Free weights are not free inference. Open source drops both a capability anchor and a price anchor: builders can self-host or buy cheaper third-party APIs; closed flagships that want premium must stay a generation ahead. Wall Street Journal reporting, relayed via Zhidx/Tencent News, is concrete: firms shift routine work to cheaper DeepSeek, Zhipu GLM and peers, cutting some scenario costs by about 95%; DeepSeek’s share on Vercel jumped from under 1% to about 17% in a month; among OpenRouter’s high-spend clients, open-source token growth ran about 4× closed models, with 500+ organizations switching from proprietary stacks.

The price war is on. The real question: whose P&L takes the margin hit first—model labs, clouds, or the application layer?

After the price anchor moves

Token list prices between some closed flagships and DeepSeek-class tiers can diverge by tens of times (parts of the coverage cite 50×+). On simple tasks, cheap models can cost a few percent of closed alternatives. Capability-wise, top closed models are still often cast as 4–6 months ahead—yet enterprises do not wait for parity; they route: hard problems to expensive models, bulk to cheap ones. Systems get thrifty and upgrade only when needed.

DeepSeek’s “third path” clarifies the commercial shape: open weights for trust and ecosystem, metered official API for cash; domestic Tongyi, Ernie, Doubao and peers saw steep API cuts over roughly a year (narratives often cite >50% collective drops), with open source blamed for pulling the market’s price expectation down. Open source does not kill APIs; it kills default closed-source premium.

Whose margins are brittle

Model layer (OpenAI / Anthropic, etc.)
Huge train/infer capex covered by high ASP and volume. Forced cuts or diversion to open weights hit revenue first; defending the gap with more compute widens losses and delays profitability—awkward next to IPO stories. Calling cost a “huge issue” is mood and structure. Contribution profit feels it first.

Official open APIs and inference clouds (DeepSeek, SiliconFlow, cloud model services)
Open weights bring traffic; traffic brings GPU bills. Ultra-low ASP attracts low-intent load that crowds peaks; official APIs also face rivals running the same open weights cheaper. Margin war becomes engineering war: inference stacks, scheduling, speculative decoding—usually not in the weight release. Whoever keeps unit token cost under price survives; survivors need not be fat, only less thin than exiters.

Traditional cloud IaaS / model platforms
The twist: model APIs race down to win developers while GPU/AI compute may reprice up on scarcity and demand—“cheaper tokens, dearer jobs” can coexist. Blended cloud margin need not fall one-for-one with API list prices. It may be a mix shift: hosted-model gross gets compressed while bare metal/AI infra sells better; or customers leave PaaS for self-hosted open weights. What breaks is the assumption that hosted models are default high-margin—not necessarily the whole cloud’s blended rate.

Application layer
Often the beneficiary: cheaper inference widens what can ship. Unless an app is glued to one expensive model without routing, margin pressure sits above them.

What the price war does not punch through

It does not erase capability gaps on hard tasks, compliance and data residency, enterprise premiums for stability, or switching costs when models sit inside office suites and cloud accounts. Buyers increasingly score price per completed task, not per token—fewer tokens on a closed model can still win. Clouds that bundle calls with storage, security, and observability can hide margin in the bundle, not on the token sticker.

Builders: default to routing and swappability, not single-vendor faith. Investors: for labs, watch ASP and cost per unit compute, not calls alone; for clouds, watch how much AI revenue is share-buying at a loss. Buyers: free weights bait; the invoice is inference plus engineering days.

Free open-weight shock punches premium narratives first, then API sellers whose unit economics cannot hold. Clouds that only match “we also have a model API” go thin; clouds that fight on compute, network, and enterprise workloads may keep the ledger intact. What gets punched through is usually the page that still prices tokens like gold ore.

Sources

• Zhidx / Tencent News on WSJ: enterprises shift to cheaper open models
• 36Kr: DeepSeek open-weight + API path and domestic API cuts
• 36Kr: compute inflation and cloud repricing