Kimi K3 Beats GPT-5.6 in Frontend Code Arena, Free Weights

Where value is actually migrating

The tell isn't that open weights caught up — it's that they did while application multiples kept climbing. Emergent raised a $130M Series C at a $1.5B valuation, a 5x markup in months on a $120M run-rate, the same week a free download topped the coding leaderboard. Those two facts don't share a room for long.

One week produced three releases that collapsed the cost to possess a frontier-class model:

Underneath, GPT-4-class inference fell from $20 to $0.40 per million tokens in 36 months, and the open-vs-closed gap closed from 8.04% to near zero. This is production reality, not a demo: open weights already route roughly a third of OpenRouter tokens.

The middle of the stack is squeezed both ways. From below, open frameworks commoditize the harness — a weekend CrewAI build reproduces most of Claude Code's planning, memory, and sandboxing, model-agnostically. From above, the next model deletes scaffolding (Anthropic's context resets became unnecessary with Opus 4.5). Hosted-inference resellers and thin API wrappers are eaten from both ends.

Where the money actually accrues

Value is barbelling to the two ends the middle can't touch: silicon and edge inference below, proprietary data, workflow lock-in, and distribution above. Between them sits a visibly underfunded wedge — orchestration and agent governance. MCP SDK downloads went from 2M to 97M monthly in 16 months, but 30+ CVEs landed in 8 weeks and only ~21% of firms report mature governance. That's the Snyk/Wiz moment of the agentic era, priced as though nobody's noticed.

The productionization gap is the tell: 79% of developers use open models but only 51% ship them (57% vs 73% for closed at enterprise scale). Whoever turns experimentation into shipped revenue takes the budget.

When frontier intelligence gets portable and cheap, value migrates to the silicon underneath and the proprietary data on top — everything between is renting a moat it doesn't own.

The constraint moved from Washington to the county board

The reversal came from local government, not from Washington. Trump's January 2025 executive order revoked the Biden-era safety guardrails, and David Sacks arrived as AI czar promising, in his words, that "we got to win." Eighteen months later that agenda is running into 300+ local data-center moratoriums and a public that has soured on AI over jobs and electricity prices. Nobody priced that into the deregulation trade.

The capital-allocation lesson here is that federal policy is not the binding constraint anymore. Local siting and power are the binding constraint, and that is the part the deregulation crowd keeps missing. Deregulation at the top means nothing if a county board blocks the substation. Any data-center position still priced on a light-touch federal buildout needs its power and permit status re-checked, and re-checked hard.

Capability itself became a liability in the same window. Anthropic's Mythos was treated as a milestone when it shipped; within roughly a year, banks and intelligence agencies were treating it as a cybersecurity threat instead, and that shift in classification is the actual story, not the model itself. The gating mechanism then generalized: in June 2026, Amazon's Andy Jassy — an investor in Anthropic — flagged model vulnerabilities to officials, which seeded a broader crackdown, and within months OpenAI shipped GPT-5.6 under a government-requested staggered release. Microsoft's security chief is now purging executives to force an AI overhaul. Time-to-market behaves like a regulated variable under the current review regime, and that could reverse if enforcement lapses.

This inverts a core diligence assumption. Wrappers and agent platforms locked to one lab's release schedule now inherit that lab's regulatory delay risk, which means model-agnostic architecture is worth a premium in this environment. Every mandated review regime also creates a demand-inelastic market for model evaluation, red-teaming, and security-audit startups. That market is forming now, and the pricing I'm seeing still assumes the light-touch 2025 story. The falsifier is straightforward: if federal review regimes lapse, or Congress moves to preempt state and local rules, that TAM shrinks fast.

One more repricing follows. Policy oscillation makes multi-year lab roadmaps, and the DCFs built on them, fragile. Or rather, more fragile than most of those DCFs currently assume. Shorten the forecast horizon on frontier-lab exposure and weight near-term revenue over long-dated model bets.

The moat in AI infrastructure is now power and permits, not the federal light touch that was supposed to carry this trade. Secure-by-design has become the growth wedge instead.