The Terminal State: Intelligence Meets Demand

Anthropic just told you the future. Not in the blog post — in the confession.

"Very expensive for us to serve, and will be very expensive for our customers to use."

That's the lab that built the most capable AI model ever created admitting they can't afford to let people use it. Claude Mythos — confirmed by Anthropic spokesperson to Fortune after a CMS leak exposed 3,000 internal files — is a new tier above Opus. Training complete. Already in limited enterprise testing. "Dramatically higher scores on software coding, academic reasoning, and cybersecurity." "Far ahead of any other AI model in cyber capabilities."

And they're delaying general release because the infrastructure doesn't exist to serve it.

This is the signal. Not the capability jump — everyone expected that. The signal is the GAP between what AI can do and what the physical plant can deliver. That gap is widening with every generation. And it tells you exactly where capital reprices over the next decade.

The Four Phases

The demand curve for intelligence has a terminal state. We're in Phase 1.

Phase 1 (NOW, 2026-2028): Supply < Demand
  Models outrun infrastructure. Mythos too expensive to serve.
  Binding constraint: PHYSICAL (power, silicon, datacenter space)
  Investable: datacenter capacity, fab, power infrastructure

Phase 2 (2028-2032): Supply ≈ Demand
  Custom silicon + efficiency close the gap.
  Binding constraint: REGULATORY (state capture, who gets access)
  Investable: quarry layer (measurement, clearing, certification)

Phase 3 (2032-2038): Supply > Demand
  Most decisions automated. Humans in loop by choice.
  Binding constraint: DESIRE (what do humans actually want?)
  Investable: platforms for co-creation, wanting/anticipation products

Phase 4 (2038+): Supply >>> Demand
  Intelligence is free. The split is complete.
  Investable: be on the right side of the split

Each phase has a different binding constraint. Each constraint has different investable vehicles. Capital accumulated during early phases buys your seat at the later ones.

Phase 1: The Inference Cost Crisis (Now)

Current Opus pricing: $5/$25 per million tokens. Mythos estimated 3-5x that. The math:

3x cost per token × growing usage = massive compute scaling pressure
Anthropic can't afford to ship broadly yet
Hyperscalers signed $22B in datacenter contracts BEFORE Mythos — they knew this was coming
Those contracts may represent floor, not ceiling

Meanwhile, Cerebras CEO Andrew Feldman laid out the disaggregated inference architecture (March 26): separate prefill (parallel, compute-bound) from decode (serial, memory-bandwidth-bound) onto specialized hardware. The tradeoff is flexibility — fixed hardware ratios break when workloads shift.

His key insight: hyperscalers win either way (fleet diversity absorbs changes). Enterprises get locked in. And disaggregation "adds to, rather than replaces" existing inference. Total hardware TAM grows. Jevons paradox in action — every approach to making inference cheaper increases demand.

The inference cost crisis doesn't have one winner. It has a hardware portfolio:

GPU (NVDA):           Low specialization, high flexibility. Expensive.
SRAM-based (Groq):    High specialization, low flexibility. Fast.
Wafer-scale (Cerebras): Medium both. Searching for market.
Weights-in-silicon (Taalas): Maximum specialization, zero flexibility. 20x cheaper for fixed models.
Custom ASIC (Etched, hyperscaler chips): High specialization per architecture.
Disaggregated:        Per-stage specialization. Only works with fleet diversity.

Hyperscalers will run ALL of these. They need datacenter SPACE and POWER for heterogeneous hardware — not just GPU farms. That's the thesis for datacenter capacity providers.

Phase 2: State Capture

Anthropic's Mythos rollout tells the story. First access goes to cybersecurity organizations and select enterprise partners. Not consumers. Not developers. Defense and security.

The state will not permit any entity to hold more power than it. When Anthropic resisted removing guardrails, they were destroyed in 48 hours. OpenAI, Google, xAI complied — and were anointed. The mechanism is three chokepoints: chips, cloud, contracts.

But state capture is temporary (3-7 year window):

Captured AI optimized for compliance gets dumber than free AI (Lysenko effect)
Open source can't be recalled — Llama is out
100x algorithmic density (Musk's claim, physics checks out) means frontier runs on consumer hardware within a decade
Government moves at 0.1x/year, technology at 10x/year

The investable play during Phase 2 is the quarry layer — measurement and clearing infrastructure that power depends on but doesn't threaten. Every AI output touching finance needs a rating, a benchmark, a credit score, a cleared transaction. AI doesn't disrupt this. It amplifies throughput.

Phase 3: Desire Is the Binding Constraint

When inference commoditizes, intelligence supply exceeds demand. Most decisions get automated. The question shifts from "can AI do this?" to "what do humans actually want done?"

Desire is chemical and embodied. AI can satisfy desire, not generate it. Five irreducible human drives survive automation:

Relative status (biological need, not resource proxy)
Competence frontier (flow at edge of skill)
Witnessed recognition (being seen)
Anticipation/wanting (dopamine = pursuit, not arrival)
Narrative coherence (meaning as constraint, not drive)

The products that satisfy these through CREATION (patron fighting with the mason about where the spire goes) survive. The products that satisfy them through CURATION (discriminator in a GAN) are self-liquidating — the more refined your taste, the faster they learn it, the sooner you're redundant.

The Terminal State: The Great Bifurcation

Super-augmented (5%):
  Direct intelligence. Set desire. Push frontier.
  Satisfy drives through WORK.
  Rock stars, founders, patrons.
  Own the directing layer.

Demand-satisfied (95%):
  Consume output. Enjoy abundance.
  Satisfy drives through consumption.
  Audience, garden-dwellers.
  Comfortable, not suffering.

This isn't dystopia. The 95% aren't suffering — abundance is real. The 5% aren't gatekeeping — they need the frontier because they can't NOT push it. Historical analog: rock stars, elite athletes, master craftsmen. Society always has a directing class that does the hard thing because they're wired for it. The split already exists. AI makes it legible.

The 5% can't be automated because:

Desire is chemical. AI satisfies it, doesn't generate it.
Direction requires choosing WHAT to optimize. AI optimizes. Humans point.
New games require "fine disregard for the rules." AI trained on existing patterns can't originate the violation.
Narrative coherence requires a self that persists and cares.

What This Means for Capital

Work backward from the terminal state:

Phase 1 owns the constraint. Datacenter capacity, fab, power. The physical bottleneck that Mythos just exposed. Hyperscaler contracts signed before Mythos → conservative relative to actual demand.

Phase 2 owns the transition. Quarry layer infrastructure. State capture beneficiaries on dislocation. Navigator teams that get acquired during the transition.

Phase 3 owns the desire infrastructure. Platforms for co-creation. "Mine > best" ecosystems (open source wins because ownership premium > capability premium). Wanting/anticipation products, not satisfaction products.

Phase 4 is the outcome. Capital accumulated during Phases 1-3 determines which side of the bifurcation you're on. The endpoint isn't wealth. It's being the kind of entity that directs intelligence rather than being directed by it.

The transition window is where all the repricing happens. By Phase 4, the game is over. Mythos just showed us we're deep in Phase 1 — the lab that built the frontier model can't afford to serve it. That gap between capability and infrastructure is where the capital goes.

The Evidence Trail

Anthropic spokesperson to Fortune (Mar 27, 2026): "Step change and the most capable we've built to date"
Leaked draft blog: "Very expensive for us to serve" — delaying general release
Feldman/Cerebras (Mar 26): Disaggregated inference adds to total TAM, hyperscalers need heterogeneous hardware
Musk/Diamandis (Jan 2026): "100x more intelligence per gigabyte from algorithmic improvements alone"
Anthropic blacklisting (Feb 2026): State capture mechanism demonstrated in real-time

Disclaimer

Entertainment and research documentation. NOT financial advice, recommendation, or offer. Documenting pattern recognition and framework development. Your capital, your decisions, your risk.