Latest / Minds, Bodies, and Terawatts / DeepSeek Unveils V4 Flagship AI Model

Transcript

DeepSeek Unveils V4 Flagship AI Model · Minds, Bodies, and Terawatts

• 0:00
  [Host] [curious] This is Minds, Bodies, and Terawatts, your daily dispatch from the edge of abundance. The Wall Street Journal reported this morning, April 24th, that DeepSeek has released V4, its long-awaited flagship model.
• 0:14
  [Guest] The Hangzhou lab is claiming 92.4 on MMLU-Pro and 88 on SWE-Bench Verified, numbers that put it inside striking distance of GPT-5 and Claude Opus 4.1 on reasoning and code.
• 0:27
  [Host] Remind me why DeepSeek specifically spooks Washington. There are plenty of Chinese labs.
• 0:33
  [Guest] Because V3 and R1 last year trained at a fraction of the cost Western labs quoted. Nvidia shed nearly six hundred billion in market cap in one January 2025 session when R1 landed.
• 0:47
  [Host] And V4?
• 0:48
  [Guest] The technical report, titled "DeepSeek-V4: Scaling Sparse Attention to Frontier Reasoning," puts training compute at roughly 8.2 million H800-hours. That's about 12 million dollars in rented compute. For comparison, GPT-4's training was widely cited north of 80 million.
• 1:05
  [Host] [serious] On restricted chips.
• 1:07
  [Guest] Stockpiled H800s and reportedly Huawei Ascend 910C silicon for the later stages. The architectural move is what matters. V4 uses a new Multi-head Latent Attention variant that cuts KV cache memory by roughly 40 percent versus V3, which is how they fit longer context on weaker hardware.
• 1:27
  [Host] So the export controls slowed them, but didn't stop them.
• 1:32
  [Guest] That's the uncomfortable read. The October 2023 and 2025 Commerce rules assumed compute was the binding constraint. The V4 paper's own framing, echoed in last week's Stanford HAI brief "Algorithmic Efficiency and the Compute Frontier," is that sparsity and distillation close the gap faster than controls can widen it.
• 1:53
  [Host] Who's buying it?
• 1:55
  [Guest] Domestic first. Alibaba Cloud and Tencent are pushing V4 into enterprise stacks today. The weights hit Hugging Face within four hours of the announcement, which guts the whole containment thesis.
• 2:08
  [Host] Pricing?
• 2:08
  [Guest] DeepSeek lists API access at 14 cents per million input tokens, 55 cents output. Anthropic's comparable tier is 15 dollars output. That's not a discount. That's a different business model.
• 2:22
  [Host] Sputnik moment or another Tuesday?
• 2:24
  [Guest] [confident] It's the question Fortune raised when Jensen Huang told Lex Fridman last month he thinks AGI is already here. If frontier capability is converging, the differentiator stops being who has the smartest model. It becomes who deploys it, under what rules, for whose benefit.
• 2:44
  [Host] Which puts Beijing in an interesting position.
• 2:47
  [Guest] V4 isn't resistance to the AI future. It's a different architecture of it, with state-aligned training data and different red lines on output. The system card documents refusal behavior on Tiananmen, Taiwan, and Xi Jinping by name.
• 3:03
  [Host] So what do enterprise buyers actually do Monday morning?
• 3:07
  [Guest] European firms run model-agnostic stacks already, so V4 slots in behind a router. U.S. government buyers are locked out by procurement rules. The real leverage point is the Global South, where price-to-performance beats provenance every time.
• 3:23
  [Host] [sighs] And sovereignty becomes a feature you shop for.
• 3:27
  [Guest] That's the Patrick Deglon framing in Unscarcity. The AI race isn't converging on one winner. It's bifurcating into competing architectures of abundance, each with its own training corpus, its own refusals, its own theory of the citizen.
• 3:43
  [Host] For more on sovereign AI and the compute frontier, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.