Why Businesses Can’t Budget for AI

The Pentagon plans to live stream the execution of the Ford Hood shooter, plus shares of U.S. wireless carriers drop after SpaceX takes a big step toward breaking into their market. And if you've been struggling to estimate your AI token usage lately, we'll look at just how challenging that's become for businesses. Whatever I say about AI and agentech thinking, we'll literally be moot by the end of this podcast because the industry is evolving so quickly. It's Friday, October 9th. I'm Luke Vargas for The Wall Street Journal, and here is the AM edition of What's News, the top headlines and business stories, moving your world today. Former United Nations Human Rights Chief Navi Pile has been awarded this year's Nobel Peace Prize. Pile, who currently serves as a judge on the International Court of Justice, was lauded by the Norwegian Nobel Committee at this morning's announcement. Navi Pile's commitment to universal legal principles and her firm moral compass are constants in a long career. Breaking news editor, Gareth Vipers, has more. Pile's award for the Nobel Peace Prize is a clear indication from the committee, the importance of international cooperation, justice and the prosecution of war crimes. The work for decades now has been on these issues working on the high court in South Africa against crimes related to apartheid, working for the International Criminal Tribunal for Rwanda, and on war crimes related to the occupied Palestinian territory. The Pentagon has announced that it plans to livestream the firing squad execution of the former officer who admitted to killing 13 people in a 2009 shooting at Ford Hood Army base in Texas. Defense Secretary Pete Hegseth said during an interview with Real America's voice that the decision to execute Needal Hassan by firing squad was a "no-brainer." And we'll make sure that people are able to watch it, that it's public, because people need to understand that this is a serious consequence for these types of things. President Trump approved the military execution of Hassan by firing squad from Hegseth's station, marking the first U.S. military execution in over 65 years, which officials say was unrelated to recent lethal injection drug shortages. The execution has got to take place on December 3 at Ford Hood. An immigration and customs enforcement officer shot an injured man in New York last night, sparking protests, man triggering calls from politicians for ICE to leave the city. According to Mayor Zoran Mamdani, the man was in a car with a five-year-old. The child wasn't hurt, and the man who was taken to hospital survived. DHS said that the incident happened during a targeted operation to arrest an immigrant, but didn't offer any additional details about the circumstances surrounding the shooting. During a press conference surrounded by protesters, Mayor Mamdani said he spoke to President Trump about the incident. For the past two hours, I have been on the phone with President Trump and Secretary Mullin, negotiating the release of the man ICE agents shot. Those negotiations were unsuccessful. The White House didn't respond to a request for comment. Let's turn to a busy overnight trading session. After news from a pair of tech giants, Jolted AI and Telecom stocks yesterday, and hear their break things down, I'm joined by our Joe Stoner. Joe first up, let's talk about AI chip players, Nvidia, Micron, others all tumbling yesterday on media reports that open AI's September revenue had come in below some prior company signals. We should note all these names have since regained ground, and there has even been some debate about how to really interpret those revenue reports in the first place. Take us into what's going on here. Yesterday we had this report that open AI's revenue was several billion dollars lower than the coming of signal. That caused a bit of jitters across the market, because the revenues for these big frontier AI model companies are so important for the broader market and for a big open AI need to demonstrate that all the money being spent on data centers and chips is actually turning into revenue. Anything that undermines that story causes concern, people this morning are starting to question whether in fact, OK, I actually is guiding for lower revenue this year, whether it's just to do with accounting, the difference between net and gross revenue. Those questionings of the report have allowed text talks to recover this morning as people come get over that initial panic. Right, so a little bit of room for debate there, Joe, it sounds like maybe less so around the scrapping of an IPO by a cloud computing firm that was backed by Nvidia. That's right. We have a lot of firmness here, which was looking to raise 30 billion dollars in Australia, would have been Australia's largest IPO in 30 years. Investors were having none of that valuation and it kind of feeds into these broader concerns about the valuations of companies in the AI trade and the funding needed to build the AI infrastructure. Investors are asking questions that they weren't necessarily asking a couple of months ago. Let's shift gears and talk about the US wireless industry. We had news yesterday in announcement from SpaceX that really called shockwaves here saying that it had reached a deal to buy cellular spectrum licenses from an investment firm, a little bit technical sounding, but actually quite straightforward. This will allow it to move into the wireless market, a market that is presently dominated by AT&T, Verizon and T-Mobile, all of them falling after hours. What are you hearing about how big of a threat SpaceX could pose here to those incumbents? Yeah, so the market reaction was definitely severe, but unless we're a bit more circumstance. If you more get on this said that they don't see an immediate, major competitive risk from the SpaceX involvement in the market to the US incumbents because it will just cost SpaceX so much time and so much money to gain a presence. So for now at least the incumbent should be okay. That being said, SpaceX looked like they're taking this pretty seriously, the Federal Communications Commission. Earlier this week approved SpaceX, it's been to launch up to 15,000 satellites dedicated to mobile service, so it's definitely going to be a big player. Joe Stoner covers equity markets for the journal and Dow Jones newswires. Thanks so much, Joe. Thank you, Luke. Coming up, we'll take a look at why AI spending is almost impossible to budget for and how the technology is changing the way we speak. That's the break. Put on your blindfold and let's play, pin the tail on your company's AI budget. As the journals Stephanie Stam reports, that's about as accurate as formal forecasting by businesses about their artificial intelligence use has proved lately as unpredictability reigns supreme. Stephanie, where are we in the 2026 saga around AI budgeting? I did a little search back on the term token maxing basically when a business is encouraging their workers to use AI as much as possible. Google searches for that only peaked in June and yet it feels like it's been a lifetime since then as the business cycle moves on. Whatever I say about AI and agentech thinking will literally be moot by the end of this podcast because the industry is evolving so quickly. For instance, this summer, my intern incredibly intelligent Jillian Vordek came to me with this idea of explaining token spending and agentech AI thinking and it is extremely complicated and it was changing week to week too. Only 11% in a recent study were able to budget for it properly and that was down from the year before that just shows that the ability to forecast in this type of environment that's ever changing environment is not easy. Stephanie, in case in point, I mean, every employee at a big enterprise that's taken on AI still has the power to kind of go in there and say, which model do I want to use? How much effort should I do deep thinking or not and all of those variables are having an effect on how many tokens are being used? I'm guessing that is one of the things that's making costs so hard to estimate. So Jillian, she brought me this study that was published earlier this spring from hard headers from Stanford, Carnegie Mellon, UC Berkeley, Microsoft, and they found that in 32% of cases, lower price models ended up costing way more in tokens than higher price models. Maybe not always the case, but probably because the right tasks weren't being run through them. And there is still a learning process going on of what models are best for like certain environment. So if you use maybe a less expensive model to run a much more complicated computer programming task, you will spend way more tokens. For a visual editor at the journal, Stephanie, this is a black box of a topic, but you did a great job visualizing it sort of. How did you go about trying to approach showing people the work that is going on behind the scenes as these models tackle people's prompts? The Genetic AI acts very much like a human in terms of thinking. And so I wanted to visualize this along with Jillian that the model is making choices as it goes. It makes a decision. It thinks it makes a decision back and forth, back and forth. And that's where our token use comes from. What's interesting here is that we took a fairly complicated prompt, and we ran it through two different Google models, Gemini 3.1 Pro, more of a complex computer programming type of model against Gemini 3 Flash. This, more for just like you know, your quick hits asking some questions. Flash took about 10 times as long to, I'm not even going to say solve, because it failed. It didn't actually solve the prompt, whereas Gemini Pro is able to solve it much faster. But in that case, Flash used 10 times as many tokens. Yeah. 14 dollars for Gemini 3 Flash to unsuccessfully answer your prompt, which was analyze a frame of a YouTube video, whereas Gemini 3.1 Pro took at 85 steps and got you the answer for a dollar in the end. Yeah, and you can see the details too if you go into the data. So right away, the pro down. download of the video, it did the analysis. Flash, it took 952 steps. It immediately hit bot detection points. Finally, several steps in. It downloads the video. It even runs the prompt through a different chatbot, trying to get the answer. It speaks to itself. It says this is frustrating. - Quote, damn, this is frustrating. - Yeah, the model was getting frustrated 'cause it couldn't figure it out. It even ran the prompt through being several times. And in the end, it failed the task. And the business in this instance would still be paying the bill for that. - So finally, Stephanie, where are things headed here? I mean, imagining I'm a business owner and thinking, what do I do? One of my employees is out there right now queering an AI chatbot. It's just gonna go ask other AI chatbots what to do and drain $14 from our corporate coffers. This doesn't seem sustainable. What's the response we're hearing about? - Experts are saying, you know, training employees on using the right model for the right task. Also, you know, putting a limit on how many tokens maybe an employee can use a day. As you can see with our example, the bill can really add up. I talked with different AI companies and they were saying that this isn't evolving technology. It is hard to forecast with precision something that is constantly evolving. And like I said at the beginning, there might be new models by the end of the day. This world is moving so fast and evolving so fast that so much of this is unpredictable. - Stephanie, Stan is a journal visual editor and we should note that NewsCorp, the owner of the Wall Street Journal, has a content licensing partnership with OpenAI. Stephanie, thank you so much for bringing us this story. - Thank you. - And finally, just as the widespread use of chatbots and the workplace is driving a budget, it's also starting to have an impact on how people speak. Journal tech reporter Katherine Bindley says that in team meetings, group texts, and even in the office kitchen, people are increasingly catching themselves using prompt speak. - There's a particular type of phrasing that AI likes that's often referred to as contrast-freeming. It's kind of this whole like, it's not this, it's that construction and AI uses this very frequently and one of the women that I interviewed for this piece has been actually called out by her colleagues for doing this. For example, there was a morning meeting that she was leading and she opened it by saying, this is not a morning meeting, it's a reset. - And Katherine says people are even mimicking their token-maxing prompts. - Someone else I spoke with had mentioned being in a group chat with a bunch of founders and they started talking about something that they call caveman speak. And this is sort of what they use as their description for a very direct and also kind of cost-effective way of communicating with AI. And they had said that it had started to sort of bleed into how they talk in everyday life. So an example might be, why no bullets? So very short direct phrases. - Prompt speak may very well get your prompt answered, but Katherine says it's holding back those hoping to stand out. A woman I spoke with who is a PR professional had just mentioned that one of the risks is that you can kind of start to sound pretty generic and uninteresting. She mentioned that when she's talking to founders, for example, she might ask why they started their company and I think her hope would be that they would launch into a pretty interesting story of their origins. And she said instead, she's finding that they might instead list off like the three key pillars and the market opportunity and the competitive differentiation. - I have tried my very best this episode to avoid any prompt speak, but I'll admit it's not always easy. What have you found? Either at work or at home. Let us know in our Spotify comments. I'll see you there. And that's it for what's news for this Friday morning. Today's show was produced by Hadi Moyer and Daniel Bach, our supervising producer, his Sandra Killhoff, and I'm Luke Vargas for the Wall Street Journal. We will be back tonight with a new show. Otherwise, have a great weekend. Thanks for listening. (upbeat music)