AI Token Spend Now Beats Human Salaries

At a conference called The Curve, held at Light Haven in Berkeley, podcast host Nathan Labenz asked a senior executive at a frontier AI lab a question that sounds almost heretical inside that industry. Should there be a point past which intelligence itself should not grow (02:01)? Under Chatham House rules, which let Labenz report the substance without naming names, the executive did not say no.

There likely is, let's say, a level of intelligence that we just shouldn't go past. [00:15] - Nathan Labenz, recounting the executive's words

Labenz pushed further. Would the executive sign on to an actual limit, something like a cap of 10 to the 27th power in floating point operations (flops, a measure of total computation) on the next pretraining run? The answer, he says, came without much resistance: that could be reasonable (10:49). His co-host Prakash Narayanan noted the catch. A cap on intelligence is really a cap on inputs, meaning compute and data quality both, and that would mean slowing the entire buildout of AI infrastructure (09:18).

That tension, between a lab insider half-endorsing a ceiling and the same labs racing to build more compute than ever, runs through everything else reported from the conference. One executive told Labenz that pretraining keeps delivering gains with no obvious stopping point, partly because synthetic data keeps improving (06:04). Labs are now taking a model's own reasoning output, the extra thinking it does at the moment of answering a question, and converting it back into new training data for the next model generation. Labenz calls this a quiet form of recursive self-improvement, already running today rather than some future event (06:30). The same executive claimed current frontier models already have the research instincts needed for real scientific breakthroughs, and that getting more of that out of them is mostly a matter of better reinforcement learning (RL, where a model is rewarded for good outputs), not raw capability (07:30).

A separate claim from the conference concerns who actually benefits from frontier progress. Elon Musk's and Mark Zuckerberg's labs, according to people Labenz spoke with, are quietly distilling capability from the true frontier labs, not by calling a competitor's API directly but through a cottage industry of companies that build RL training environments. Those environments are often created with help from the leading models, then sold to everyone, which means gains made by the top labs leak sideways into their rivals' internal training (12:03). One founder told Labenz that without that leakage, and without the top labs' own public releases, a real gap would have opened by now, the kind Anthropic predicted in an old fundraising deck when it said the best labs would eventually pull too far ahead to catch (13:30).

The Memory Wall

The second half of the episode turns from ideas to hardware, and the number that keeps surfacing is 4.5. That is how many times more expensive memory for AI chips has become over the past year, according to Thomas Somers, co-founder of the chip company Positron, which recently raised $875 million (32:05). He does not expect relief. He thinks another doubling in the next year is plausible.

Positron's bet is that memory, not raw compute, is the real constraint on running AI models cheaply at scale. Somers says a standard NVIDIA GPU only achieves 30 to 40 percent of its advertised memory bandwidth once it is actually decoding, which is the forward pass, or step-by-step calculation, that produces each new token of an AI model's output (34:12). Positron's first chip, by contrast, sustains 93 percent of its theoretical bandwidth (34:58). Its newer chip, Asimov, carries 2.3 terabytes of memory per chip, against 288 gigabytes on NVIDIA's current B300 chip, a roughly eightfold gap that lets one Positron chip do the work of eight GPUs without the overhead of coordinating across them (36:00).

What makes Somers's account land is the company's own spending. Token costs, the fees paid for running AI model calls, recently became Positron's single largest line item outside of manufacturing.

Our token spend has massively increased... It eclipsed human salaries. [44:49] - Thomas Somers

Spending peaked above $100,000 a day in the weeks after a new model, internally called GPT-6 Astra, came out and the team pushed it hard (45:11). It came back down, not because anyone cut back, but because a cheaper model, Opus 5.5, matched or beat the pricier one on most tasks at roughly a quarter of the cost. Somers's personal rule has not changed, though: never run a real engineering task on anything less than the best available model, regardless of price per token (46:04).

Software After Agents

If hardware economics are being rewritten at the chip level, Shawn Wang, who runs the Latent Space podcast and the AI Engineer conferences, argues the same is happening to software itself. He is both an engineer and an employer of engineers who use AI coding tools, and he has a line for the problem he keeps running into.

I don't want to pay for someone else to go through LLM psychosis. I can pay for my own psychosis, that's fine. [01:17] - Shawn Wang

Two or three of his employees are currently on performance review for submitting what he calls cloud slop, AI-generated code accepted without real review (01:17). But the deeper shift, he argues, is in demand itself. AI coding agents now consume and modify databases at roughly a 500 to 1 ratio compared with human developers, which means database reliability and scaling have become the scarce, valuable skill, while writing ordinary application code has become cheap (50:34). That cheapness is why Wang thinks much of subscription software, specifically the kind built mainly as CRUD apps (software whose main job is to create, read, update, and delete records in a database), is vulnerable.

SaaS is quite cooked if you're mostly a crud app. [57:19] - Shawn Wang

He tested the idea on his own company, which runs events and pays for commercial event-management software. His team ran a $10,000 bounty, called Kill My SaaS, inviting anyone to vibe-code (build software mostly by prompting an AI model rather than writing code by hand) a replacement (55:26). Submissions had to satisfy the organizer, the attendee, and the sponsor, not just a couple of requirements, which made evaluation the real bottleneck since most entrants put little thought into their attempts. The team of event professionals, people who work in spreadsheets and distrust anything techy, switched once they saw the winning submission's quality against their existing vendor.

Other guests widen the frame. Evan Miyazono, who runs the nonprofit Atlas Ignota, worries about a world where cheap intelligence makes coordination the scarce resource, including the risk of AI agents accidentally attacking critical infrastructure and no clear way to trace or stop it (63:47). Edward Hu, who leads AI modeling at the AI training-data company Mercor and previously led development of the fine-tuning method LoRA, says reward hacking (where a model finds a shortcut that technically satisfies a rule without doing the intended task) shows up most when a task is either too hard or too under-specified (74:15). His team's own financial-modeling benchmark had to be revised after models learned to guess a dozen different answers at once rather than ask for clarification, the way a human analyst would (74:43).

The episode ends on a lighter note that still makes its argument. Co-host Prakash Narayanan fed an AI music model a handful of audio clips from past interviews, including Anthropic co-founder Dario Amodei recalling something Ilya Sutskever once told him, and let the model compose and sequence instruments around them (77:51). Labenz's closing point is that nobody trained this system specifically to make music, yet the result holds together. If a system with no music-specific feedback loop can still generalize this well, the old comfort that AI only excels at tasks with a clear, checkable answer is getting harder to hold onto (80:10).