From accelerating drug discovery to automating patent searches, AI is popping up in a growing number of R&D contexts. Two of the top stories this week had a direct NVIDIA connection — first, Elon Musk’s announced what he terms the world’s largest GPU cluster, featuring 100,000 of its H100 GPUs, designed to power his new AI venture, xAI. Second, Nvidia’s stock took a nosedive, shedding a record-breaking $279 billion in market value in a single day, raising questions about the sustainability of the current AI investment frenzy and potentially impacting future R&D budgets. Also in this week’s roundup are two significant developments related to OpenAI: a report that the company is considering a monthly subscription cost of up to $2,000 for access to its cutting-edge language models, and news that Ilya Sutskever, OpenAI’s former Chief Scientist, has launched a new AI safety-focused startup with $1 billion in funding. These developments underscore the rapid pace of change and the high stakes involved in the race to develop and deploy advanced AI technologies.

Musk claims Colossus, world’s Largest GPU cluster with 100,000 GPUs, is operational

Sources: The Information (September 4, 2024), Fortune (September 3, 2024)

Tesla and X CEO Elon Musk announced that xAI’s “Colossus,” a massive AI training cluster comprising 100,000 Nvidia H100 GPUs, is now online. Founded in March 2023, xAI is an artificial intelligence company founded to take on current rivals such as OpenAI, a company Musk also cofounded in 2015. Colossus, based in Memphis, Tennessee, makes use of a liquid-cooled system. The supercomputer reportedly took just 122 days to build and uses a single RDMA fabric for interconnectivity. Musk claims it is the “most powerful AI training system in the world” and notes that it will be instrumental in developing xAI’s large language models, including Grok-3. He aims to double its size to 200,000 GPUs and incorporate 50,000 advanced H200 chips in the near future, with the goal of creating “the world’s most powerful AI by every metric” by December 2024. The news is not without controversy. The Information notes that Musk has a history of exaggerating technical announcements. In addition, Meta CEO Mark Zuckerberg has hinted at plans to purchase 350,000 Nvidia H100s.

OpenAI reportedly mulls premium LLM subscriptions for top-tier models

Source: Reuters/The Information (September 5, 2024)

OpenAI is reportedly considering launching higher-priced subscription tiers for its upcoming large language models (LLMs), including its rumored models “Strawberry” and “Orion” that are said to have enhanced reasoning capabilities. The Information reported that internal discussions have explored prices potentially reaching as high $2,000 per month, a 100-fold jump from the current $20 monthly fee for ChatGPT Plus, which costs incrementally more for the business version of the same service. To clarify, The Information did not say that OpenAI had plans to charge premium enterprise customers $2,000 monthly, but that this figure was mentioned in early internal discussions. This news comes as OpenAI experiences strong adoption of its business products, ChatGPT Enterprise, Team, and Edu, surpassing 1 million paid users, according to Reuters.

Departed OpenAI co-founder launches safety-focused AI startup with $1B in funding

Source: Reuters (September 4, 2024)

Deep learning trailblazer Ilya Sutskever, who was the former Chief Scientist at OpenAI, has co-founded a new AI safety startup, Safe Superintelligence (SSI) that has secured $1 billion in funding with a valuation of $5 billion, according to Reuters. SSI aims to develop AI systems that are both powerful and safe, surpassing human capabilities while ensuring alignment with human values. Sutskever’s departure from OpenAI followed a tumultuous period involving the temporary ousting and subsequent reinstatement of OpenAI CEO Sam Altman. Since Altman was reinstated in March 2024, several core members of OpenAI’s safety team have departed the company.

Nvidia’s stock faces $279B loss in one day, prompting concerns over AI investment mania

Sources: Reuters (September 3, 2024), Business Insider (September 4, 2024)

As reports emerged that Nvidia received a subpoena from the U.S. Department of Justice related to an antitrust investigation (which the company denies), its stock experienced a record-breaking single-day loss of $279 billion in market value on September 4. This drop, the largest ever for a U.S. company, also points to questions about the rapid growth of AI. While the company managed to beat investor expectations when announcing its Q2 fiscal 2025 earnings on August 28, 2024, with record quarterly revenue of $30.0 billion (15% YoY), investors were not impressed. Investors cited several reasons for concerns: One, a slowing growth rate. While NVIDIA’s revenue was up 122% year-over-year to $30 billion, the 16% growth from Q1 to Q2 hinted at a slowdown compared to previous quarters. Investors have grown used to NVIDIA consistently beating expectations by wide margins. Consequently, the narrower beat this quarter (4.1% above estimates) less impressive, as Yahoo Finance noted.

TIME Magazine unveils 2024 list of 100 most influential people in AI

Source: TIME (September 5, 2024)

TIME Magazine released its second annual TIME100 AI list, highlighting the 100 most influential individuals shaping the field of artificial intelligence. The list features a diverse mix of figures, from tech CEOs like Sundar Pichai (Google) and Satya Nadella (Microsoft) to AI safety advocates like Meredith Whittaker (Signal) and researchers like Ilya Sutskever (Safe Superintelligence — formerly OpenAI). TIME considered several criteria in its picks, including individuals’ impact, innovation, leadership, and public engagement. Other luminaries named in the list include Nvidia CEO Jensen Huang, Google DeepMind CEO Demis Hassabis, and Amazon’s head scientist of artificial general intelligence Rohit Prasad, and OpenAI CEO Sam Altman.

Salesforce CEO declares “hard pivot” to autonomous AI agents

Source: Fortune (September 5, 2024)

Salesforce CEO Marc Benioff aims to rebuild Salesforce around autonomous AI agents through a new platform called Agentforce. Other notable tech companies such as Meta are also betting big on AI agents in the long term. Meta CEO Mark Zuckerberg has predicted that the world would eventually have potentially more AI agents than there are people in the world. In Salesforce’s case, the company is making what Benioff terms as as a “hard pivot,” positioning Agentforce as a different entity than traditional AI chatbots. The platform is designed to allow users to build and deploy AI agents that can make decisions and automate customer service tasks. He plans to rebrand the upcoming Dreamforce conference as “Agentforce” and aims to get thousands of customers using the platform by early 2025. Some analysts have questioned Salesforce’s ability to maintain its edge in the generative AI era. Yet Benioff remains optimistic, underscoring the importance of tapping the company’s longstanding experience in CRM and enterprise software. He claims that major customers like ADP, OpenTable, and Wyndham are already using Agentforce.

Google DeepMind unveils AlphaProteo, an AI system for designing novel protein binders

Source: Google DeepMind Blog (September 5, 2024) 

Google DeepMind has introduced AlphaProteo, a new AI system designed to create novel, high-strength protein binders for use in biological and health research. The organization notes that the technology demonstrates superior binding success rates and strengths compared to existing methods. Potential applications include drug development and biotechnology. The platform could also lead to an improved understanding of various diseases. In creating AlphaProteo, Google DeepMind trained it on hefty datasets of protein information from the Protein Data Bank (PDB) and more than 100 million predicted structures from AlphaFold (DeepMind’s protein structure prediction AI) system. The company has verified its abilities. In experimental tests against seven diverse protein targets, including viral proteins (like the SARS-CoV-2 spike protein) and proteins associated with cancer and autoimmune diseases, AlphaProteo achieved high success rates in generating binders that successfully bound to their intended targets. For example, 88% of the candidate binders for the viral protein BHRF1 bound successfully. AlphaProteo has also demonstrated the ability to design proteins with 3 to 300 times stronger binding affinities than existing methods. AlphaProteo can also create  new protein binders for VEGF-A. Going forward, DeepMind plans to continue improving AlphaProteo’s capabilities and explore its potential in drug design through Isomorphic Labs, another Alphabet subsidiary.

Broadcom’s AI successes fail to outweigh sluggish sales elsewhere

Source: Bloomberg (September 5, 2024)

Chipmaker Broadcom’s strong growth in AI-related products, projected to hit $12 billion in revenue this year and exceeding expectations, still ran into roadblocks. For one thing, investor concerns about slower growth in other segments of its business were a central hurdle. Nevertheless, CEO Hock Tan expressed optimism about a recovery in non-AI chip sales, thanks to a 20% increase in bookings. While the company’s stock dipped following a disappointing overall sales forecast, dropping as much as 10% in its biggest intraday decline in a month, its long-term prospects, especially in the AI chip market, appear promising, Tan said. The CEO anticipates a shift towards custom in-house AI chip designs, a trend that could benefit the company as it assists companies in producing their own chips. Despite the turbulence, Broadcom’s recent $69 billion acquisition of VMware buttressed growth, with software sales hitting $5.8 billion. On the hardware side, Apple remains a notable customer, with next-generation iPhones expected to boost wireless revenue by 20% in the fourth quarter. Tan also signaled that Broadcom would focus on integrating VMware rather than pursuing new semiconductor acquisitions in the near future. Shown here is the company’s BCM88890 Jericho3AI Machine Learning Ethernet switch. Broadcom’s Jericho3-AI fabric provides high-performance Ethernet for large-scale AI clusters of up to 32,000 GPUs.