Google's TPUs are now running in orbit, Samsung is paying to put Mistral's models inside its chip factories, and Cloudflare trained its own decision models to route agent traffic in milliseconds. Anthropic opened a $100M academy for deployment engineers, HPE landed a $1.2B rack order for AMD hardware, and Airbnb says AI now writes most of its code. Seven stories, sources at the bottom of each.
Google's Project Suncatcher has TPUs running in orbit
Google launched its first Project Suncatcher prototype on October 1 aboard a SpaceX Transporter-18 rideshare from Vandenberg. The fridge-sized satellite was built with Planet Labs and carries four Trillium TPUs, the equivalent of a Cloud TPU v6e-4 slice. Ground controllers made contact the same day, and the spacecraft is running short Gemini inference jobs in 15-minute bursts before its radiators catch up.
— Google Research · NPR
The constraint is heat, not power. In a vacuum there is no air to carry heat away, so every watt the chips burn has to leave through radiator panels. The satellite flies a dawn-dusk sun-synchronous orbit around 550 km up, where solar panels stay in near-continuous sunlight. Google estimates a panel in this orbit can collect up to eight times the usable solar energy of a comparable installation on the ground.
Radiation is the second gate. Earlier lab tests at UC Davis used a 67 MeV proton beam, and a rerun with corrected shielding showed a higher memory error rate than first reported. Even so, Google says the Trillium silicon showed no hard failure through a cumulative dose above what a five-year mission expects, which keeps long-running inference in orbit plausible. The next flight, planned for 2027, sends two purpose-built satellites to test a laser inter-satellite link, after a bench demo pushed 1.6 Tbps between a single transceiver pair.
The economics remain the hardest part. Google's peer-reviewed paper in Joule puts the break-even point at launch costs below $200 per kilogram, which by its own modeling would take roughly 1,800 Starship launches over ten years carrying 370,000 tons of payload. The long-term vision is a constellation of 81 satellites flying in tight formation as an orbital data center. That is a decade-scale research bet, not a product, but the first silicon is now in space and reporting back.
🔗 NPR · Tech Times
Cloudflare trained its own decision models for agent workflows
Cloudflare released Clef and Clef-flash on October 1, the first models trained by its Workers AI team. They are decision models, not chatbots: instead of generating text, they read an input state plus a set of typed questions and return a probability for every allowed answer. An agent gets a structured choice it can act on immediately, like route the ticket, block the request, or escalate to a human.
— Cloudflare
The speed numbers are the pitch. In Cloudflare's own benchmark runs, Clef posted a median latency of 209.3 ms and Clef-flash 38.8 ms, against 524.1 ms for Jev, the decision model from Typesafe AI that defines the category. Across ten decision benchmarks the Clef models lead seven, though Jev still scores higher on two, including When2Call. Clef is built on a frozen Qwen3.8-27B and Clef-flash on Qwen3.5-9B; both score all permitted answers in parallel instead of decoding token by token, and both stay compatible with Jev's API so existing callers can switch without a rewrite.
Weights are open under Apache 2.0 on Hugging Face, and both models run hosted on Workers AI across Cloudflare's network. There is a 64k context window and a vision encoder, so the model can classify screenshots and pages too. Cloudflare also opened an RL fine-tuning service: customer traffic flows through AI Gateway into a training dataset, Containers provide the RL sandbox, and the fine-tuned model redeploys on Workers AI. Its own threat intelligence team already uses Clef to categorize websites, cutting a classification step from 4.7 seconds on a general model to 2.2 seconds.
For agent builders the takeaway is architectural. Routine routing decisions are moving off the general LLM and onto small, fast, calibrated classifiers that sit in the request path, with the frontier model reserved for the steps that need reasoning.
Samsung is paying to put Mistral's models inside its chip business
Samsung Electronics and Mistral AI announced a strategic partnership to build customized AI models for chip design and manufacturing. The plan covers chip design, process optimization, and engineering workflows, with models tailored to run on Samsung's own infrastructure. For a fab, on-premises deployment matters: design data and process parameters are the most sensitive assets a chipmaker owns.
— Mistral AI · Korea Herald
The deal follows Samsung's lead investment in Mistral's €3 billion Series D, the largest equity round ever completed by a European technology company, at a post-money valuation above €21 billion. EQT's Scaleup Europe Fund and PSG Equity co-led, with Advent, BlackRock, and the Grand Duchy of Luxembourg joining, and ASML, NVIDIA, and a16z returning. Two rounds in a row, the biggest check has come from a semiconductor company, ASML last year and Samsung this year.
The strategic logic runs both ways. Samsung gets a sovereign, open-weight AI partner it can deploy inside its fabs without data leaving its walls, plus a way to tune models on its own design and manufacturing data. Mistral gets compute capital and a flagship industrial customer as it builds toward its own European data centers. Mistral now operates in 20 countries serving more than 125 enterprises including Airbus, ASML, and HSBC, and its CFO has pointed to roughly $1 billion in annual recurring revenue by year end. Sovereign AI has moved from policy talk to a procurement line item, and Mistral is currently the only company selling the full open stack for it.
Anthropic commits $100M to train 10,000 deployment engineers
Anthropic launched Claude Frontier Academy on October 2, backed by a $100 million commitment, with a target of 10,000 Frontier Deployed Engineers by the end of 2027. Engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk are in the first cohorts, which are already running in San Francisco, New York, and London. Participation is by nomination, not open enrollment.
— Anthropic · CNBC
The first program, the Frontier Deployed Engineer Residency, copies the medical training model. Engineers start with a multi-day in-person course taught by Anthropic engineers, work through a simulated enterprise deployment from use-case selection through security review to handover, and finish with a graded practical. Those who pass earn a Resident Engineer badge and move into a 12-week residency leading a real Claude project at their own organization, with a second assessment before the full Frontier Deployed Engineer credential, first expected in early 2027.
The bet underneath is about where AI value actually gets captured. Anthropic argues that inside its customers, a small group of deeply skilled people drives an outsized share of what AI delivers, and that nobody has trained that talent at depth inside customers and partners. The program builds on the Claude Partner Network, where professionals across 46,000 firms have earned over 175,000 certifications. CNBC also frames it against Anthropic's expected IPO: Reuters reported this week, citing a leaked prospectus, that the company generated almost $4.6 billion in revenue last year against an operating loss of over $8 billion. Training the deployment layer is product strategy for an enterprise franchise, not a philanthropy program.
HPE lands a $1.2B first order for AMD Helios racks from Vultr
HPE announced a $1.2 billion order from Vultr, the largest privately-held cloud infrastructure company, to deploy AMD Helios AI Rack systems across Vultr's US data centers. It is the first customer order for the platform, which pairs AMD compute with HPE's own scale-up networking, a combination HPE has been assembling since it closed the Juniper acquisition.
— HPE · ET Telecom
Each Helios rack integrates 72 AMD Instinct MI455X GPUs with EPYC Venice CPUs, Pensando Vulcano AI NICs, and ROCm software, designed for trillion-parameter training and high-volume inference. The networking layer is the interesting part: six HPE Juniper QFX5252 scale-up Ethernet switch trays per rack connect all 72 GPUs over standards-based Ethernet, supporting open fabrics like UALink over Ethernet rather than a proprietary interconnect. Direct liquid cooling and HPE deployment services round out the stack.
The order came alongside a raised outlook. HPE now expects its networking segment to grow at a high-teens compound annual rate from fiscal 2026 through 2029, up from a prior 5 to 7 percent forecast, and the guidance move sent shares up nearly 6 percent in premarket trading. Vultr gets a second accelerated architecture alongside the NVIDIA GB300 NVL72 systems it ordered through HPE earlier in the year. The signal for the rack-scale market is that open Ethernet scale-up fabrics are now shipping at billion-dollar scale, not just on slide decks.
Dyna's Taku robot folded laundry for an hour, uncut
Dyna Robotics released Dyna-2.1 at the end of September, pairing a new semi-humanoid robot called Taku with an agentic stack built around its Dyna-2 world-action model. The demo is an uncut, one-hour commercial laundry workflow: the robot loads washers and dryers, folds towels, stacks them on shelves, and recovers from its own mistakes the whole way through. No cuts, no hidden teleoperation.
— Dyna Robotics · Robot24
The hardware choices are deliberate. Taku has a human-sized upper body with two 7-degree-of-freedom arms, a folding lower body that reaches low shelves and high racks, and four steerable wheels instead of legs. Dyna's argument is that the workflows it targets need reach and manipulation precision across stations, not humanlike walking. A vision-language orchestrator picks the next move at roughly 1 Hz, the Dyna-2 world-action model works at about 10 Hz, and a whole-body controller drives joints and wheels at up to a kilohertz, with the model trained on about a million hours of human and robot video.
The most honest part of the release is the reliability math. Dyna shows that a laundry cycle chains roughly 79 subtasks, so at a 95 percent per-step success rate the chance of finishing a full cycle unassisted is 1.7 percent. That is why the company optimizes mean time between interventions instead of single-task success rates, and why an hour of uncut autonomy matters more than a dozen perfect short clips. Taku is rolling out to hotels, laundromats, and restaurants, with restaurant chain Din Tai Fung already using earlier Dyna robots for napkin folding. Claims about commercial reliability remain company numbers for now, but the workflow-first framing, complete workflows rather than single skills, is becoming the standard Physical AI pitch.
🔗 DYNA Robotics · Robot24
Airbnb says AI now writes 60% of its code
Airbnb CTO Ahmad Al-Dahle said in a Latent Space interview published October 2 that about 60 percent of the company's code is now AI-authored. The company shipped nearly 80 percent more features and improvements year over year, and average pull-request throughput per engineer is up roughly 1.6x. Al-Dahle joined Airbnb in January after leading generative AI at Meta through the Llama years.
— Airbnb CTO · Latent Space
The velocity gain came mostly from deleting process, not adding tools. Product, design, and engineering teams now work directly from working prototypes instead of passing requirements documents and Figma files between groups, with the code itself as the artifact everyone reasons about. Internally, an organizational context graph called Everest, built on LLMs, embeddings, and AI retrieval, lets generalist engineers navigate specialist codebases. The payoff shows in partner integrations: grocery delivery took eight to nine months to build, and the airport pickup service that reused those patterns took about six weeks.
Airbnb runs as a multi-model shop with at least ten customized models in production, doing most post-training and reinforcement learning on open models while reserving frontier models for reasoning-heavy work. Customer support was the first user-facing deployment, and roughly half of tickets are now resolved by AI, matching the nearly 45 percent figure in Q2 results, with safety-related tickets deliberately kept human. Next up is automating on-call: asynchronous agents in containers that spin up when a monitoring alert trips and open remediation PRs for human review. One policy stands out against the throughput numbers: every engineer must be able to explain what they shipped, even when AI generated the code. That is the same lesson every fast-moving team is about to learn.
🔗 Latent Space interview via Pivot News
KD Agentic · AI Daily Digest
Top comments (0)