Sushi 🍣 + cocktails 🍸 + good company. Join 500 Global for Investor Sushi Night: Japan during #SFTechWeek. An intimate evening for venture investors to connect with peers, joined by Japanese founders participating in the J-StarX Silicon Valley Extended Program, who are spending the Fall in San Francisco and getting to know the local tech ecosystem. 🇯🇵✨ No pitches, no structured intros, and no fundraising agenda. Just an opportunity to meet interesting people, and enjoy great food and conversation. 🗓 Thu, Oct 8 · 6 - 8pm PDT 📍 San Francisco 🎟 Free · Approval required · Request to attend: be.500.co/investorsushinight… @JETRO_jgc @JETRO_info

🔥

The slope of progress changes when the right people end up in the same room. 500 Fellowship Batch 37 is open. 4 months in SF, Nov 30 → Apr 9. Founders can receive up to $50K with potential for up to $1M. Apply by Oct 2: be.500.co/flagship37

🔥

For @trans_celestial, AI infrastructure doesn't stop at the data center. As AI compute and data centers move into orbit, Co-founder & CEO @jharohit believes we can no longer rely on communications technology developed more than 100 years ago to move ever-growing amounts of data. His answer: take the light inside fiber optic cables and run it wirelessly. Transcelestial's laser communications systems use AI to precisely point and track lasers across ultra-long distances, creating high-bandwidth links between satellites and back to Earth. The bigger vision is to build the connectivity backbone for the space economy, supporting satellite networks, orbital data centers, and eventually AI and robotics deeper into space. Great to see Rohit share this vision on NYSE Live. Full conversation below 👇🏻 @vishalharnal @khailee @AriefJohan @atxfrench

🚀🚀🚀 Full story on @GilmourSpace: capitalbrief.com/article/gil…

Most AI models are trained to be assistants. @humansand is exploring a different category: AI designed to model people. The team introduced Persimmon, a large-scale user model designed to simulate how people behave in multi-turn, multi-user conversations, including how differently people communicate, share information, and change over time. Starting from NVIDIA’s 550B-parameter Nemotron 3 Ultra, Persimmon was mid- and post-trained on a diverse collection of public conversations between people, using thousands of Blackwell-generation GPUs on the team’s recently built cluster. In testing, Persimmon’s conversations were significantly harder to distinguish from human conversations than those generated by frontier assistant models prompted to simulate people. It’s an interesting direction for AI that can better understand differences between people and explore how they may respond as situations unfold. Congratulations to co-founders @gharik, @ericzelikman, @YuchenHe07, @noahdgoodman, and the humans& team! More on Persimmon, including limited research preview access: persimmon.humansand.ai/blog/… @christine_tsai @TonyW @vishalharnal @atxfrench @leoclee

For founders without large content teams, video creation can still be a heavy lift. @chin_jlyc has created @videoclaw , a video creation agent that's simple enough, yet powerful enough Start with an idea, reference link, or raw footage, then use the agent to script, generate, and refine launch videos, demos, ads, and more. Congrats on the launch, @chin_jlyc 🚀

Access to reliable compute is becoming a key differentiator for AI companies running production workloads. Singapore-founded Aolani is partnering with FriendliAI to supply the GPU cloud infrastructure supporting its growing inference services globally and increasingly across Asia. Purpose-built for AI workloads, Aolani combines GPU infrastructure with orchestration, automation, and lifecycle management to help AI-native companies deploy and manage compute more efficiently. For FriendliAI, that infrastructure will support growing customer demand for fast, reliable inference across open-weight and custom AI models. Excited to see Aolani becoming an infrastructure partner to AI companies building for global demand. Full story here: finance.yahoo.com/technology… @vishalharnal @khailee @ariefjohan @atxfrench

Congrats @fin_ai on now being officially part of the @salesforce Ohana 🌺 🙌

Algolia is now available as a native integration in the Vercel Marketplace, putting AI-powered search directly where applications are actually being built. Developers can add @algolia without creating another account, managing another set of credentials, or introducing a separate billing relationship. And with Agent Studio supporting the Vercel AI SDK, that extends to conversational search that can interpret intent and determine the appropriate indices, filters, and search parameters. Algolia already handles more than 1.75 trillion queries annually across 18,000+ businesses. Integrations like this become even more powerful as sophisticated AI capabilities become building blocks developers can readily incorporate into their applications. Full story here: algolia.com/about/news/algol… @christine_tsai @TonyW @vishalharnal @atxfrench @leoclee

Sakana AI’s new Fugu Max and Fugu Ultra v2 models are their strongest evidence yet for the power of orchestration. Fugu coordinates a diverse pool of models, dynamically selecting the right combination for each task. Fugu Max delivers performance within striking distance of elite models, with output pricing 40–60% lower than several competing models, while Fugu Ultra v2 achieved the best or joint-best result on 5 of 8 challenging benchmarks. Notably, Ultra v2 achieved those results without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool. We’re excited to see @SakanaAILabs keep pushing orchestration forward, and demonstrating how smarter coordination between models can open up new paths to both capability and efficiency. More details here: sakana.ai/fugu-max-release/ @christine_tsai @TonyW @vishalharnal @atxfrench @leoclee

Talkdesk and Microsoft are expanding their partnership, allowing enterprises to put existing Azure commitments toward Talkdesk CXA through the Azure Marketplace. What stands out is how @Talkdesk is approaching automation. CXA is designed to move beyond bots that hand work back to humans, coordinating specialized AI agents to access enterprise data and complete workflows across existing business systems. Just as importantly, companies don’t need to rebuild their contact center around AI. CXA works across cloud, hybrid, and on-premises environments, bringing more capable automation into the technology environments enterprises already operate. Full story here: talkdesk.com/news-and-press/… @christine_tsai @TonyW @vishalharnal @atxfrench @leoclee

🚀

Persimmon is a genuinely different idea: a model of how people actually talk, not another assistant. Proud to support @humansand on this launch with DeepCluster, a dedicated NVIDIA Blackwell cluster we deploy and operate. Excited to see where it goes. deepinfra.com/deepcluster

It's official: DroneDeploy is now part of @procoretech. Today we closed our acquisition by Procore, joining a team building the operating system for construction. Read more from Procore CEO Ajei Gopal: procore.com/blog/procore-acq…

Kudos @gharik @ericzelikman @YuchenHe07 @noahdgoodman and the entire @humansand team for launching Persimmon, the first large-scale user model designed to be more, well, human. It is built to capture the messiness and range of human interaction & behavior. @500GlobalVC is honored to be along for the ride. Plus, persimmons are one of the best fruits of all time. 🟠🎉

Google’s TPU Economics 👀 Google GCP CEO Thomas Kurian @ThomasOrTK just said that their payback period on AI servers is <2 years, and half that on their own silicon. 😊 (at @GoldmanSachs today) But it’s not clear whether he means revenue payback or profit/cash-flow payback. 🤔 Let’s assume he means revenue payback, the more realistic of the two in my opinion. 😐 I’d also interpret the statement as payback for just the servers and not the entire datacenter. This means that they monetize TPU clusters at about $30M per facility MW. How do I get there?🤓 Currently, the list price for Ironwood ranges from $5.40/hr (with a 3 yr commitment) to $12/hr on demand. Google says 10MW of IT/pod power can support 9,216 TPU chips. However, this likely does not include the other energy needs (chillers, pumps, transformers, lighting, etc.) So assuming a very efficient modern Google AI data center with 1.10-1.15 PUE (power usage effectiveness), it means about 800-840 Ironwood TPUs per facility MW. Assuming 820 TPUs per MW, at the 3 yr per hour rate of $5.40, that’s $39M/MW at 100% utilisation. I likely wouldn’t lower the utilization assumption because everyone is sold out everywhere. 😮 But I’d lower the rate for these two reasons: (1) TK paired his statement with a mention of 5 yr contracts, and (2) it’s possible volume enterprise contracts (eg for Anthropic) will be at a further discount with a longer commitment. Here there is not list price to use for 5 yr contracts, so extrapolating a diminishing returns curve on the discount from the listed schedule (from on-demand to 1 yr to 3 yrs) suggests $4.50/hr for 5 year commitments (16% discount from the 3 year $5.40/hr rate). Add on top an additional 10% discount for scaled customers and you get $4.08/hr. Thats how I estimate the ~$29M revenue per facility MW of TPUs. TKs statement also has presents some suggestions on margin. Alphabet depreciates their servers and networking equipment over 6 years. With 2yr/1yr payback on AI servers/TPU respectively, it means every $2 server investment produces $1/$2 per year respectively. Depreciation for AI servers generally consume 33% of each revenue dollar ($2 investment divided by 6 years depreciation as percentage of $1 per year) whereas it consumes 17% of each revenue dollar for TPUs ($2 investment divided by 6 years as percentage of $2 per year). So there’s a 16 percentage point gross margin advantage for TPUs. TPU revenue covers the capex in one year for assets that are depreciating over six years. Of course there are operating costs, but once you get long term commitments, it becomes an attractive asset. No wonder Google is spending on capex. This whole post is speculative, but the margin part especially so. 😝 But it will be interesting to build out this model with more information to come. Tell me what I missed!