IOSF @IOSFVC

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?

keyboard_arrow_left Previous keyboard_arrow_right Next

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.

@elonmusk Do we need a pilot license?

JEV: AI That Chooses, Not Chats @typesafeai's Jev doesn’t write answers. It picks an option, scores a decision, or returns a yes/no probability. That makes AI cheap and fast enough to act inside apps without waiting for a prompt.

@typesafeai Full Report： x.com/IOSFVC/status/…

IOSF @IOSFVC

2 weeks ago

x.com/i/article/2110…

3 0 4 3K 2

@typesafeai

x.com/i/article/2110…

@alexandr_wang first law of creative thermodynamics. second law: every remix compiles cheaper than the last.

Zuckerberg was ready to bet $100–200B to Muse before the model existed🤯. And Muse Wonder shipped TODAY. Takeaways from @collision sit-down with Zuck & @alexandr_wang: 1. Muse was trained for personal agents, not coding, the opposite of where the industry was racing. 2. Memory is the bet: #Muse Wonder is built to remember what you told it weeks or months ago. 3.Ignore public benchmarks. @alexandr_wang: chasing them outsources your product thinking to whoever made a viral benchmark. 4. Compute costs ~$50–60B upfront per gigawatt and ~$15B/yr to run. If Muse fails, the compute can be resold or redirected. 5. Talent density over headcount. A day earlier on @davidsenra’s podcast, @alexandr_wang split businesses into two kinds: diamond mines (most bets fail, a few return 1,000,000x) and skyscrapers (you grind out a few % of efficiency at a time). Put the two conversations together and Meta’s strategy is clear: a diamond bet on top of a skyscraper. A small team chases the asymmetric upside, while hard assets cap the downside. Up the AI stack, from compute to agent infra to agent apps, the mix shifts. Know which game you’re in before you pick your investors. #IOSFVC #MuseWonder #Meta

We co-organized the Workshop on Efficient Reasoning at #COLM2026 in San Francisco this Friday 🧠 A full day on making reasoning models faster, cheaper, and deployable at scale, from test-time scaling and RL to agents, world models, VLMs, and KV cache efficiency. Grateful to our speakers: Yuandong Tian (ex-Meta), Niklas Muennighoff (Stanford), Jiajun Wu (Stanford), Kai-Wei Chang (UCLA), Sewon Min (UC Berkeley), Junjie Hu (Wisconsin) and many more! Thanks to everyone who joined in person and online 🙌 🔗 wdlctc.github.io/efficient-reas…

keyboard_arrow_left Previous keyboard_arrow_right Next

x.com/i/article/2108…

What a night. Thanks to @glazecl for hosting a dinner with some of the sharpest AI builders and researchers we know! At the table: · Head of GTM at @fal · Engineering Manager at @nvidia · First author of the Seedance prototype · @McKinsey AI · @Meta AI scientist working on ads · One of @togethercompute 's first 20 hires, and Meta Llama 4 agent training lead · Founder of @novita_labs, a top-3 provider on @OpenRouter · UCSB full professor specializing in LLM training, @amazon AI Lab Scholar · Pioneer in GPU circuit simulation The people training models, running inference and building the silicon, all at one table. More gatherings soon!

x.com/i/article/2102…

x.com/i/article/2100…