In late March 2026, entrepreneur Alex Finn posted a viral warning on X about Anthropic's leaked Claude Mythos model. He argued that "$2,000/month Ultra tiers" were coming, that intelligence was becoming a luxury good, and that a permanent divide between AI haves and have-nots was inevitable. The post struck a nerve—because it articulated a fear many investors share.
But the data tells a completely different story. And for those who understand the underlying economics, this misperception creates real alpha.
The Core Thesis
AI inference costs are following the most predictable deflationary curve in technology history. The perception of "escalating prices" is a surface-level misread that creates systematic mispricing across public and private markets. Understanding this curve—and its second-order effects—is where the real investment edge lies.
Part I: The Numbers That Matter
Price Per Capability, Not Raw Token Price
The critical error in the "AI is getting expensive" narrative is confusing sticker price with price per capability. Yes, GPT-4 launched at $30 per million input tokens. Yes, new frontier models often debut at premium pricing. But this misses the fundamental dynamic.
Epoch AI's landmark analyses (updated through early 2026) quantify what's actually happening:
for fixed capability
annual deflation
observed on some benchmarks
becomes accessible
Concrete example: A task requiring 43 million output tokens on o4-mini in April 2025 needed only 5 million tokens on GPT-5.2 by December 2025—a 3× real cost drop in eight months after adjusting for pricing changes.
The pattern is consistent across every model generation:
- o1 was "too expensive" at launch → now available in free tiers
- Claude Opus was unaffordable 12 months ago → Sonnet 4.5 delivers equivalent value at 1/10th the price
- Mythos will launch expensive (leaked docs call it "very expensive to serve") → expect 6-18 months until democratization
💡 The Investor's Edge
Markets systematically overweight the current price of frontier AI and underweight the trajectory. This creates opportunities in both directions: companies perceived as "locked out" of AI are undervalued, while AI lab valuations often bake in monopoly-like pricing power that won't materialize.
What's Driving the Deflation?
The cost curve isn't magic—it's the predictable result of four compounding forces:
| Driver | Mechanism | Impact Timeline |
|---|---|---|
| Algorithmic Efficiency | Better architectures, distillation, reasoning optimizations | Continuous (every 3-6 months) |
| Hardware Improvements | Newer GPUs cheaper per FLOP, inference-optimized chips | Annual cycle (H100 → H200 → Blackwell) |
| Competition | DeepSeek, Meta, Alibaba, Mistral driving prices toward marginal cost | Accelerating (2026 is the breakout year) |
| Open-Source Catch-Up | Llama 4, Qwen, DeepSeek-V3 at near-frontier on key benchmarks | 6-12 month lag to frontier, compressing |
The result: Epoch AI projects 5-10× yearly capability-cost reductions holding through at least 2027-2028. This is not speculation—it's the continuation of trends visible in every major model release.
Part II: The Market Misperceptions
Misperception #1: "VC Subsidies Are Masking True Costs"
The "house of cards" narrative is partially correct: OpenAI, Anthropic, and others are burning capital to acquire users at below-profitable pricing. Some analysts warn of 3-10× price normalization for frontier APIs as subsidies end and IPO pressures mount.
But efficiency gains are outpacing the subsidy burn rate. The math:
- If costs are 40-50% below sustainable pricing today (the bear case)
- And capability-cost falls 5-10× annually
- Then within 12-18 months, current subsidized pricing becomes profitable at current capability levels
The "correction" will be smaller and shorter than skeptics expect—because the deflation curve keeps running.
Misperception #2: "Frontier Access = Competitive Moat"
The most damaging misperception for investors: believing that access to Mythos-tier models creates durable competitive advantage for early adopters.
Reality: The frontier advantage window is 6-18 months, compressing with each generation. Companies building moats around "we have the best AI" will see those moats evaporate faster than they can monetize them.
"A skilled user with today's $20/month model outperforms a novice with Mythos." — Common refrain in responses to Finn's viral post
The durable moats are in data, distribution, workflow integration, and execution speed—not model access.
Misperception #3: "AI Labs Will Capture Most of the Value"
The prevailing market thesis: OpenAI, Anthropic, and Google will capture outsized returns as AI becomes essential infrastructure.
The counter-thesis (where the alpha lies): AI inference is commoditizing faster than cloud computing did. Open-source is more advanced relative to proprietary than at any point in cloud history. The value capture will flow to:
- Application layer companies that turn AI into specific, valuable workflows
- Data moat companies with proprietary training/fine-tuning advantages
- Infrastructure picks-and-shovels (chips, energy, cooling) with genuine supply constraints
- Incumbents in regulated industries who can deploy AI with existing trust/compliance infrastructure
Part III: The Investment Framework
Timeline and Probability Weighting
Based on the evidence, here's the most probable timeline:
| Timeframe | Development | Confidence |
|---|---|---|
| 2026-2027 | Mythos-level intelligence via $20-50/month plans or free distilled tiers | High (85%+) |
| 2027-2028 | Open-source rivals run locally on consumer hardware for near-zero marginal cost | High (80%+) |
| 2028+ | AI intelligence "too cheap to meter" for most uses—like cloud storage today | Medium-High (70%) |
| Ongoing | Ultra-agentic enterprise/government applications stay premium | High (90%) |
Risk/Reward Matrix
⚠️ Key Risks
- Energy constraints: Data center power demand up 267% in some areas. If grid capacity doesn't scale, compute scarcity persists.
- Regulatory compute caps: Government intervention limiting AI training/inference could freeze the cost curve.
- Algorithmic plateau: If efficiency gains slow dramatically, current pricing becomes structural.
- Oligopoly consolidation: If 2-3 labs dominate and coordinate pricing, commoditization slows.
✅ Opportunity Vectors
- Application layer: Companies turning cheap AI into valuable workflows capture margin the labs can't.
- Infrastructure: Genuine supply constraints in chips, data centers, and energy.
- Incumbents with data: Healthcare, finance, legal—regulated industries with proprietary data.
- Open-source ecosystem: Companies building on open weights (lower platform risk).
Sector-by-Sector Analysis
Overvalued (Relative to Thesis)
- Pure-play AI labs at current valuations: Pricing in monopoly-like margins that won't materialize
- "AI wrapper" startups: Thin value-add on top of APIs that will commoditize
- GPU-heavy hardware plays: Already pricing in multi-year demand that assumes slower efficiency gains
Undervalued (Relative to Thesis)
- Traditional enterprises with AI adoption potential: Market pricing them as "disrupted" when they're actually beneficiaries
- Vertical SaaS with domain data: Can fine-tune/embed AI in ways that create durable advantage
- Energy/utility infrastructure: Genuine bottleneck as AI demand explodes
- Education and workforce development: The real constraint is adoption, not access
Part IV: The Macro Picture
GDP and Productivity Impact
Goldman Sachs, McKinsey, and others project AI adding $7-25 trillion annually to global GDP. Even conservative models (MIT's Daron Acemoglu) see 1-2% U.S. GDP boost over a decade from profitable AI deployments.
The productivity J-curve is real: initial slow adoption, then rapid gains as costs fall and integration matures. We're still in the early, slow phase—which is why the opportunity exists.
📊 The Abundance Paradox
As AI commoditizes, total inference spend is skyrocketing—AI now drives 55% of some cloud budgets. But per-task costs keep falling. This is classic abundance economics: the pie grows faster than prices fall, creating value for everyone in the chain. The winners are those who can use the abundance, not those trying to restrict it.
Inequality Dynamics
The honest assessment:
- Short-term (2026-2027): Early frontier access creates velocity advantages. Well-funded startups test 5-10× more hypotheses. Some winner-take-most dynamics accelerate.
- Medium-term (2028-2030): Open-source and efficiency erode moats. History of tech (internet, smartphones, cloud) shows rapid diffusion once costs drop. AI literacy becomes the new middle-class skill.
- Policy wildcard: Without intervention, skill-biased technical change widens wage gaps. With education reform, UBI pilots, or public AI infrastructure, it narrows them.
What Would Change the Thesis?
Monitor these signals for thesis invalidation:
- Efficiency gains slow to <2× annual: Current 5-10× gains are the engine. Dramatic slowdown changes everything.
- Open-source gap widens: If proprietary models pull away from open alternatives, commoditization stalls.
- Coordinated pricing by dominant labs: Oligopoly behavior could sustain premiums longer.
- Energy/compute crisis: Genuine scarcity would override efficiency gains.
Part V: Actionable Positioning
For the Individual Investor
- Avoid pure AI lab exposure at current multiples. The commoditization thesis suggests their margin compression will exceed market expectations.
- Look for "AI beneficiary" incumbents. Traditional companies that can deploy AI to reduce costs or expand capabilities—but aren't priced as AI plays.
- Infrastructure with genuine constraints. Energy, cooling, data centers—but be selective. Chips are likely already fully priced.
- Build personal AI fluency. The real arbitrage is using cheap tools effectively while others wait for "better" AI. This compounds.
For the Entrepreneur
- Build on open-source. Lower platform risk, better economics, and the performance gap is closing faster than most realize.
- Focus on workflow, not model. The value is in turning AI output into specific outcomes—not in model access.
- Move fast. The 6-18 month frontier advantage window means speed is the moat. Ship imperfect products with current AI rather than waiting for perfect AI.
- Bet on adoption infrastructure. Training, templates, integrations—the bottleneck is human, not technical.
For Policy Watchers
The bipartisan U.S. House AI Task Force and global discussions emphasize workforce reskilling and energy planning. Watch for:
- Compute access as public good (changes the open-source dynamic)
- Energy infrastructure investment (removes a key constraint)
- Antitrust focus on chip/data center chokepoints
- Education reform at scale (shifts the adoption bottleneck)
Conclusion: The Gap That Matters
The leaked Mythos episode is a symptom of hype cycles, not destiny. Intelligence is not becoming permanently expensive—it is following every transformative technology before it: expensive at the cutting edge, ubiquitous shortly after.
The most probable future is one of broad access, accelerated innovation, and societal abundance, provided we prioritize literacy and equitable infrastructure over fear of temporary premiums.
The gap that matters most isn't between Sonnet and Mythos. It's between those using AI effectively today and those waiting for perfect, cheap intelligence tomorrow. The data says: don't wait.
For investors, the alpha lies in understanding this curve—and positioning before the market catches up.
Key Takeaways for Investors
- 40-50× annual cost deflation for fixed AI capability is the dominant trend
- 6-18 month frontier advantage windows mean model access isn't a durable moat
- Value capture flows to applications, data, and infrastructure—not labs
- The real bottleneck is adoption—human workflows, not AI capability
- Position for abundance, not scarcity