Latest / Elon Musk Podcast / OpenAI trades tokens for startup equity

Transcript

OpenAI trades tokens for startup equity · Elon Musk Podcast

• 0:00
  Open AI is giving $2,000,000 worth of artificial intelligence
• 0:05
  tokens to all 169 startups in the current Y Combinator batch
• 0:11
  in exchange for equity. Right.
• 0:13
  Which I mean, they aren't handing out cash.
• 0:15
  They're trading raw compute power for actual ownership
• 0:18
  stakes in these young companies. Yeah.
• 0:20
  And that brings up this fundamental problem for anyone
• 0:23
  trying to build a business right now.
• 0:24
  You know, when a startup accepts millions in infrastructure
• 0:27
  instead of traditional funding, are they securing their future
• 0:30
  or are they just like signing away their independence to a
• 0:33
  single platform? Yeah, that is exactly the
• 0:36
  tension founders are dealing with.
• 0:37
  Sam Altman actually calls this specific strategy token maxing.
• 0:41
  Token maxing right? Yeah, it's an approach where
• 0:44
  startups and their engineers just, you know, maximize their
• 0:47
  use of AI models and credits to accelerate product development
• 0:50
  and their internal workflows. You just saturate your entire
• 0:53
  oeration with AI because the cost is temorarily 0.
• 0:56
  Exactly. But to understand what token
• 0:59
  maxing actually does to a company's architecture, we kind
• 1:02
  of have to look at what a token functionally is.
• 1:04
  Right, It's the basic unit of text that these large language
• 1:07
  models process and generate. Like if you're building an app
• 1:10
  in English, 1 token is roughly 4 characters.
• 1:14
  Which works out to about 3/4 of a word.
• 1:16
  Yeah. So a prompt of say, 75 words is
• 1:19
  going to cost you about 100 tokens.
• 1:22
  But that linguistic constraint really alters the math, because
• 1:26
  that 4 character estimate only applies to English.
• 1:30
  Oh right, because of the training data.
• 1:32
  Exactly. Tokenization is inherently
• 1:35
  linguistically biased based on how these models were trained.
• 1:38
  So non-english text produces a much higher token to character
• 1:42
  ratio. Yeah, if you're a founder
• 1:44
  building a product for the Japanese or Arabic markets, the
• 1:47
  same amount of semantic meaning requires significantly more
• 1:50
  tokens. So you basically pay a tax for
• 1:53
  not operating in English. You really do. 2 prompts of
• 1:57
  identical length and meaning are going to generate entirely
• 1:59
  different bills depending on the language.
• 2:01
  And for a startup trying to build multilingual products,
• 2:04
  that discrepancy completely alters their burn rate.
• 2:06
  Right. It's a fundamental planning
• 2:08
  assumption that changes how far that $2,000,000 actually goes.
• 2:11
  Which is wild because, I mean the face value of the offer
• 2:14
  seems staggering at first glance, $2,000,000 per startup
• 2:18
  across a cohort of 169 companies.
• 2:22
  Yeah. That totals over $330 million
• 2:25
  based on retail pricing. But Open AI is definitely not
• 2:28
  spending $330 million to make this happen.
• 2:30
  No, not at all. Those credits are priced at the
• 2:32
  retail rate they charge developers for API access.
• 2:36
  Open AI is actual marginal cost to run those API calls in their
• 2:40
  data centers. It's just a fraction of that
• 2:42
  retail price. Right, It's kind of like an
• 2:44
  airline buying equity in a tech startup using frequent flyer
• 2:48
  miles. That's a perfect analogy.
• 2:50
  The face value looks enormous on a balance sheet, but the actual
• 2:54
  cost to the provider is minimal because the flights are already
• 2:57
  scheduled and the seats are empty.
• 2:59
  Right. They're taking excess server
• 3:00
  capacity and effectively turning it into venture capital.
• 3:02
  It's pretty brilliant, yeah. But the billing complexities for
• 3:05
  the startups using these tokens, you know, they go beyond just a
• 3:09
  flat retail price. Because token costs behave like
• 3:12
  metered compute, and they're split into these 4 main buckets
• 3:16
  that founders have to track. Right, you have input tokens,
• 3:19
  which is the text sent to the model.
• 3:21
  So that includes the visible prompt, the user types, the
• 3:24
  system messages running in the background.
• 3:26
  Yeah, and the conversation history of the app keeps alive
• 3:28
  to maintain context. Then you have output tokens,
• 3:32
  which is the text generated by the model.
• 3:34
  And cache tokens too, right? Yeah.
• 3:36
  Cache tokens involve reused conversation history, which cost
• 3:40
  less. Yeah.
• 3:41
  And then there are reasoning tokens.
• 3:42
  Right for the advanced models. Exactly.
• 3:44
  They use reasoning tokens internally while they process a
• 3:47
  complex request before they produce the final answer.
• 3:51
  You essentially pay for the models internal scratchpad.
• 3:53
  And output pricing is really where the rate card spread gets
• 3:56
  unforgiving. Oh, absolutely.
• 3:58
  Generating text requires significantly more computational
• 4:01
  power than just reading it. Across the current Open AI
• 4:05
  lineup, output tokens commonly cost 4 to 8 times as much as
• 4:09
  input tokens. Yeah, if you look at the GT54
• 4:13
  short context ricing, it sits at a six times output premium.
• 4:17
  And models like nano, mini, GPT 5, they maintain an 8 times
• 4:21
  ratio. Which completely changes the
• 4:24
  definition of product optimization for a developer.
• 4:27
  Right, because if you want to lower your cloud bill, trimming
• 4:29
  a few words from a user's prompt barely does anything.
• 4:33
  Exactly. Capping the answer length or
• 4:35
  enforcing A strict Jason data schema saves you far more money.
• 4:39
  Because the real cash is burned on the output side.
• 4:41
  Yeah, and if a startup's finance team groups all the charges
• 4:45
  under a single line item for AI spend, they completely lose
• 4:49
  visibility. They won't know if they're
• 4:50
  overpaying for input context or just bleeding money on overly
• 4:55
  verbose outputs. And the mechanism these startups
• 4:57
  use to accept these tokens is an uncapped safe.
• 5:01
  Right, simple agreement for future equity.
• 5:03
  It's a standard Silicon Valley instrument, but that uncapped
• 5:06
  part changes the risk profile entirely.
• 5:09
  It's highly risky for a founder. Usually a protective investment
• 5:12
  instrument includes a floor evaluation.
• 5:14
  Right. A floor sets a minimum post
• 5:16
  money valuation to guarantee the founder isn't diluted too
• 5:19
  severely if the next funding round is small.
• 5:21
  Let me put some specific numbers to that just to show how a floor
• 5:25
  actually functions. Say an investor puts $500,000
• 5:30
  into your startup using a safe with a $10 million floor
• 5:33
  valuation. The target ownership percentage
• 5:36
  implied there is 5%. If your startup thrives and
• 5:40
  raises money later in an equity round at a $15 million
• 5:43
  valuation, the SAFE converts at that higher valuation.
• 5:47
  So the investor stake dilutes down to roughly 3.7%.
• 5:51
  Right. And they're prorator right
• 5:54
  triggers, allowing them to buy more shares to maintain their
• 5:57
  target 5% if they want to put more money in.
• 5:59
  But the floor protects the founder when things go wrong,
• 6:03
  like if that same startup struggles and raises money at a
• 6:05
  $7.5 million valuation. The safe still converts at the
• 6:09
  $10 million floor. Exactly.
• 6:11
  It converts at a price higher than the new money investors are
• 6:13
  paying. So the investor still gets their
• 6:15
  5%, but the founder is protected from extreme dilution.
• 6:19
  Because the investor share is capped by that floor, but an
• 6:23
  uncapped safe operates without that safety net.
• 6:26
  Right, there is no floor open. AI's ownership slice depends
• 6:30
  entirely on whatever valuation the startup receives at its next
• 6:35
  priced round. Usually the Series A.
• 6:37
  So if a startup uses these credits, build something
• 6:40
  incredible and hits a $100 million valuation at that next
• 6:45
  round. Open AI gets about 2% of the
• 6:47
  company for that $2,000,000 token grant.
• 6:50
  But the problem happens when the startup fails to reach escape
• 6:53
  velocity. Yeah, if they burn through their
• 6:55
  token budget without successfully building a product
• 6:58
  that commands a high valuation. They've still surrendered
• 7:00
  equity. Exactly.
• 7:01
  They've diluted their equity, yeah, and they have zero actual
• 7:05
  cash in the bank to extend their runway.
• 7:07
  Right. They trade equity for
• 7:09
  infrastructure, fail to increase their valuation and dilute
• 7:12
  themselves heavily. And they're left with a smaller
• 7:14
  piece of a struggling company and they still need to raise
• 7:17
  capital just to make payroll. We really have to look at the
• 7:19
  behavioral psychology driving founders to accept this specific
• 7:23
  deal despite those risks. Yeah, because startups naturally
• 7:26
  operate under extreme resource scarcity.
• 7:29
  Cash is tight, time is tight. And that constant pressure
• 7:33
  physically alters how a founder processes information.
• 7:37
  It completely overwhelms the prefrontal cortex, which is, you
• 7:40
  know, the executive center of the brain responsible for long
• 7:43
  term planning and risk assessment.
• 7:45
  Right, so the offer creates this false sense of relief for an
• 7:48
  overwhelmed executive function. By eliminating the immediate
• 7:52
  need to budget for server costs, it frees up mental bandwidth.
• 7:56
  They delay the financial pain. They're basically creating
• 7:59
  future control for present comfort.
• 8:00
  Exactly, it's a classic symptom of chronic stress induced
• 8:04
  decision paralysis. Yeah, when you're just worried
• 8:06
  about keeping the lights on next month, giving up two percent of
• 8:09
  a theoretical future valuation feels like a harmless
• 8:12
  abstraction. And there's an evolutionary
• 8:14
  psychology angle here as well. I mean, the human brain evolved
• 8:17
  to reward social inclusion and conformity, right?
• 8:20
  Hearing Sam Altman and Y Combinator in the same sentence
• 8:24
  triggers a little neural chemical response.
• 8:26
  It produces dopamine and oxytocin for these founders.
• 8:29
  Yeah, being accepted into this highly visible group reinforces
• 8:33
  behaviors that lead to group alignment.
• 8:35
  You just naturally want to do what the rest of the tribe is
• 8:37
  doing. And when Y Combinator general
• 8:39
  partners endorse the deal publicly, it acts as an
• 8:42
  amplifier for that social proof. Right, the decision shifts
• 8:45
  entirely away from financial logic.
• 8:47
  It just becomes about a desire to be part of the winning
• 8:50
  faction. Founders see their peers taking
• 8:53
  the deal, and they subconsciously mimic them.
• 8:56
  And that mimicry is driven by mirror neurons in the premotor
• 8:59
  cortex. They activate when we observe
• 9:02
  the actions of others, making us subconsciously copy them.
• 9:05
  Which is the biological mechanism behind how trends
• 9:08
  spread through a cohort? More adoption leads to more
• 9:10
  perceived legitimacy. And that legitimacy leads to
• 9:13
  more adoption. The fear of missing out on
• 9:16
  access to these cutting edge models feels to the brain like a
• 9:19
  primal threat of exclusion. Really reframes the risk
• 9:22
  completely. Due to reciprocity bias,
• 9:25
  founders perceive giving up the equity not as a financial loss
• 9:28
  but as an investment in a relationship.
• 9:30
  Right. Open AI gave them something they
• 9:32
  perceive as highly valuable, so they feel the psychological
• 9:35
  compulsion to give something valuable back.
• 9:38
  But we're looking at a situation where the true cost isn't just
• 9:42
  the monetary value of the equity.
• 9:44
  Right. The real cost is the
• 9:46
  architectural independence they give up.
• 9:48
  It's like building your company's headquarters on land
• 9:51
  someone else owns, using materials they supply and
• 9:54
  Justice agreeing they can change the rent whenever they want.
• 9:56
  Exactly. By accepting the infrastructure,
• 9:59
  startups surrender that independence entirely.
• 10:02
  Because every single API call feeds open AI's ecosystem.
• 10:06
  Yeah. Over a few months, the startup's
• 10:08
  entire code base grows around Open AI's specific API.
• 10:12
  'S the data structures flow naturally into their systems.
• 10:14
  The engineering teams become highly specialized in optimizing
• 10:18
  for open AI's tools. And once that lock in happens,
• 10:20
  switching to a competitor becomes this huge engineering
• 10:23
  task. You aren't just using a vendor
• 10:25
  for a service anymore, you have become a dependent node in
• 10:28
  someone else's network. Right.
• 10:30
  If Open AI decides to raise prices later, or if they they
• 10:32
  change their data usage policies, the startup has no
• 10:35
  choice but to comply. Because ripping out the core
• 10:37
  intelligence of your product to switch to an open source model
• 10:40
  would stall product development for months.
• 10:43
  And this lock in strategy becomes very clear when you look
• 10:46
  at the broader compute market and how capital is flowing at
• 10:49
  the macroeconomic level. Oh, definitely.
• 10:51
  Compute resources are increasingly viewed as a primary
• 10:54
  currency. They're being traded for equity
• 10:57
  and leverage just like Fiat currency.
• 10:59
  The numbers moving around the space completely validate that,
• 11:02
  like Enthropic is paying SpaceX $1.25 billion a month for
• 11:08
  compute power. Right.
• 11:09
  And that deal runs for years. It's designed to scale up
• 11:12
  capacity in massive data centers.
• 11:14
  The tech industry is literally treating compute as capital, and
• 11:18
  startups are just caught in the crossfire of these giants trying
• 11:21
  to establish monopolies by controlling the infrastructure.
• 11:24
  Layer so startups have to defend themselves against this dynamic.
• 11:28
  The only way to survive that kind of dependency is with a
• 11:31
  strict token budgeting framework.
• 11:32
  Yeah, a founder needs to understand exactly how pricing
• 11:35
  works beyond the headline numbers Open AI puts on their
• 11:38
  website. Because the pricing cliffs are
• 11:40
  really where startups split out. Take the GPT 5.4 model as an
• 11:44
  example. Right, the published short
• 11:46
  context rate only applies below a threshold of 272,000 input
• 11:51
  tokens. And below that line, it costs
• 11:53
  $2.50 per million input tokens and $15 per million output
• 11:58
  tokens. But crossing that 272,000 input
• 12:02
  token threshold is a critical pricing event.
• 12:05
  Yeah, once a single request crosses it, Open AI charges a
• 12:08
  long context schedule for the entire session.
• 12:11
  Your input costs immediately doubled to $5 per million tokens
• 12:15
  and the output cost jumped to $22.50 per million tokens.
• 12:19
  Right, so if a user generates a request with 250,000 input
• 12:23
  tokens and 20,000 output tokens, it costs you about $0.92.
• 12:28
  But if that same user keeps the conversation going and generates
• 12:31
  a request with 300,000 input tokens and the exact same 20,000
• 12:34
  output tokens. The cost jumps to nearly $2.00.
• 12:37
  A modest 20% increase in input tokens pushes the entire request
• 12:42
  onto a much more expensive schedule, literally doubling the
• 12:46
  cost of the interaction Which. Is why a technical founder needs
• 12:48
  to actively manage those thresholds.
• 12:51
  They should really route low stakes tasks like you know,
• 12:54
  ticket triage, basic classification or summarizing
• 12:57
  small documents to cheaper models like GT5 Mini.
• 13:00
  And they have to reserve premium models only for tasks where the
• 13:04
  reasoning depth directly impacts revenue, compliance or user
• 13:07
  retention. Exactly.
• 13:09
  Treating that $2,000,000 is free money to burn indiscriminately
• 13:12
  on the most expensive models is a critical error in judgment.
• 13:15
  Yeah. Trading equity for
• 13:18
  infrastructure credits might solve an immediate cash flow
• 13:20
  problem for these founders, but it creates a permanent
• 13:23
  architectural dependency on a single provider.
• 13:26
  Right. And if compute power is truly
• 13:28
  replacing cash as the new global currency for technology
• 13:31
  companies, you really have to wonder how startups that choose
• 13:34
  to remain entirely independent will manage to compete against
• 13:37
  peers who have millions in subsidized infrastructure.
• 13:40
  If you're not subscribed yet, take a second and hit follow on.
• 13:43
  Whatever app you're using helps us keep making this.
• 13:46
  We appreciate you being here. Also check out our YouTube
• 13:50
  channel for more business and tech updates.
• 13:52
  There's a link in the description.