain uncertain.

“If a Level 5 vehicle is equipped with sufficient inputs like radar, LiDAR, and computer vision, it could learn both good and bad behaviours to navigate through our roads,” he noted. However, he questioned the necessity of such advanced vehicles in India’s context, adding, “Would you spend so much for a Level 5 vehicle when hiring a driver is easier and more cost-effective here?”

Given the safety concerns and the rise in road rage incidents in the city, the case for self-driving vehicles and robotaxis such as Tesla and Waymo becomes increasingly compelling.

As Bengaluru’s Metro and suburban rail networks expand, the city’s focus appears to be on improving public transportation. For now, the dream of fully autonomous vehicles navigating Namma Bengaluru’s roads remains more of a futuristic vision than an immediate reality.

Citing the USA and Canada, where the shortage of drivers for heavy goods vehicles, such as long trailers and trucks, has accelerated the push for Level 4 and Level 5 automation, in regions where there is no shortage of the same, investing in Level 5 autonomous vehicles might not be justifiable. 

“I don’t think the adoption will be that much because one, it’s very cost prohibitive. Second, I think, it’s easier to hire a driver and have a level 3 which is a driver assistant, giving more importance to road safety rather than automation,” he said. 

While emphasis lies on four-wheelers, Anucheth also believes that if two wheelers have better adoption of tech, accident deaths can significantly come down. “If you look at the vulnerable group in Bengaluru city, about 72% are two-wheelers,” he said, attributing 2-wheelers as the cause of maximum accidents. 

Autonomous Vehicles on the Rise

Recently, at CES 2025, NVIDIA chief Jensen Huang said, “I predict
that this [autonomous vehicles] will likely be the first multi-trillion dollar robotics industry.” 

At the same event, NVIDIA released Cosmos, a platform built to accelerate the development of autonomous vehicles and robots. The platform consists of work foundation models (WFMs), video tokenisers, guardrails and other features that will help developers build models without the dependency of real-world data.

Tesla, NVIDIA are all heavily investing in autonomous vehicle production and abilities. Tesla’s fully autonomous robotaxi is being positioned as a safety vehicle, allowing one to travel without the risk of any form of accidents or mishaps. 

Not just vehicular safety, but passenger safety is also becoming a rising concern. Recently, a woman passenger jumped out of a moving autorickshaw run by a when she sensed the driver was taking a deviated path. With robotaxis, safety can be easily met. Something Musk had positioned as the predominant goal for his vehicles. 

Interestingly, Indian players are also working on fully autonomous vehicles, with some claiming level 5 autonomy. AI and robotics startup Swaayatt Robots is building level 5 autonomy and has raised $7 million at a valuation of $175 million. Similarly, Kochi-based Rosh AI has been testing their fully autonomous vehicles for more than a decade.

Taiwan-based electronics manufacturing giant Foxconn is teaming up with American chipmaker NVIDIA to develop humanoid robots in Kaohsiung City, Taiwan.  

Speaking at the Kaohsiung Smart City annual meeting, Foxconn chairman Young Liu revealed the company’s ambitious plans to integrate NVIDIA’s advanced software and hardware technologies to develop these humanoids, as reported by Focus Taiwan. 

The collaboration marks a significant leap in Foxconn’s efforts to diversify its portfolio beyond contract manufacturing in electronics. Liu also highlighted plans to collaborate with Taipei and Keelung in smart city projects, enhancing app development and sovereign AI capabilities.  

The shift toward humanoid robotics aligns with the company’s financial outlook. Liu forecast consolidated sales to exceed NT$7 trillion ($213 billion) in 2025, driven by growing demand for AI servers and robotics applications. 

AI servers accounted for 40% of the company’s server revenue in 2024 and are projected to rise to 50% by 2025, underscoring their role as a key growth driver.  

NVIDIA’s Role 

NVIDIA CEO Jensen Huang has championed the role of robotics in the AI revolution, stating that humanoid robots, alongside self-driving cars, represent the next frontier of innovation. 

With the humanoids market expected to reach $38 billion in the next two decades, creating quality datasets for humanoids to use for imitation learning becomes extremely tedious and time-consuming. 

Huang announced the NVIDIA Isaac GR00T Blueprint at the CES stage, with 14 humanoid robots standing in the background. This blueprint is a simulation workflow for synthetic motion generation, enabling developers to create large datasets for training humanoids using imitation learning.

Jensen called this “the ChatGPT moment of general robotics.”

Source: X

As announced on its official blog, users can now use the Apple Vision Pro to capture human actions in a digital twin. The robot mimics the action in simulation and records it for use. 

Central to this vision is NVIDIA’s forthcoming Jetson Thor computing system, expected to debut in early 2025. “Building foundation models for general humanoid robots is one of the most exciting problems to solve in AI today,” Jensen said.

Built on NVIDIA’s cutting-edge Blackwell architecture, Jetson Thor is a compact AI superchip boasting 208 billion transistors and featuring a high-performance CPU cluster and integrated safety processors.

Meeting Real-World Application

Foxconn and NVIDIA’s partnership is emblematic of the broader trend toward embodied AI, integrating perception, cognition, and action into physical entities. 

Advanced datasets, such as AgiBot World’s humanoid manipulation trajectories and large-scale AI models like Robot Era’s ERA-42, are accelerating the development of versatile, intelligent robots.  

David Friedberg from The All-In Podcast also predicted that 2025 will be the year of AI robots. He said, “I think this is going to be the year where we’re all going to look at humanoid robots and autonomous systems and be like, ‘Oh my God, I can’t believe this is here.’”

The Road Ahead

With a rich history of deploying robots, such as its proprietary ‘Foxbots’, to automate production lines, Foxconn is positioning itself to expand its robotics expertise into humanoid systems. 

Now, the company plans to extend its reach into sectors like healthcare by introducing humanoid robots capable of more sophisticated interactions and functionalities.  

Previously, the company also partnered with NVIDIA to leverage its EV ambitions and proposed building Taiwan’s fastest AI supercomputer with NVIDIA’s flagship Blackwell. 

In India, the company has partnered with HCL Group for semiconductor operations and announced Project Cheetah, another Foxconn facility for the manufacture and assembly of EV components, as announced by CM Siddaramaiah. This adds to the list of Karnataka Foxconn projects after Project Elephant.

As Foxconn and NVIDIA join forces, the evolution of humanoid robotics stands poised to transform industries, offering new possibilities in manufacturing, healthcare, and beyond. 

With advanced hardware, extensive datasets, and innovative AI models, the partnership signals a future where robots not only mimic human behaviour but seamlessly integrate into everyday life, redefining the boundaries of AI and robotics. 

The era of AI agents has officially begun. Making way for them, NVIDIA chief Jensen Huang predicted that in the future, an organisation’s IT department would evolve into an ‘HR department for AI’. It would be responsible for onboarding, managing, and maintaining a new generation of AI agents. 

At the ongoing Consumer Electronics Show (CES) 2025, Huang said, “In a lot of ways, the IT department of every company is going to be the HR department of AI agents in the future. Today, they manage and maintain a bunch of software from the IT industry; in the future, they will maintain, nurture, onboard, and improve a whole bunch of digital agents and provision them to the companies to use.”

He added that these AI agents will work along with human employees, offering unprecedented capabilities in automation and efficiency across industries. Speaking to a captivated audience, Huang explained how specialised AI agents will become integral to companies, performing tasks ranging from customer service to complex problem-solving.

“AI agents are a multi-trillion dollar opportunity,” he said.  

“Jensen Huang’s CES 2025 keynote wasn’t just about breakthroughs—it was a glimpse into how AI agents will shape the future. From physical AI that reasons, plans, and acts, to tools like Cosmos and Project DIGITS, NVIDIA is building the foundation for AI agents to integrate seamlessly into our lives and industries,” said RagaAI founder Gaurav Agarwal. 

Far away at the Microsoft AI Tour in Bengaluru, Microsoft chief Satya Nadella said that “building agents should be as simple as creating a spreadsheet”. He introduced a no-code platform called Copilot Studio that allows users to create new agents based on their needs.

“Think of AI as a co-pilot for your work. It’s the UI for AI,” Nadella said, illustrating the role it will play as an interface between employees and the AI. He gave the example of an AI agent in a healthcare setting, describing a scenario where a doctor prepares for a tumour board meeting, and the AI creates the agenda, prioritises cases, and takes detailed notes during the discussion.

Nadella also unveiled Copilot Actions, which allows users to create cross-application workflows that connect people, data, and tasks across the Microsoft 365 ecosystem.

Likewise, OpenAI chief Sam Altman recently predicted that AI agents could enter the workforce by 2025. “We believe that, by 2025, we may see the first AI agents join the workforce and materially change the output of companies,” Altman wrote in a recent blog post.

Meanwhile, Google published a comprehensive whitepaper exploring the development and functionality of AI agents. Last December, the company launched Gemini 2, which it said will have agentic capabilities. 

Too Soon? 

Google’s senior product manager, Logan Kilpatrick, feels that it will take at least another year before AI agents become a reality. “2025 is the year of AI vision capabilities going mainstream; 2026 will be agents,” he said. 

“There’s a ~12-month capabilities-to-wide-scale-production gap. Most vision use cases work now but aren’t widely deployed. Agents still need a little more work for billion-user-level scale,” Kilpatrick added.

According to a recent report, it could take OpenAI some time to launch AI agents. This is because the company is concerned about prompt injection, a type of attack where a large language model is tricked into following instructions from a malicious user.

Huang may be right. It looks like the primary responsibility of enterprise IT teams will be to ensure that the agents are safe to use and do not have access to data they are not supposed to have.

In an interview with AIM, Okta customer identity CTO Bhawna Singh spoke about the growing need to authorise AI agents. “A platform is needed to handle both authentication and authorisation, making sure not all data is accessible to the agent,” she said. 

She explained that since these AI agents interact with each other, it is essential that they have the right data access. “We need to make sure these agents are verified,” she said.

Similarly, NVIDIA NeMo is helping companies onboard and train their AI agents, mimicking the process of onboarding a new employee. “Nemo is essentially a digital employee pipeline where companies can provide feedback, define company-specific vocabulary, and set guardrails on the behaviour of these agents,” Huang explained. 

Recently, AI startup Composio launched AgentAuth, a product that efficiently integrates AI agents with third-party tools and APIs. It supports a variety of authentication protocols, including OAuth 2.0, OAuth 1.0, API keys, JWT, and Basic Authentication. 

The platform also integrates over 250 widely used apps and services, catering to diverse needs such as customer relationship management (CRM) systems and ticketing platforms.

“The biggest problem that people face while building agents is connecting them to reliable tools. For example, if someone builds a sales agent, they would need to connect it with CRMs like Salesforce, HubSpot, etc,” said Karan Vaidya, Composio chief, in an exclusive interview with AIM. 

Vertical AI Agents

The AI agents market, valued at $5.1 billion in 2024, is projected to soar to $47.1 billion by 2030. Just as companies today rely on SaaS services, in the near future, they will hire specialised AI agents to meet their needs. Employees will widely use autonomous agents to perform tasks like attending meetings, making summaries, drafting emails, and translating meetings live.

“The early winners in LLM-based solutions might just be general-purpose platforms. Over time, vertical AI agents will emerge. It’s like how, in the box software world, the early vendors were just trying to convince people to use software… As the market matures, it will get more sophisticated, and vertical solutions will become dominant players,” said Jared Friedman, group partner at Y Combinator, in a recent podcast with YC president Gary Tan.

Salesforce chief Marc Benioff describes AI agents as digital labour. “I am the CEO of a company that manages agents and humans, and I have a digital labour platform at my disposal to augment my support, sales, service, and marketing,” said Benioff. 

In India, Freshworks unveiled a new version of Freddy AI, an autonomous agent that resolved 45% of customer support requests and 40% of IT service requests (in beta).

However, as AI agents become increasingly common, selecting the right sectors for their implementation will be crucial. “Departments such as sales, marketing, and finance usually have well-established software systems like CRM, ERP, analytics dashboards, etc., so they can plug AI agents directly into these data pipelines,” said Ramprakash Ramamoorthy, director of AI research at Zoho and ManageEngine.

Besides SaaS companies, several Indian startups are also building AI agents. Bengaluru-based AI startup KOGO AI, founded by Praveer Kochhar and Raj K Gopalakrishnan, is developing AI agents and solutions to simplify workflows and improve productivity for businesses. The company recently launched an AI agent store.“We are currently building an agent that can look at a database and actually think like a data scientist or a business analyst, generating extremely intelligent questions,” Kochhar said in a recent podcast with AIM.

Leading chipmaker NVIDIA unveiled Project DIGITS, a new small supercomputer, at the Consumer Electronics Show (CES) 2025.  It is aimed at AI researchers, data scientists, and students across the world.  The supercomputer provides access to the GB10 Grace Blackwell Superchip, which Jensen Huang, CEO of NVIDIA, calls a ‘super secret chip’. 

The GB10 features a 20-core NVIDIA Blackwell GPU, one of the most powerful AI hardware systems available today. The CPU was built in collaboration with MediaTek. 

Project DIGITS features 128 GB of unified memory and offers storage options of up to 4TB. Similar to a typical computer, DIGITS requires only a standard electrical outlet to operate. It operates on a Linux-based NVIDIA DGX operating system. 

The supercomputer can also run up to 200 billion parameter LLMs locally, and if you have two of them, NVIDIA says you can link them up to run AI models double the size.

what??

NVIDIA just dropped Project DIGITS, a $3,000 personal AI supercomputer that’s small enough to look like a Mac Mini but packs 1,000x the power of your average laptop.

Handles AI models with up to 200 BILLION parameters.

This is incredible.. pic.twitter.com/z4JOeFD2JI

— el.cine (@EHuanglu) January 7, 2025

DIGITS allows users to deploy AI models on the NVIDIA DGX cloud and leverage all the tools present inside NVIDIA’s AI Enterprise software platform. For instance, you can fine-tune models on the NeMo framework and build agents on NVIDIA Blueprints and NIM microservices. 

The higher performance and portability increase the cost. DIGITS will be available in May of this year and will cost a whopping $3000. 

“AI will be mainstream in every application for every industry. With Project DIGITS, the Grace Blackwell Superchip comes to millions of developers,” said Huang. 

“Placing an AI supercomputer on the desks of every data scientist, AI researcher, and student empowers them to engage and shape the age of AI,” he added. 

At first glance, there’s no doubt that Project DIGITS is entering the Mac Mini territory, at least in terms of the form factor. “I must have the NVIDIA Project DIGITS for my home lab. 128GB pooled RAM, 4TB storage, the size of a Mac mini, Running DGX OS, a Linux-based OS. The coming years are going to be wild in the world of AI and robotics,” said Jamie Madden, a machine learning developer on X.

In December of last year, NVIDIA introduced the Jetson Orin Nano Super Developer Kit, a compact generative AI supercomputer now priced at $249, down from $499. According to the company, it offers enhanced performance with 67 INT8 TOPS, marking a 70% improvement over its predecessor, alongside a memory bandwidth of 102GB/s, which is a 50% increase. 

At CES 2025, NVIDIA unveiled Cosmos, a platform built to speed up the development of physical AI systems, including autonomous vehicles and robots. The platform includes generative world foundation models (WFMs), video tokenisers, guardrails, and an accelerated data processing pipeline to help developers create and refine AI models with reduced reliance on real-world data.

Cosmos is available under an open model license on Hugging Face and the NVIDIA NGC catalogue. Fully optimised NVIDIA NIM microservices will follow, with enterprise support provided through the NVIDIA AI Enterprise software platform.

Speaking at CES, NVIDIA CEO Jensen Huang said, “The ChatGPT moment for robotics is coming. Like large language models, world foundation models are fundamental to advancing robot and AV development, yet not all developers have the expertise and resources to train their own. We created Cosmos to democratise physical AI and put general robotics in reach of every developer.”

The Cosmos models can generate physics-based videos using inputs such as text, images, and sensor data, enabling their use in applications like video search, synthetic data generation, and reinforcement learning. 

Developers can customise the models to simulate industrial environments, driving scenarios, and other specific use cases. NVIDIA also introduced NeMo Curator, an accelerated video processing pipeline that can process 20 million hours of video in 14 days, and Cosmos Tokeniser, a visual data compression tool.

“Data scarcity and variability are key challenges to successful learning in robot environments,” said Pras Velagapudi, chief technology officer at Agility Robotics. “Cosmos’ text-, image-, and video-to-world capabilities allow us to generate and augment scenarios for a variety of tasks that we can use to train models without needing as much expensive, real-world data capture.”

Major robotics and transportation companies, including Agile Robots, XPENG, Waabi, and Uber, have begun adopting Cosmos for their AI development. 

Uber CEO Dara Khosrowshahi said, “Generative AI will power the future of mobility, requiring both rich data and very powerful compute. By working with NVIDIA, we are confident that we can help supercharge the timeline for safe and scalable autonomous driving solutions for the industry.”

In addition to Cosmos, NVIDIA introduced the Llama Nemotron large language models and Cosmos Nemotron vision language models, developed for enterprise use in sectors including healthcare, finance, and manufacturing.

At CES 2025, NVIDIA CEO Jensen Huang launched new Nemotron models, including the Llama Nemotron large language models (LLMs) and Cosmos Nemotron vision language models (VLMs), to improve agentic AI and boost enterprise productivity.

The Llama Nemotron models, built on Llama foundation models, allow developers to create AI agents for applications like customer support, fraud detection, and supply chain optimisation.

“Llama 3.1 is a complete phenomenon, with the downloads reaching 650,000 times. It has been derived and turned into other models, about 60,000 different models. It is singularly the reason why every single enterprise and every single industry has been activated to start working on AI,” said Huang.

“We realized that the Llama models could really be better fine-tuned for enterprise use, so we fine-tuned them using our expertise and capabilities and turned them into the Llama Nemotron suite of open models,” he added. 

The Nemotron families will be offered in Nano, Super, and Ultra sizes to suit deployment needs, from low-latency real-time applications to high-accuracy data center use cases. Optimised for computing efficiency and accuracy, these models support agentic AI tasks like instructions for following, coding, and math.

“Agentic AI is the next frontier of AI development, and delivering on this opportunity requires full-stack optimization across a system of LLMs to deliver efficient, accurate AI agents,” said Ahmad Al-Dahle, vice president and head of GenAI at Meta.

 NVIDIA announced that the models will be available as downloadable resources or as microservices for deployment across various computing platforms, including data centers and edge devices. Llama Nemotron and Cosmos Nemotron models will be available soon on build.nvidia.com, Hugging Face, and through the NVIDIA Developer Program. 

Enterprise-grade deployments will be supported via the NVIDIA AI Enterprise platform on accelerated cloud and data center infrastructure.

NVIDIA’s Cosmos Nemotron models extend AI capabilities to vision and video tasks, allowing agents to analyse and respond to images and videos. These tools aim to support industries like autonomous systems, healthcare, retail, and media. 

NVIDIA also unveiled Cosmos world foundation models for physics-aware video generation in robotics and autonomous vehicle applications.

NVIDIA NeMo microservices allow enterprises to customise these models for specific domains and workflows.

Leading AI platform providers, such as SAP and ServiceNow, have backed the Nemotron models. SAP plans to incorporate them into its Joule platform to improve enterprise user productivity, while ServiceNow seeks to utilise the models for AI agent services across various industries.

The models are built using NVIDIA’s NeMo platform for distillation, pruning, and alignment, ensuring high accuracy and throughput across various hardware configurations. NVIDIA NeMo Retriever allows integration with enterprise data, boosting model functionality through retrieval-augmented generation capabilities.

At CES 2025, NVIDIA unveiled the GeForce RTX 50 Series, the latest GPUs developed for gamers, creators, and developers. Using the Blackwell architecture, the RTX 50 Series delivers improved performance and efficiency through AI-driven innovations like neural shaders and DLSS 4.

The RTX 50 series, revealed during NVIDIA CEO Jensen Huang’s keynote at CES includes the RTX 5090, 5080, 5070 Ti, and 5070. The GeForce RTX 5090 will be available starting January 30 for $1,999, while the other models are priced between $999 and $549, with launches scheduled for February.

The flagship GeForce RTX 5090 GPU offers a significant leap in power, featuring 92 billion transistors and over 3,352 trillion AI operations per second (TOPS). According to NVIDIA, the RTX 5090 GPU doubles the performance of its predecessor, the RTX 4090.

“Blackwell, the engine of AI, has arrived for PC gamers, developers, and creatives,” said Jensen Huang, founder and CEO of NVIDIA. “Fusing AI-driven neural rendering and ray tracing, Blackwell is the most significant computer graphics innovation since we introduced programmable shading 25 years ago.”

The GPUs are equipped with fifth-generation Tensor Cores and fourth-generation RT Cores, enabling advanced capabilities such as RTX Neural Shaders and DLSS 4. The latter features Multi Frame Generation, which uses AI to generate up to three frames per rendered frame, boosting frame rates by up to 8x over traditional rendering.

NVIDIA announced that the RTX 50 Series is now available for laptops, providing desktop-level performance in a portable form. With NVIDIA Max-Q technology, battery life is extended by up to 40% without compromising on design or performance. Laptops featuring RTX 50 Series GPUs will roll out in March.

To improve realism in game characters and environments, NVIDIA introduced RTX Neural Faces and other AI-driven features. RTX Neural Faces uses generative AI to create high-quality, stable digital faces in real time, while RTX Mega Geometry allows for up to 100x more ray-traced triangles in scenes.

The GPUs also power autonomous game characters through NVIDIA ACE technologies. These characters can perceive, plan, and act, creating dynamic interactions in games such as PUBG: BATTLEGROUNDS and the upcoming InZOI.

RTX 50 Series GPUs support FP4 precision, optimising AI image generation and enabling generative AI models to run efficiently on local hardware. New tools, such as NVIDIA Broadcast’s Studio Voice and Virtual Key Light, further enhance creative workflows.

NVIDIA is one of the most influential hardware giants in the world. Apart from its much sought-after GPUs, the company also provides free courses to help you understand more about generative AI, GPU, robotics, chips, and more. 

Most importantly, all of these are available free of cost and can be completed in less than a day. Let’s take a look at them.

1. Building RAG Agents for LLMs

Building RAG Agents for LLMs course is available for free for a limited time. It explores the revolutionary impact of large language models (LLMs), particularly retrieval-based systems, which are transforming productivity by enabling informed conversations through interaction with various tools and documents. Designed for individuals keen on harnessing these systems’ potential, the course emphasises practical deployment and efficient implementation to meet the demands of users and deep learning models. Participants will delve into advanced orchestration techniques, including internal reasoning, dialog management, and effective tooling strategies.

In this workshop you will learn to develop an LLM system that interacts predictably with users by utilising internal and external reasoning components.

Course link: https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-FX-15+V1

2. Accelerating Data Science Workflows with Zero Code Changes

Efficient data management and analysis are crucial for companies in software, finance, and retail. Traditional CPU-driven workflows are often cumbersome, but GPUs enable faster insights, driving better business decisions. 

In this workshop, one will learn to build and execute end-to-end GPU-accelerated data science workflows for rapid data exploration and production deployment. Using RAPIDS-accelerated libraries, one can apply GPU-accelerated machine learning algorithms, including XGBoost, cuGraph’s single-source shortest path, and cuML’s KNN, DBSCAN, and logistic regression. 

More details on the course can be checked here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+T-DS-03+V1

3. Generative AI Explained

This self-paced, free online course introduces generative AI fundamentals, which involve creating new content based on different inputs. Through this course, participants will grasp the concepts, applications, challenges, and prospects of generative AI. 

Learning objectives include defining generative AI and its functioning, outlining diverse applications, and discussing the associated challenges and opportunities. All you need to participate is a basic understanding of machine learning and deep learning principles.

To learn the course and know more in detail check it out here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-NP-01+V1

4. Digital Fingerprinting with Morpheus

This one-hour course introduces participants to developing and deploying the NVIDIA digital fingerprinting AI workflow, providing complete data visibility and significantly reducing threat detection time. 

Participants will gain hands-on experience with the NVIDIA Morpheus AI Framework, designed to accelerate GPU-based AI applications for filtering, processing, and classifying large volumes of streaming cybersecurity data. 

Additionally, they will learn about the NVIDIA Triton Inference Server, an open-source tool that facilitates standardised deployment and execution of AI models across various workloads. No prerequisites are needed for this tutorial, although familiarity with defensive cybersecurity concepts and the Linux command line is beneficial.

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+T-DS-02+V2/

5. Building A Brain in 10 Minutes

This course delves into neural networks’ foundations, drawing from biological and psychological insights. Its objectives are to elucidate how neural networks employ data for learning and to grasp the mathematical principles underlying a neuron’s functioning. 

While anyone can execute the code provided to observe its operations, a solid grasp of fundamental Python 3 programming concepts—including functions, loops, dictionaries, and arrays—is advised. Additionally, familiarity with computing regression lines is also recommended.

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+T-FX-01+V1/

6. An  Introduction to CUDA

This course delves into the fundamentals of writing highly parallel CUDA kernels designed to execute on NVIDIA GPUs. 

One can gain proficiency in several key areas: launching massively parallel CUDA kernels on NVIDIA GPUs, orchestrating parallel thread execution for large dataset processing, effectively managing memory transfers between the CPU and GPU, and utilising profiling techniques to analyse and optimise the performance of CUDA code. 

Here is the link to know more about the course – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+T-AC-01+V1

7. Augment your LLM Using RAG

Retrieval Augmented Generation (RAG), devised by Facebook AI Research in 2020, offers a method to enhance a LLM output by incorporating real-time, domain-specific data, eliminating the need for model retraining. RAG integrates an information retrieval module with a response generator, forming an end-to-end architecture. 

Drawing from NVIDIA’s internal practices, this introduction aims to provide a foundational understanding of RAG, including its retrieval mechanism and the essential components within NVIDIA’s AI Foundations framework. By grasping these fundamentals, you can initiate your exploration into LLM and RAG applications.

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:NVIDIA+S-FX-16+v1/

8. Getting Started with AI on Jetson Nano

The NVIDIA Jetson Nano Developer Kit empowers makers, self-taught developers, and embedded technology enthusiasts worldwide with the capabilities of AI. 

This user-friendly, yet powerful computer facilitates the execution of multiple neural networks simultaneously, enabling various applications such as image classification, object detection, segmentation, and speech processing. 

Throughout the course, participants will utilise Jupyter iPython notebooks on Jetson Nano to construct a deep learning classification project employing computer vision models. 

By the end of the course, individuals will possess the skills to develop their own deep learning classification and regression models leveraging the capabilities of the Jetson Nano.

Here is the link to know more about the course – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-RX-02+V2

9. Building Video AI Applications at the Edge on Jetson Nano

This self-paced online course aims to equip learners with skills in AI-based video understanding using the NVIDIA Jetson Nano Developer Kit. Through practical exercises and Python application samples in JupyterLab notebooks, participants will explore intelligent video analytics (IVA) applications leveraging the NVIDIA DeepStream SDK. 

The course covers setting up the Jetson Nano, constructing end-to-end DeepStream pipelines for video analysis, integrating various input and output sources, configuring multiple video streams, and employing alternate inference engines like YOLO. 

Prerequisites include basic Linux command line familiarity and understanding Python 3 programming concepts. The course leverages tools like DeepStream, TensorRT, and requires specific hardware components like the Jetson Nano Developer Kit. Assessment is conducted through multiple-choice questions, and a certificate is provided upon completion. 

For this course, you will require hardware including the NVIDIA Jetson Nano Developer Kit or the 2GB version, along with compatible power supply, microSD card, USB data cable, and a USB webcam. 

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+S-IV-02+V2/

10. Build Custom 3D Scene Manipulator Tools on NVIDIA Omniverse

This course offers practical guidance on extending and enhancing 3D tools using the adaptable Omniverse platform. Taught by the Omniverse developer ecosystem team, participants will gain skills to develop advanced tools for creating physically accurate virtual worlds. 

Through self-paced exercises, learners will delve into Python coding to craft custom scene manipulator tools within Omniverse. Key learning objectives include launching Omniverse Code, installing/enabling extensions, navigating the USD stage hierarchy, and creating widget manipulators for scale control. 

The course also covers fixing broken manipulators and building specialised scale manipulators. Required tools include Omniverse Code, Visual Studio Code, and the Python Extension. Minimum hardware requirements comprise a desktop or laptop computer equipped with an Intel i7 Gen 5 or AMD Ryzen processor, along with an NVIDIA RTX Enabled GPU with 16GB of memory. 

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+S-OV-06+V1/

11. Getting Started with USD for Collaborative 3D Workflows

In this self-paced course, participants will delve into the creation of scenes using human-readable Universal Scene Description ASCII (USDA) files. 

The programme is divided into two sections: USD Fundamentals, introducing OpenUSD without programming, and Advanced USD, using Python to generate USD files. 

Participants will learn OpenUSD scene structures and gain hands-on experience with OpenUSD Composition Arcs, including overriding asset properties with Sublayers, combining assets with References, and creating diverse asset states using Variants.

To learn more about the details of the course, here is the link – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-FX-02+V1

12. Assemble a Simple Robot in Isaac Sim

This course offers a practical tutorial on assembling a basic two-wheel mobile robot using the ‘Assemble a Simple Robot’ guide within the Isaac Sim GPU platform. The tutorial spans around 30 minutes and covers key steps such as connecting a local streaming client to an Omniverse Isaac Sim server, loading a USD mock robot into the simulation environment, and configuring joint drives and properties for the robot’s movement. 

Additionally, participants will learn to add articulations to the robot. By the end of the course, attendees will gain familiarity with the Isaac Sim interface and documentation necessary to initiate their own robot simulation projects. 

The prerequisites for this course include a Windows or Linux computer capable of installing Omniverse Launcher and applications, along with adequate internet bandwidth for client/server streaming. The course is free of charge, with a duration of 30 minutes, focusing on Omniverse technology. 

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+T-OV-01+V1/

13. How to Build Open USD Applications for industrial twins

This course introduces the basics of the Omniverse development platform. One will learn how to get started building 3D applications and tools that deliver the functionality needed to support industrial use cases and workflows for aggregating and reviewing large facilities such as factories, warehouses, and more. 

The learning objectives include building an application from a kit template, customising the application via settings, creating and modifying extensions, and expanding extension functionality with new features. 

To learn the course and know more in detail check it out here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-OV-13+V1

14. Disaster Risk Monitoring Using Satellite Imagery

Created in collaboration with the United Nations Satellite Centre, the course focuses on disaster risk monitoring using satellite imagery, teaching participants to create and implement deep learning models for automated flood detection. The skills gained aim to reduce costs, enhance efficiency, and improve the effectiveness of disaster management efforts. 

Participants will learn to execute a machine learning workflow, process large satellite imagery data using hardware-accelerated tools, and apply transfer-learning for building cost-effective deep learning models. 

The course also covers deploying models for near real-time analysis and utilising deep learning-based inference for flood event detection and response. Prerequisites include proficiency in Python 3, a basic understanding of machine learning and deep learning concepts, and an interest in satellite imagery manipulation. 

To learn the course and know more in detail check it out here – https://courses.nvidia.com/courses/course-v1:DLI+S-ES-01+V1/

15. Introduction to AI in the Data Center

In this course, you will learn about AI use cases, machine learning, and deep learning workflows, as well as the architecture and history of GPUs.  With a beginner-friendly approach, the course also covers deployment considerations for AI workloads in data centres, including infrastructure planning and multi-system clusters. 

The course is tailored for IT professionals, system and network administrators, DevOps, and data centre professionals. 

To learn the course and know more in detail check it out here – https://www.coursera.org/learn/introduction-ai-data-center

16. Fundamentals of Working with Open USD

In this course, participants will explore the foundational concepts of Universal Scene Description (OpenUSD), an open framework for detailed 3D environment creation and collaboration. 

Participants will learn to use USD for non-destructive processes, efficient scene assembly with layers, and data separation for optimised 3D workflows across various industries. 

Also, the session will cover Layering and Composition essentials, model hierarchy principles for efficient scene structuring, and Scene Graph Instancing for improved scene performance and organisation.

To know more about the course check it out here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-OV-15+V1

17. Introduction to Physics-informed Machine Learning with Modulus 

High-fidelity simulations in science and engineering are hindered by computational expense and time constraints, limiting their iterative use in design and optimisation. 

NVIDIA Modulus, a physics machine learning platform, tackles these challenges by creating deep learning models that outperform traditional methods by up to 100,000 times, providing fast and accurate simulation results.

One will learn how Modulus integrates with the Omniverse Platform and how to use its API for data-driven and physics-driven problems, addressing challenges from deep learning to multi-physics simulations.

To learn the course and know more in detail check it out here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-OV-04+V1

18. Introduction to DOCA for DPUs

The DOCA Software Framework, in partnership with BlueField DPUs, enables rapid application development, transforming networking, security, and storage performance. 

This self-paced course covers DOCA fundamentals for accelerated data centre computing on DPUs, including visualising the framework paradigm, studying BlueField DPU specs, exploring sample applications, and identifying opportunities for DPU-accelerated computation. 

One gains introductory knowledge to kickstart application development for enhanced data centre services.

To learn the course and know more in detail check it out here – https://learn.nvidia.com/courses/course-detail?course_id=course-v1:DLI+S-NP-01+V1

The story was updated on 2nd Jan, 25 to reflect the latest courses and correct the URLs to them.

NVIDIA, the world’s leading chipmaker company, completed the acquisition of Run:ai, an Israel-based AI infrastructure platform, on Monday. NVIDIA reportedly paid $700 million after overcoming several hurdles posed by the US Department of Justice (DOJ).

Earlier in August, the US DOJ initiated an antitrust investigation on the acquisitions in lieu of the impact it can have on the competition. However, in October, the European Commission reviewed the deal and “unconditionally” approved it, citing that it may not have any concerning effects in the European Economic Area. 

However, the delays only benefited Run:ai, and its employees. Given NVIDIA’s stock price surge, founders and employees are set to be retained and will be paid $200 million in NVIDIA shares – reportedly double the initial amount. 

That said, Run:ai is also open-sourcing its software, which currently runs only on NVIDIA GPUs. “While our colours will change from pink to green, our commitment to our customers, partners, and the broader AI ecosystem remains unwavering,” the company stated. 

As for NVIDIA, the company is ending the year on a high note. The company achieved unprecedented success in 2024 and became the world’s most valuable company with a $3.5 trillion market cap, surpassing Apple. 

However, a less talked-about development is NVIDIA’s acquisitions throughout the year. Out of the 23 companies NVIDIA has acquired to date, six of them were completed in 2024 alone. 

Jensen Huang’s Big, Happy Family

Interestingly, Run:ai is the second Israeli company to be acquired by NVIDIA in 2024. Earlier in May, the company announced the acquisition of Deci for $300 million. Post-acquisition, the company was dissolved as an independent corporate entity. 

Deci’s deep learning acceleration platform helped developers improve their performance by up to 15 times without compromising accuracy. 

Soon after, NVIDIA acquired Shoreline.io in June, reportedly for $100 million. Shoreline, founded by a former AWS executive, developed a platform for automating processes to fix computer system incidents.

The company then acquired Brev.Dev, a San Francisco-based AI/ML development platform for training and deploying models on the cloud. In September, NVIDIA acquired OctoAI, a healthcare startup based in Seattle, US, reportedly for $250 million.

A few weeks ago, they completed the acquisition of VinBrain, another healthcare startup owned by the Vietnamese conglomerate Vingroup. NVIDIA also set up an R&D centre in the country. 

From software development to healthcare, AI is catching up with every other sector, and NVIDIA wants to be at the forefront of it all. 

In a recent podcast episode, CEO Jensen Huang said that breakthroughs in AI will include “everything from quantum computing to quantum chemistry. “Every field of science is involved in the approaches that we’re talking about…Nothing will be left behind. We’re going to take everybody with us.” 

In addition to acquisitions, NVIDIA participated in several investment activities, the notable ones being OpenAI’s $6.6 billion funding round and Elon Musk’s xAI’s $6 billion Series C funding round. 

That said, NVIDIA has also had a fruitful year when it comes to partnerships in India. 

‘NVIDIA is AI in India’

In 2024, Huang demonstrated a significant focus on India. At the NVIDIA AI Summit in 2024 in Mumbai, Huang called NVIDIA “AI in India” and praised the country’s AI and cloud infrastructure. 

“By the end of this year, we will have nearly 20 times more compute here in India than just a little over a year ago,” he added. 

A few weeks ago, Yotta Data Services introduced Rudra, a program that provides startups and developers access to NVIDIA’s Hopper GPUs and their AI Enterprise software suite. Startups and independent software vendors (ISVs) enrolled in the NVIDIA Inception and Connect programs can receive cloud credits of up to $50,000 on Yotta’s Shakti Cloud.

Companies like Infosys, TCS, Tech Mahindra, and Wipro announced a collaboration with NVIDIA. For instance, TCS and Wipro use NVIDIA’s NIM Agent Blueprints for various sectors in their services. Infosys, on the other hand, incorporated NVIDIA’s AI Enterprise into their Topaz suite, which helps businesses integrate generative AI into their workflows. 

In October, NVIDIA also announced the Nemotron-4-Mini Hindi model, which Tech Mahindra was the first to implement in their Indus 2.0 Model. 

Huang also interacted with Reliance Industries Chairman Mukesh Ambani and said that India is key to NVIDIA’s vision. At Jamnagar, Reliance is preparing to build a data centre capable of supporting 1GW of power, with the potential to expand multiple gigawatts at a single location. The facility is said to use NVIDIA’s latest Blackwell chips.

Earlier in July, NVIDIA also announced a collaboration with six Indian PC builders – The MVP, Ant PC, Vishal Peripherals, Hi-Tech Computer Genesys Labs, XRig, and EliteHubs, to launch ‘Made in India’ PCs equipped with RTX AI. 

Amazon Web Services (AWS) is preparing to take on NVIDIA as a strong contender as its 2015 acquisition of Annapurna Labs – an Israeli startup whose name was inspired by the Annapurna mountain range in the Himalayas – is proving to be an advantage. At the AWS re:Invent in Las Vegas, the cloud giant announced new chips, including Trainium2, Graviton 4 and Inferentia.

AWS claims that Trainium2 offers 30-40% better price performance than the previous generation of the graphics processing unit (GPU) based Elastic Compute Cloud (EC2) instances. Customers like Anthropic, Databricks, Adobe, Qualcomm, poolside, and even Apple are already on board. 

“Today, there’s really only one choice on the GPU side, and it’s just NVIDIA,” said Matt Garman, CEO at AWS. “We think that customers would appreciate having multiple choices.”

It is worth noting that AWS recently invested $4 billion in Anthropic, making it the primary cloud provider and training partner. The company also introduced Trn2 UltraServers and the next-generation Trainium3 AI training chip. 

AWS is working with Anthropic to build Project Rainier – a large AI compute cluster powered by thousands of Trainium2 chips. This will help Anthropic develop its models, including optimising its flagship product Claude, to run on Trainium2 hardware.

“This cluster is going to be five times the number of exaflops as the current cluster that Anthropic used to train their leading set of Claude models that are out there in the world,” Garman added. 

On the other hand, OpenAI plans to partner with Taiwan Semiconductor Manufacturing Company (TSMC) and Broadcom to launch its first in-house AI chip by 2026. Meanwhile, OpenAI is also banking on NVIDIA’s Blackwell GPU architecture to scale its o1 model and test time compute. 

Notably, Anthropic CEO Dario Amodei, in a recent podcast, said that the cost of training AI models today can reach up to $1 billion. While models like GPT-4 cost approximately $100 million, he predicts that within the next three years, training costs could escalate to $10 or even $100 billion.

Advantage Trainium? 

According to Garman, Trainium2 delivers 30-40% better price performance than current GPU-powered instances. The new TRN2 instances come equipped with 16 custom-built chips interconnected via NeuronLink, a high-speed and low-latency interconnect. This configuration provides up to 20.8 petaflops of compute from a single node. 

The company also introduced Trn2 UltraServers, which combine four Trn2 servers into a single system and offer 83.2 petaflops of compute power for better scalability. These servers feature 64 interconnected Trainium2 chips. For comparison, NVIDIA’s Blackwell B200 is expected to provide up to 720 petaflops of FP8 performance with a rack of 72 GPUs.

Few expected Apple to use AWS Trainium2 to train its models. Benoit Dupin, Apple’s senior director of machine learning and AI, revealed how deeply the company relies on AWS for its AI and ML capabilities. Dupin accredited the decade-long partnership with AWS for enabling Apple’s innovations like Siri, iCloud Music, and Apple TV. “AWS has consistently supported our dynamic needs at scale and globally.”

Apple has leveraged AWS’s solutions, including Graviton and Inferentia chips. It achieved milestones like a 40% efficiency boost by migrating to Graviton instances. Dupin also teased early success with AWS Trainium2 chips, which could deliver a 50% leap in pre-training efficiency.

“We knew that the first iteration of Trainium wasn’t perfect for every workload, but we saw enough traction to give us confidence we were on the right path,” Garman revealed.

Trainium chips leverage the Neuron SDK, which efficiently optimises AI workloads. It supports deep learning inference and training and seamlessly integrates with TensorFlow and PyTorch while avoiding closed-source dependencies. However, it still faces challenges from NVIDIA’s CUDA.

Switching from NVIDIA to Trainium requires hundreds of hours of testing and rewriting code – a barrier few companies want to cross. Acknowledging this challenge internally, AWS called CUDA the single biggest reason customers stick with NVIDIA.

Meanwhile, Amazon’s cloud rivals, Microsoft and Google, are working on their own AI chips to reduce their reliance on NVIDIA. Google recently announced the general availability of Trillium, its sixth generation of Tensor Processing Unit (TPU). “Using a combination of our TPUs and GPUs, LG AI Research reduced inference processing time for its multimodal model by more than 50% and operating costs by 72%,” said Google chief Sundar Pichai during the recent earnings call.

In a similar way, numerous companies today are competing for NVIDIA’s chip market share, including AI chip startups such as Groq, Cerebras Systems, and SambaNova Systems. 

The AI chip market is projected to hit $100 billion in the upcoming years, and AWS is pouring billions of dollars into Trainium to stake its claim. However, beating NVIDIA’s dominance won’t be an easy feat.