Skip to content

Podcast

The Race to Put AI Agents Everywhere

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • Nvidia Signals Demand at Unprecedented Scale
    • Nathaniel Whittemore argues Jensen Huang’s trillion-dollar forecast matters less as accounting math than as proof Nvidia sees sustained demand for $500 billion annual sales.
    • He notes only Walmart and Amazon are near that revenue scale, making Nvidia’s projected growth historically unmatched. Transcript: Nathaniel Whittemore I’m old enough to remember when a trillion dollar market cap was a big deal. And now here we are, AI is booming, and NVIDIA CEO Jensen Huang has kicked off the company’s annual GTC conference with a massive prediction that the company will see a trillion dollars In revenue between now and 2027. At every GTC, Jensen’s keynote, which is planned but not fully scripted, is the big event. This one was no exception. It was two and a half hours long, totally jam-packed with big announcements. We got confirmation of the new Grok-powered server focused on inference. The new rack-mounted system will combine 256 Grok chips with 72 NVIDIA Rubin GPUs, delivering 35 times the inference efficiency of current generation Blackwell chips, with the system Expected to ship in the second half of this year. Jensen also unveiled a new Gen.AI system that can enhance video game graphics on the fly. Called DLSS5, the technology combines traditional graphics with an AI filter to create stable photorealistic graphics. Being able to produce this effect at runtime on consumer hardware is a big breakthrough that could significantly change the way video games are made. For my claw fans out there, there is a new entrant into the open claw category, which we’ll cover in the main episode. But ultimately, while the keynote had many big moments, none grabbed headlines like Jensen’s massive revenue forecast. Late last year, Huang said that he expects $500 billion in sales in 2026. On Monday, he doubled the forecast to a trillion, stating, I believe that computing demand has increased by 1 million times in the last two years. It’s the feeling that we all have. It’s the feeling every startup has. Now, some tried to downplay the forecast, noting that it merely combines two financial years at $500 billion apiece, meaning it’s not so much a material change. Bloomberg analyst Kunjan Sobani wrote, The update should ease fears of a pullback in 2027 as Rubin enters the cycle, although it may also reset market expectations higher and raise The bar again. This feels to me to be slightly missing the bigger picture. Jensen has now signaled that Nvidia can see enough demand to drive $500 billion in annual sales. This would more than double revenue from the past year. In fact, the list of companies with a half a trillion in annual sales is just Walmart and Amazon, with Saudi Aramco falling slightly short. If Huang’s forecast is correct, it will be completely unparalleled growth for a company of anywhere near NVIDIA’s size. Remarking on the event, Josh Kale wrote, Demand doubled his demand forecast to a trillion dollars, announced data centers in space, and closed the show with a robot singing country Music. This is NVIDIA’s world. Everyone else is just renting ComputeNet. (Time 0:01:02)
  • Meta’s Nebius Deal Shows Compute Hunger
    • Meta’s $27 billion Nebius deal shows AI labs will absorb any compute capacity they can find, not just hyperscaler infrastructure.
    • Nathaniel Whittemore says the simpler explanation is industry-wide capacity constraint, with neoclouds becoming meaningful beneficiaries. Transcript: Nathaniel Whittemore On that front, Meta has signed a $27 billion deal with Nebius. Nebius, which is similar to CoreWeave and NScale, operates smaller AI data centers than their hyperscaler counterparts. This often includes differentiated chips or full-stack support for model training or specialized inference. Nebius’ new deal with Meta spans five years, and this is in addition to a $3 billion deal signed by Meta in November. Nebius plans to deploy NVIDIA’s new Verorubin chips on Meta’s behalf. The chips are expected to be available in the second half of this year, with Nebius powering on the new cluster early next year. Now, while possible that Meta is turning to Nebius for specialized data center management, the simpler explanation is just that the entire industry is capacity-constrained right Now, and that Meta, like all the other AI labs, is gobbling up all the available data centers they can get their hands on. That includes partnering with the neoclouds to take any capacity they can offer. The deal, also represents a phase shift for the smaller end of the data center industry. Nebius is one of the larger neoclouds, yet they only had a little over a billion dollars in revenue last year. Meaning for my math friends out there, this is an order of magnitude larger than all the business they’ve done so far. AI infrastructure continues to scale up at a massive pace, and the neoclouds seem to be getting their slice of the action. One area of infrastructure build-out that has been a little bit, shall we say, beleaguered is the OpenAI Stargate effort. The company has now appointed new leaders to oversee their revamped and restructured Stargate. Now, over the last couple of months, we’ve heard all sorts of things about Stargate. We learned that the joint venture with Oracle and SoftBank never really got off the ground, and more recently that OpenAI was walking away from expansion plans at the flagship site In Aveline, Texas. That reporting also suggested that the Stargate name would be attached to all data centers operated by OpenAI, rather than only their own site developments. Now the information reports that the structure of the new look Stargate division has been put in place. Former Intel executive Sachin Khati will oversee the division, which consists of three distinct teams. One team will work on technical data center design, another on commercial partnerships with various cloud providers and chip manufacturers, and the third will be responsible for On-the management of facilities. Previously, OpenAI’s infrastructure teams were organized by project rather than role and reported up to President Greg Brockman, meaning this restructuring could represent a More specialized and dedicated in-house team being put in place. Reporting also confirms that OpenAI is less concerned about ownership of data centers and more willing to lease in order to scale up compute. This would comport with basically everything else we’re seeing in the industry, where all of the fancy and fiddly efforts are kind of flowing by the wayside in order to just get access To as much compute as possible. (Time 0:03:34)
  • Chinese Labs Are Rebalancing Open Source
    • Chinese AI labs appear to be splitting strategy between lightweight open models for distribution and stronger closed models for enterprise monetization.
    • Alibaba reorganized Quen under a token-focused business unit, while Z.ai released GLM-5 Turbo as closed source despite its open-source reputation. Transcript: Nathaniel Whittemore The first story is that Alibaba has restructured their AI organization in a shift, it seems, to maximize profits. Rumors were swirling earlier this month that a big move was in the works as three senior researchers left the Quinn team. The departures included technical lead Justin Lin, who is credited with shepherding Quinn from its first training run to becoming one of the most popular open-source models. Speculation at the time was that Alibaba was shifting focus from pure research to driving AI-related revenue through their first-party API. Some wondered if this shift would herald the end of open-source Quen models. According to a memo cited by Bloomberg, the restructuring is now complete. The Quen research team has been folded into a new division that also includes consumer-facing apps and AI-related products like the Quark smart glasses. The new division is called the Alibaba Token Hub and will be directly led by CEO Eddie Wu. Wu wrote in the memo, Bloomberg writes that the restructuring, quote, signals the company’s clear emphasis on monetizing AI. The division’s name is a direct reference to the units of computing that companies charge users. Bloomberg Meanwhile, another Chinese startup, Z.ai, has released a faster, cheaper version of their leading model, but they are keeping it closed source. The new model is called GLM-5 Turbo, and offers similar performance to GPT 5.2 at a cost that’s closer to Gemini 3 Flash. The speed boost is arguably a bigger deal, with the model optimized for running open-claw tasks like tool use and long-chain execution. ZAI said the model would be released as closed-source, but that its capabilities would be folded into future open-source releases. VentureBeat wrote that the decision is emblematic of a broader shift in the Chinese market. They suggest that Chinese labs are adopting an approach where lightweight open-source models are used to boost distribution and generate goodwill among developers, while more Powerful models are delivered as proprietary systems aimed at generating enterprise sales. Writes VentureBeat, This, I think, is a trend that is worth keeping an eye on. Corinne on X wrote, ZAI has been the loudest open source voice in AI for two years. They just released their first closed source model. That one decision tells you more about where the industry is heading than any benchmark. By the way, for those of you who are just listening and not watching, the picture that ZAI chose to release the model with is a glowing lobster riding a horse. Nathan Lambert, who just wrote an interesting essay on this topic, wrote, We’re in the era when the cost of building LLMs is skyrocketing and the why for releasing them Ethan Lambert, Who just wrote an interesting essay on this topic, wrote, Definitely a trend worth watching, but for now, that is going to do it for the headlines. (Time 0:06:29)
  • OpenClaw Marked AI’s Second Moment
    • Nathaniel Whittemore frames Q1’s biggest shift as agents becoming genuinely viable, with OpenClaw representing AI’s new systems-access capability set.
    • He says the breakthrough is not chat quality but agents actually getting work done through machine access and persistent experimentation. Transcript: Nathaniel Whittemore We’re coming up on the end of Q1, and as part of that, I’ve been working on a big quarter two state of AI report. As you might expect, maybe the key story of Q1 was OpenClaw, not even just because of OpenClaw itself, but because of what it represented. I think you can look at OpenClaw as the instantiation of the new capability set that shifted around the end of last year, and which has really come to the fore this year. It’s what I called on yesterday’s episode, AI’s second moment, and refers to this idea that agents are actually at this point viable, and that people are in the midst of a million experiments Right now giving agents systems access, building new types of systems to have agents interact, and especially, and as we’ll talk about today, solving some of the key challenges of Agents to make sure that they can diffuse across the entire business world. Part of the specific catalyst for today’s show is NVIDIA CEO Jensen Huang’s speech at their annual GTC event yesterday, where Jensen said explicitly, every software company in the World needs to have an open claw strategy, and where he began to show off their enterprise-grade version of the software. Now, even before this, the clawfication of the world was well underway. Kevin Simbach from Delphi Labs recently wrote a post about all of the different variations and competitors, and started by claiming that OpenClaw opened the door. Kevin writes, technical experiments that produced nothing more than timeline slob. After OpenClaw and with the advent of Opus 4.5 and 4.6, agents became accessible, just a telegram message away, always on, actually doing helpful things, and kickstarting a new generation Of digital opportunities. OpenClaw quickly proved two things at once. People don’t want AI chat, they want to get work done, and giving an LLM broad access to your machine and or personal info is both insanely useful and mildly terrifying. (Time 0:12:37)
  • The Computer Is Becoming the Agent Interface
    • The emerging product thesis is that users do not want chatbots; they want agents embedded in the full canvas of their computer and workflows.
    • Nathaniel Whittemore links OpenClaw, Notion agents, and Perplexity Computer to the same shift toward systems deeply integrated with files, apps, and context. Transcript: Nathaniel Whittemore So, as he writes, the last month has been a weird kind of Darwinism, with builders shipping faster than slot posters, security people screaming into the void, and a growing cohort of People saying, oh crap, this is actually going to rewrite how software and digital businesses work. And yet, as Kevin acknowledges, not everyone is sold on OpenClaw itself, and there has been a mad race to build or update alternatives. A bunch of them, like Nanobot, ZeroClaw, PicoClaw, or Nanoclaw, are all attempts to reduce the overall complexity down to some specific useful feature set, and then there’s others Like OpenFang, Hermes, Moltis, and IronClaw that are all trying to bring security to it through self-hosting. Yet if that represents one end of the spectrum of the clawification of AI, on the other hand you have a huge number of companies, some that were AI native, some that weren’t AI native, Offering up what are effectively their own versions of OpenClaw. In other words, agents that are deeply integrated and integratable with some key set of systems and personal context. At the end of February, Notion introduced custom agents, which have a lot of features in common with OpenClaw, and also all of the context that comes from integration with Notion, where Many companies are running all of their information. And of course, we also got Perplexity Computer. Perplexity Computer is a very full-throated reimagining of Perplexity from the ground up into a complete problem-solution design system capable of spinning up complex systems Of agents and sub-agents to get things done and build things that people want. In the couple weeks since Perplexity released Computer, they’ve also released Computer for Enterprise, which can operate from within Slack and which also has direct connections, They claim, to more than 400 applications. And they also even got on the Mac Mini part of the theme with their launch of Personal Computer, which they call an always-on local merge with Perplexity Computer that works for you 24-7. Getting philosophical, Perplexity CEO Arvind Srinivas wrote a long post about why the AI is the computer. In it, he argues, AI models are becoming so capable that the products built around them have been bottlenecked for showing their true potential. The chat UI is good for answers and agents are good for individual tasks. Meanwhile, the UI for entire workflows has always been the computer. Effectively, what Arvind is arguing is that the full potential of agentic systems requires the complete canvas of what your computer offers, bridging from local files to cloud systems And beyond. (Time 0:14:22)
  • Desktop Agents Are Chasing Real Business Tasks
    • Manus and Adaptive both launched desktop-style agent systems that act on a local machine instead of staying trapped in the cloud.
    • Adaptive’s example targets a hardware store owner dropping a spreadsheet of 47 products for automatic entry into Square. Transcript: Nathaniel Whittemore Menace, which was purchased by Meta in December, was one of the early leaders throughout 2025 in general purpose agents. This week they announced a new Manus desktop app, the key feature of which they called My Computer. Very much picking up on the new design pattern, they write, it’s your AI agent now on your local machine. The use cases they point to include organizing thousands of unsorted photos, renaming hundreds of invoices, building desktop apps in Swift entirely on your computer with no code Written manually, combining with existing connectors to create seamless automated workflows, and creating local routines with personal projects, agents, and scheduled tasks. In the blog post, without naming OpenClaw, they acknowledge the realization of the need to be able to bridge from cloud cloud My computer, then, is a way to close that gap. Now, one interesting thing about the Manus announcement is that they’re thinking a little bit ahead in terms of the specific opportunities that come with desktop. For example, doing something that I haven’t seen from a lot of the other competitors, they’re actually pushing the idea of building fully working Mac apps, not just cloud-based applications That other people would use. Cedric Chi writes, CloudCode, Cowork, OpenClaw, Codex, and Manus all seem to be converging on the same idea. The agent lives on your machine. The second related announcement yesterday came from Adaptive. They wrote, Introducing Adaptive Computer. We put AI inside of an always-on personal computer that it uses to get work done. Schedule agents, create software, automate anything. By the end of this year, they write, AI agents will use more software than humans do. You won’t be the one clicking the button or browsing the webpage. Your agent will. That requires a new kind of computer. We built one. Most business software they continue has the same problem. Someone has to sit there and operate it, moving data, updating records, filling out forms. That someone is usually you. The example they gave, interestingly, is the real-world business example of a hardware store owner who has 47 new products in a spreadsheet and needs them to get added to Square. Adaptive says drag the file into Adaptive, tell it what you want, and it handles the rest. Out of scope of this particular show, but I think it’s super interesting that you’re seeing these very bleeding-edge tech companies trying to appeal to the hardware store owner use Case. They then go on to pitch their secret sauce, which they call encoded memory. They write, what makes Adaptive different is what happens after. It encodes what it learned, how Square works, how your catalog is organized, and how you prefer things to be done. So the next week, when you ask for a daily sales report at 8pm, it builds the agent, schedules it, from Square data that it already knows. Now anytime there’s a new launch, it tends to be pretty hard to get good signal from Twitter at this point because so much of the discourse is either AI bots or undisclosed paid tweets. But Ole Lemon did write of a good experience that he recently had through Adaptive. The example he gave was automating YouTube AI research. Basically, his argument is that YouTube has a ton of really great videos on in-depth AI systems that are extremely up-to and current with the moment, but there is a ton to filter through That makes it hard to sit around and browse to get the diamonds in the rough. The prompt he gave Adaptive was, analyze YouTube videos about AI and clawed workflows from the last 24 hours that have at least 10,000 views, pull the full transcripts, extract the Top three most tactical and actionable workflows, and send me a daily email report every morning. (Time 0:16:39)
  • NemoClaw Targets the Enterprise Security Gap
    • Nvidia’s NemoClaw tries to solve the main blocker to enterprise agents by wrapping OpenClaw in sandboxing, access controls, and policy guardrails.
    • Nathaniel Whittemore says it adds privacy and security while staying model- and hardware-agnostic, making enterprise deployment more plausible. Transcript: Nathaniel Whittemore The third and maybe biggest OpenClaw and agent-related announcement yesterday, however, came from NVIDIA. The context for that quote we heard at the beginning about every company needing an OpenClaw strategy was the setup for Jensen introducing NemoClaw. Now, functionally, this is not actually a standalone agent, but rather a software toolkit built on top of the OpenClaw project. OpenClaw creator Peter Steinberger wrote yesterday, been so much fun cooking OpenShell and NemoClaw with the NVIDIA folks. Huge step towards secure agents you can trust. So what this is, is basically an approach that adds privacy and security to OpenClaw instances by giving them an isolated sandbox to work in. The agent can still access resources as necessary, but the NemoClaw stack formalizes access control. Specifically, it integrates into policy-based security and other guardrails to theoretically allow it to operate safely within enterprises. NemoClaw is model and hardware agnostic, and allows users to choose between cloud and local models. Encapsulating this whole shift, Jensen Huang said, OpenClaw gave the industry exactly what it needed at exactly the time, just as Linux gave the industry exactly what it needed at Exactly the time, just as Kubernetes showed up at exactly the right time, just as HTML showed up. It made it possible for the entire industry to grab onto this open source stack and go do something with it. Now what’s been interesting about the response is that for most, although not for all, this hasn’t been a jump the shark or jump the lobster moment. Instead, people have been pretty enthusiastic about what NVIDIA is trying to do. Kevin Simbach again writes, excited to dig into NemoClaw. Have spent a good bit of my career in enterprise. I’ve been pretty vocal about OpenClaw not being enterprise ready. But the concept of an agentic workforce is a killer and enterprises are going to want it, so this may be what really kicks it off. Tristan Rhodes writes, I’ve been avoiding OpenClaw and waiting for it to mature. There have been countless variation in forks along the way, but NVIDIA is the most valuable company in the history of the world. Does that mean NemoClaw becomes the dominant variation of OpenClaw? Eric Su wrote an entire X article called NVIDIA just solved the One Problem Blocking AI Agents, of course all about the security concerns. (Time 0:20:04)
  • Enterprise Buyers Are Already Testing OpenClaw
    • In Nathaniel Whittemore’s EnterpriseClaw program, participants split roughly evenly between learning OpenClaw and using agent-building tools like Codex, Cursor, and Clauded Code.
    • He treats that demand as evidence enterprises want to learn the platform even before OpenClaw is fully enterprise-ready. Transcript: Nathaniel Whittemore Now one thing I will say that’s been interesting from our own experience, regular listeners know we have two different OpenClaw related things going on right now. Claw Camp is an open, free, self-directed program that walks people step-by through setting up their own OpenClaw and giving them access to a community of other builders who can help Them along the way that at this point more than 7,000 people have signed up to participate in. EnterpriseClaw, meanwhile, is a managed six-week executive sprint that’s meant to help individual enterprise leaders and teams from enterprises get that same sort of learning But in a much more in-depth and supported way. Now, as part of EnterpriseClaw, we gave people the choice to either use OpenClaw or do a generic version of agent team building using Clawed Code, Codex, Cursor, etc. And interestingly, it’s about half and half in terms of who wanted to learn on OpenClaw versus who wanted to use other systems. Meaning that even in the pre-Enterprise grade OpenClaw world, there is still demand for figuring out how to use this platform, which I think is certainly validation of everything That Jensen is saying. Now, Robert Scoble had an interesting note from the NVIDIA GTC Expo Hall that was actually more about OpenAI than it was about NVIDIA. (Time 0:22:03)
  • OpenAI Is Refocusing on Enterprise and Coding
    • OpenAI is narrowing from many product bets toward enterprise productivity and coding, treating the competitive moment as a continuing code red.
    • The shift shows up in Codex subagents and rapid API adoption, with GPT 5.4 reaching 5 trillion daily tokens within a week. Transcript: Nathaniel Whittemore He writes, visiting the Expo Hall shows you why OpenAI is changing strategy. All the big booths are enterprise. The biggest news here is how NVIDIA is bringing OpenClaw to the enterprise. Which brings us to another important story from yesterday. The Wall Street Journal reports that OpenAI is done with side quests and will refocus on nailing a core business which is now more than ever refocused on enterprise encoding. The journal reporting states that CEO of Applications, Fiji Simo, has delivered a wake-up call within the company, pointing out that their do-everything strategy has reduced their Lead on the competition. Simo told staff last week, We cannot miss this moment because we are distracted by side quests. We really have to nail productivity in general, and particularly productivity on the business front. Now, this is of course a big shift away from Sam Altman’s traditional management approach, which he described as betting on a series of startups within the company. That led to a fairly dizzying array of product bets, including the Sora app, the Atlas browser, and the yet-to Johnny Ive device, just to name a few. As basically everyone on AI Twitter has done, the journal compared that approach to Anthropic’s very narrow strategy built around agentic coding and the way that that expands into Broader sets of knowledge work for the enterprise. Now, it’s not new that OpenAI has decided to refocus efforts on similar themes, that’s been the big story since GPT-5 was released and Codex came out, but there clearly seems to be a new Urgency. Interestingly, according to Simo, the Code Red from last year is not over. Last week, she told staff, we are very much acting as if it’s a Code Red. And while a lot of people are speculating around what might get the axe because of that, for example, the much maligned ads approach, every day it seems we get some new announcement around Codex and their larger coding suite. The most recent, and the one that we got yesterday, and that I think is coherent with all of these qualification themes, is the native integration of sub-agents into Codex. The OpenAI Developers account writes, You can accelerate your workflow by spinning up specialized agents to keep your main context window clean, tackle different parts of a task In parallel, steer individual agents as work unfolds. LLM Junkie M. Will writes, in the next Codex update, multi-agents will get a massive flexibility upgrade. Hey Codex, when you implement this plan, I want you to delegate all of the lower complexity tasks to GPT 5.3 Spark subagents. Instead of needing to create 100 different custom agent roles for different situations, you can just prompt your agent to spawn whatever model or reasoning level you want, with only Natural language. Emmanuel DiPietro went through some use cases for the subagent system, things like code review where he argues you could have one agent per concern, test coverage with one subagent Writing tests, another checking edge cases, and another validating, etc, etc. And it’s clear that even though the foot is still firmly on the gas, the shift in OpenAI strategy seems to be bearing some fruit. OpenAI president Greg Brockman wrote yesterday, GPT 5.4 has ramped faster than any other model we’ve launched in the API. Within a week of launch, 5 trillion tokens per day, handling more volume than our entire API one year ago, and reaching an annualized run rate of 1 billion in net new revenue. Sam Altman showed a chart of Codex usage being very aggressively up and to the right, adding the Codex team are hardcore builders and it really comes through in what they create. No surprise, all the hardcore builders I know have switched to Codex. Responding to the news about OpenAI shifting focus, Duane on X writes, I actually thought OpenAI were already doing a good job focusing on coding. Codex is amazing for coding. One area where they absolutely fail is UI. GPT 5.4 can’t design to save its life even if you have super detailed skill to guide it. It has zero taste. And for what it’s worth, I talked about this on my operator show, this has very much been my experience to the point where I can’t just give Codex guidelines, I literally have to give it The actual design files from Claude for it to copy exactly. Although my experience with Codex when it comes to actually building has been really good. (Time 0:23:07)
  • Q2 Will Test How to Productize Agents
    • Nathaniel Whittemore sees Q2 as a sprint to productize agents, moving from experimentation into broader enterprise diffusion.
    • He expects the key uncertainty to be complexity: OpenClaw broke out despite being hard to use, so teams will test many usability bands. Transcript: Nathaniel Whittemore Summing all this up, if Q1 was a realization that agents are here, and a mass, wide-scale experimentation with the form factors and design patterns introduced by OpenClaw, Q2 is set Up to be an absolute sprint to productize those agents and get them ready for broader diffusion, especially within the enterprise. One thing that I will be watching closely is how much old patterns of productization, where conventional wisdom was all about simplifying things for wider audiences, still hold given That the breakout was this incredibly complex system in OpenClaw. I’m not sure I know where the right complexity band is going to be, or if it’s going to be a spectrum of different types of complexity for different users, but I can guarantee that just About everything that can be tried will be tried in the quarter to come. (Time 0:26:38)