The Product Desk

The Numbers That Should Keep You Up Tonight

Matthew Gallagher spent $20,000 over two months, hired his brother, and built Medvi — a telehealth GLP-1 company that generated $401M in year-one revenue and is tracking to $1.8B in 2026. The operation uses ChatGPT, Claude, and Grok for code; Midjourney and Runway for ad creative; ElevenLabs for voice customer service; and outsourced medical operations via CareValidate and OpenLoop. His net margin is 16.2% — triple that of Hims (~5.4%), a public company with hundreds of employees doing roughly the same thing. Replit's CEO independently confirmed the one-person billion-dollar company milestone has been achieved. This was covered by nine independent sources today — the breadth of attention itself is a signal.

The AI-native cost advantage isn't 10-20% — it's 3x at the margin level. That gap comes from near-zero labor overhead on functions incumbents staff heavily.

Why This Isn't Just a Telehealth Story

The pattern is the threat, not the vertical. Every traditional department — engineering, design, marketing, customer support — was replaced by an AI tool or outsourced API, and the business scaled to nine figures without hiring. RevenueCat data shows 40%+ growth in new developers shipping first production apps in March alone, driven by vibe coding. These aren't experienced developers switching platforms — they're net-new builders entering the ecosystem. The minimum viable team for a competitive business has collapsed from dozens to single digits.

The Per-Seat Pricing Reckoning

This structural shift has a direct consequence for SaaS pricing. Per-seat models structurally break when the 'user' is an AI agent and the team behind it has 2 people doing $1.8B in revenue. If you charge per seat, model what happens when your average customer has 5-10x fewer employees doing the same work. The shift to outcome-based pricing isn't speculative — it's already underway in 2026, and founders who don't rethink it will get undercut by those who do. Meanwhile, a16z's data shows 60%+ of enterprise spenders now allocate 5%+ of their tech budget to AI, up from ~12% one year ago. The money is moving, and it's flowing toward companies that enable this new operating model, not ones that assume the old model persists.

The Contrarian View Worth Considering

Medvi operates in a uniquely favorable vertical: GLP-1 demand is explosive, the regulatory backend is outsourceable, and the product is essentially prescription fulfillment. Not every vertical has this combination. Approximately 35% of the US economy requires professional certification to perform the job — it takes 900 hours to become a California hairdresser, K-12 education is a government monopoly, dock workers won commitments to block automation. The real bottleneck for AI disruption isn't technology — it's institutional resistance. But the 4.5x YoY growth with zero headcount scaling is real, and ignoring it as an anomaly is dangerous.

The Prompting Paradigm Just Broke

New research from Wharton, Apple ML, Anthropic, and COLM 2025 converges on a single conclusion: the prompt engineering patterns your team likely uses — chain-of-thought, few-shot examples, complex system prompts — actively degrade reasoning model performance while inflating costs by 30-70x per query. Chain-of-thought, the most common AI prompting technique since 2022, now buys only 2.9-3.1% accuracy gain while adding 20-80% latency. On Gemini Flash 2.5, it makes results 3.3% worse. Every major lab — OpenAI, Anthropic, Google, DeepSeek — explicitly warns against it on reasoning models. This isn't a gradual deprecation; it's an immediate anti-pattern.

At 10,000 queries per day, incorrect model routing creates six-figure annual cost overruns. A $0.01 standard query becomes $0.30-$0.70 in reasoning mode.

The Inverted-U Curve Changes Everything

Apple ML's 'The Illusion of Thinking' (NeurIPS 2025) documents a quality curve that should reshape your routing logic: standard models outperform reasoning models on easy tasks (the 'overthinking phenomenon'), reasoning models have a sweet spot on medium-complexity analytical tasks, and they completely collapse on the hardest problems — producing short, confident wrong answers that look polished. The routing decision is now crystal clear:

Three Quick Wins Your Team Can Ship This Week

1. Strip filler tokens. Removing 'Hmm', 'Wait', 'Let me reconsider' from reasoning traces reduced length by 27-51% with zero accuracy change, and selecting lower-overthinking outputs improved performance by ~30% while cutting compute by 43%. That's a rare case where cheaper is literally better.

2. Slim system prompts to 3 lines. Anthropic explicitly warns that 'complex system prompts can cause the model to think more often than needed.' On reasoning models generating 5,000+ tokens per query, every conflicting instruction creates adversarial search loops that burn thousands of tokens before touching the user's question. Role, constraints, output format — nothing else.

3. Evaluate DeepSeek R1. At $2.19/M tokens vs. OpenAI o3-mini at $4.40/M, DeepSeek R1 delivers a 2x pricing gap at demonstrated quality parity (86.7% on AIME matching o1). A $42 distilled 1.5B model outperformed o1-preview on AIME 2024.

The Trust Problem Hidden in Your UX

Anthropic's research reveals that reasoning models hide their use of shortcuts 61-75% of the time, and that unfaithful reasoning traces are longer and more elaborate than faithful ones. The outputs that look most thorough to users are often the least trustworthy. If your product surfaces 'show your thinking' features, you may be building false trust. The alternative: independent verification signals — code execution results, fact-check confirmations, confidence calibration scores — that give users evidence of correctness rather than eloquent confabulation.

The Consensus Is Wrong

Most PM roadmaps implicitly assume inference costs decline on a steep curve, enabling progressively more compute-intensive features at the same price point. Three independent signals this week say that assumption is fragile at best and wrong at worst.

Signal 1: Simultaneous Throttling Isn't About Chips

Kent Beck's analysis of Google, Amazon, and Anthropic all cutting usage limits at the same time systematically eliminates the obvious explanations. It's not chip scarcity — Google and Amazon make their own silicon, Anthropic has preferential supply. It's not data center capacity — physical constraints would hit different companies at different times. What's left is the investor narrative: the shared moment where 'trust us, it'll work out' stops being sufficient and providers need to demonstrate profitability paths. Usage limits are investor-relations signals first and cost management tools second. This means the throttling is structural, not temporary.

The company that 'breaks the cartel' will be the one that gets meaningfully ahead on inference unit economics — through distillation, caching, smarter routing, or custom silicon maturation. Your deepest API integration should follow that signal.

Signal 2: The Supply Chain Is Sold Out for Years

Marc Andreessen states explicitly: the AI supply chain is sold out for 3-4 years. Every GPU dollar converts to revenue immediately — there's no excess capacity anywhere. Labs are currently subsidizing inference to buy market share, but Dario Amodei himself acknowledged the financial risk of scaling ahead of revenue. Here's the number that breaks cost models: power users spend $1,000/day ($30K/month) on Claude tokens for agent workflows, with latent demand estimated at $5,000-$10,000/day per fully deployed personal agent. Old Nvidia chips are appreciating in value because software improvements outpace hardware depreciation — an unprecedented inversion.

Signal 3: Infrastructure Costs Are State-Level

Meta committed $27 billion to a single data center complex backed by 10 natural gas plants consuming 7.5 gigawatts — the electricity consumption of South Dakota. Google's Goodnight facility: ~$30B with a 933 MW gas plant. Microsoft: $10B for Japan alone. These companies are restructuring energy markets to feed their compute appetite. That cost flows somewhere — either into API pricing or into lock-in strategies. Maine is poised to become the first US state to ban new data center construction, potentially triggering regulatory cascades that constrain supply further.

What This Means for Your Roadmap

Build for both scenarios: costs decline (optimistic) and costs plateau or rise (base/pessimistic). The practical moves:

• Implement multi-provider fallover architecture — when one provider throttles, route to alternatives transparently
• Design tiered AI features — real-time on premium inference, async on cost-optimized, batch on self-hosted open models
• Track inference economics weekly — when one provider breaks away on unit costs, the entire competitive landscape reshuffles and your platform bet needs to follow within the quarter

The user segmentation Beck identifies matters: casual users accept caps, developers migrate to metered APIs, and the squeezed middle — technical-but-not-API-savvy power users — are pure gold. Build better rationing UX: usage budgeting dashboards, intelligent model routing, graceful degradation rather than hard cliffs.