DeepSeek went Huawei. Bezos went Slack.

Capital and architecture diverged this week. China decoupled inference from US silicon. $11.1B flowed to labs explicitly betting against the LLM paradigm Sequoia and Nvidia helped fund. And Project Prometheus, reportedly Bezos-backed, closed $10B to buy what nobody scraped. The market does not have a thesis right now — it has positions across four paradigms, hedging because nobody knows what wins. Three things shifted this week that change how to read the next 90 days. Start with the chip story.

DeepSeek launched V4 on April 24. The model itself is competitive but unremarkable. The deployment story is the news.

V4 is the first Chinese frontier model officially optimised for Huawei's Ascend hardware — for inference. Within 72 hours, ByteDance, Tencent, and Alibaba reportedly placed emergency orders for Ascend 950 chips. Supply constraints are now expected through H2 2026.

The non-obvious detail: Huawei optimisation covers inference, not training. China has cracked the deployment layer of domestic hardware, not the training layer. The distinction matters — the inference market is roughly 4× the training market today, and IDC projects inference demand surpasses training globally by 2030.

Read the order book. Three of China's largest tech companies are locking in domestic inference capacity right now, ahead of the May 5 expiry of DeepSeek's 75% developer discount. They are not waiting for V5. They are not testing. They are placing orders the size of full-year capex at six-month notice.

What this means operationally: the cost of running Chinese AI applications on US silicon was already a strategic vulnerability across the Strait. After this week, it is also a competitive disadvantage. Workloads served on Ascend hardware in mainland China will run cheaper, faster, and with no export-control overhang. The closer your product is to inference-heavy use cases — search, recommendation, generative content — the more material this gap becomes.

The smart operator move: if you ship to a Chinese user base, audit hardware exposure inside 30 days. If you sell US-silicon inference into APAC, your addressable market just got smaller, possibly permanently. If you hold the supply chain (NVDA, AMAT, ASML), the question is no longer whether Chinese inference revenue stays onshore — it is how fast the same dynamic plays out for training.

One further question: what does Western policy do? Export controls remain locked on the training stack. Inference sits outside the current regime. By H2 2026 most pundits expect a controls-broadening attempt; until then, Chinese inference compute is uncontrolled. Operators planning hardware roadmaps through 2027 should price in a 30–50% probability of expanded inference controls and a 50%+ probability that mainland Chinese workloads stay onshore regardless.

China spent five years convincing itself it had a chip problem. This week it convinced its three biggest tech firms it has a chip answer.

NVDA closed Tuesday at $896, down 3.4% on the week as the China inference signal landed. AMD traded sideways — its data-centre exposure to mainland China is structurally lower. Huawei is unlisted, but the SSE 50 semis sub-index ran 7.1% on the week, with SMIC (981.HK) up 11.2%.

Macro overlay: the dollar weakened against CNH on Wednesday following the State Council's confirmation of expanded Made-in-China-2025 tax credits for fab-scale equipment. Treasury 10-year yields drifted to 4.18% on softening April PCE. Risk-on tape, but AI hardware leadership rotated decisively from US to APAC names mid-week.

One spread to watch: the NVDA premium over SMIC compressed roughly 200bps on the week. If that compresses further into Friday's NVDA Q1 print, the buy side is already pricing the inference shift. If it widens, the print is a buyable dip. Consensus EPS sits at $0.67 — the question is whether mainland China revenue guidance acknowledges what just happened or punts to next quarter.

We think the divergence is the story. Capital is flowing simultaneously into the LLM consensus (Anthropic, OpenAI), explicit anti-LLM bets (Ineffable), data-moat plays (Prometheus), and architectural-bridge bets (MIT–IBM). That is not a market with a thesis. It is a market hedging because nobody knows what wins.

Forwarded this? Subscribe to The Model → Free, every Wednesday.

Read the signal. Skip the noise.