transcript

Speaker 1:
[00:01] Today on The AI Daily Brief, how headless agents will change software and work. Before that, in the headlines, the compute competition heats up. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. OpenAI has accelerated their ambitious roadmap for scaling inference. In an X-post, they said they plan to deploy 30 gigawatts of compute by 2030. Now, during the Stargate announcement at the beginning of 2025, OpenAI announced their massive 10 gigawatt target by the end of the decade. Meaning, for those of you who are sitting there doing the math, they are tripling their medium-term compute goals. To give a sense of the scale, Epic AI estimated that total global AI data center capacity reached 30 gigawatts at the end of last year. That figure includes both the power use for chips and ancillary systems like cooling and networking, so it's not entirely clear this is an apples-to-apples comparison. 30 gigawatts also happens to be roughly peak power demand for the entirety of New York State. OpenAI, meanwhile, says that they are already well on their way. They said that they tripled their compute supply last year, going from 0.6 gigawatts to around 1.9 gigawatts. OpenAI also said that they've identified, whatever that means, more than 8 gigawatts already. Now, for those of you who feel like, sure, I know why this is important, but it's not really the part of AI that impacts me. This year has shown exactly why it actually does affect all of us. The rise of agentic work this year has brought a huge inference crunch. Most observers believe that Anthropic is straining under a wave of new demand, though they've yet to discuss that issue in public. Instead, we're seeing a bunch of weird things that end up feeling like missteps that could all be attributed to just simply not having enough compute and power to serve as much of their AI as people want. Hader writes, Right now compute is everything. Anthropic does not have enough of it, which is why Opus Performance is degrading. OpenAI felt the pressure in 2025, especially after the Ghibli wave, which pushed SAM to lock in long-term compute. Until there is a breakthrough in model architecture or chip design, this cycle will continue. Now whether or not that's true, it is certainly the case that OpenAI is positioning themselves as the startup with ample compute. During the rollout of ChatGPT images on Tuesday, President Greg Brockman remarked, Really incredible what you're now able to do with a little bit of compute. At the moment, it might be a few subtle jabs, but OpenAI pretty clearly wants customers to know that they have ample capacity to accommodate any clawed refugees out there. At the same time, SemiAnalysis is calling out another crucial bottleneck in energy supply. In a classic vague post on X, they wrote, 100 gigawatts under contract, 10 gigawatts of capacity left through 2030, pricing up double digits, competitor literally stopped taking orders, and they generated more free cash flow in 90 days than the prior 365. This market is the tightest it's been in decades and nobody's talking about it. Most believe this referred to GE Vernova, one of the few suppliers of gas turbines required for co-located power generation. GE Vernova stock was up 13.7% on Wednesday after they delivered a blowout earnings beat. They reported $17.44 in earnings per share, smashing the consensus forecast of $1.67. Below the headline, they also discussed a massive increase in their backlog, with new orders rising 71% last quarter to bring that backlog to $163 billion. They're only forecasting $45 billion in revenue this year, so it's likely their entire capacity for the remainder of the decade is spoken for unless they expand. TLDR, everywhere you look, there are massive bottlenecks showing up for scaling compute. Consequently, one thing you're not hearing anymore is that data center construction is a bubble that could end in a capacity glut. The rapid adoption of token-hungry agents has shown that the AI industry hasn't come close to meeting demand and likely needs to accelerate the build out. Nate Silver wrote, It feels like the bubble argument might fully shift to AI is a bubble because compute is too expensive to meet staggeringly high user demand for compute intensive advanced models. Now staying in these waters for a moment, Google has unveiled a new generation of their TPUs for training and inference. The eighth generation TPUs will feature two different chips for the first time. One will be optimized for model training while the other will deliver more efficient inference. Until recently, AI data centers have always been an integral purpose. AI demand didn't support inference-only clusters, so chips were designed to support both training and inference. Over the past year in particular, demand for inference has massively outstripped demand for training to the point that dedicated inference clusters make a lot of sense. Thomas Kurian reinforced this point saying, People increasingly are specializing how they are deploying AI infrastructure, whether it's for training or inference. Now it's not entirely clear that the two chips will have different chipsets or if the components are the only difference. Google said that the inference chip will be designed to maximize memory bandwidth, reducing latency for agentic tasks, while the training chip will feature larger compute throughput and more scaled up bandwidth. More indications that the industry is headed towards the separation of training from inference include last month when NVIDIA unveiled their first collaboration with their recently acquired chip startup Grok, their forthcoming Rubin generation will feature a server configuration with additional Grok chips that is designed to optimize inference, and we've also seen OpenAI recently sign a deal with Cerebris, whose chips are only useful for delivering fast inference. While it's a big move for Google, semiconductor analyst Patrick Moorhead thinks the chips will only have a modest impact on the industry. He writes, This is not Google taking on NVIDIA. It's Google looking for optionality and using TPU primarily for its own services. One more note on Google, although we'll have more on their announcements from Next in the main, Sundar Pichai has corrected the record, stating that three-quarters of Google's code is now generated by AI. There's been a significant amount of discussion over recent weeks about Google falling behind in coding, leading to reports we covered earlier this week that co-founder Sergey Brin is back leading a strike team to rectify the situation. As part of that reporting, the information resurfaced a statistic from Google's February earnings call that half of their code was now written by coding agents. Last year, that would have seemed like a lot, but the information contrasted that against comments from cloud code creator Boris Cherny, who has said numerous times that pretty much 100 percent of Anthropics code is now written by agents. Apparently, Pichai took that personally because in a blog post discussing the Cloud Next conference, he wrote, We've been using AI to generate code internally at Google for a while. Today, 75 percent of all new code at Google is now AI generated and approved by engineers up from 50 percent last fall. Lastly today, two stories in the Openweight AI world. OpenAI has released a new Openweights model tuned for a very particular use case. The new model is called OpenAI Privacy Filter and is designed specifically for detecting and redacting personally identifiable information in text. The concept is that users can run the model locally, drop in their sensitive text, and receive a fully redacted output without any data leaving their system. Privacy Filter is an ultra-small model with just 1.5 billion parameters and 50 million active parameters per query. Still, it delivers state-of-the-art performance on a privacy filtering benchmark, achieving a 97% score. OpenAI didn't discuss which model it was built on top of. However, they did mention that the model is an alternate version of a model they built for internal use. They wrote, This release is part of our broader effort to support a more resilient software ecosystem by providing developers practical infrastructure for building with AI safely, including tools and models that make strong privacy and security protections easier to implement from the start. Now I'm certainly going to be watching to see if this idea of releasing small open-source micro models tuned for particular tasks is something the AI labs pursue more broadly or whether this is just a one-off. Lastly today, speaking of AI labs focused on open models, according to reports, Mistral might be joining Elon Musk's AI super team. Earlier in the week, SpaceX announced they had signed a huge deal with Cursor, which might include an acquisition later this year. Business Insider now reports that Mistral could join in, forming a three-way partnership on whatever XAI is cooking up. Sources said that Musk had floated the idea of a collaboration as a way to take on rivals like Anthropic. XAI has already snapped up a little bit of talent from Mistral in a roundabout way. Last month, they hired a founding team member from Mistral who had since left to join Thinking Machines Lab. As always, the Elon Rorschach test is alive and well, with people who like Elon seeing this as a fearsome combination that would put XAI in a central role and others being sceptical of whether this is enough to really get back in the game. Nothing is confirmed at the moment, so for now, it's really all just about speculating what the team up might be called. Sathow writes Groxtral, LaGroque. That, however, is going to do it for the headlines. Next up, the main episode. Welcome back to the AI Daily Brief. Today, we are talking about the growing phenomenon of headless agents, really headless software in general. What people mean when they use this term headless agents is software that does things without having a user interface. Now, for a while, people have been talking about the idea that increasingly AI agents would be doing things on our behalf and that whatever combination of tools and searching and internet access that they need to get the job done that they're trying to do will inevitably be quite different from the type of interfaces that people would need to do the same things. That creates a whole bunch of challenges and also a whole bunch of new opportunities. Chief among the challenges is the fact that many different types of software providers are now going to have to support entirely different categories of quote-unquote users as what their human users and their agentic users need will be fundamentally different. Over the last week or so, we've gotten a set of announcements that all relate to this new Headless paradigm and also give us a good sense of where enterprise agents are right now. The first news that we're covering in this trend came about a week ago as Mark Benioff and Salesforce announced their new Headless 360 product. Not burying the lead at all with that name, right? Benioff tweets, Welcome Salesforce Headless 360, no browser required. Our API is the UI. Entire Salesforce and Agent Force and Slack platforms are now exposed as APIs, MCP and CLI. All AI agents can access data, workflows and tasks directly in Slack, voice or anywhere else with Salesforce Headless 360. Faster builds, agentic everything. So again, the whole concept here is the decoupling of agentic work from human interfaces. Right, Salesforce? For 25 years, using Salesforce meant working inside Salesforce. A customer service rep opened a console, clicked into a case and manually updated its status. A human navigating a platform to get work done. But in the agentic enterprise, humans aren't the only ones doing the navigating. Agents are too, and they don't go to a browser or click through UIs. They call APIs, invoke MCP tools and run CLI commands directly. Though, two and a half years ago, we made a decision, rebuild Salesforce for agents. Instead of burying capabilities behind a UI, expose them so the entire platform will be programmable and accessible from anywhere. If your platform requires humans to click through UIs or write code directly to make progress, it is not ready for the agentic enterprise. The goal, they say, is that, quote, work that used to pull people out of the conversation now happens inside it. Not just text back and forth, but approvals, decisions, rich data, full workflows all surfaced inside the channels where your people already are. That's what it means for the conversation to be the interface. Now, of course, taking a step back, it makes sense why Salesforce, with their Slack platform, is particularly keen on making the conversational interface of work more robust, given how much of that conversation happens on their Slack product. Certainly, it sounds like they're already seeing just how integral to this agentic evolution Slack is, writing that custom agents on Slack have grown 300 percent since January. Now, to give context for just how complete a change they're thinking about, they start their announcement article with a quote from Salesforce co-founder Parker Harris. Why should you ever log in to Salesforce again? With apologies to my Salesforce and Agent Force listeners, Salesforce announcements don't always get a ton of attention. That's not just Salesforce, that's basically any company that's not named Anthropic or OpenAI at this point. This one, however, got a ton of chatter, with people immediately grokking the significance of the trend outside of Salesforce itself. Writes Jacob Hampson, This is the future, every app headless, with the head being your chat or message or agent interface of choice. Now, the conversation about headless agents and headless software was starting even before that, but you could really start to feel it pick up and move away from just developers into more mainstream knowledge work circles after this announcement. Speaking of announcements, we got another enterprise agent related announcement yesterday, this time from OpenAI. Not content to sit back and let this just be the week of GPT images too, OpenAI announced what they're calling Workspace Agents. The idea is that organizations can set permissions and controls, and then individuals and teams in those organizations can spin up their own agents. These agents are meant to do a lot of the things that make up the day-to-day of knowledge work. Responding to messages, preparing reports, things like that. They are cloud-based so they can work even when you're not on ChatGPT, and they are designed for team use from the beginning so that once someone builds them, it can be shared with an organization and used in shared spaces like ChatGPT Enterprise or in Slack. Now, in some ways, they are an obvious evolution of custom GPTs. In fact, in many ways, these workspace agents are kind of what I think people imagined custom GPTs would be. It was obviously just a little bit too early for this full breadth of use cases to really be useful with that earlier version. Simon Smith writes, These are like GPTs on steroids, like mini open clause. You give them skills, tools, and files so they can do a lot more work. You can add them to channels like Slack so you can communicate with them there. You can run them on schedule so you can automate work. And big bonus, you can add memory so that they can remember past actions, preferences, and other information as you work with them. I think Simon's framing of them as sitting between custom GPTs and the complexity of open clause style agents is a pretty good shorthand. OpenAI gave five examples of workspace agents that Teams internally had built. A software reviewer to review employee software requests, including checking them for things like adherence to approved tools and policies. A product feedback router agent that could monitor Slack or other support channels and turn feedback into prioritized tickets. Weekly metrics reporter agents that can pull data on a regular schedule and turn that into shareable charts and summaries. And they also talked about outreach agents, which is something that I feel like at this point, every company that's trying to be at least even a little bit agentic has thought about. IE agents that do things like research inbound leads, update your CRM, et cetera. And the last example they gave was a third party risk manager that could research vendors assessing a bunch of signals and ultimately produce a report with recommendations. Workspace agents also has templates in functional areas like finance sales and marketing. Now when Simon was describing all these features they have access to, because they are powered by Codex in the cloud, which means they have access to the same sort of files, code, tools and memory that Codex does. This is also why they are a difference in kind, not just a difference in scale from custom GPTs. As OpenAI puts it, agents do more than answer a prompt, they can write or run code, use connected apps, remember what they've learned and continue work across multiple steps. Now Aaron Levy from Box connected this directly to the idea of headless software. After the announcement he wrote, This is probably the biggest news yet in software going headless and will bring knowledge work agents to the masses. The new ChatGPT agents have access to any of the tools and data you want to work with and complete coding and tool use available to them. This is precisely what agents will start to look like for knowledge work. You'll be able to spin them up in the foreground or background to help augment work. Big opportunity right now for headless platforms and for all the new builders and designers of these agents in the enterprise. Still, we are not even halfway through the enterprise agent announcements that we've gotten this week. Microsoft's latest agentic offering is called Hosted Agents. And this is basically their version of Anthropix recently announced Claude Managed Agents. CEO Satya Nadella writes, Every agent will need its own computer. And with new hosted agents in Foundry, every agent gets its own dedicated enterprise grade sandbox with durable state, built-in identity and governance, and support for any harness or framework. Basically, the idea is that Microsoft is now allowing Foundry customers to run agents on their, Microsoft's infrastructure. The agent gets its own configurable sandbox with a persistent file system, essentially creating a safe and observable environment for an agent to operate within. Users can bring their own harnesses, orchestration systems and execution environments to ensure the agent can operate exactly how it needs to. Now, as a side point, you can see Microsoft leaning to their new multi-vendor platform. In their announcement post even, they wrote, This flexibility is the point. Unlike platforms that force one model and one harness, Foundry is multi-model and multi-harness by design. Run models from OpenAI, Anthropic, Metamistral and more. Bring any orchestration framework. No lock-in. Remember, this is Microsoft saying that. A time traveler from the 90s would be gobsmacked. And then there was Google. This week was Google's Cloud Next conference, and agents were 100% the star of the show. CEO Sundar Pichai said that agents, and particularly enterprise agents, will be the key to AI monetization efforts. The central release is a unified platform called Gemini Enterprise, or in full, the Gemini's Enterprise Agent Platform. The platform seems to be functionally a rebrand and relaunch of Vertex AI, which is Google's agent design and orchestration tool. Google has added a new set of governance and security features to allow enterprises greater control over their agent deployments, and Google Cloud CEO Thomas Kurian described that this platform update was a response to shift in customer behavior. He said, There's definitely a strategic shift as the models become much more sophisticated. The primary use case of Vertex AI recently shifted from old style machine learning to a sudden explosion in users building their own custom AI agents. Now, alongside the new look Gemini Enterprise, Google has filled out a bunch of necessary features to make their agentic offering complete. They now have Data Agent Kit, which provides a set of agent skills and plugins to make building agents a more streamlined process, Knowledge Catalog, which provides a unified context engine for the entire enterprise. In Google Workspace, the company has added Gemini for Google Slides. The agent can now be steered to follow your brand or corporate style as well as working from your templates. And there were literally dozens of other features or products announced, with the clear overall goal being to deliver everything one could possibly need to deploy and operate enterprise-grade agents at scale. Silicon Angle argued, The real story isn't AI, it's the control plane. Google, they writes, is not launching AI features. Google is trying to build the operating system for the agentic enterprise. The thesis is basically that if model performance is converging and becoming a commodity, and if inference is getting cheaper, the control layer or agentic operating system becomes the new land grab. Silicon Angle continues, The shift now underway across the enterprise stack moves in a clear arc, from systems of record to systems of engagement to systems of execution. Salesforce, Workday and ServiceNow built their empires on the first. The web and mobile era built fortunes on the second. The third, systems that actually do the work, is the prize Google is reaching for at Next. So let's take a step back and try to contextualize all of these announcements. First let's turn back to Erin Levy again who writes, Software going headless is inevitable in a world where agents use the tools 100x more than people do. And the reality is for a lot of software, this is actually a huge boon to potential use cases for these platforms. Software business models have largely been predicated on selling to the number of seats that are in the company in a given function. And the usage of your software is constrained by how much people can do in a given day. This means that your technology is often vastly underutilized relative to what it can actually power for the customer. Enter agents. Agents can work 24x7, run in parallel and string together work across systems. This is a big deal because now the agent can do far more than people ever could with these tools. Instead of reviewing contracts one by one, the agent will review all of them. Instead of manually moving data between marketing systems and across campaigns, the agent will let you run 10x more of them. Instead of being rate-limited in the client onboarding process by human steps, agents accelerate these. Here's the key point that Aaron makes, that people are just starting to discuss now. Agents, he writes, end up using these underlying platforms far more than people ever did, which opens up use cases that the platform couldn't go after before. Now, not every software market has the same amount of positive sum use cases between people and agents. But I'd argue that a significant portion of systems of record, for instance, can be used far more than they are today. Your Salesforce data can be leveraged 100x more to do vastly more customer targeting and sales automation. Your documents can be turned into structured data and analyzed for insights and knowledge to automate other workflows, and so on. Now, of course, you have to find a way to make this all commercially attractive, but it's not hard to picture the revenue from API and agent consumption on these platforms becoming a rich component of revenue streams over time. Seats for the people, consumption for the agents. Lots of upside here for the companies that embrace this trend. Now, many others jumped on the business model implications here. Vibe Marketer JB writes, Salesforce going headless is bigger than people realize. Software has been priced per seat for decades. The entire business model assumes a person logs in, clicks around, and gets value from a dashboard. Agents don't log in. They make API calls. So what happens to per seat pricing when the primary user of your platform isn't a person? When one company runs 50 agents that each make more API calls in a day than the entire sales team makes in a month. Every SaaS company is about to face this question. Salesforce just forced it into the open by going fully headless. The ones that figure out agent native pricing first will own the next cycle. The ones still charging per seat while agents do the work will get left behind. Now, interestingly, this idea that the per seat model was on its last legs has been around for the last few years. Even before ChatGPT, there was a lot of consternation around this between these SaaS providers and their customers. In the wake of ChatGPT and the beginning of more advanced generative AI, you saw many companies start to try to at least experiment with outcome based pricing or just non-seat based models, but their users were still primarily human and nothing really quite stuck. What's different about this time is that you actually have a different category of user, where I think different types of natural business models, like these consumption based models or whatever we figure out works, actually fall naturally out of the new user behavior pattern. One of the big speculator debates people are having is who captures this new type of value. Is it horizontal providers like OpenAI and Anthropic, the old systems of record like Salesforce, vertical agents that deal with specific functions and use cases, and there are debates on all sides of this. Simon Smith writes, Yes, value is shifting from UI to agents and the data and infrastructure to support them, but it's not clear to me that vertically specific Salesforce is best positioned to capture this shifting value. The AI labs are building vertical agnostic agent infrastructure, and there are many capable data and infrastructure providers in the market already. Akash Gupta provides the counterargument. He says, The math for mid-tier SaaS is brutal. Once the agent layer works across tools, enterprise buyers stop paying $30 a seat for AI and five different products when one ChatGPT seat runs the same workflow. The real winner, he continues, is whoever owns the data access layer. Salesforce, ServiceNow, and Workday sit on the schemas every agent has to read from. They just became the toll road. And yet others are seeing this shift to Headless as an opportunity for more, not less, usage of the existing platforms. Matthew Kobach writes, SaaS has taken a beating in the public markets lately as a response to AI. There's an interesting counterposition that Headless SaaS becomes more valuable with AI, not less. Software has a learning curve and a human bottleneck. But if I could have a dozen capable agents working inside any given software, that software becomes more valuable, not less. So one argument for why the SaaS apocalypse won't play out exactly as investors have thought is that reducing the friction in agents using tools means that agents will use more tools. Seat price might go down, but consumption will go up, and there's actually room to make even more money. The other thing that some are thinking about, though, is that some of the SaaS tools that are used for human coordination today might be naturally positioned for agent coordination tomorrow. Ivan Burazin, for example, writes, Atlassian is still undervalued. Everyone thinks Jira and Confluence will get vibe-coded away. Instead, they'll build headless versions, API first, same product, no UI, for agents. Agents need standardized project management just like humans do. Stock is underperforming because people misunderstand what's happening. Ivan has actually been thinking a lot about this shift to headless software. In another tweet from back at the end of March, he wrote, Agents need headless tools to work efficiently. Basically, APIs and programmatic access built for autonomous systems. Headless IDEs, headless terminals, headless everything. And then just this week, he wrote, Most infracompanies are still not solving for what agents will need. They're not considering the agent as the primary consumer. They are creating agents that make human lives easier by automating more of our use cases. In the next two years, agents will completely take over and make comparisons, purchases and other real-life decisions with consequences. The alpha will shift to optimizing the agent's life better after you give it the prompt. HubSpot founder Darmesh Shah agrees. He wrote, Every B2B software company is or should be building a headless version of their product, one that can be used by agents. But headless doesn't mean brainless. You don't just wrap your existing APIs into an MCP server and call it a day. The companies that succeed in the agentic era are those that take a thoughtful approach to designing an agentic user experience. Yes, that will likely involve APIs, MCPs and CLIs. But the difference will be in the ergonomics of the interface. We need to figure out how agents actually want to use our products and platforms. Because if all they wanted to do was use them like humans do, we have computer use for that. Startup Ideas podcast host Greg Eisenberg sees nothing but opportunity here. He argues that there's a trillion dollars up for grabs for agent-first startups, as every SaaS company follows Salesforce and goes headless within 18 months. He argues that a new category of agent-native startups will emerge that treat the Salesforce's and HubSpot's and Workday's, as in his words, dumb backends. The startup is the agent, he says, the SaaS is just the database. And yet I will point out, for those of you who are sitting there feeling like, what role do humans have in this? I would point you to another news report from earlier this week around OpenAI working with firms including Accenture, Capgemini and PwC to help sell codecs to businesses. Aaron Levy again writes, The real world will need a ton of help actually getting agents going in the enterprise. Companies have legacy tech stacks they need to modernize, data and tons of fragmented tools, knowledge that isn't captured or digitized, and change management needed to actually utilize agents effectively. And they all have to do this while actually running their business day to day, unlike startups. This is why there is so much opportunity for companies, software or services, to actually deploy agents in specific domains and workflows. This remains a big opportunity for both existing service providers, but also tons of new startups as well. Every new technology wave produces a new era of consulting firms that can deliver on that technology. That's also why the FDE model is going to be alive and well for a long time, because companies will want to have their vendor actually help drive the change management and implementation for their new workflows. The people aren't going away, far from it. So friends, I think headless is probably a word you will hear a lot more on this show as these trends evolve. For now though, that is going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace!