Apr 23, 2026 · 45m · invest-like-the-best
The Supply and Demand of AI Tokens | Dylan Patel Interview · Invest Like The Best
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Invest Like The Best, Dylan Patel of SemiAnalysis breaks down the explosive growth of AI token demand, the economic shift toward cheap automated execution, severe semiconductor supply chain bottlenecks, and the macro dynamics shaping frontier AI labs.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Patrick holds 17.9% of the talking time here. How this is scored →
Patrick as informed peer 4.4 Guest teaching 7.1 Guest disagreement 3.1 Patrick pushing back 1.7
speaking balance: gold is Patrick, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 29:29 Permanent underclass warning
Dylan issues a stark, uncompromising warning that anyone who does not aggressively utilize and capture value from tokens will be relegated to a permanent economic underclass.
Hardest push from Patrick ▶ 4:56 Challenging runaway token budgetsPatrick pushes back on exponential token spending, asking when a business owner must hit the brakes and switch to cheaper commodity models.
Biggest teaching moment ▶ 34:05 Semiconductor lead time realitiesDylan refutes standard assumptions about fast-clearing supply chains by detailing structural multi-year fabrication lead times for DRAM and specialized PCB materials.
Patrick holds their own ▶ 13:20 In-flight frontier model upgrade demandPatrick illustrates firsthand enterprise user psychology by describing his instant unwillingness to use Opus 4.6 the moment Opus 4.7 became available.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Patrick as informed peer | Guest teaching | Guest disagreement | Patrick pushing back | Why |
|---|---|---|---|---|---|---|
| SemiAnalysis's Skyrocketing Token Spend and Claude Code Adoption | 4 | 7 | 2 | 1 | Patrick sets up the premise regarding Dylan's internal token usage and validates the rapid trajectory. Dylan educates Patrick and listeners on concrete internal workflow transformations, from reverse-engineering chip material overlays to calculating phantom GDP. | |
| AI Strategy, Commoditization, and Energy Grid Modeling | 5 | 6 | 3 | 3 | Patrick probes how a business owner manages runaway token costs versus cheaper models. Dylan counters with the existential necessity of moving fast and details how his team rapidly mapped the entire US energy grid to outcompete traditional legacy analytics firms. | |
| Sponsor Mid-Roll: Ramp, WorkOS, and Rogo | 5 | 5 | 2 | 3 | Segment includes mid-roll sponsor reads followed by Patrick asking if well-capitalized funds might bypass third-party research by building in-house tooling. Dylan explains why agile specialization keeps external research vendors ahead. | |
| Macro Token Economics: Anthropic's Revenue and Margin Explosion | 4 | 8 | 3 | 1 | Patrick prompts Dylan on token macroeconomics. Dylan reveals financial breakdowns of Anthropic's exploding ARR and expanding 72%+ gross margins due to intense pricing power and compute scarcity. | |
| Demand for Frontier Intelligence and Token Arbitrage | 4 | 7 | 3 | 2 | Patrick shares his own experience desiring the bleeding-edge Opus 4.7 over 4.6. Dylan unpacks the dynamics of enterprise frontier access, token efficiency economics, and how higher per-token intelligence lowers overall task cost. | |
| Accelerated Model Release Cadences and Execution Shifts | 4 | 7 | 2 | 1 | Patrick asks Dylan about his reaction to Mythos benchmarks. Dylan explains how implementation costs have collapsed, compressing the frontier model release cycle from months down to weeks. | |
| Restricting Frontier Access and Societal Impacts | 4 | 8 | 4 | 2 | Patrick questions whether Dylan feels genuine fear or merely grappling with uncertainty. Dylan describes selective frontier deployment and the emerging concentration of AI leverage among elite institutions. | |
| Physical AI, Few-Shot Robotics, and Token Demand Expansion | 4 | 6 | 3 | 1 | Patrick brings up robotics as a potential secondary demand driver. Dylan explains why current VLA models are sample-inefficient and predicts few-shot pre-trained robotics models within 6 to 18 months. | |
| Pre-Training Scaling Laws and Compute Efficiency | 5 | 8 | 3 | 2 | Patrick asks what Mythos signals about pre-training scaling laws. Dylan breaks down compute efficiency curves, chip architectural scaling, and the structural capacity divergence between Anthropic and OpenAI. | |
| Avoiding the 'Permanent Underclass' Through Token Utilization | 5 | 7 | 5 | 2 | Dylan makes the provocative claim that failing to maximize token usage relegates individuals to a permanent economic underclass, laying out a three-part framework for capturing AI value. | |
| Sponsor Mid-Roll: Vanta and Ridgeline | 5 | 8 | 4 | 2 | Following sponsor reads, Patrick asks why supply chains have not rapidly expanded to meet hardware shortages. Dylan walks through severe upstream bottlenecks in DRAM, TSMC CapEx, and raw materials. | |
| Unsung Bottlenecks: FPGAs and CPU Demand in Reinforcement Learning | 5 | 8 | 2 | 1 | Patrick asks about non-GPU hardware components. Dylan details the surge in CPU demand required for reinforcement learning evaluation environments and application deployment. | |
| Quantifying Tokenomics and Tracking 'Phantom GDP' | 4 | 7 | 2 | 1 | Patrick asks what unknown metrics Dylan seeks most. Dylan explains the difficulty of quantifying phantom GDP and tracking the diffuse macroeconomic value generated by downstream token adoption. | |
| Predicted AI Backlash, Public Perception, and Rebranding | 4 | 7 | 5 | 2 | Patrick asks what to expect in three months. Dylan bluntly predicts public protests against AI lab leaders, critiquing executive media appearances and public communication strategies. |
Statements from this episode (34)
Patel: SemiAnalysis spends $7M on Claude Code vs $25M payroll
“Across a firm, we're spending seven million dollars a year now on Claude code at the current rate versus our salary expense being in the neighborhood of twenty-five million dollars.”
Patel: SemiAnalysis AI spend could exceed total payroll by year-end
“And if this trajectory continues, then, you know, we'll spend more than a hundred percent by the end of the year”
Patel: Companies will lay off staff as Claude Code multiplies productivity
“I think other folks will start to reckon with the fact that, huh, If this person can do the work of five to 10 to 15 people using quad code, then all of a sudden I should probably cut people.”
Patel: SemiAnalysis built SEM chip material mapping app using Claude tokens
“One person on the team, they've been able to spend with a couple thousand dollars of Claude tokens, they've been able to create this application that is GPU accelerated, runs on a server that we have at CoreWeave, and anytime we send it an image, it's able to …”
Patel: AI can currently automate roughly 3% of BLS tasks
“The BLS has this entire Bureau of Labor Statistics has this entire, like, set of, like, 2000 tasks, And so he did that with AI, which ones can be done by AI, which ones cannot and grading them across a rubric, you know, about three percent are doable now with …”
Patel: Solo economist's AI project would have taken 200 economists a year
“And he's like, dude, this would have taken the team of 200 economists a year.”
Analyst mapped entire US power grid using $6,000 daily AI tokens
“He was spending like 6000 dollars a day. It was an insane amount, but he scraped every single power plant in the U S every single transmission line above a certain voltage. And created this entire mapping of the entire US grid, as well as a lot of demand sourc…”
Patel: Anthropic's ARR has surged to between $35B and $40B
“Anthropic has gone from nine billion revenue to what they're at, 35, forty billion. Now probably by the time this airs, 40, forty-five billion. Who knows? ARR.”
Patel: Anthropic gross margins have a floor of 72%
“So ultimately what they've done, even if you assume all incremental compute they've gotten has gone towards inference, their margins are at a floor of 72%. In reality, some of that incremental compute they've got probably went to research and development and m…”
Patel: Low-value SaaS startups will soon get priced out of tokens
“But the shitty SaaS startup and SF who is using Claude to generate, you know, their software product is not necessarily actually creating a ton of value, and therefore they're gonna get priced out of tokens soon enough.”
Patel: Anthropic's Mythos is potentially the biggest capability jump in two years
“Mythos is potentially the biggest step up in model capabilities in like two years.”
Patel: Opus 4.6/4.7-level AI inference will be 100x cheaper in a year
“Current 4.6 Opus or 4.7 Opus to your models a year from now, my spend for the same exact quality of the model would probably be like 70 K. I bet you it'll be a hundred times cheaper.”
Patel: Anthropic Mythos is cheaper in most tasks than Claude 4.6 Opus
“Anthropic Mythos is more expensive as a model, but it spends a lot less tokens to do the thing, and therefore it is actually cheaper in most tasks than four, six Opus, because it's just way more efficient, even though each individual token is smarter.”
Patel: AI shifts value from implementation to idea selection and capital
“Your ability to implement something is not actually that important. Your ability to choose the correct idea for AI to implement, and then your ability to sell that idea, or sell what the AI has implemented, is what matters. Your ability to garner capital towar…”
Patel: Frontier AI models will face increasingly restricted deployment
“Where they only release, meet those to certain companies for cyber, that's just going to be something that continues. Models will have less broad and less broad deployment.”
Patel: Anthropic could double Opus pricing and users would still pay
“They could double their pricing on Opus, and I would continue to pay, and I bet most users would continue to pay. I bet that wouldn't solve their humongous capacity problem that they have.”
Patel: VLA models will probably not scale for robotics due to data inefficiency
“Robots, currently the robot models VLAs Vision Language Action Models, which is very popular right now, is probably Not going to be the thing that ultimately scales beyond. They are inefficient in data. And we can't scale the data for them fast enough.”
Patel: Robotics few-shot learning breakthroughs will arrive in 6 to 18 months
“And so I think in the next six to 18 months, we'll start seeing real breakthroughs in robotics that enable few shot learning, i.e. There's a pre-trained robot model, and now there's a robot that you have hired or bought or whatever. You showed a few examples a…”
Patel: Mythos proves that AI pre-training scaling laws still hold
“But ultimately, yes, Mythos is a significantly larger model. It's proof that the scaling laws still work. Everything about it shows that the trend line continues of models. More compute into model makes model better.”
Patel: Google and Anthropic are not heavy users of training GPUs
“Google and Anthropic are not heavy, heavy users of GPUs on the training side, but OpenAI, they'll, they'll start having their new class of models.”
Patel: OpenAI is procuring Trainium AI chips from Amazon
“Now they're getting Tranium as well from Amazon.”
Patel: Tier-two and tier-three AI labs will sell out of tokens
“It's pretty clear even the tier two lab is going to be sold out of tokens, let alone the tier one lab. The tier one lab will have better margins, but the tier two lab will be sold out and probably the tier three lab will also be close to sold out.”
Patel: Frontier Model Value Is Outpacing Infrastructure Serving Capacity
“Economic value that the best model can deliver is growing faster Then our ability to actually serve those tokens to people via the infrastructure. And so this gap will continue to grow and the model labs will continue to have expanding margins until people in …”
Patel: GPU useful life is reaching 7 to 8 years, not under 5
“There's people who have argued GPUs full lives are less than five years. Complete nonsense. There are clusters now re-signing three or four year old hopper clusters re-signing for three or four more years. There's a 100 clusters that are re-signing for another…”
Patel: ASML is sold out and constrained by Carl Zeiss capacity
“You see ASML, Is completely sold out, and they need Carl Zeiss to expand faster.”
Patel: Incremental memory fab capacity will not arrive until late 2027 or 2028
“Even if they wanted to build as fast as possible, it doesn't come till 28 early, late 27 at best.”
Patel: DRAM prices will double or triple from current levels
“DRAM will double or triple from here still, because That's how much capacity is required, and they have to steal capacity from somewhere else, and the only way to steal capacity from somewhere else in a capitalist economy is demand destruction via higher prici…”
Patel: TSMC could reach $100B in annual CapEx by 2028
“Sincerely, they may spend a hundred billion dollars on capex in twenty-twenty-eight, and people, like, just can't fathom that”
Patel: Next-Generation AI Racks Require 120 FPGAs per Rack
“There's a project we did on FPGAs, and it turns out there's a 120 FPGAs per, per next generation rack.”
Patel: CPUs are completely sold out driven by reinforcement learning demand
“CPU-wise, all these reinforcement learning environments plus all the slop code you and I are generating that is now running on some, you know, Vercel instance or whatever it is or some AWS instance or some bucket that we've spun up, all of that requires CPU, a…”
Patel: RL Simulation Environments Run on CPUs, Not GPUs or ASICs
“So the environments can get more and more complex, and those environments run on CPUs. They don't run on GPUs. They don't run on ASICs. The ASICs run the model,”
Patel: Sam Altman's residence was targeted with Molotov cocktails twice
“You look at the comments of news articles where Sam Altman had a Molotov cocktail thrown at his house twice in like two weeks.”