Weights, Watts, and Vibe-Engineering

The headline this week sounds like a trade-war press release: China has publicly released models that claim to tango with the best of the West. But the real story isn't a capability sprint so much as a reallocation of where value accrues — from guarded weights to the bricks-and-mortar (and kilowatt-hour) economics of serving intelligence.

What matters now is not only which lab trained the fanciest transformer but who owns the inference line: the data centers, the cheap electricity, the integrated product that folds a model into a customer's workflow. That’s the frame Ben Thompson offers in Stratechery — a sober reminder that the industry’s headline metrics (parameter counts, leaderboard ranks) are flashy, but the business lives in marginal costs and token economics. See: Ben Thompson’s essay.

If you work on LLM integration, this is practical: your procurement decision will soon resemble buying power and latency, not virtue-signaling model cards. If you’re an investor, the question is which firms turn model advantage into durable product and infrastructure moats — and which are a glorified weights repo waiting for a cheaper fork.

Moonshot’s Kimi K3 (claimed 2.8 trillion parameters) and Alibaba’s Qwen 3.8 burst onto the scene promising broad access to frontier capability rather than a gated API. That matters because openness accelerates experimentation — and copying. As The Verge notes, Moonshot and Alibaba are weaponizing public weights as a strategic advantage: openness becomes a multiplier for developer adoption and geopolitical influence. See: The Verge.

Ben Thompson reframes this as an economics story: training is a fixed, headline cost; serving is the recurring, revenue-tied business. Critically, he argues that while “tokens are not a commodity,” the intelligence you extract per token is drifting toward commodity economics for many tasks. The practical corollary: price/performance in inference — token efficiency, memory, and latency — will increasingly determine who profits. Read: Stratechery.

A few concrete numbers to stop the rhetorical breath-holding:

• Kimi K3 pricing (reported): ~$3 / 1M input tokens and $15 / 1M output tokens; Sol sits higher at $5 / $30 per million respectively. (Stratechery)
• Open models are proliferating: GLM 5.2, Kimi, Qwen 3.8 — each reduces the monopoly rent available to frontier labs. (EmergingTrajectories)

Open weights also lower the barrier to distillation — using a closed Western model to teach an open Chinese model — which Thompson highlights as a recurring structural edge for China. That’s not conspiracy theory; it’s simple leverage: if you can use a teacher model to compress capability into a cheaper-to-serve student, your COGS collapses faster than your rival can whisper “safety review.” See: Stratechery.

The fight is less about logits and more about electrons. EmergingTrajectories lays out the arithmetic: inference costs (compute + power + hosting) dominate margins, so firms that own more of the stack — data centers, bespoke chips, real estate, even generation — convert scale into asymmetric economics. If you own the watt, you make money when the world asks for another inference call. See: EmergingTrajectories.

This is why Anthropic’s position looks exposed. Ethics-first branding and high per-task cost (Fable 5’s pricing, per analysts, is materially higher than rivals) are defensible if customers value safety enough to pay. But when open-weight competitors drive the price-performance frontier downward, those margins compress unless Anthropic turns Fable into a sticky product rather than a priced-per-token commodity. The strategic choices are blunt but clear:

• Build your own fixed-cost infrastructure (data centers, chips) and make margins via utilization; or
• Productize the model into differentiated, high-margin services (agents, enterprise integrations); or
• Lean on regulation and perceived safety as a de facto moat.

None are trivial. Owning chips and power is capital-heavy; productizing requires product discipline many model-first labs lack; regulation is a two-edged sword that can entrench incumbents — or legislate everyone to parity. Read the economics sketchbook: EmergingTrajectories and Stratechery.

If you look where money flows, you see sentiment catching up with product pushes. Business Insider reports a resurgence of interest in OpenAI secondary shares after the company previewed new models and enterprise tools — GPT-5.6 Sol, Terra, Luna, and the Codex agent are helping narratives shift from ‘misses’ to ‘rebound.’ According to the piece, OpenAI’s implied valuation is sitting around $933 billion, while Anthropic’s has been nearer $1.2 trillion. See: Business Insider.

Investors are voting with an important, noisy instrument: the secondary market. Dave McClure notes a recent uptick in buyers, and Sam Altman’s public messaging — admitting past failures, promising better quarters — plays well to momentum-trading buyers. As Altman wrote on X:

“we did not have our best last 12 months ever... but we are about to have our best 12 months to date.” (Business Insider)

But markets are not destiny. The open-weight push from China and the cost arbitrage of distillation are tail risks to the valuation story: if commoditization of intelligence accelerates, the premium for being an exclusively frontier, high-cost provider shrinks. Secondary-market gyrations are useful signals, but they’re noisy — the durable value will be in product, infrastructure ownership, and real enterprise adoption patterns.

Josh Bleecher Snyder’s essay, "Claude Is Not a Compiler," pushes a corrective to the ‘models-as-accelerators-of-pure-code’ story. Claude isn’t merely better at emitting Python; it operates vertically — strategy, architecture, design tradeoffs, and code — enabling what Snyder calls vibe-engineering: converging on durable decisions across layers without the usual handoffs. The result is not perfect automation, but a faster, more coherent engineering process. Read: exe.dev.

This matters because it shows where real economic value can hide when weights are cheap: in the glue that turns a model into a reliable product. If every lab can produce a 2–3x cheaper token, the winner will be the team that turns that token into predictable outcomes at scale — multi-agent workflows, production specs, tests, monitoring, operational playbooks. Snyder’s DNS case study is instructive: the model didn’t just write code, it helped design the architecture, tests, and deployment checks that actually reduced outages to zero. That’s product work, not model alchemy.

So reconcile two trends:

• Commodity pressure on model inference and the democratizing effects of open weights (China’s Kimi, Qwen, GLM 5.2); and
• The rising importance of verticalized product processes and agentic systems that wrap models into durable customer value (vibe-engineering, enterprise agents).

If you’re building an LLM company, your checklist is simple and ugly: own latency and costs where possible, ship product features that reduce configuration and cognitive load, and document decision-making so customers can audit and repeat outcomes. The edge is no longer just the best perplexity score — it’s the repeatable engineering loop that turns model output into lowered incidence of human toil.