Pulse

Market thesis / Jul 13, 2026 / 4 min

Goldman Wrote the Menu Washington Fights

On July 11, Goldman Sachs initiated coverage on Zhipu and named DeepSeek and ByteDance as China's top AI model picks — the same week OpenAI, Anthropic, and Google united against Chinese distillation and Beijing weighed locking Qwen and GLM inside China.

Thesis Goldman's July 11 initiation on Hong Kong-listed Zhipu — with a Neutral rating, HK$1,880 target, and $110 billion implied valuation — just told institutional clients to price the Chinese discount stack as investable infrastructure while Washington treats the same labs as national-security threats and IPO-week vendors need Americans paying $30-per-million-token Sol rates.

Goldman Sachs just handed CIOs a shopping list for Chinese frontier models — Zhipu, DeepSeek, ByteDance — three days after Washington's top labs formed a distillation defense pact and one week after Reuters reported Beijing may lock those same weights inside China.

The split: Wall Street is normalizing what geopolitics still treats as contraband.

What Goldman said

  • On July 11, analyst Ronald Keung's 50-page report initiated coverage on Hong Kong-listed Zhipu (Knowledge Atlas) with a Neutral rating and an HK$1,880 price target — roughly 15% above that day's close.
  • Goldman named DeepSeek and ByteDance as its other top Chinese model picks. DeepSeek and ByteDance are private; Zhipu is the only public equity proxy.
  • The bank pegged Zhipu's implied valuation at $110 billion and estimated independent Chinese model companies collectively exceed $200 billion.
  • On foundational text, Goldman ranked Zhipu GLM5.2 and DeepSeek strongest on pricing power and cost. On video, ByteDance's Seedance leads — reportedly above $2 billion ARR with 70% gross margins.
  • Goldman maintained Buy on MiniMax (HK$860 target) and Kuaishou.

Why the math matters

  • Goldman's framework scores pricing power, cost advantage, and financial strength — not patriotism.
  • Chinese high-end models price around $1 per million tokens. U.S. frontier APIs run $4–$8 — Chinese stacks at roughly 10–25% of American list prices.
  • DeepSeek V4-Flash API pricing: $0.14 input / $0.28 output per million tokens. GPT-5.6 Sol: $5 / $30.
  • Goldman forecasts China's AI model API and subscription revenue growing from ¥35 billion in 2026 to ¥879 billion by 2030 — daily token consumption rising from 35 trillion to 460 trillion.
  • MiniMax derives 60–70% of revenue overseas. Goldman calls international expansion — especially outside the U.S. — the biggest upside.

The politics Goldman didn't price

  • In April, OpenAI, Anthropic, and Google began sharing distillation intelligence through the Frontier Model Forum after accusing Chinese labs of systematically copying frontier outputs.
  • Anthropic told Congress that operators linked to Alibaba's Qwen ran 28.8 million fraudulent Claude exchanges across ~25,000 accounts (April 22–June 5).
  • On July 7, Reuters reported Beijing held Commerce Ministry meetings on restricting overseas access to frontier Chinese models — including Qwen, Doubao, and Z.ai's GLM — and treating AI leaks as national-security crimes.
  • Goldman noted Microsoft is considering hosting DeepSeek inside Copilot as a low-cost option — with inference running inside Azure, per the Wall Street Journal.
  • AWS Bedrock and Google's Gemini Enterprise platform already host Chinese open-weight models.

The enterprise shift underneath

  • Goldman describes the buyer mood moving from "token maximization" to ROI-first deployment — the same week Satya Nadella warned enterprises they pay twice for AI (cash plus proprietary know-how).
  • A Jellyfish engineering study Goldman cites found heavy enterprise AI users consume 10× the tokens but deliver only 2× the output.
  • Open-source economics remain messy: Alibaba Cloud can host GLM5.2 without paying Zhipu. Goldman expects a shift toward open-weight plus community licenses with revenue sharing — MiniMax's M-series is the template.

The catch in the coverage

  • Goldman gave Zhipu a Neutral rating despite the bullish price target — signaling valuation risk even as capability closes in.
  • The bank warns disclosed ARR understates real deployment when weights ship free on third-party clouds.
  • DeepSeek announced peak/off-peak pricing for V4 from mid-July — peak rates off-peak, blended around $0.35 (Pro) and $0.12 (Flash) per million tokens.

Convina's view: Goldman just did what establishment banks do when a technology stops being exotic — it wrote the menu. That does not resolve the fight. Washington is building distillation defenses and export guest lists; Beijing is drafting outbound model controls; and U.S. labs need IPO-week multiples on $30-per-million-token APIs while Goldman tells clients the durable trade may be the 90%-cheaper stack hosted on Azure anyway. Buy the discount. Price the sovereignty tax separately.

Research Signals

https://www.cnbc.com/2026/07/12/goldman-sachs-picks-its-favorite-chinese-ai-models.html https://www.techechelon.com/post/goldman-sachs-names-zhipu-deepseek-and-bytedance-as-top-chinese-ai-model-picks https://www.htx.com/news/goldman-sachs-in-depth-report-who-will-be-the-long-term-winn-VDvgPaUf/ https://www.anthropic.com/research/detecting-and-preventing-distillation-attacks https://www.businesstimes.com.sg/international/global/beijing-looking-curbing-overseas-access-chinas-top-ai-models-sources https://www.scriptbyai.com/gpt-gemini-claude-pricing/