The Board Room

The Medvi Signal

Medvi, a telehealth company selling GLP-1 weight loss drugs, hit $401M in year one and is tracking $1.8B in year two with two employees — the founder and his brother — on $20K in starting capital. The AI tool stack: ChatGPT, Claude, and Grok for code; Midjourney and Runway for ad creative; ElevenLabs and custom agents for customer service. Regulated functions (doctors, pharmacy) are outsourced to CareValidate and OpenLoop. Net margin: 16.2%, triple competitor Hims. Replit's CEO independently confirmed the one-person billion-dollar company has been achieved.

This isn't a SaaS tool with theoretical multiples — it's a healthcare company moving product at scale with virtually no human overhead.

Why This Is Replicable, Not Anomalous

The Medvi playbook has three transferable components. First, the healthcare value chain had componentized itself — doctor networks, pharmacy fulfillment, and shipping are available as services. Second, AI handles the high-volume customer-facing operations (support, advertising, code) at near-zero marginal cost. Third, the founder exploited the gap between market demand (GLP-1 drugs) and regulatory enforcement speed. This pattern repeats in any industry where regulated or specialized functions can be accessed via APIs while AI handles everything else.

Seven independent sources converged on the same conclusion this cycle. Marc Andreessen's Latent Space appearance framed it as the $15B a16z thesis: "founder + AI superpowers" will displace professionally managed companies. a16z data shows enterprise AI budget reallocation went from 12% to 60%+ in twelve months, creating the demand environment. Google's Gemma 4 under Apache 2.0 means the next Medvi founder won't even pay for API calls if willing to self-host. Nvidia's latest MLPerf results show software-only optimizations doubled AI throughput on existing hardware.

The Sources Agree on the Threat — But Disagree on Timing

There is healthy skepticism. The $1.8B figure carries a 0.75 confidence score from source analysis — it may be $800M-$1B. But the structural signal is identical even at the low end: $400M+ per employee is unprecedented by 1000x over the most efficient SaaS companies. Multiple sources note that Medvi's model works best in high-margin, digitally-deliverable products in exploding markets — a condition that limits but does not eliminate the transferability.

What Makes You Defensible — And What Doesn't

Strip away the telehealth context and the transferable insight is an uncomfortable audit. Functions that are defensible: proprietary data assets, genuine network effects, regulatory moats built over years, deep customer relationships. Functions that are not: operational complexity, institutional knowledge embedded in process, large engineering teams translating requirements into code. As one source put it: "If your competitive advantage is primarily coordination capacity rather than innovation capacity, Andreessen is betting against your organizational model."

Intuit provides the counter-model: its AI agents hit 85% repeat usage by keeping humans involved — the most important enterprise adoption data point this cycle. The lesson isn't to eliminate people. It's to ruthlessly identify which coordination functions AI can absorb and begin the transition before a competitor demonstrates it can be done at 1/100th your cost.

The Commoditization Event

Google released Gemma 4 under Apache 2.0 — the most permissive open license — with zero MAU limits, zero usage restrictions, and four size variants from Raspberry Pi to leaderboard-competitive 31B. The dense 31B model matches Kimi K2.5 (744B total) and GLM-5 (1T total) on benchmarks despite being 20-30x smaller. LM Arena puts it on the Pareto frontier at ELO 1441. Simultaneously, Alibaba's Qwen3.6-Plus matches Claude Opus 4.5 on SWE-bench coding benchmarks with 1M-token context. Nine independent sources converged on the same conclusion: the model layer has crossed the commodity threshold.

The moat is no longer 'which model you use' — it's how effectively you orchestrate, integrate, and deploy agents at scale.

Google's Android Playbook for AI

This is not generosity — it's strategy. By making the model layer free and excellent, Google accomplishes three things: (1) undercuts OpenAI's per-token revenue by making equivalent capability free at inference compute cost; (2) drives developer adoption toward the Google ecosystem (GCP, TPUs); (3) potentially secures what could be the most valuable consumer AI distribution deal ever — powering Apple's 'New Siri' with Gemma 4 edge models. The E2B model runs on devices with 5GB RAM. The 26B MoE variant activates just 3.8B parameters per forward pass. Every proprietary AI company's pricing power took a structural haircut this week.

Where Value Is Migrating

The breakout signal is Hermes Agent — Nous Research's open-source agent harness — which multiple developers are publicly migrating to from OpenClaw, citing better stability on long tasks. The architecture includes pluggable memory (7+ backends), credential rotation pools, and autonomous skill creation. LangChain shipped Claude Code → LangSmith tracing in the same cycle. Harrison Chase declared memory can't remain behind proprietary APIs. The pattern is clear: value is migrating from model layer to harness/orchestration layer at inflection pace.

The model-harness training loop — where teams capture traces from agent runs, fine-tune open models on those traces, and create compounding improvement — is the new playbook. Axolotl's release claiming 15x faster and 40x less memory for MoE+LoRA training with immediate Gemma 4 support makes this practically implementable today.

The Inference Economics Catch

Sources diverge on cost trajectory. Google, Amazon, and Anthropic simultaneously throttled AI usage limits despite different supply chains — Kent Beck's analysis proves the constraint is investor narrative, not compute scarcity. One source warns inference costs may plateau or rise as labs stop subsidizing. Meanwhile, reasoning models require 15-30x more tokens per query, creating 30-70x cost overruns when queries are misrouted. Apple ML research shows reasoning models actually perform worse on low-complexity tasks. The 'AI gets cheaper forever' assumption is breaking — intelligent routing between model tiers is now a six-figure infrastructure decision.

The Platform Crisis

GitHub has degraded to approximately 90% availability — roughly 2.5 hours of daily degradation. Three major incidents in February-March 2026 reveal systemic architectural failures: database saturation from AI agent traffic on Feb 9, a failover triggering incorrect security policies on Feb 2, and a failover-induced Redis failure on Mar 5. The root cause: Claude Code traffic alone grew 6x in three months, and GitHub's stateful infrastructure was designed for human-scale interaction patterns that AI agents have overwhelmed.

We are entering an era where the primary consumers of developer infrastructure are not humans but AI agents — and the entire toolchain must be rearchitected for that reality.

The Governance Vacuum Above It

Microsoft absorbed GitHub into its AI group, eliminated the CEO position after Thomas Dohmke's departure, and left internal factions (Azure, Microsoft AI, legacy GitHub) competing for control. GitHub Copilot fell from undisputed market leader to third place behind Claude Code and Cursor — the clearest signal yet that integrated AI strategies lose to best-of-breed in fast-moving markets. Mitchell Hashimoto's recommendation: shut down Copilot, acquire Pierre Computer, cut 50% of product lines, reorient entirely around agentic code lifecycles.

The Broader Infrastructure Strain

GitHub is not an isolated case — it's a leading indicator. Google, Amazon, and Anthropic all throttled usage simultaneously despite fundamentally different supply chains, confirming the constraint is financial sustainability, not engineering capacity. Meta committed $27B to a single data center (Hyperion) requiring 7.5 GW of gas-powered electricity — more than South Dakota consumes — adding 12.4M metric tons of CO₂ annually, a 50% increase over Meta's entire 2024 footprint. Google abandoned its own climate commitments for a $30B gas-powered AI data center. The infrastructure trilemma has crystallized: fast, cheap, or clean — pick two.

What This Means for Your Platform

If you operate any API, SaaS, or infrastructure product, GitHub's February 9 database saturation incident is your future if you don't invest in horizontally scalable stateful infrastructure now. A startup called Pierre Computer (Code.storage) claims 65x GitHub's repo creation throughput for agent workloads and reported 9 million repos created in 30 days from AI agents. Whether Pierre specifically succeeds matters less than whether the market validates that agent-scale infrastructure is a distinct category.

The human bottleneck reinforces the infrastructure story. Simon Willison — one of the most productive developers in the ecosystem — reported that orchestrating four parallel coding agents is "mentally exhausting by mid-morning." Community consensus settled at 2-4 parallel sessions as the cognitive ceiling. The productivity curve isn't 'more agents = more output'; it's 'better orchestration tooling = more effective agent supervision.' Invest in observability and session management, not just raw compute.