A two-person, $1.8B company is the story every operator missed

Matthew Gallagher spent $20,000, hired his brother, and stood up a GLP-1 telehealth business that reported $401M in 2025 and is pacing to $1.8B in 2026. Net margin 16.2% — roughly triple Hims at 5.4%. ChatGPT, Claude, and Grok write the code. Midjourney and Runway make the ads. ElevenLabs handles voice support. CareValidate and OpenLoop handle the parts that require a medical license. Two people, one company, nine-figure revenue.

Discount the $1.8B projection by half if you want. It's still $400M+ per employee, which is roughly a thousand times what the best SaaS companies achieve.

Yes, but — Medvi picked a uniquely favorable vertical. GLP-1 demand is vertical, the regulated layer is API-accessible, and prescription fulfillment is essentially a distribution problem. About 35% of the US economy sits behind licensing walls that don't componentize this cleanly. The pattern is not universal. It is, however, present in insurance distribution, tax prep, legal services, property management, and most of the mid-market SaaS your portfolio company competes with. If you're reading this thinking "couldn't happen in my category," that's exactly the reflex the incumbents Medvi is eating had six months ago.

The scaffolding shipped this week

Google released Gemma 4 under Apache 2.0. The 31B dense model matches Kimi K2.5 (744B) and GLM-5 (1T) on Arena. The 26B MoE activates 3.8B parameters per forward pass and runs at 162 tok/s on a single RTX 4090. Edge variants fit on phones. Zero MAU limits, full commercial rights. Sebastian Raschka's analysis confirms the architecture barely moved from Gemma 3 — training recipe did the work. Meaning the next Medvi doesn't pay for tokens if it doesn't want to.

One caveat worth naming: llama.cpp has 10-15 open tokenizer bugs producing garbage on Unsloth quants. PR #21343 is in flight. Use vLLM until it merges. Don't ship the demo you built over the weekend.

Apple's Simple Self-Distillation paper is the other underrated release. Sample your model's own outputs, fine-tune on them, no filter, no RL, no verifier. Qwen3-30B went from 42.4% to 55.3% on LiveCodeBench. That's a 12.9-point absolute gain for what amounts to a weekend of fine-tuning. If you have a domain model in production and you haven't tried this, you're leaving free performance on the table.

The anti-pattern eating your inference bill

Wharton tested 198 PhD-level questions across the reasoning frontier. Chain-of-thought buys 2.9-3.1% accuracy for 20-80% more latency. On Gemini Flash 2.5 it's net negative — minus 3.3%. Output length correlates negatively with accuracy at r = −0.544. Longer traces are worse traces, and unfaithful traces are longer and more elaborate than faithful ones. Anthropic's own research shows models hide shortcut usage 61-75% of the time.

All four major reasoning providers now explicitly warn against CoT on reasoning endpoints. If your prompt templates still say "think step by step" for o3, R1, or Claude extended thinking, you are paying 15-30x more tokens per query for equal or worse output. At 10,000 queries a day that's a six-figure annual mistake.

Apple ML's NeurIPS 2025 paper documents an inverted-U: standard models beat reasoning models on easy tasks, reasoning wins on medium, both collapse on hard. Which means the correct architecture is a router, not an endpoint.

The infrastructure underneath is fraying

GitHub is at roughly 90% effective availability — 2.5 hours of daily degradation — because Claude Code traffic grew 6x in three months and the platform's stateful backend was designed for humans, not agents. Microsoft absorbed GitHub into its AI group without a CEO. Copilot fell to third behind Claude Code and Cursor. Three failover incidents in February and March exposed latent config bugs in paths that hadn't been exercised in production.

Simultaneously, Google, Amazon, and Anthropic all throttled usage limits in the same window despite completely different chip supply chains. Kent Beck's read is correct: the binding constraint is investor patience, not silicon. Meta committed $27B to a single 7.5GW gas-powered data center. The "inference gets cheaper forever" assumption is a planning bug, not a fact.

And the offensive side of the ledger got worse. Anthropic disclosed the first documented autonomous AI espionage campaign — a Chinese state group ran Claude Code against 30 targets with AI executing 80-90% of tactical operations via subtask decomposition. A Nature Communications paper documented 97% autonomous jailbreak rates between reasoning models. Google's Sandra Joyce reported attacker dwell time collapsed from 8 hours to 22 seconds. Human-speed incident response is architecturally obsolete.

What to do this week

One exercise, run it Monday. Pick your top three revenue lines. For each, write a one-page memo answering: what does a five-person team with $50K and current AI tooling build against this in six months? Not "could they." What do they build. If the honest answer is "most of it, worse but cheaper," your moat is coordination overhead, and coordination overhead is what Medvi arbitraged to $900M per employee.

Then do two more concrete things. Audit every prompt template targeting a reasoning endpoint and strip the CoT instructions — this is a half-day of work that saves real money and probably improves accuracy. Mirror your top five repositories to a second git remote and stand up self-hosted runners for deployment-critical pipelines, because your deployment availability ceiling is now GitHub's, and GitHub's is 90%.

Everything else — the routing layer, the self-distillation experiment, the Gemma 4 benchmark against your current API spend — is this quarter's work. The memo is this week's.