IOSF

For more information, please check hyperbolic.ai/blog/a1-fundin…

Come and join us!

🍽️IOSF Founder Table Vol. 1 | From Inference to RSI Great AI conversations don’t always happen on stage. Sometimes, they happen around a table. For our first gathering in San Francisco, researchers from OpenAI, xAI & Google DeepMind joined founders and investors to exchange ideas on inference economics, world models, recursive self-improvement (RSI), and AI agents. A few takeaways: 🔶 Inference: Customer retention and sustainable margins matter as much as growing demand. 🔶 World models: Impressive demos are just the beginning. Real use cases drive adoption. 🔶 RSI: Faster research doesn’t always mean better decisions. 🔶 AI agents: Greater autonomy makes oversight and trust more important. 💬 Interested in joining a future gathering? DM us or leave a comment! 🔗 Learn more about our events: iosf-gatherings.vercel.app #IOSF #FounderTable #AIResearch

🚀 From GPU to Token, join us in San Francisco! Together with @hyperbolic_labs and SourceM3D, we’re bringing the AI compute ecosystem together for an evening of conversations and connections following SGLang Summit! ⚡From data centers and GPU clouds to inference and AI applications, let’s connect the people powering the next generation of AI. Expect guests from SGLang, NVIDIA, GMI Cloud, Apple, Google, Amazon & more! 📅 Nov 12 | 6–9 PM PST 📍 San Francisco, CA 🎟️ RSVP: luma.com/xubn83xh Let’s connect, exchange ideas, and make things happen! 🤝 #AIInfrastructure #GPU #SGLang

China AI Weekly | Oct 28 – Nov 4 Where the capital moved this week: 💰 @deepseek_ai raised $15.9B, China's largest startup round ever. It alone lifted October global VC to $59B (+45% YoY); AI took 66% of all dollars. 🔌 Compute is going domestic. @deepseek_ai reportedly ordered ~160K Huawei Ascend chips (~$2.5B) for V5 (unconfirmed). Nvidia's China AI chip share is down to ~34%. 📱 @Apple Intelligence launches in China in December on @Alibaba_Qwen . $BABA +5%. 🤖 @deepseek_ai previewed a multi-agent "super agent" and shipped Creator Mode — going straight at the coding-assistant market. ⚖️ NDRC is drafting full-lifecycle AI safety standards; a financial regulator warned AI investment is "overly concentrated." 📉 Counter-signal: Chinese AI stocks are sliding, and private benchmarks show a ~12-pt gap to US frontier models. Our read: mega-rounds at the top, compression everywhere else. Distribution now beats leaderboard rank. Follow @IOSFVC for the weekly China AI → capital signal.

For AI agents, price per token is the wrong number. @deepseek_ai V4.1 Flash: $0.003 per million input tokens on a cache hit vs $0.15 fresh (off-peak). That's 50×. Our 60-second video explains why this matters for agents, how DeepSeek got here, and the caveats 👇

China’s AI race is shifting from model size to data control. But who are the people actually shaping that shift? As high-quality public training data becomes harder to scale, the teams responsible for sourcing, training data, and evaluation are becoming increasingly central to AI development. A closer look at 8 major Chinese AI companies reveals the people behind this transition: • @BytedanceTalk : Yinglei Wang oversees AI Data & Safety, while Chenggang Li leads pretraining data within Seed. • @TencentGlobal : Shunyu Yao has taken on leadership across foundation models, AI infrastructure, and AI Data. • @AlibabaGroup : Minggang Wu, Zhiwei Yang, Ze Xu, and Zenan Meng cover training data, expert data, evaluation, and embodied AI data. • @meituan : Qiong Huang is reportedly responsible for training data within its LongCat model team. • @xiaohongshu : Former DeepSeek researcher Guanting Chen leads pretraining data. • @deepseek_ai : Yishi Pu and Yukun Li are key figures in its training data operations. • @moonshot AI: Yangyang Liu is a key contributor to its multimodal data and evaluation research. • @ZhipuAI AI: Xiaohan Zhang has played a central role in building its data annotation infrastructure. The bigger picture: data is no longer just a supporting resource. The people deciding what models learn, how data is produced, and how progress is measured are becoming increasingly influential. As China’s AI industry moves beyond scaling models, these data teams are worth watching. Explore the map below. 📷 Based on reporting by Z Finance. Roles reflect reported information and may change.

China's humanoid boom: the industry will win. Most companies won't. 150+ companies. ~50 mass-production models. Sub-5% margins on entry units. Don't bet on who wins — back the layers that win either way. 🧵

Robots are still in school. Valuations already graduated. 74% of a leading Chinese humanoid maker's shipments went to universities. Funding: RMB 93.5B in H1 2026, ~5x YoY. Unitree: ~61% off its post-IPO high. The industry is in Phase 3. Capital is paying for Phase 6. 📷

China's embodied AI industry is still in its capitalization phase, but capital is pricing it as if robots already earn their keep. Domestic embodied AI funding hit RMB 93.5B in H1 2026, about 5x year on year. Valuations have already peaked. Unitree opened 629% up on its STAR Market debut, then fell about 61% from its high within two months. September humanoid funding dropped about 73% from August. Three points from @0xjacobzhao's new piece in our Physical AI 2026 series: → Don't wait for AI-style earnings pass-through. About five hyperscalers buy compute, and they buy defensively. Thousands of plant managers buy robots, and they always have an alternative: people. → Overheating hurts OEMs directly. Compute, tactile sensing, simulation assets and scenario operations aren't priced off robot ASP. → The first real signal will come from a single vertical where unit economics work and replicate: payback under 18–24 months, plus repeat purchases. For founders, the pricing anchor is moving from scarcity to revenue quality. Lead with buyer-calculated ROI, manual-intervention rates and cross-year repeat orders, not shipment counts. xitter.twitterapi.xyz/IOSFVC/status/21151084…

IOSF is doubling down as @hyperbolic_labs announces its latest round. How we saw it: In 2024 the open question wasn't whether inference would grow. It was whether the newest GPUs would just obsolete everything that came before. Our read was they wouldn't. Inference was going to keep climbing, and older cards would slide down the cost curve into new buyers and new workloads, not get thrown away like last year's phone. The hard part wasn't owning more GPUs. It was about making heterogeneous compute supply unified, liquid, productive, and durable. That's the layer Hyperbolic built. Making messy supply easy to access, manage, and orchestrate. Two years later that call still looks right. And a new demand is showing up on top of it: hedging compute. Long GPU books want price protection, not just capacity. Indices are easy; physical settlement isn't. Hyperbolic is rare in being able to do both: take multi-year supply, break it into short-term product customers actually buy, and sit underneath the hedge as the party that can deliver the hardware. We're proud to continue our partnership with @zjasper and the @hyperbolic_labs team on turning compute into a liquid market.

Today AI has raised over $200M. Is it China's answer to Meta's Muse? The bigger story isn't just funding. It's the shift from AI that waits for prompts to AI that remembers, anticipates, and acts. But can that become lasting value?

GPU rent is the crude. AI tokens are the gasoline. But who hedges the margin in between? For the same AI model, prices can vary up to 4× across providers. GPU futures alone won't solve that risk. So who writes the grade? Hedging Intelligence | Part 2 📷

One AI provider sells 25 billion tokens a day. Yet there's no market to hedge its token revenue. Token prices were expected to keep falling. Now they're moving both ways. If AI tokens become a commodity, who takes the other side? Hedging Intelligence | Part 1 👇

@deepseek_ai just disclosed ~$38 of revenue per GPU-hour on its V4.1-Flash cluster, at off-peak pricing. A top-tier GPU rents for ~$2–4/hr. That’s a 10–19x spread, and it comes from software, not silicon. The disclosure came with DeepInfer, the inference engine behind DeepSeek’s own production API, now open-sourced under Apache-2.0 with no revenue gate. One codebase runs on NVIDIA Blackwell or Huawei Ascend 950, and you switch between them with a single env var. Where the spread comes from: – Speculative decoding with a confidence head checks ~30% fewer draft tokens per step – A prefill technique cuts total prefill compute nearly in half – Built on vLLM and DeepSeek’s own kernels (FlashMLA, DeepGEMM, DeepEP) In inference, utilization is the product. What this means for compute investors: 1/ The GPU-hour is a commodity. Tokens per GPU-hour is the moat. Two operators renting identical hardware can run very different businesses. 2/ The margin sits in the serving layer, and DeepSeek knows it. KV cache management, prefill/decode disaggregation and load balancing were all kept out of the release. Open-sourcing the engine moves the hard part onto whoever deploys it. 3/ Hardware-agnostic serving loosens lock-in at the inference layer. If the same production engine runs on Ascend 950, non-NVIDIA silicon becomes a cost line, not a science project. Caveats: every figure here is DeepSeek-reported, the published benchmarks ran only on NVIDIA, and nobody has tested it independently yet. Outside estimates put inference gross margin around 80–95%, before R&D, networking and people.