Dylan Patel — AI Token Supply & Demand Investment Thesis

Source: The Supply and Demand of AI Tokens | Dylan Patel Interview, Invest Like The Best (Patrick O'Shaughnessy), April 23, 2026.

The Framework: Tokconomics Meets Physical Supply Chains

Patel reframes AI investing around supply and demand for tokens: frontier capability creates surplus value per token versus sticker price, while silicon, memory, power, and packaging remain slow to expand. SemiAnalysis lives this demand curve internally—enterprise Claude usage scaled to ~$7M/year run-rate versus ~~$25M salary expense (~~25%+ of salary spend on coding agents)—so margins at labs and scarcity across the hardware stack are observable, not hypothetical.

Axis	Patel’s lens	Investable implication
Demand	Token value >> token price for serious adopters; usage concentrated among enterprises with leverage	Frontier labs + hyperscaler resale capture pricing power until compute catches up
Lab economics	Inference-heavy growth + rationing lifts gross margins rapidly when demand floors throughput	Revenue trajectories can outrun silicon receipts
Supply	Memory, logic capacity, WFE, niche materials—long lead times, prepayments, rising ASPs	Sellers of scarce inputs capture upside until greenfield fabs qualify
Second-order compute	RL “environments” and deployed apps chew CPU, not GPU	Server CPU sockets tight alongside accelerator hunger

Investment Thesis #1: Frontier Labs Sit on Extreme Inferencedriven Margin Expansion

The argument: Anthropic scaled revenue roughly $9B → $35–45B ARR (with monthly increments Patel pegs around $10B) while compute did not scale proportionally—so incremental inference economics reportedly dominate. Even assuming minimal cuts to research silicon, Patel estimates ≥72% gross margins today versus leaked docs earlier in the year showing ~30-something percent.

"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. It may be higher than 72% gross margins."

Bears anchor on GPU sticker shock; bulls should anchor on rationed throughput + enterprise contracts + rate-limit bargaining, which behaves like pricing power in a physical commodity shortage.

Trigger: Published gross-margin reconciliation vs. accelerator receipts; sustained throttle/re-tiering on frontier SKUs without churn.

Names: Anthropic (private — thematic); silicon lessors (Oracle ORCL, CoreWeave CRWV, Microsoft Azure, Amazon AWS with Trainium) that bank incremental rent while labs expand output.

Investment Thesis #2: “Ideas Cheap / Execution Easy” Compresses Release Cycles and Widens the Moat on Good Ideas

The argument: Implementation cost collapse lets teams test more hypotheses per month—model release pacing compresses (Patel cites ~2 months versus ~6 months historically) and non-technical operators ship production tooling (SemiAnalysis examples: GPU-accelerated chip reverse-engineering overlays, nationwide grid mapping in weeks vs. legacy 100-person shops on decade-old products).

"Historically has been very difficult. If I have an idea now I have to implement it. Implementing is hard. Now ideas are there. Implementation is very easy. It's expensive but it's very easy."

Generic “AI wrapper” startups get priced out of frontier tokens; winners pair capital + distribution + customer lock-in with aggressive model adoption.

Trigger: Divergence between headcount growth and output at information services firms; enterprise spending on coding agents exceeding 20–30% of payroll in more S-curves like SemiAnalysis.

Names: Microsoft (MSFT) (Copilot/GitHub ecosystem), Amazon (AMZN) (Bedrock + enterprise reach), quality data/API vendors feeding agent workflows.

Investment Thesis #3: Supply Is the Long Pole—Memory Pricing Has Multiple Doubles Ahead

The argument: DRAM/HBM fabs grow low–mid double digits annually; meaningful new bytes do not arrive until 2027–28 even with urgent responses. Patel expects DRAM ASPs to “double and triple again” because wafer reallocations require demand destruction via price, not rationing coupons.

"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."

Consensus “already long memory” ignores the years required to convert greenfield fabs—tightness extends through the current frontier ramp.

Trigger: Quarterly HBM/DRAM blended ASP guides; capex pulls from consumer electronics corroborating smartphone unit pressure Patels’ team has flagged elsewhere.

Names: Micron (MU); SK Hynix (HXSCF); (Samsung thematic, thinner US listing).

Investment Thesis #4: TSMC Super-Cycle → WFE Whiplash Still Underpriced

The argument: TSMC guided ~$56B capex this year with SemiAnalysis tracking ~$57.4B since January and pathing toward ~$100B annual fab spend ~2028. That ripples through ASML, Applied Materials, Lam Research, and downstream component vendors (Patel cites MKSI as an example).

"What people aren't focusing on is what does that mean next year and what does that mean the year after? … TSMC is going to spend hundred billion on capex … Maybe 28."

Investors watch leading-edge GPU wafer counts; tool install + subcomponent queues (lasers, glass, copper-clad laminates) remain the stealth gating items.

Trigger: TSMC rolling 3-year capex communications; ASML backlog vs. TSMC CoWoS additions (use tool shipment milestones, not gossip).

Names: TSMC (TSM); ASML (ASML); Applied Materials (AMAT); Lam Research (LRCX).

Investment Thesis #5: RL Training and Agent Deployment Create a Parallel CPU Bottleneck

The argument: RL environments (file edits, CAD hooks, physics sims) execute on CPU hosts; deployed model output also lands in CPU-bound serving paths. Patel: “CPUs are completely sold out and demand is skyrocketing.” He also notes ~120 FPGAs per next-gen AI rack, raising auxiliary silicon intensity.

"Those environments run on CPUs. They don't run on GPUs. They don't run on ASICs."

Capital allocators overweight GPU counts in training headlines; socket + networking + control-plane silicon becomes the hidden throttle for RL-heavy shops.

Trigger: Hyperscaler disclosures on server CPU backlog; FPGA lead times embedded in accelerator rack BOMs.

Names: Intel (INTC) (general-purpose server CPU leverage in tight market); FPGA-exposed suppliers (thematic watch on AMD/Lattice/Xilinx-class BOMs—not all US-listed cleanly).

Investment Thesis #6: Model Tiers Will Fracture—Liquidity Flows to Whoever Ships the Next “Exponential” Capability

The argument: Linear diffusion of today’s frontier could absorb ~$100B/yr economy-wide spend on a fixed model generation, but exponential spend needs the next capability jump—which Anthropic cannot unilaterally serve if silicon remains finite. Tier-2 labs still sell out: “pretty clear even the tier 2 lab is going to be sold out of tokens.” That sustains elevated pricing for OpenAI, Google, and neocloud lessors once they match capability.

"Economic value that the best model can deliver is growing faster than our ability to actually serve those tokens to people via the infrastructure."

Narrative leadership (Mythos / Opus 4.7 timing) ≠ durable exclusivity if compute pooling elsewhere closes the capability gap and margins compress for the follower.

Trigger: Cross-lab benchmark releases vs. simultaneous neocloud spot price spikes; evidence of Mythos-class models broadening beyond cyber verticals.

Names: Alphabet (GOOG) (TPU/cloud + model stack), OpenAI partners (MSFT, ORCL, CRWV), Amazon (AMZN) (Trainium + OpenAI linkage per discussion).

The Ecosystem Map (What Patel Is Seeing)

SemiAnalysis posture: ~$7M/yr Anthropic enterprise spend; coding agents consuming $6k/day spikes during grid build; reverse-engineering & macro “phantom GDP” research now one-person workloads.
Frontier model access: Mythos described as largest capability step in ~two years, 5–10× token list price, selectively distributed; Opus 4.7 launch same interview window.
Hardware stack callouts: H100 remarketing/pricing strength, lengthening Hopper renewals (3–4 year extensions), copper-clad laminates tight, optics prepayments.
Robotics (speculative bridge to future tokens): VLAs limited; expect 6–18 month window for few-shot pretrained robot policies that re-accelerate physical-world data demand.

Key Risks

Socio-political backlash: Patel forecasts large-scale protests against major labs within ~3 months as revenue scale makes AI tangible; vandalism/incidents already surfacing against executives.
Model hoarding / inequality: Preferential enterprise access concentrates advantage; regulators or public opinion could impose export-style controls on APIs.
Measurement error: “Phantom GDP” complicates macro signals—hard to reconcile official output with token-driven deflation.
Commodity glut cycle: Historic gluts followed shortages; aggressive memory/fab builds post-2028 could invert spreads.
Security/cyber dual-use: Mythos-class models withheld partly for offensive capability—misuse or breaches could force product delays.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	Nvidia (NVDA)	GPUs carry extended useful life, rising spot/contract rents, labs still throughput-limited	Blackwell/Hopper renewals above 3-year norms; sustained >70% GM at parent
2	TSMC (TSM)	Path to ~$100B/yr capex funds leading-edge AI logic + packaging	Multi-year CoWoS/logic spend guide; supplier pull-ins
2	SK Hynix (HXSCF) / Micron (MU)	DRAM/HBM ASPs have multiple doubles ahead; new bytes mostly 2027–28	Blended memory margin guide; smartphone demand destruction pass-through
2	ASML / AMAT / LRCX	TSMC capex whip accelerates WFE absorptive capacity	Tool backlog vs. shipment linearity; service intensity
3	CoreWeave (CRWV)	Neocloud leverage for incremental OpenAI/Core-style multi-year contracts	Contract duration & renewal pricing vs. spot
3	Oracle (ORCL)	OpenAI-scale GPU commitments sit alongside other hyperscalers—financing flexibility	AI infrastructure revenue disclosures & capex pacing

Monitoring Checklist

Anthropic (or peers) gross margin vs. accelerator opex — Confirms inference pricing power thesis vs. silicon cost.
Enterprise token throttle wait times — Persistent queueing = pricing still below market-clearing levels.
DRAM spot/contract settlements — Second leg of 2–3× move from here validates memory scarcity path.
TSMC 3-year capex glide path — Approaching $100B run-rate triggers WFE/order backlog revisions.
Server CPU backlog / lead times — Validates RL-environment thesis independent of GPU headlines.
Anti-AI protest frequency & policy responses — Societal backlash could cap deployment velocity or invite usage taxes.

Bottom Line

Frontier inference economics can look “software-like” gross margins (>72%) while silicon stays finite, because throughput rationing—not marginal COGS—clears the market when tokens generate multiples of their sticker price.
Ideas scaling faster than execution costs reshuffles winners toward organizations with distribution + capital, not whoever buys the cheapest subscription SKU.
Memory is mid-inning: expect another multi-step DRAM repricing before 2028 new supply, with consumer electronics paying the congestion tax.
TSMC→WFE whip is approaching a ~$100B annual capex imagination window—second-derivative plays (tooling, critical subcomponents) still lag big-cap SOX narratives.
RL + agent deployment makes CPU socket scarcity a first-class bottleneck—watch this as the “hidden” companion trade to GPU counts.