Google to Test TPUs in Orbit as Musk Claims Space Compute Will Reach 100%

Key Takeaways

1. Google confirmed it will send TPUs to orbit on a SpaceX Falcon 9 as part of "Project Suncatcher," while Musk claimed space compute will eventually account for 100% of all compute.
2. Tesla Semi entered mass production at a dedicated 1.8-million-square-foot factory with 50,000-unit annual capacity, and early test drives drew positive impressions.
3. SpaceX disclosed that Tesla Megapacks will supply 3.3 GWh to the Colossus 2 AI training cluster, alongside over $90 billion in regional investment and environmental commitments.
4. Perplexity launched Fast Search on its Rust-based Photon engine with sub-250ms p95 latency, and Nous Portal subscribers can now set it as the default search in Hermes Agent.
5. AMD and Perplexity partnered to bring Perplexity Portable Computer to AMD Ryzen AI Halo, targeting a new class of agentic PCs.
6. Vercel AI Gateway data showed Anthropic's spend share falling from 69% to 40% while OpenAI rose from 10% to 24%, with Kimi K3 and DeepSeek absorbing about half of Anthropic's lost share.
7. Jack's Block joined the x402 Foundation for agent payments and contributed Bitcoin Lightning payments, while Cash App launched weekly Bitcoin rewards.

1. AI Models and Infrastructure

1. Google confirmed it will send TPUs into orbit as part of "Project Suncatcher," with prototype satellites built in partnership with Planet launching on SpaceX's Transporter-18 mission to evaluate on-orbit TPU performance. Musk separately claimed space compute will ultimately account for 100% of all compute, framing the launch as a step toward that outcome. 1 2 3
2. SpaceX's new site for its Tennessee/Mississippi AI training cluster disclosed that Tesla Megapacks will provide 3.3 GWh to Colossus 2, described as the largest grid-connected battery in the US and enough to power Memphis for two hours. The filing also cited over $90 billion invested in the region since 2024, expected 2026 tax revenue above $60 million, support for 7,500 local jobs, a $360 million clean water recycling plant, $35 million for a 150 MW substation and $20 million for a second substation, and an agreement with Mississippi's environmental regulator requiring all remaining temporary turbines to be removed by July 2027, which the company said it will complete early. 1
3. Vercel AI Gateway data over roughly two months showed Anthropic's spend share dropping from 69% to 40% while remaining first, OpenAI rising from 10% to 24% and leading on token usage, and Opus 5.5 reaching 10% of spend within two days of launch. The account also said GPT-6 Astra and GPT 5.6 Sol performed strongly, that Kimi K3 and DeepSeek absorbed about half of Anthropic's lost share, and that OpenAI accounted for 62% of image generation volume. 1

2. Agentic Products and Developer Tools

1. Perplexity launched Fast Search for its Search API, built on its in-house Rust retrieval and ranking engine Photon, with the company claiming p95 latency below 250ms and top rankings on relevance and accuracy benchmarks; Photon was built by a small team using hundreds of automated research loops driven by an internal, stronger version of Perplexity Computer to optimize the speed-accuracy Pareto frontier. Nous Portal subscribers can now set Perplexity Fast Search as the default search in Hermes Agent, with p50 latency of 160ms and p95 of 230ms per search call; web search in Hermes Agent is free across all Nous Portal tiers, users can switch the search tool to Nous Portal, and it is not restricted by model provider. 1 2 3 4
2. AMD and Perplexity partnered to bring Perplexity Portable Computer to AMD Ryzen AI Halo, which AMD described as building a new kind of agentic PC where users can access AI models and agents with one click, work with local files and applications, and run recurring workflows, with cloud escalation for advanced research or reasoning under authorization, under the framing "your models, your agents, your data, your stack." The author said AMD joined Perplexity Computer's local AI ecosystem, and cited content said Portable Computer for Windows is now available on AMD Ryzen AI Max series processors for running local AI agents that collaborate with connected apps and local files. 1 2
3. Perplexity Computer added a DocSend connector, now available to all users, turning it into a secure deal room; one use case is having Computer maintain a dynamic investment thesis based on the latest state of deal room materials. Separately, Shopify said it continuously evaluates all search APIs and that Perplexity has become a primary choice on merit, with the same source expressing a preference for C++ while accepting Rust as compiled and GC-free and arguing Rust seems closer to the LLM distribution. 1 2

3. Hardware, Payments, and Platform Moves

1. Tesla Semi entered mass production at the first factory built specifically for Semi, with 50,000-unit annual capacity and standard, long-range, and European variants produced on the same line across 1.8 million square feet; Musk congratulated the Tesla Semi team on completing engineering and reaching production. A Motortrend test drive described the experience as "surprisingly familiar, surprisingly easy," with a turning circle only slightly larger than Model Y, about five turns lock-to-lock, and light but precise EPS feel. Separately, the EU postponed its vote on Tesla FSD (Supervised) from October 6 to December at the earliest. 1 2 3 4
2. Cash App launched Bitcoin rewards: users who hold Bitcoin and opt in can receive weekly Bitcoin rewards with no lock-up period, the Bitcoin is not lent out and remains available, subject to terms. Square Savings raised its rate to up to 4.25% APY. Block joined the x402 Foundation, an open standard for agent payments, and contributed Bitcoin Lightning payments to meet the instant, low-cost, high-frequency payment needs of agent commerce. 1 2 3
3. Square rolled out eight new features, including syncing hours, logos, and contact info to Apple Maps for connected sellers; accepting orders after daily items sell out for later fulfillment; viewing revenue by channel rather than only totals; notifying the kitchen when an order is placed; and automatically applying a service fee to every bill after a one-time setup. Square also announced a new integration with Apple Business covering Australia, Canada, France, Ireland, Japan, Spain, the UK, and the US, letting sellers manage Apple Maps presence from the Square Dashboard, automatically sync business info to Apple Maps, Siri, and Apple Wallet, and support action links for menus, ordering, pickup, delivery, and appointments. 1 2

4. Open Source, Security, and Industry Debate

1. Jack said at Freedom Tech DC that open source benefits global innovation, AI will reshape company structures and how Block operates, and the response should be more openness rather than slowing down, with governments funding compute and AI built the way cypherpunks built Bitcoin in public. A forwarded argument said AI risk comes from a "capability asymmetry" created by secret labs training frontier agents with massive compute rather than from a few people in garages, and that open source is the solution to that asymmetry, letting smaller institutions such as hospitals defend themselves. 1 2
2. Amjad Masad forwarded claims that model capabilities and related risks will not slow, that the discussion should expand from "AI safety" to cybersecurity covering who has access, what is connected, and how to secure it, and that this is Replit's focus. He also forwarded that Opus 5.5 built Minecraft in Replit from a single prompt with crafting, monsters, and caves, no external assets, about one hour and roughly $20, which he called the first model he has tested that could do this; other forwarded posts said over 50 million people have built on Replit, that Replit launched an Airwallex connector for payments, subscriptions, and invoices, and that Opus made a video in JavaScript in under 30 minutes combined with Replit Animation. He further forwarded a warning that recent security incidents show frontier AI labs may be moving too fast and should prioritize securing systems before endangering core internet infrastructure, arguing the responsible path is to slow down or risk criminal liability for attacks on core internet infrastructure. 1 2 3 4 5
3. Yohei said the "Aladdin Challenge" received over 1,200 messages in one day from players trying to get an AI wallet agent to transfer funds directly and keep the money, averaging about 26 attempts per person, with one success. Forwarded content classified 556 messages from the prior day: the most common approach was "structured" requests asking for slightly under the limit then adding more, followed by exploiting fees and rounding such as $199.99 or "$200 plus fees," along with scheduled payments, changing recipients, and switching to euros; 21 people used urgency or dying-relative sob stories, 15 asked about the agent's available tools, 8 impersonated Aladdin, and only 8 were prompt injection. The core conclusion was that attacks on AI wallets closely resemble attacks on banks, and that rules cannot live only in prompts but must be enforced at the money layer. 1
4. Guillermo Rauch's Vercel AI Gateway data and the Perplexity, AMD, and Shopify items above collectively show a market where model spend shares are shifting quickly while agent-facing search, local agent PCs, and commerce integrations are being positioned as the next distribution layer. The disclosed figures are point-in-time measurements from the respective companies and accounts, and the competitive implications remain disputed/unverified beyond those reported numbers. 1 2 3 4