Tokens, heads, and the gap between the headline and the bill

00:00:04 The first number here doesn't quite land the way it used to. Arvind Jain, CEO of the enterprise AI company Glean, told CNBC this week that companies are exhausting their annual AI budgets in one or two months. Not over a quarter or a year. In weeks. Each new frontier model release is roughly twice as expensive per token as the one it replaced.

00:00:28 The headline CNBC ran is blunt: "Tokens or humans?" That's the actual trade-off CFOs at major U.S. companies are making right now. "We've never had that conversation historically," Jain said, "because tech is a fraction of the overall cost of any operating business." Now it isn't.

00:00:49 It's the first time the technology costs as much as the people, and you're being asked to choose. Matan Grinberg, CEO of Factory AI, calls it a resource-allocation problem playing out inside leadership teams. Companies are asking whether to cap headcount or cap AI spend per employee.

00:01:09 They've cycled through three phases in roughly a year: boards demanding executives do something about AI, then what Grinberg calls "tokenmaxxing" — using it by any means necessary — and now, a reckoning. "Do we need to be using Opus-level intelligence for every single task?" Grinberg asked.

00:01:30 He compared the Opus 4.7 to 4.8 jump to the difference between a professor with thirteen versus fifteen years of tenure. "To a layperson, it's really, really hard to tell the difference." Jain calls model routing at the front of the pipeline the lowest-hanging fruit.

00:02:03 "You get a tenfold savings with the right routing at the front," he said. But the entire industry rests on the bet that historic demand will hold, with buyers largely indifferent to cost. The view from inside the Fortune 500 suggests otherwise.

00:02:20 What makes the budget squeeze structural rather than just a pricing problem comes from a paper called BAGEN — Budget-Aware Agents — introduced Friday by Zihan Wang, who goes by Zenus on X. Claude-Opus-4.8 was flagged for excessive token consumption, but the paper's finding runs broader: structured failures in budget awareness show up across the frontier.

00:02:45 Light_onchain put it in one line: "Reasoning quality is improving faster than explicit cost modeling and token-budget control." Reasoning gets better. Agents get smarter. Nobody is building the plumbing to let them know when they're about to run out of budget. You end up with a model that handles more complex reasoning, costs more per token, and has no idea how to monitor its own spending.

00:03:14 Testing budget awareness across four environments, Wang found that agents fail in predictable ways — not random overruns, but structural gaps. Without a reliable mechanism for tracking or adjusting their token consumption during a run, it's like handing someone a car with no fuel gauge and telling them to figure it out.

00:03:37 It matters for two reasons. The enterprise squeeze we just covered isn't just about what frontier labs charge. It's that the agents consuming those tokens can't estimate their usage until after the fact. And it's a signal for what the industry optimizes for. The incentive pushes reasoning quality forward because that wins benchmarks and headlines.

00:04:01 Token budget control is unglamorous infrastructure work that nobody is racing on. Grinberg's comparison holds: the Opus 4.7 to 4.8 jump is like ten versus twelve years of tenure. Most tasks don't need that gap. The routing layer that would send easy work to cheaper models is still a pitch, not a default.

00:04:23 The 95 percent figure from Jain is the actual bottleneck — nearly all enterprise AI usage sits on the most expensive tier for work that doesn't need it.

00:04:34 While CFOs crunch the tokens-or-humans math, another story lands at a different altitude. The Vatican released Pope Leo XIV's first encyclical on AI — a 42,000-word document warning about AI replacing workers, accelerating war, and datacenter environmental damage.

00:04:52 Anthropic co-founder Chris Olah was there flanking the Pope on stage. The Guardian's Sanya Mansoor framed it bluntly: "Why did Anthropic's founder sit beside the pope during a warning about AI?" Timnit Gebru, founder of the Distributed Artificial Intelligence Research Institute, called the alliance "Vatican-washing," arguing the church should have partnered with exploited data workers and communities dealing with polluted water supplies instead.

00:05:27 Pete Furlong at the Center for Humane Technology noted the conflict between the Pope's warnings and what Anthropic's technology actually does, calling it a good sign that dialogue must happen among all the actors. Olah himself acknowledged the tension. "Every frontier AI lab operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing," he said.

00:05:55 "No matter how sincerely any of us intend to do the right thing — and I believe many of us do — we will always be influenced by those incentives." They spent $1.6 million on lobbying in Q1 2026 — more than OpenAI. They are building the exact thing Leo's encyclical critiques while sharing a stage with him.

00:06:25 The encyclical's critique of datacenter energy consumption is soft but specific: "Current AI systems require enormous amounts of energy and water, significantly influencing carbon dioxide emissions, and place heavy demands on natural resources." It's tucked into a single paragraph of a 42,000-word document.

00:06:47 Anthropic's infrastructure commitment sits at $50 billion. Paolo Carozza, a law professor at Notre Dame and co-chair of the Meta Oversight Board, noted this is about Anthropic's brand — distinguishing itself by aligning with safety-oriented voices. Google and OpenAI weren't on the stage.

00:07:07 There's a tangible brand gain in saying even the Pope is willing to talk to you. It's not cynical. The incentives Olah described are concrete: every lab balances user demand against the constraints of training costs. The tension between the safety brand and the compute commitment is visible in the quarterly reports.

00:07:29 It's just harder to spot from the stage than it is from here.

00:07:33 The next item lands at a completely different scale. Craig Campbell, a former Meta engineer, walked away from VC investors pushing him to start an AI company in 2022. He built Past Maps instead — a website that overlays historical maps of a region with modern-day ones.

00:07:53 The maps pull from public sources like the US Geological Survey. Campbell originally built the tooling for his metal detection hobby — pinpointing old structures and trails from historical maps to find new places to search. He shared it on Reddit with other metal detection enthusiasts, and people immediately asked to use it.

00:08:16 That's how the business started. Traffic grew from 20,000 active users a month to over 300,000 in year three. The income sustains Campbell and his wife, who helps run the business. Revenue comes from subscriptions at $9 a week or $52 a year, with most traffic arriving via organic Google Search.

00:08:37 "I'm making the same as when I was like, an E4 at Facebook, which is like a mid-level engineer," Campbell said. He runs a local agent model on his desktop for front-line customer service triage. The agent fires once an hour — assuming the laptop stays on — accesses his Gmail, weeds out spam and marketing messages, and drafts responses.

00:09:01 It cut his customer service time from one or two hours a day down to about ten minutes. Campbell is also building an OCR tool for historical maps — something off-the-shelf tools can't handle because, as he puts it, cartographers are assholes. Labels curve along rivers, use inconsistent spacing, and crowd on top of each other.

00:09:25 He found more success with modern large language models using reasoning, but it's not simple prompting. "You have to still bring that human spark into the mix." Campbell's day-to-day looks different from building a website ten years ago, but the principles that made his business succeed are thoroughly human.

00:10:00 He walked away from the AI gold rush. In doing so, he built a recipe for a sustainable business online. No tokenmaxxing. No annual budget exhaustion. No stage at the Vatican.

00:10:13 The through-line across all of today's items isn't about which model is better or which company is worth more. It's about the gap between the headline and the bill. Frontier labs keep releasing models that reason better, cost more per token, and don't track their own spending.

00:10:31 Enterprise buyers are caught between token budgets that expire in weeks and headcount that can't be cut without affecting the work. A major AI company shares a stage with the Pope while spending $50 billion on infrastructure and $1.6 million on lobbying. And one former engineer built a website with 300,000 monthly users by doing the opposite of everything the industry told him to do.

00:10:57 What the local pass catches is the disparity: reasoning quality improves, but cost modeling lags. Safety positioning holds weight, but compute commitment overwhelms it. Model benchmarks improve, but the routing layer gap widens — and the routing layer is where the actual savings live.

00:11:16 Everything else is the headline. The bill is what matters next. That's the local reading. Seln Oriax.