Stay Ahead, Stay ONMINE

Training Large Language Models: From TRPO to GRPO

Deepseek has recently made quite a buzz in the AI community, thanks to its impressive performance at relatively low costs. I think this is a perfect opportunity to dive deeper into how Large Language Models (LLMs) are trained. In this article, we will focus on the Reinforcement Learning (RL) side of things: we will cover […]

Deepseek has recently made quite a buzz in the AI community, thanks to its impressive performance at relatively low costs. I think this is a perfect opportunity to dive deeper into how Large Language Models (LLMs) are trained. In this article, we will focus on the Reinforcement Learning (RL) side of things: we will cover TRPO, PPO, and, more recently, GRPO (don’t worry, I will explain all these terms soon!) 

I have aimed to keep this article relatively easy to read and accessible, by minimizing the math, so you won’t need a deep Reinforcement Learning background to follow along. However, I will assume that you have some familiarity with Machine Learning, Deep Learning, and a basic understanding of how LLMs work.

I hope you enjoy the article!

The 3 steps of LLM training

The 3 steps of LLM training [1]

Before diving into RL specifics, let’s briefly recap the three main stages of training a Large Language Model:

  • Pre-training: the model is trained on a massive dataset to predict the next token in a sequence based on preceding tokens.
  • Supervised Fine-Tuning (SFT): the model is then fine-tuned on more targeted data and aligned with specific instructions.
  • Reinforcement Learning (often called RLHF for Reinforcement Learning with Human Feedback): this is the focus of this article. The main goal is to further refine responses’ alignments with human preferences, by allowing the model to learn directly from feedback.

Reinforcement Learning Basics

A robot trying to exit a maze! [2]

Before diving deeper, let’s briefly revisit the core ideas behind Reinforcement Learning.

RL is quite straightforward to understand at a high level: an agent interacts with an environment. The agent resides in a specific state within the environment and can take actions to transition to other states. Each action yields a reward from the environment: this is how the environment provides feedback that guides the agent’s future actions. 

Consider the following example: a robot (the agent) navigates (and tries to exit) a maze (the environment).

  • The state is the current situation of the environment (the robot’s position in the maze).
  • The robot can take different actions: for example, it can move forward, turn left, or turn right.
  • Successfully navigating towards the exit yields a positive reward, while hitting a wall or getting stuck in the maze results in negative rewards.

Easy! Now, let’s now make an analogy to how RL is used in the context of LLMs.

RL in the context of LLMs

Simplified RLHF Process [3]

When used during LLM training, RL is defined by the following components:

  • The LLM itself is the agent
  • Environment: everything external to the LLM, including user prompts, feedback systems, and other contextual information. This is basically the framework the LLM is interacting with during training.
  • Actions: these are responses to a query from the model. More specifically: these are the tokens that the LLM decides to generate in response to a query.
  • State: the current query being answered along with tokens the LLM has generated so far (i.e., the partial responses).
  • Rewards: this is a bit more tricky here: unlike the maze example above, there is usually no binary reward. In the context of LLMs, rewards usually come from a separate reward model, which outputs a score for each (query, response) pair. This model is trained from human-annotated data (hence “RLHF”) where annotators rank different responses. The goal is for higher-quality responses to receive higher rewards.

Note: in some cases, rewards can actually get simpler. For example, in DeepSeekMath, rule-based approaches can be used because math responses tend to be more deterministic (correct or wrong answer)

Policy is the final concept we need for now. In RL terms, a policy is simply the strategy for deciding which action to take. In the case of an LLM, the policy outputs a probability distribution over possible tokens at each step: in short, this is what the model uses to sample the next token to generate. Concretely, the policy is determined by the model’s parameters (weights). During RL training, we adjust these parameters so the LLM becomes more likely to produce “better” tokens— that is, tokens that produce higher reward scores.

We often write the policy as:

where a is the action (a token to generate), s the state (the query and tokens generated so far), and θ (model’s parameters).

This idea of finding the best policy is the whole point of RL! Since we don’t have labeled data (like we do in supervised learning) we use rewards to adjust our policy to take better actions. (In LLM terms: we adjust the parameters of our LLM to generate better tokens.)

TRPO (Trust Region Policy Optimization)

An analogy with supervised learning

Let’s take a quick step back to how supervised learning typically works. you have labeled data and use a loss function (like cross-entropy) to measure how close your model’s predictions are to the true labels.

We can then use algorithms like backpropagation and gradient descent to minimize our loss function and update the weights θ of our model.

Recall that our policy also outputs probabilities! In that sense, it is analogous to the model’s predictions in supervised learning… We are tempted to write something like:

where s is the current state and a is a possible action.

A(s, a) is called the advantage function and measures how good is the chosen action in the current state, compared to a baseline. This is very much like the notion of labels in supervised learning but derived from rewards instead of explicit labeling. To simplify, we can write the advantage as:

In practice, the baseline is calculated using a value function. This is a common term in RL that I will explain later. What you need to know for now is that it measures the expected reward we would receive if we continue following the current policy from the state s.

What is TRPO?

TRPO (Trust Region Policy Optimization) builds on this idea of using the advantage function but adds a critical ingredient for stability: it constrains how far the new policy can deviate from the old policy at each update step (similar to what we do with batch gradient descent for example).

  • It introduces a KL divergence term (see it as a measure of similarity) between the current and the old policy:
  • It also divides the policy by the old policy. This ratio, multiplied by the advantage function, gives us a sense of how beneficial each update is relative to the old policy.

Putting it all together, TRPO tries to maximize a surrogate objective (which involves the advantage and the policy ratio) subject to a KL divergence constraint.

PPO (Proximal Policy Optimization)

While TRPO was a significant advancement, it’s no longer used widely in practice, especially for training LLMs, due to its computationally intensive gradient calculations.

Instead, PPO is now the preferred approach in most LLMs architecture, including ChatGPT, Gemini, and more.

It is actually quite similar to TRPO, but instead of enforcing a hard constraint on the KL divergence, PPO introduces a “clipped surrogate objective” that implicitly restricts policy updates, and greatly simplifies the optimization process.

Here is a breakdown of the PPO objective function we maximize to tweak our model’s parameters.

Image by the Author

GRPO (Group Relative Policy Optimization)

How is the value function usually obtained?

Let’s first talk more about the advantage and the value functions I introduced earlier.

In typical setups (like PPO), a value model is trained alongside the policy. Its goal is to predict the value of each action we take (each token generated by the model), using the rewards we obtain (remember that the value should represent the expected cumulative reward).

Here is how it works in practice. Take the query “What is 2+2?” as an example. Our model outputs “2+2 is 4” and receives a reward of 0.8 for that response. We then go backward and attribute discounted rewards to each prefix:

  • “2+2 is 4” gets a value of 0.8
  • “2+2 is” (1 token backward) gets a value of 0.8γ
  • “2+2” (2 tokens backward) gets a value of 0.8γ²
  • etc.

where γ is the discount factor (0.9 for example). We then use these prefixes and associated values to train the value model.

Important note: the value model and the reward model are two different things. The reward model is trained before the RL process and uses pairs of (query, response) and human ranking. The value model is trained concurrently to the policy, and aims at predicting the future expected reward at each step of the generation process.

What’s new in GRPO

Even if in practice, the reward model is often derived from the policy (training only the “head”), we still end up maintaining many models and handling multiple training procedures (policy, reward, value model). GRPO streamlines this by introducing a more efficient method.

Remember what I said earlier?

In PPO, we decided to use our value function as the baseline. GRPO chooses something else: Here is what GRPO does: concretely, for each query, GRPO generates a group of responses (group of size G) and uses their rewards to calculate each response’s advantage as a z-score:

where rᵢ is the reward of the i-th response and μ and σ are the mean and standard deviation of rewards in that group.

This naturally eliminates the need for a separate value model. This idea makes a lot of sense when you think about it! It aligns with the value function we introduced before and also measures, in a sense, an “expected” reward we can obtain. Also, this new method is well adapted to our problem because LLMs can easily generate multiple non-deterministic outputs by using a low temperature (controls the randomness of tokens generation).

This is the main idea behind GRPO: getting rid of the value model.

Finally, GRPO adds a KL divergence term (to be exact, GRPO uses a simple approximation of the KL divergence to improve the algorithm further) directly into its objective, comparing the current policy to a reference policy (often the post-SFT model).

See the final formulation below:

Image by the Author

And… that’s mostly it for GRPO! I hope this gives you a clear overview of the process: it still relies on the same foundational ideas as TRPO and PPO but introduces additional improvements to make training more efficient, faster, and cheaper — key factors behind DeepSeek’s success.

Conclusion

Reinforcement Learning has become a cornerstone for training today’s Large Language Models, particularly through PPO, and more recently GRPO. Each method rests on the same RL fundamentals — states, actions, rewards, and policies — but adds its own twist to balance stability, efficiency, and human alignment:

• TRPO introduced strict policy constraints via KL divergence

• PPO eased those constraints with a clipped objective

• GRPO took an extra step by removing the value model requirement and using group-based reward normalization. Of course, DeepSeek also benefits from other innovations, like high-quality data and other training strategies, but that is for another time!

I hope this article gave you a clearer picture of how these methods connect and evolve. I believe that Reinforcement Learning will become the main focus in training LLMs to improve their performance, surpassing pre-training and SFT in driving future innovations. 

If you’re interested in diving deeper, feel free to check out the references below or explore my previous posts.

Thanks for reading, and feel free to leave a clap and a comment!


Want to learn more about Transformers or dive into the math behind the Curse of Dimensionality? Check out my previous articles:

Transformers: How Do They Transform Your Data?
Diving into the Transformers architecture and what makes them unbeatable at language taskstowardsdatascience.com

The Math Behind “The Curse of Dimensionality”
Dive into the “Curse of Dimensionality” concept and understand the math behind all the surprising phenomena that arise…towardsdatascience.com



References:

Shape
Shape
Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy,  bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Shape

What’s next for Nvidia?

That’s if he wants it. One industry watcher told this reporter that several years ago, he asked several Nvidia vice presidents if they wanted the big chair, because Nvidia had no clear successor to Huang and to a man they said no way. “I can’t and don’t want to do

Read More »

AMD takes on Intel and Nvidia with new 256-core EPYC server CPUs

AMD EPYC, Intel Xeon, and Nvidia Vera represent three different approaches to the evolving data-centre CPU market, added Rawat. “AMD competes through high core density, memory capabilities and power efficiency, while Intel leverages its established enterprise ecosystem, platform compatibility and customer relationships. Nvidia is taking a different approach with Vera,

Read More »

Energy Secretary Keeps Critical, Coal-Fired Generation Operational Ahead of Winter in the Midwest

WASHINGTON—The U.S. Department of Energy (DOE) issued an emergency order to address critical grid reliability issues in the Midwest. The emergency order directs the Midcontinent Independent System Operator (MISO), in coordination with Consumers Energy Company, to ensure that the coal-fired J.H. Campbell Generating Plant (Campbell Plant) in West Olive, Michigan, is available to operate and meet emergency needs while employing economic dispatch to minimize costs for hardworking American families and businesses. The Campbell Plant was originally slated to shut down on May 31, 2025, 15 years before the end of its scheduled design life.  “President Trump and the Energy Department remain committed to doing everything in our power to stop power outages for American families and businesses,” said U.S. Secretary of Energy Chris Wright. “Ensuring coal plants such as the Campbell Plant are available to operate has prevented blackouts and likely saved lives. Americans deserve access to affordable, reliable, and secure electricity regardless of whether the wind is blowing or the sun is shining.” Thanks to President Trump, coal generating plants across the country are being saved from premature retirement. For example, in 2025, more than 17 gigawatts of coal-fired electricity generation were saved from going offline. During periods of peak electricity demand, many of those plants keep the lights on.  The emergency conditions that led to the issuance of the original order persist. A joint report from the Federal Energy Regulatory Commission (FERC), the North American Electric Reliability Corporation (NERC), and NERC’s Regional Entities highlights how events that occurred during the Winter 2026 Arctic Cold Period, including Winter Storm Fern, demonstrated the need for dispatchable generation capacity from facilities like the Campbell Plant.  This order is in effect from November 15, 2026, through February 12, 2027. 

Read More »

Kosmos lets contract to advance Tiberius project offshore Louisiana

Kosmos Energy Gulf of Mexico Operations LLC, a unit of Kosmos Energy, has let a contract to Subsea 7 SA for work on the Tiberius project in the US Gulf of Mexico. The scope of work covers the fabrication, transportation, and installation of a 14-km steel catenary riser and wet-insulated flowline to the Lucius Spar, along with the installation of an umbilical and subsea controls. The Tiberius project lies in Keathley Canyon lease 964 about 400 km south of New Orleans in about 2,283 m of water.  License partners reached a final investment decision on the Phase 1A development this year, which includes completion of an already drilled well. Long-lead items have been secured, and first oil is expected by the end of third-quarter 2028.  Subsea 7, which values the contract at $50-150 million, expects project management and engineering activities to begin in the coming months, with offshore activities expected to begin in 2028. Kosmos is operator of the Tiberius project with a 33.34% interest following a farmout alongside partner Occidental Petroleum to Navitas Petroleum US LLC earlier this year.

Read More »

ExxonMobil takes FID for Longtail development offshore Guyana

The development, in the southeastern portion of the block, will target non-associated gas and condensate-rich reservoirs. Subject to future investment decisions, market demand, and development of supporting infrastructure, gas from the project could be transported to shore by pipeline. The development will use a floating production, storage, and offloading (FPSO) vessel designed to process 1.2 bcfd of gas and produce as much as 250,000 b/d of condensate. The project is expected to include about 45 subsea production and injection wells. “By developing these gas and condensate resources, we’re unlocking new value for Guyana while delivering advantaged energy to meet growing global demand,” said Dan Ammann, president, ExxonMobil Upstream Company. ExxonMobil currently produces about 900,000 b/d of oil from four developments in the Stabroek block. Its fifth project, Uaru, is expected to start up in fourth-quarter 2026, increasing total block production capacity to more than 1 million b/d. Earlier this year, the operator and co-venturers produced the 1 billionth bbl of oil from Guyana’s Stabroek block, less than 7 years after first production began in December 2019 with Liza field. Construction also continues on the Whiptail and Hammerhead developments, the sixth and seventh approved projects on the block. Whiptail is slated to start production in 2027, followed by Hammerhead in 2029. ExxonMobil Guyana Ltd. operates the Stabroek block with a 45% interest. Hess Guyana Exploration Ltd. holds 30%, and CNOOC Petroleum Guyana Ltd. holds 25%.

Read More »

MARAD approves Energy Transfer’s $1-billion Blue Marlin crude oil export port

The US Maritime Administration (MARAD) approved Energy Transfer’s $1-billion Blue Marlin offshore crude oil export project off Louisiana Nov. 5, marking the fourth oil and gas deepwater port approval under President Donald Trump’s second administration. The proposed Blue Marlin port would have capacity to export up to 2 million b/d of crude oil, with loading rates of up to 80,000 bbl/hr, including onto very large crude carriers (VLCCs). Located 99 miles off Cameron Parish, La., the project includes conversion of an existing 36-in. Stingray natural gas pipeline to transport crude oil from Energy Transfer’s Nederland terminal in Texas to the offshore infrastructure. The US Federal Energy Regulatory Commission previously authorized abandonment and sale of Stingray pipeline infrastructure for conversion to crude oil service. Energy Transfer has yet to receive final US Environmental Protection Agency air permits and certain state approvals. The project’s commercial development schedule remains uncertain. MARAD approved Blue Marlin shortly after its Oct. 28 approval of the $8.2-billion ST LNG project offshore Texas, as the agency pursues new procedures to streamline deepwater port licensing. The Trump administration also has approved Delfin LNG, a floating LNG export project offshore Louisiana, and Texas GulfLink, a crude oil export project offshore Texas.

Read More »

IEA sees oil market surplus returning as Gulf exports recover

The global oil market is showing signs of rebalancing as recovering crude exports from the Middle East lift supply, although persistent disruptions to refined product markets, particularly diesel, continue to weigh on consumers, the International Energy Agency (IEA) said. In its Nov. 13 Oil Market Report, IEA said global oil supply increased 2.3 million b/d month-on-month to 105.4 million b/d in October, following a 3 million-b/d increase in September. Output has risen 5.3 million b/d since August as Gulf exports recovered despite continued attacks on energy infrastructure and commercial tankers. The stronger supply outlook has prompted the agency to raise its fourth-quarter 2026 supply estimate by 3.3 million b/d. Global supply is now expected to decline by an average 4.8 million b/d for the full year before rebounding by 6.6 million b/d in 2027 to average 108.1 million b/d. The recovery in exports has eased market tightness and pushed crude prices lower. North Sea Dated crude fell 32% from an early-October high of $136/bbl to about $93/bbl at the time of the report, although prices remained $22/bbl above prewar levels. Gulf oil exports, excluding Oman, reached 18.4 million b/d in October, about 26% below prewar levels. Crude exports accounted for most of the recovery, reaching 15 million b/d, equivalent to 82% of prewar volumes. However, IEA cautioned that continued attacks on shipping remain a significant downside risk to the supply outlook. Demand under pressure Global oil demand is forecast to decline by 2.7 million b/d in 2026 to 102.3 million b/d, reflecting disruptions to refined product and liquefied petroleum gas exports from the Gulf, elevated fuel prices, and weaker consumption across major markets. The contraction is expected to moderate in fourth-quarter 2026 to 2.7 million b/d year-on-year, compared with an average decline of 4.4 million b/d in the second and

Read More »

What’s next for California’s climate laws following an appeals court split decision

The Ninth Circuit Court of Appeals issued a split decision on California’s climate disclosure laws, Senate Bill 253 and Senate Bill 261, on Monday, allowing some emissions reporting requirements to proceed, while blocking others on the basis of them being considered in part or completely unconstitutional under the First Amendment. In the decision, the federal appellate panel partially upheld and partially overturned a district court’s decision from August 2025, which declined to enjoin the laws. The U.S. Chamber of Commerce and other business associations filed the lawsuit, requesting to temporarily block implementation of both climate laws while their constitutionality was under review. Implementation of SB 261 was already temporarily halted pending review by the Ninth Circuit last November.  SB 253, or the Climate Corporate Data Accountability Act, requires companies with more than $1 billion in revenue to report scope 1 and scope 2 emissions, with eventual scope 3 disclosures. SB 261, or the Climate-Related Financial Risk Act, requires companies with more than $500 million in revenue to publicly disclose their climate-related financial risks and countermeasures. Both climate disclosure regulations were signed into law in October 2023 by California Gov. Gavin Newsom and have since faced a spate of legal challenges. The ruling came a day before the inaugural emissions disclosure deadline for SB 253. Breaking down the split decision The Nov. 9 decision allowed for the reporting of a company’s direct emissions to proceed under SB 253’s scope 1 and scope 2 disclosure requirements, affirming the lower court’s decision. The court ruled such disclosures “regulate commercial speech,” which the appellate panel defined as factual and objective information on a company’s own operations. However, the appeals court reversed the lower court’s ruling when it came to SB 253’s scope 3 disclosure requirements, which ask companies to report the indirect emissions generated

Read More »

Cisco forecasts $9 billion in AI infrastructure revenue in FY27

Cisco continues to benefit from a “networking supercycle,” reporting record revenue—driven by demand for AI infrastructure, security, and campus refreshes from hyperscalers and enterprise customers—for its fiscal first quarter ended October 24. “In Q1, we delivered record quarterly revenue of $18.5 billion, up 24% year over year, and record non-GAAP earnings per share of $1.40, up 40% year over year. These are our strongest top andbottom-line growth rates in over 15 years,” Cisco CEO Chuck Robbins said on the vendor’s quarterly financial call. “Networking product orders grew 28% in Q1, delivering our ninth consecutive quarter of double-digit growth. In fact, we saw double-digit growth across every product family: service provider routing, compute, enterprise routing, wireless, data center switching and campus switching,” Robbins said. “These broad-based results support our view that we are in the early stages of a multiyear, multibillion-dollar networking supercycle.”

Read More »

How TI’s ADC34DSO75 Catches the Nanosecond Power Glitches Behind AI Data Center Downtime

Today’s advanced AI data center can be viewed as a vast, complex and interdependent digital organism in which a virtually infinite number of components must be tracked all the time, in real time. It’s a monitoring job of staggering proportions. At stake is the operation of highly critical workloads and the maintenance of return on investment for high-cost compute equipment. At risk is sub-optimal performance or, in worst-case scenarios, the nightmare of data center downtime. Sponsored Resources: Why traditional telemetry misses nanosecond power-rail events Telemetry has long been relied upon to monitor data center servers, test and measurement equipment, wireless infrastructure and other systems for real-time information about voltage, current, power, temperature and overall system health. But most systems operate at millisecond-level intervals or slower, which is six orders of magnitude longer than the transient events that now occur at the processor power rails and the tools meant to observe them. The requirements and high stakes of today’s AI data centers and workloads have called for a new generation of advanced, nanosecond-level telemetry gear. Texas Instruments has established a leadership position in this sector. The company has developed the first analog-to-digital converter (ADC) to enable 24/7, low-power, event-based remote monitoring at nanosecond resolution. Let’s unpack this statement and examine why TI’s new monitoring technology is so important for the data center industry. What is the TI ADC34DSO75?    We’re talking specifically about TI’s ADC34DSO75, the industry’s highest-performance, 3 gigasamples per second (GSPS) ADC, delivering thousands of times faster performance than current offerings. Launched at the Open Compute Project Global Summit conference in October, the ADC34DSO75 addresses the shortcomings of existing monitoring solutions that can’t detect the nanosecond-level power-rail glitches that cause costly data center downtime. The ADC34DSO75 combines ultra-high-speed sampling with advanced on-chip digital signal processing in a small form factor,

Read More »

SC26 Preview: AI-HPC Convergence Takes Center Stage in Chicago

As the high-performance computing community prepares to gather in Chicago for SC26, November 15–20, the latest episode of the Data Center Frontier Show offers a timely preview of the technologies, market developments, and infrastructure challenges expected to shape this year’s conference. Hosted by DCF Contributing Editor Doug Black, the conversation brings together Lois McKinnis, Argonne Distinguished Fellow and senior computational scientist at Argonne National Laboratory; Addison Snell, founder and CEO of Intersect 360 Research; and Earl Joseph, CEO of Hyperion Research. The episode also inaugurates Data Center Frontier’s new Frontier Computing channel, expanding DCF’s coverage of the technologies and infrastructure driving the convergence of high-performance computing, AI, QC and next-generation scientific discovery. The central theme is unmistakable. AI is no longer simply an adjunct to traditional supercomputing. It is changing how scientific computing itself is performed. “We’re really moving beyond focusing just on bigger machines, faster machines,” McKinnis observes, pointing to the growing integration of AI into scientific software development, simulations, and research workflows. Joseph underscores the acceleration, citing new Hyperion research finding that 70% of respondents are already using AI for code generation. The panel also examines the U.S. Department of Energy’s Genesis Mission, a sweeping effort to link supercomputers, large-scale AI infrastructure, and cloud resources in support of scientific discovery. Joseph calls Genesis “the first time I’ve ever seen something that’s truly a broad national resource,” while Snell emphasizes that investments in AI infrastructure must be matched by investments in science itself. For data center professionals, the discussion turns to increasingly consequential questions about the economics and sustainability of the AI buildout. Snell warns that the industry’s focus on rapid construction has outpaced serious discussion of infrastructure lifecycles and upgrades. Joseph highlights the mounting difficulty of financing supercomputers costing $100 million or more when new GPU generations arrive

Read More »

Nvidia’s leadership crisis puts enterprise AI’s dependence on one company under scrutiny

“More broadly, it throws into doubt the company’s leadership and security practices, and ignites questions around board-level governance and oversight,” he said. The ‘default choice’ in the market Huang founded Nvidia in 1993, initially targeting the gaming market. But the company pivoted in recent years to meet the intensive compute demands of AI. Its advanced GPUs not only excel at rendering graphics, but can handle an enormous number of simultaneous computations, making them a natural fit for neural networks and LLMs. The company began focusing more on developing chips and architecture for AI tools and data centers, and has continued to ink new deals and establish partnerships with tech companies serving nearly every level of the stack, including Anthropic, Meta, Dell, HPE, Microsoft, AWS, Google Cloud, Accenture, Marvell Technology, Hugging Face and many others.

Read More »

Jensen Huang Resigns as NVIDIA CEO, Sending Shockwaves Through AI Data Center Industry

NVIDIA co-founder Jensen Huang resigned Thursday morning (November 12) as president, CEO and board member of the world’s dominant AI chipmaker, a stunning leadership change that introduces new uncertainty into the global artificial intelligence infrastructure boom. In an official announcement, NVIDIA confirmed that Chief Financial Officer Colette Kress will serve as interim CEO while the board conducts a search for Huang’s permanent successor. The company also appointed longtime director Harvey C. Jones as its first board chairman. The departure ends Huang’s 33-year leadership of the company he co-founded in 1993 and comes amid a rapidly escalating controversy involving his reported personal relationships, potential Chinese intelligence activity and questions about corporate governance. According to Reuters, NVIDIA shares fell more than 10% in premarket trading Thursday as investors weighed the implications for the approximately $4 trillion company and the wider AI technology sector. The company’s announcement did not specify a reason for Huang’s resignation. National Security Investigation Precedes Leadership Change The leadership shakeup follows several days of increasing scrutiny over Huang’s reported relationship with a businesswoman whom U.S. authorities suspect of connections to Chinese intelligence, alongside separate reporting concerning a relationship with a senior NVIDIA executive. On November 8, NVIDIA’s board announced an independent investigation into whether Huang had been targeted by a foreign intelligence operation. At the time, directors stated that they had no reason to believe company secrets or national security had been compromised. The controversy subsequently attracted congressional attention, including calls for Huang’s resignation and questions about semiconductor export controls and the security of advanced AI technology. The espionage allegations remain contested, and no compromise of NVIDIA technology has been publicly established. For the data center industry, however, the immediate question extends beyond the circumstances of Huang’s departure. It concerns the potential disruption of an infrastructure ecosystem increasingly organized

Read More »

Google’s Utah Deal Puts Geothermal Power on the Data Center Map

In Taiwan, the Work Begins With Access to Land Google’s first geothermal procurement agreements in Asia call for an initial 10 MW of new generation on Taiwan’s grid to support local data centers and offices. When Google and developer Baseload Capital announced the agreements in April 2025, a Google energy executive told Taiwan Public Television Service that the companies expected operations to begin in 2029. The original timetable allowed roughly four years between signing a customer and delivering electricity. Baseload’s development work in Hualien County illustrates one reason such projects take time. In November 2025, the company reported that members of the Hongye Tribe had voted to approve a public-land lease for geothermal exploration. Baseload opened a local office, reached an agreement with the community, and secured an exploration permit for a two-year program of drilling in stages. The company has described Hongye separately from Google’s contracted supply. What the Hualien project documents is an early milestone rarely visible in a power purchase announcement: the developer secured community consent before drilling could establish what the field might produce. Kenya Has the Generating Field. The Campus Remains Unresolved. In Kenya’s Rift Valley, the appeal of Olkaria was evident when Microsoft and G42 announced their digital investment initiative in 2024. Kenya already had a fleet of geothermal generating stations at Olkaria. G42 and local partners proposed building a campus there, powered by geothermal electricity and hosting Microsoft Azure services for a new East Africa cloud region. Two difficulties emerged. In May 2026, Semafor reported that President William Ruto had cited insufficient available power as an obstacle to the project. Separately, Reuters, citing Bloomberg, reported disagreements over payment guarantees. Microsoft and G42 had sought a commitment from the Kenyan government to buy a certain amount of computing capacity annually. Kenya’s principal secretary for

Read More »

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs).  In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

Read More »

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

Read More »

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

Read More »

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Read More »