DeepSeek Secures 16,000 Huawei Chips for AI Scaling and Projects $1 Billion in Annual API Revenue
In a leaked investor presentation, DeepSeek CEO Liang Wenfeng said the company has secured two Atlas supercomputing clusters containing 16,000 of Huawei’s 950DT chips for future model training, aiming to reduce reliance on Nvidia hardware. He added that an inference margin of roughly 85% would generate enough revenue to fund research, with $1 billion in annual API sales sufficient to turn the company cash-flow positive.
Wenfeng outlined a development path focused on continuous learning rather than simply scaling parameter counts, noting the company held about 20,000 Nvidia Hopper-equivalent accelerators as of May. He cautioned that limited chip manufacturing keeps Chinese firms roughly 12 to 18 months behind U.S. competitors, though he argued Nvidia’s CUDA software ecosystem is losing its dominance as new programming tools streamline hardware migration. The comments, reported by Yicai, underscore the company’s commitment to open-source model releases despite navigating domestic supply constraints.
From the sources (23 posts)
@poezhao0605DeepSeek closed a round of more than $7bn. Barely a month later, it began talks for a new round at a $71bn pre-money valuation. Its management has signaled that frontier research takes precedence over near-term revenue. My latest ⬇️
@teortaxestexFull transcript of Wenfeng's investor conference call. He's a lot more forceful here. He is insistent that open sourcing is *the* agenda. If you're not on board, go buy more Zhipu. And crucially, he says restraint is *the only way to surviv
@teortaxestexDeepSeek as of early June had 20K "H-equivalent" units (H100). Wenfeng intends to spend everything in 6 months, "basically all NVIDIA". «If we could convert all the money into GPUs, we'd convert every cent without hesitation—and we're willi
@teortaxestexLOTS OF ALPHA FROM WENFENG HERE He expects he won't be able to spend >20B RMB on compute in 2026 «The smartest people—maybe less than 50% stay in China» «The largest current model activates ≈800B parameters» «I'd need about 50K GB300s, o
@teortaxestexOn domestic compute: bullish within a year. «Domestic AI chips have no problems in hardware or ecosystem—the only problem is insufficient production capacity» «Previously, domestic GPU adaptation had a problem called poor ecosystem… The mo
@teortaxestexOn Huawei: «we participate deeply in Huawei's ecosystem» «Huawei gives us capacity for about 16,000 GPUs; internet giants might get over a 100K… but this may already be all the capacity Huawei has.» «So we can't count on training the next b
@teortaxestex950s vs Nvidia: «when V3 trained, it still used NVIDIA GPUs, but no longer used NVIDIA's ecosystem… As long as I redo this whole process on Huawei GPUs, it's done. I think this might be a historic mission» «Huawei 950 supernode can fully su
@teortaxestexLiang Wenfeng believes that the comprehensive gap in AI between China and the US is 12-18 months, just like Kai-Fu Lee says and Dario hopes; and can be shrunk to 3, with surpassing in some few key areas. @scaling01 @zephyr_z9
@teortaxestexThis is extraordinary. This is what I expected of China that gets serious about compute independence. A 14nm chip that has Hopper+ level utility. Yes it runs HOT, but it gets the job done. 6.4TB/s now. 20TB/s – Rubin level – in 2027. Electr
@teortaxestexRT @teortaxesTex: Liang Wenfeng believes that the comprehensive gap in AI between China and the US is 12-18 months, just like Kai-Fu Lee sa…
@poezhao0605The detail worth sitting with: China's flagship domestic GPU company gets its best inference numbers by pooling its own chips with foreign ones. Import substitution, built on top of the imports.
@teortaxestex«So we simply won't consider competing with the US at that scale now… when we have more resources later, we'll push to 150B, 156B, or 250B activation scale.» «You could force-train a model that big, but you couldn't do sufficient research»
@firstsquawkDEEPSEEK CEO WENFENG: ADVANCEMENTS ‘SUCH AS TILELANG ARE EXPECTED TO RAPIDLY LOWER THE BARRIERS TO ENTRY CREATED BY THE CUDA ECOSYSTEM’
@reutersFounder says DeepSeek prioritises AGI over profit, likely to keep top models open-source, Yicai reports
@reutersFounder says DeepSeek prioritises AGI over profit, likely to keep top models open-source, Yicai reports
@rnaudbertrandRT @poezhao0605: Notes from a DeepSeek investor meeting leaked today (reported by Sina Tech). Liang Wenfeng rarely speaks in public. Three…
@kakashiii111That’s very true, and it’s why Jensen is deeply concerned about losing clients in Southeast Asia, and that's one of the reason he gives them as much allocation as they want, flooding them with supply, so they won’t even consider Huawei unle
@poezhao0605One line from the transcript: “When Silicon Valley says scaling has hit a ceiling, that is for Silicon Valley. We in China are nowhere near that point.” Liang believes in scaling. He wants more compute.
@techmemeIn a leaked four-hour investor talk, DeepSeek's Liang Wenfeng says the main US-China gap is compute access, Nvidia's CUDA moat is disintegrating, and more (@gaoyingshi / Inside China) (Visit Techmeme dot com for the link and full context!)
@zephyr_z9RT @teortaxesTex: Full transcript of Wenfeng's investor conference call. He's a lot more forceful here. He is insistent that open sourcing…
@deredleritt3rKey quotes from the Liang Wenfeng investor presentation: 1. He's convinced that current AI models can't reach AGI and we need continual learning: "Just like CoT—after CoT reached its ceiling, it already surpassed the most top-tier humans
@zephyr_z9Production volume for 950DT is around 400k this year
@zephyr_z9Some interesting details from the leaked Liang Wenfeng investor call: - Inference margins are around 85% (sixfold profit) - They only had around 20k Hopper equivalents till May - $1B in API revenue is enough to turn the company cash flow po