Synthesis
Token spend: value versus volume
Judging AI spend by value or margin created per token rather than by volume, and treating heavy frontier-model spend as an experiment budget that CFOs will eventually scrutinise.
Across 18 episodes the hosts agree on a warning more than a method: do not mistake token volume for value. Sam Lessin says to measure margin and profit, not spend, and illustrates it with his own Cursor commits and web-service bills that were “not super valuable”. Dave Morin and Brit Morin say to measure productivity or efficient use per token, and Jess Lessin describes “token maxing”, ranking staff by tokens used. Reported anecdotes (an Uber CFO using a full annual token budget in 15 days; Uber’s CTO versus COO on ROI) show enterprises struggling to see returns, and Sam argues the lack of measurable ROI is a “weird moat for a while” until spend becomes core and CFOs take over. Personal figures are given as ranges (Sam: hundreds to a few thousand dollars a month; Brit’s 11-year-old about $1,500 in a month). Almost all numbers are self-reported or secondhand, and Dave’s Anthropic-versus-open-source cost framing sits beside his OpenClaw ties.
Key points
- Measure value, margin or productivity per token, not volume spent. — 4 episodes. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 21:58)
- Enterprise ROI on AI spend is unclear; reported cases include Uber’s budget and CIOs doubting efficiency gains. — 3 episodes. (one of the hosts · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 19:44) (Amir Efrati · Amir Efrati (May 2026) ▶ 19:30)
- When spend moves from experimental to core, CFOs take control and behaviour changes; until then heavy spend is an experiment budget. — 3 episodes. (Sam · Every CTO Is Quietly Switching (Jul 2026) ▶ 45:10)
- Individual hosts report running their own bills of hundreds to thousands of dollars a month, and cannot compare the counterfactual. — 4 episodes. (Sam · Google I/O Just Killed SEO? Ag (May 2026) ▶ 28:31)
- Audit recurring agent jobs and over-logging for hidden spend. — 2 episodes. (Brit · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 29:38)
- Compare agent cost with the cost of hiring an engineer, not with zero (Dave’s CRM app example). — 1 episode. (Dave · Agentic AI Will Break the Inte (Jan 2026) ▶ 08:56)
- Tokens are not comparable units, so measuring frontier versus open-source share is suspect (Sam). — 1 episode. (Sam · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 27:32)
Where guests agree
- The hosts treat raw token volume as a poor success metric. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 21:58)
Where guests disagree
Do you care enough about model cost to switch? ⚠️
- Sam: he does not care enough about cost to switch and will pay for slightly better accuracy, though enterprises will optimise. (Sam · Open Source AI Wins. Now Labs (Jul 2026) ▶ 08:00)
- Dave: he moved utilitarian agent uses to a cheap open-weight model to avoid frontier prices. (Dave · Open Source AI Wins. Now Labs (Jul 2026) ▶ 06:04)
Tactics
- Audit recurring agent jobs for hidden token spend. — Brit found some recurring jobs spent far more than others and cut them back in a batch. (Brit · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 29:38)
- Run many agents at once to drive token usage and throughput — Dave says token spend climbs when you run north of 10 agents at once; core engineers run 50 to 100 agents with continuous integration. (Dave · Google I/O Just Killed SEO? Ag (May 2026) ▶ 36:27)
- Move from enterprise AI plans to open models via a router — Scott Belsky says startups find enterprise plans about 10x more expensive, notice they use tokens for mundane work and switch to local open models behind a router. (Scott Belsky · Peter Deng & Rachel Holt (Aug 2026) ▶ 25:13)
- Route to the most efficient model to control token spend — Use a routing service that picks the most efficient model so finance leaders can control token spend. (Dave · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 20:50)
Stories & examples
- An 11-year-old’s token bill: Brit says her 11-year-old token-maxed to about $1,500 in a month. Lesson: Token spend scales fast without limits. (Brit · Allbirds Pivoted to AI Data Ce (Apr 2026) ▶ 32:09)
- Sam’s Cursor commits: Sam committed more lines of code with Cursor in a month than the previous summer, but says the diffs carry no economic value. Lesson: Activity is not value; he is sceptical millions will find economically valuable personal software. (Sam · Apple, OpenAI & Why Nobody Mak (Feb 2026) ▶ 31:19)
- Postgres, not tokens, was the biggest cost increase: Looking at the bill for the infrastructure he runs for himself since building many apps, Sam found the Postgres bill had gone up most, not token spend. Lesson: Storage and databases are where the leverage is as agents make stored data useful. (Sam · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 42:27)
- Offline CRM iOS app in two hours: Dave had always wanted an iPhone app for Offline’s custom internal CRM; with the agent he built it in about 2 hours over a weekend for about $150 of tokens, and colleagues now use it. Lesson: Compare agent token cost with hiring an engineer (tens of thousands of dollars and months). (Dave · Agentic AI Will Break the Inte (Jan 2026) ▶ 08:48)
- Sam’s CPU spend: Sam says he spends about $2,000 a month on CPU rather than GPU running agent tools, and would pay thousands more to be faster. Lesson: Agent users want more computers, not just more models. (Sam · Matt Mazzeo (Sep 2026) ▶ 37:39)
Numbers
- Increase in token consumption when agents are deployed in an enterprise: 3000 percent (host’s framing, approx.) (Dave · Amir Efrati (May 2026) ▶ 20:39)
- Relative price of Anthropic tokens vs open source: 100 times (host’s framing, approx.) (Sam · Why SpaceX’s $1.75T IPO Makes (Jun 2026) ▶ 21:19)
- Token cost of building the Offline iOS CRM app: 150 dollars (host’s framing, approx.) (Dave · Agentic AI Will Break the Inte (Jan 2026) ▶ 08:48)
- Monthly token bill for Brit’s 11-year-old: 1500 dollars (host’s framing, approx.) (Brit · Allbirds Pivoted to AI Data Ce (Apr 2026) ▶ 32:09)
- Sam’s monthly AI bill: 3000 dollars per month (‘a few thousand’) (host’s framing, approx.) (Sam · Chinese AI Model GLM 5.2 Beati (Jun 2026) ▶ 28:58)
- Token spend rule reported by Dave for one team: 1000 dollars per day (host’s framing, approx.) (Dave · Why OpenAI Can’t Win The Decen (Feb 2026) ▶ 50:02)
Caveats & warnings
- Share of a token budget an Uber CFO is said to have used in 15 days (secondhand, ‘Uber CFO or something’): 100 % of annual token budget (self-reported, approx.) (one of the hosts · Elon Musk vs OpenAI Trial, Goo (May 2026) ▶ 19:44)
- Token spend some people claim, which Sam finds hard to believe: 30000 dollars (host’s framing, approx.) (Sam · Google I/O Just Killed SEO? Ag (May 2026) ▶ 35:58)
- Raised versus spent on inference by each of four labs, as relayed by Dave: 900 million dollars spent of one billion raised (spoken ‘900’) (host’s framing, approx.) (Dave · Anthropic’s Bet on Coding Is W (Mar 2026) ▶ 20:49)
- Sam says measurement of open-source versus frontier usage is really suspect because tokens are not comparable. (Sam · SpaceX Hit $2.5 Trillion. Now (Jun 2026) ▶ 27:32)
- Sam admits that in practice he just gives Anthropic his money and cannot measure the counterfactual of other models. (Sam · Google I/O Just Killed SEO? Ag (May 2026) ▶ 28:09)
How much to trust this
- Evidence: 18 episodes, 3 independent speakers (the four hosts and unidentified host speech count as one voice, the show; guests counted individually).
Blueprints that use this
Sources
- (01)
Open Source AI Wins. Now Labs Are Running to Washington |Thinking Machines, Johnny Ive, Realtime API
- (02)
Elon Musk vs OpenAI Trial, Google Cloud Surge, Meta’s Blocked Acquisition, Anthropic Winning
- (03)
Allbirds Pivoted to AI Data Centers | SpaceX $1.5T IPO, Anthropic $800B, $175M Seed Rounds
- (04)
Apple, OpenAI & Why Nobody Makes Money in AI | The SaaS Apocalypse
21:58 into “Apple, OpenAI & Why Nobody Makes Money in AI | The SaaS Apocalypse”
- (05)
AI Can Change Everything and Still Be a Terrible Business + Peptide 101 from Matt Mazzeo
- (06)
Why OpenAI Can't Win The Decentralized AI Future (OpenClaw, Apple's Win, X.AI Exodus)
11:02 into “Why OpenAI Can't Win The Decentralized AI Future (OpenClaw, Apple's Win, X.AI Exodus)”
- (07)
Agentic AI Will Break the Internet (ClawdBot, OpenAI & the AI Demand Delusion)
8:48 into “Agentic AI Will Break the Internet (ClawdBot, OpenAI & the AI Demand Delusion)”
- (08)
Normal People Don't Want AI Agents (And Silicon Valley Can't Accept It)
19:08 into “Normal People Don't Want AI Agents (And Silicon Valley Can't Accept It)”
- (09)
The Risk Of Securitizing GPUs Like Long-Lived Infrastructure & More - A MOL Special Team Edition
- (10)
SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.
27:32 into “SpaceX Hit $2.5 Trillion. Now Watch What Elon Buys Next.”
- (11)
Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web
26:46 into “Google I/O Just Killed SEO? Agentic Search, Universal Cart, and the New Web”
- (12)
Chinese AI Model GLM 5.2 Beating Frontier Models | Meta Glasses, Polymarket Scandal, AI Talent War
- (13)
Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)
20:50 into “Why SpaceX's $1.75T IPO Makes No Financial Sense (And Doesn't Have To)”
- (14)
Anthropic’s Bet on Coding Is Working (OpenAI Shopping Pivot, A16Z’s Top 50 List, $1B Tennis Channel)
- (15)
Every CTO Is Quietly Switching Off Anthropic and OpenAI | GLM 5.2, Vox Americanus, Claude Fable 5