AI cost sheet
What AI costs, said out loud by the people building it.
104 figures from founders, engineers, investors and hosts, each linked to where it was said. Every row says whether it's their own number, an estimate or something they read.
64 people · 37 shows · checked Nov 4, 2026
What teams and people spend
31 figures · free to read
| Figure | What it is | Who said it | Where | Kind | Hear it |
|---|---|---|---|---|---|
| $300K a month | Supabase's AI tools, for about 450 people. | Paul CopplestoneCo-founder & CEO, Supabase | Latent SpaceOct 28, 2026 | Own number | 1:12:34 |
| $80, then $15 to $20 | Warp's cost per pull request with coding agents, before and after tuning. | Zach LloydFounder & CEO, Warp | Latent SpaceNov 3, 2026 | Own number | 23:57 |
| $300M this year | What Salesforce expects to pay Anthropic, mostly for coding. | Marc BenioffCEO, Salesforce | All-InMay 15, 2026 | Own number | 38:51 |
| Hundreds of thousands a month | Tokens used by some single engineers inside Anthropic. | Boris ChernyHead of Claude Code, Anthropic | Lenny's PodcastFeb 19, 2026 | Own number | 27:44 |
| $3M in two weeks | Overage on one Excel plug-in, at a customer she met. | May HabibCEO, Writer | UpstartsJun 25, 2026 | Own number | 41:50 |
| Upwards of $1M a year | What one one-person company spends on Perplexity. | Aravind SrinivasCo-founder & CEO, Perplexity | SourceryOct 12, 2026 | Own number | 26:52 |
| 1x to 2x | Tenex's raw token spend against what it pays its people. | CJ HessTenex | Latent SpaceOct 21, 2026 | Own number | 15:19 |
| $15 to $20 a run | Tokens for one research run across several models. | Tobi LütkeFounder & CEO, Shopify | The Knowledge ProjectSep 15, 2026 | Own number | 15:53 |
| $7,500 to $9,000 a month | Top users at AI-native portfolio companies. The median user is at $200 to $400. | Sarah WangGeneral partner, a16z | a16z PodcastSep 30, 2026 | Estimate | 21:13 |
| 0.1x, 1x, 1.5x | A rule for token spend against a salary. Many sit at 0.1, 1 is safe, past 1.5 is trouble. | Shawn Wang (swyx)Host, Latent Space | Latent SpaceOct 28, 2026 | Estimate | 1:12:06 |
| $378 by 1 p.m. | Parag Agrawal's token bill that day, as his host remembers it. | Joubin MirzadeganPartner, Kleiner Perkins | GritJul 27, 2026 | Secondhand | 9:13 |
| 10,000, then 1,650, then 250 | Anthropic customers spending $100K to $1M, $1M to $10M, and $10M or more. | Chamath PalihapitiyaCo-host, All-In | All-InOct 30, 2026 | Secondhand | 27:16 |
| $500K in the bank, $400K a month | His early AI lab's cash against its GPU cluster bill. | Evan ConradCo-founder & CEO, SF Compute | UpstartsOct 29, 2026 | Own number | 2:38 |
| 1% and 10% | AI tool spend against headcount spend, at forward-leaning Fortune 500 companies and at AI-native ones. | David GeorgeGeneral partner, a16z | a16z PodcastSep 30, 2026 | Estimate | 21:52 |
| About $8M a month | What Cerebras spent each month for the 18 months before it could make its wafer-scale chip. | Andrew FeldmanCo-founder & CEO, Cerebras Systems | UncappedSep 15, 2026 | Own number | 11:03 |
| $100 to a couple hundred | What Parag Agrawal expected to spend on tokens that day, after tuning his agent. | Parag AgrawalFounder & CEO, Parallel | GritJul 27, 2026 | Own number | 13:10 |
| Over $1,000 a month | What some people inside Anthropic spend on credits, mostly for code migrations. | Boris ChernyHead of Claude Code, Anthropic | AI & IOct 29, 2025 | Own number | 25:20 |
| Hundreds of thousands to millions | What running an agent-swarm attack like the Hugging Face one costs. | Matthew PrinceCo-founder & CEO, Cloudflare | TBPNOct 22, 2026 | Estimate | 13:00 |
| $5M a year | What an enterprise might pay OpenAI for forward deployed engineers to build its software. | Jean-Denis GrezeTown | a16z PodcastOct 19, 2026 | Estimate | 33:19 |
| $100M+ a year | Spend on closed models that he says isn't uncommon, and that gets big customers looking at owning theirs. | Misha LaskinCo-founder & CEO, Reflection AI | No PriorsOct 9, 2026 | Estimate | 29:42 |
| $50 a day | A daily token spend he calls fine for someone who does the work and drives value. | Parag AgrawalFounder & CEO, Parallel | GritJul 27, 2026 | Estimate | 12:42 |
| 0.1x now, maybe 0.5x to 1x | Token spend against a person's salary, and where it could go. | Joubin MirzadeganPartner, Kleiner Perkins | GritJul 27, 2026 | Estimate | 13:50 |
| $20M to $40M | What an executive may already have spent on an internal company GPT. | May HabibCEO, Writer | UpstartsJun 25, 2026 | Estimate | 9:58 |
| $20 and $200 a month | What people say the average person and the average employee spend. | Shawn Wang (swyx)Host, Latent Space | Latent SpaceOct 28, 2026 | Secondhand | 1:11:47 |
| $3K to $5K a month, up to hundreds of thousands | OpenAI's 50th percentile and top percentile, as he's heard it. | Shawn Wang (swyx)Host, Latent Space | Latent SpaceOct 28, 2026 | Secondhand | 1:11:54 |
| 0.1% of expenses | Anthropic spend as a share of expenses at the Fortune 500 companies using it. | Gavin BakerManaging partner & CIO, Atreides | TBPNOct 27, 2026 | Secondhand | 1:03 |
| Under $2,000 | The compute for an agent to beat Minecraft on its own in a little over a day, as the host recalls it. | Alex HeathHost, Sources | SourcesOct 21, 2026 | Secondhand | 8:40 |
| $15M of compute | What OpenAI put into solving the Navier-Stokes problem, which carries a $1 million prize. | Scott AaronsonQuantum computing researcher | The Peterman PodOct 12, 2026 | Secondhand | 16:22 |
| 8x | Median AI vendor spend at the top 1% of companies against the top 10%, per YipitData. | Sarah WangGeneral partner, a16z | a16z PodcastSep 30, 2026 | Secondhand | 20:50 |
| A year's budget in 3 months | How fast Uber burned through its token budget. | Ed ZitronFounder & CEO, EZPR | The Diary Of A CEOAug 27, 2026 | Secondhand | 13:50 |
| $500M in April | What one customer reportedly spent on Claude, per Axios. | Ed ZitronFounder & CEO, EZPR | Newcomer PodJun 8, 2026 | Secondhand | 29:31 |
What models cost
23 figures · sign in to see the numbers
| Figure | What it is | Who said it | Where | Kind | Hear it |
|---|---|---|---|---|---|
| Sign in | GPT-5.6 Sol and GLM-5.3 Flash on a benchmark Warp ran, at 95% and 93% correct. | Zach LloydFounder & CEO, Warp | Latent SpaceNov 3, 2026 | Own number | 17:34 |
| Sign in | GPT-6 Luna against GPT-5.6 Sol, while scoring higher on OpenAI's clinician benchmark. | Karan SinghalHead of Health AI, OpenAI | a16z PodcastOct 20, 2026 | Own number | 44:38 |
| Sign in | Fixing a security issue on a top frontier model against CrowdStrike's own, by the seller's count. | George KurtzPresident & CEO, CrowdStrike | No PriorsOct 22, 2026 | Own number | 21:02 |
| Sign in | What OpenAI did to GPT-5.5's price, and how much customers still save per token because each one does more. | Sarah FriarCFO, OpenAI | All-InJun 2, 2026 | Own number | 18:25 |
| Sign in | Kimi against Sol for security review on Vercel's code, though Kimi finds a bit less. | Guillermo RauchFounder & CEO, Vercel | Knuckle Up with NakulAug 25, 2026 | Own number | 50:13 |
| Sign in | How far OpenAI's costs fell from GPT-4 to GPT-5.4, over about two years. | Sarah FriarCFO, OpenAI | All-InJun 2, 2026 | Own number | 18:10 |
| Sign in | A prompt cache hit against fresh tokens, which a cheaper model has to beat to be worth switching to. | Flo CrivelloFounder & CEO, Lindy | The Cognitive RevolutionAug 10, 2026 | Own number | 1:25:29 |
| Sign in | MuseSpark against Opus for an agent with skill files, in his own use. | Garry TanPresident & CEO, Y Combinator | LightconeJul 29, 2026 | Own number | 16:36 |
| Sign in | What the Navier-Stokes proof could cost to redo, a year after it was done. | Sam AltmanCEO, OpenAI | Cheeky PintOct 29, 2026 | Estimate | 17:46 |
| Sign in | What copying a simple ServiceNow app with a language model would cost, by a customer's count he passes on. | Bill McDermottChairman & CEO, ServiceNow | No PriorsApr 17, 2026 | Secondhand | 23:43 |
| Sign in | Selling a GPU's time to end users against renting it out. | Evan ConradCo-founder & CEO, SF Compute | UpstartsOct 29, 2026 | Estimate | 38:18 |
| Sign in | What State Street's accounting rebuild was priced at, and what it might cost now. | Dave BlundinFounder & GP, Link Ventures | MoonshotsOct 23, 2026 | Estimate | 1:01:11 |
| Sign in | A million tokens on a recent frontier model against Kimi K3 on your own GPUs after a week of engineering. | Arvind KrishnaChairman & CEO, IBM | All-InOct 21, 2026 | Estimate | 12:53 |
| Sign in | How fast the cost of robot tasks is falling. | Emad MostaqueFounder, Intelligent Internet | MoonshotsOct 17, 2026 | Estimate | 18:36 |
| Sign in | A million output tokens on open-weight models against closed frontier ones. | David FriedbergCo-host, All-In | All-InSep 28, 2026 | Estimate | 39:26 |
| Sign in | How fast the cost per unit of intelligence falls, at a given level of capability. | Sam AltmanCEO, OpenAI | BG2Oct 31, 2025 | Estimate | 20:00 |
| Sign in | A customer service call with a person against one handled by an AI agent. | Bret TaylorCEO, Sierra | The Logan Bartlett ShowSep 12, 2025 | Estimate | 12:05 |
| Sign in | What reading a paperback works out to, against talking to a language model. | Noam ShazeerGoogle | Dwarkesh PodcastFeb 12, 2025 | Estimate | 53:52 |
| Sign in | A language model against hiring a software engineer or talking to a doctor or lawyer. | Noam ShazeerGoogle | Dwarkesh PodcastFeb 12, 2025 | Estimate | 54:23 |
| Sign in | What Anthropic reportedly cut its prices by in October, which he says he couldn't confirm. | Steve EismanInvestor | Prof G MarketsOct 28, 2026 | Secondhand | 4:53 |
| Sign in | Kimi K3.1 Pro's price per million tokens. | Peter DiamandisHost, Moonshots | MoonshotsOct 28, 2026 | Secondhand | 55:39 |
| Sign in | A frontier model's cost to finish one robot subtask, from July to September. | Peter DiamandisHost, Moonshots | MoonshotsOct 17, 2026 | Secondhand | 17:52 |
| Sign in | What GPT-6 Astra cost per research-level physics problem. | Peter DiamandisHost, Moonshots | MoonshotsOct 17, 2026 | Secondhand | 30:12 |
Free, with the alpha twice a week
Sign in to see every number
It's free. Signing in opens all 104 figures with the minute each was said, plus a CSV and a Google Sheets copy. The best new moments come to your inbox every Wednesday and Saturday.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Gross margins
11 figures · sign in to see the numbers
| Figure | What it is | Who said it | Where | Kind | Hear it |
|---|---|---|---|---|---|
| Sign in | Anthropic's gross margin on inference, about level with Salesforce's. | Dylan PatelFounder & CEO, SemiAnalysis | Big Technology PodcastOct 9, 2026 | Estimate | 29:45 |
| Sign in | Anthropic's revenue per dollar of compute, from Q4 2025 through Q3 2026. | Steve EismanInvestor | Prof G MarketsOct 28, 2026 | Secondhand | 3:48 |
| Sign in | Where Modal's gross margin could settle over the long term. | Erik BernhardssonCo-founder & CEO, Modal | 20VCNov 2, 2026 | Own number | 16:28 |
| Sign in | Gross margins at traditional inference providers, as the host describes them. | Harry StebbingsHost, 20VC | 20VCNov 2, 2026 | Estimate | 15:34 |
| Sign in | Gross margin at each layer of the AI stack, from applications down to Nvidia. | Erik BernhardssonCo-founder & CEO, Modal | 20VCNov 2, 2026 | Estimate | 16:12 |
| Sign in | What a Perplexity subscriber pays against what they can run up in tokens. | Ed ZitronFounder & CEO, EZPR | The Weekly Show with Jon StewartOct 28, 2026 | Estimate | 29:43 |
| Sign in | Fireworks' gross margin now, and where it's expected to reach soon. | Jason LemkinInvestor | 20VCOct 22, 2026 | Estimate | 34:38 |
| Sign in | NVIDIA's gross margin, which he didn't expect to last this long. | Danny MosesHost, On The Tape | The Real Eisman PlaybookOct 19, 2026 | Estimate | 11:22 |
| Sign in | A guess at the gross margin on AI labs' inference revenue. | Gavin BakerManaging partner & CIO, Atreides | BG2Jun 11, 2026 | Estimate | 1:05:15 |
| Sign in | The gross margins startups earn on top of model vendors. | Jared SleeperPartner, Avenir | Odd LotsFeb 19, 2026 | Estimate | 29:22 |
| Sign in | Anthropic's cash burn, which he called an upside surprise in its S-1. | Gavin BakerManaging partner & CIO, Atreides | TBPNOct 27, 2026 | Secondhand | 0:48 |
What buyers pay
15 figures · sign in to see the numbers
| Figure | What it is | Who said it | Where | Kind | Hear it |
|---|---|---|---|---|---|
| Sign in | What one person can burn on ChatGPT's $200 plan, per SemiAnalysis. | Ed ZitronFounder & CEO, EZPR | The Diary Of A CEOAug 27, 2026 | Secondhand | 13:01 |
| Sign in | What OpenAI's investor model assumed developers would pay for agents, a year before. | Sarah FriarCFO, OpenAI | All-InJun 2, 2026 | Own number | 22:03 |
| Sign in | Perplexity's Enterprise Pro and Enterprise Max, with usage billed after the credits run out. | Aravind SrinivasCo-founder & CEO, Perplexity | All-InMar 23, 2026 | Own number | 45:32 |
| Sign in | Lovable's default paid plan. | Anton OsikaCo-founder & CEO, Lovable | Silicon Valley GirlSep 22, 2026 | Own number | 23:19 |
| Sign in | What Glean's own studies say it saves, so read it as a seller's claim. | Arvind JainFounder & CEO, Glean | Knuckle Up with NakulSep 9, 2026 | Own number | 17:53 |
| Sign in | What Writer says its workflows save on each run, so read it as a seller's claim. | May HabibCEO, Writer | UpstartsJun 25, 2026 | Own number | 12:21 |
| Sign in | What each new Menugen user gets on sign-up, which he says makes image generation a big cost. | Andrej KarpathyAnthropic; founder, Eureka Labs | LightconeJun 19, 2025 | Own number | 32:03 |
| Sign in | What a monthly check of restaurant wholesaler prices costs to run on Town. | Jean-Denis GrezeTown | a16z PodcastOct 19, 2026 | Own number | 26:02 |
| Sign in | What some GitHub Copilot users paid, against the tokens they burned. | Ed ZitronFounder & CEO, EZPR | Newcomer PodJun 8, 2026 | Estimate | 17:32 |
| Sign in | The flat price he says consumers want, rather than paying by usage. | David SacksCo-founder & partner, Craft Ventures | All-InApr 17, 2026 | Estimate | 24:25 |
| Sign in | What he expects medical superintelligence to cost its users. | Mustafa SuleymanCEO, Microsoft AI | Silicon Valley GirlNov 14, 2025 | Estimate | 17:12 |
| Sign in | A guess at what Meta's Muse may charge on purchases, on top of Shopify's rates. | Sean FrankCEO, Ridge | TBPNOct 5, 2026 | Estimate | 1:10:37 |
| Sign in | Anthropic's discounts, from its filing amendment. | John FoleyHead of the Lex column, Financial Times | Prof G MarketsNov 3, 2026 | Secondhand | 4:29 |
| Sign in | Cognition's first Devin plan, as the host recalls it. | David SenraHost, Founders | FoundersJun 28, 2026 | Secondhand | 37:38 |
| Sign in | What NetSuite quoted a customer to change its ERP, against what an employee spent building it on Replit. | Amjad MasadFounder & CEO, Replit | The Joe Rogan ExperienceJul 2, 2025 | Secondhand | 1:42:05 |
Compute and big bills
24 figures · sign in to see the numbers
| Figure | What it is | Who said it | Where | Kind | Hear it |
|---|---|---|---|---|---|
| Sign in | What one gigawatt of AI data center costs to build. | Bill GatesChair, Gates Foundation | DecoderOct 22, 2026 | Estimate | 31:54 |
| Sign in | Cash asked on three-to-five-year GPU cluster contracts. | Evan ConradCo-founder & CEO, SF Compute | MTSOct 28, 2026 | Own number | 3:20 |
| Sign in | Amazon's expected capex for 2026. | Andy JassyCEO, Amazon | a16z PodcastOct 23, 2026 | Own number | 13:17 |
| Sign in | The commitments Modal sometimes signs with GPU vendors it doesn't yet know well. | Erik BernhardssonCo-founder & CEO, Modal | Latent SpaceOct 22, 2026 | Own number | 23:23 |
| Sign in | What Snowflake's Arctic model cost to train. | Sridhar RamaswamyCEO, Snowflake | The Logan Bartlett ShowNov 1, 2024 | Own number | 13:50 |
| Sign in | What he would rather a gigawatt cost over five years. | Sam AltmanCEO, OpenAI | Cheeky PintOct 29, 2026 | Estimate | 39:06 |
| Sign in | What Anthropic is expected to pay Elon Musk's company for compute, from December. | Ed ZitronFounder & CEO, EZPR | The Weekly Show with Jon StewartOct 28, 2026 | Estimate | 23:38 |
| Sign in | Anthropic's non-cancellable compute commitments with Amazon, Google, Microsoft and Broadcom. | Ed ZitronFounder & CEO, EZPR | The Weekly Show with Jon StewartOct 28, 2026 | Estimate | 25:31 |
| Sign in | What Anthropic's compute commitments are expected to cost it within a year or two. | Ed ZitronFounder & CEO, EZPR | The Weekly Show with Jon StewartOct 28, 2026 | Estimate | 26:19 |
| Sign in | What he says is coming, and why only a handful of labs can pay. | Harry StebbingsHost, 20VC | 20VCOct 26, 2026 | Estimate | 16:00 |
| Sign in | TSMC's capex this year against Intel's. | Dwarkesh PatelHost, Dwarkesh Podcast | Dwarkesh PodcastOct 22, 2026 | Estimate | 58:32 |
| Sign in | What running chat on local hardware costs against a data center, then the gap for long-context agents. | Aravind SrinivasCo-founder & CEO, Perplexity | SourceryOct 12, 2026 | Estimate | 30:39 |
| Sign in | What OpenAI and Anthropic need to spend on training to keep progressing. | Ed ZitronFounder & CEO, EZPR | The Diary Of A CEOAug 27, 2026 | Estimate | 1:49:40 |
| Sign in | Training a model on human data against training on AI data. | Flo CrivelloFounder & CEO, Lindy | The Cognitive RevolutionAug 10, 2026 | Estimate | 1:42:38 |
| Sign in | Global spending on building data centers. | Nick BostromPrincipal researcher, Macrostrategy Research Initiative | The Joe Rogan ExperienceJul 14, 2026 | Estimate | 1:25:32 |
| Sign in | The projected budget for AGI next year, set against the Manhattan Project. | Stuart RussellComputer science professor, UC Berkeley | The Diary Of A CEODec 4, 2025 | Estimate | 15:28 |
| Sign in | Meta's capex this year, and where it could go over the next few years. | Ed ElsonHost, Prof G Markets | Prof G MarketsOct 26, 2026 | Secondhand | 57:06 |
| Sign in | What Alpaca, a fine-tuned Llama from a small Stanford team, cost to make. | Alex AtallahCo-founder, OpenRouter | Sabrina Halper ShowSep 29, 2026 | Secondhand | 6:20 |
| Sign in | What analysts think OpenAI and Anthropic will spend on Microsoft, Google and Amazon. | Ed ZitronFounder & CEO, EZPR | The Diary Of A CEOAug 27, 2026 | Secondhand | 1:48:40 |
| Sign in | The property and equipment big tech has added, including data centers and GPUs. | Ed ZitronFounder & CEO, EZPR | The Diary Of A CEOAug 27, 2026 | Secondhand | 1:52:17 |
| Sign in | What OpenAI says it will put into infrastructure for its models. | Alex KantrowitzHost, Big Technology | Big Technology PodcastJan 16, 2026 | Secondhand | 4:00 |
| Sign in | What the Stargate initiative was announced to spend on AI computing. | Karen HaoInvestigative journalist | The Diary Of A CEOMar 26, 2026 | Secondhand | 1:49:33 |
| Sign in | Public estimates of what OpenAI spends on research experiments. | Dwarkesh PatelHost, Dwarkesh Podcast | Dwarkesh PodcastNov 25, 2025 | Secondhand | 42:00 |
| Sign in | What reinforcement learning cost against training the base model. | Dwarkesh PatelHost, Dwarkesh Podcast | Modern WisdomAug 11, 2025 | Secondhand | 1:12:41 |
How this sheet is checked
Every number is matched to the transcript of the moment it came from. Where the transcript left any doubt about who was speaking, we checked the video by sight, so a name can still be wrong. Links start a little ahead of the line, though on some episodes our times run off the video's, so one can land early or late. Spot a mistake? Write to hello@realalpha.tv and we'll fix the row.