the Hundred-Model War Ended in Career Changes, Not Casualties — Facts: Jiazi Guangnian classified every WAIC 2026 exhibitor from official expo records: the LLM/GenAI track held steady at 130 companies year-over-year, but 83 (63.8%) now carry Agent/AI-application as their primary label and only 18 still present as "foundation model" companies (11 in AIGC/video); caveat pinned: this is an expo-label census, not an industry census, and no list of the 18 was published — but precisely because booth signs are self-chosen, it's the most honest census there is; the number trajectory: 79 models above 1B params (May 2023, MOST), 238 (Oct 2023, cited by Robin Li), ~305 (Apr 2024, NBD), 1,509 cumulative published (July 2025, official, incl. industry models) — ever more released, ever fewer willing to call themselves foundation-model companies. Judgment: (1) elimination rarely means bankruptcy — it means career change: 01.AI abandoned its trillion-param training plan, merged most of its pretraining/infra team into Alibaba Cloud and pivoted to enterprise apps; Baichuan stopped general-base training and went all-in on medical (Baichuan-M4); of the "six tigers," only Zhipu, MiniMax, Moonshot and StepFun remain at the general-foundation table — telecom-style consolidation (her June 21 thesis: the model layer is carriers, not platforms) never kills small carriers, it converts their licenses into other businesses; (2) survivors sort into four tiers with entirely different survival logic: state capital (DeepSeek's July 15 registry filing seats the National AI Industry Investment Fund as the only outside shareholder with voting rights and no lockup while commercial investors take 5-year lockups with no votes — the state holds votes, the market holds checks; ~¥50B round at ~$52B post, with round-two chatter at $71B pre already circulating); listing pipelines (Zhipu's HK listing + ¥31.4B HKD placement + STAR-board tutoring, MiniMax's IPO + ¥16B HKD refinancing — 18C refinancings now exceed the IPOs themselves); big-tech internal (Doubao at 180 trillion daily tokens, Hunyuan's share jumping 0→8.7% — internal models needn't earn independently); and vertical cash flow (Kling's ~$3B raise at ~$18B post ≈ 36x its ~$500M annualized revenue); that leaves Moonshot as the only pure-blood independent: Kimi K3 (July 16 — 2.8T params, 1M context, open weights promised by July 27 — the largest open-weights model once released) is a handsome technical flag, but $3/$15 pricing runs 2-3x Zhipu's GLM-5.2 ($1.4/$4.4), ~$300M ARR stands against a $31.5B pre-money ask, and the release landed one day before WAIC — technology, pricing and fundraising narrative are no longer separable: foundation models are a capital industry now, not a startup industry; (3) primary-market read: the venture window for new general foundation models is formally closed — new money has two destinations: later rounds of the surviving foundation players (in essence pre-IPO or consortium tickets, priced on exit paths and state positioning, not technical odds) or the layer above; and the 83 Agent-badged companies are the new hundred-team war — the same shakeout replays within two years (fewer than 14% of the 130 still fly the foundation flag), so the diligence question for an Agent company is no longer "how strong is your model" but "what will your booth sign say in two years." Rollback: the 18 is one firm's expo-label count — trend, not precision; regulator registrations keep rising (988 services by end-June) so consolidation is happening only at the foundation layer while the application layer still inflates; falsifier: if within a year a new player outside the big-tech/five-strong set genuinely enters the foundation table on a novel architecture (MiniCPM's edge route is half a candidate), "window closed" gets downgraded; and K3 is two days old — community verdicts will drift. Watch: next year's WAIC label census; whether the "foundation five" (Alibaba, ByteDance, DeepSeek, StepFun, Zhipu) consolidates further (Kai-Fu Lee bets on three); the two-year survival rate of the 83.