Analog Devices says intelligence is leaving the cloud. NVIDIA is about to unveil chips the world hasn't seen. DeepSeek just locked American chipmakers out of its next model, and the Dow had its worst day since April 2025. The frontier isn't just accelerating, it's restructuring. We've been building something that doesn't look like software. It looks like a contact in your phone. Last weekend, he called a formation top before the geosteering team did. Meet Derrick in the Field Report.
Read time: ~5 min
What’s new and interesting in AI/ML this week.
Alibaba's Qwen team released the Qwen3.5 Small series — four open-source models from 0.8B to 9B parameters, built for phones, laptops, and edge hardware. The 9B outperforms OpenAI's GPT-OSS-120B on graduate-level reasoning benchmarks. The 0.8B runs on an iPhone 17. All are natively multimodal (text, image, video) and Apache 2.0 licensed. The tagline: "More intelligence, less compute." ADI called it physical intelligence. Alibaba just open-sourced it.
Ayar Labs closed a $500M Series E co-led by NVIDIA and AMD to scale optical interconnect technology — replacing the copper wiring between AI chips with light. As GPU clusters grow to hundreds of thousands of processors, the bottleneck is no longer compute but the speed at which chips can talk to each other. Ayar's co-packaged optics promise lower latency and dramatically less power draw at data center scale. When both NVIDIA and AMD co-lead the same round, they're not placing a bet — they're buying the plumbing.
Perseverance drove 807 feet across Jezero Crater on a route planned entirely by generative AI — no human drivers involved. JPL used vision-language models to analyze orbital imagery, identify hazards like boulder fields and sand ripples, and chart a safe path. The AI's route was validated against 500,000 telemetry variables in a digital twin before commands were sent to Mars. An AI that reads terrain imagery, identifies geological features, flags hazards, and plans a path through unfamiliar ground 140 million miles from the nearest human operator. Sound familiar?
Analog Devices: 2026 Is the Year Intelligence Gets Physical
Analog Devices' VP of Edge and Enterprise AI published five predictions for 2026, and the throughline is blunt: AI is leaving the cloud. The company is betting that "physical reasoning models", systems that learn from real-world signals like vibration, sound, magnetics, and motion, will migrate from data centers to the edge, powering what ADI calls "Physical Intelligence." Not chatbots. Not dashboards. Systems that perceive, reason, and act locally, sensitive to local physics and without recourse to centralized servers.
The most relevant prediction: a new class of "micro-intelligences" will emerge, compact models with deep reasoning across narrow domains, running at the edge. ADI describes them as the middle ground between rigid programmed AI and sprawling foundation models; fluid, adaptive, task-specific, and capable of orchestrating specialized agents in the field.
ADI calls it Physical Intelligence, systems that perceive, reason, and act at the edge. The industry that should be leading this shift is the one that's been collecting real-time data in harsh environments for decades. The wellsite has always been the edge. The question was never whether the intelligence would move there, it was what it would look like when it did.
Quick hits worth your attention this week.
DeepSeek is preparing to release V4, a multimodal model handling text, image, and video. In a break from standard practice, they've withheld it from US chipmakers including Nvidia for performance optimization. UBS says MiniMax's M2.5 already rivals Claude Opus 4.6 at one-tenth the price. The cost curve for frontier AI just got steeper for everyone betting on American hardware.
NVIDIA's GTC 2026 expects 30,000 attendees and 1,000+ sessions. Jensen Huang's keynote on March 16 is expected to unveil next-gen GPU architecture and agentic AI infrastructure. Nvidia also invested $2B each in photonics companies Lumentum and Coherent to bolster AI chip interconnects. If you're building on GPU compute, the next two weeks set the roadmap.
Taalas is building chips that embed AI models directly into the silicon itself—no software inference stack, no GPU overhead. At $169M in backing, they're betting that for specific high-throughput inference tasks, purpose-built hardware will crush general-purpose GPUs. Think: real-time wellbore analysis at the edge, where milliseconds matter and cloud latency kills.
AI capability is now doubling approximately every seven months, and experts say the trend is accelerating. The problem: existing benchmarks saturate so fast that measuring progress is becoming harder than making it. When your measurement tools break before the thing you're measuring slows down, you're in uncharted territory.
ASML is developing new scanner systems and lithography tools because AI chips have physically outgrown current manufacturing. The chips are getting so large that existing EUV lithography can't pattern them in a single pass. When the manufacturing process itself becomes the bottleneck, every downstream cost—from training runs to inference—gets repriced.
A Photo, a Text, and a Formation Top Called 25 ft Ahead of Prognosis
Derrick is a contact in your phone. Not an app. Not a dashboard. Not a chatbot behind a login screen. He's a messaging-native agent with full access to your live WITSML stream, your geo prognosis, your offset well data, your knowledge base, even your real-time geosteering project. He operates in threaded replies and group chats, handles photos, and already has the context before you ask.
This weekend, an engineer texted Derrick a photo of cuttings at 7,369 ft. Here's the thread.
Fewer Rigs, Record Output: The Permian's Quiet Efficiency Revolution
The Permian Basin produced 6.6 million barrels per day in 2025, a record, while running 30% fewer rigs than its 2018–2019 average. The EIA forecasts production will hold near that level through 2026 even as rig counts continue to slide below 300. Goldman Sachs attributes this to a simple trend: every stage of a well's building cycle is now 20–50% faster than it was in 2019, with average time from rig to production dropping by a third to 63 days.
The math is counterintuitive. Fewer rigs. Fewer people. More barrels. The Permian now accounts for 51% of all onshore Lower 48 production and 65% of tight oil production growth. Wolfcamp alone produces 3.4 million barrels per day, equivalent to every other non-Permian tight oil play combined. The basin isn't just America's most productive oil patch. It's becoming a case study in what happens when an industry learns to do more with less before it has to.
But the efficiency story has a ceiling. The EIA's January 2026 outlook warns that WTI is forecast to fall to $52/bbl in 2026 and $50/bbl in 2027, below the $61–62/bbl breakeven that Permian operators reported in the Dallas Fed survey. Drilling productivity gains can offset a lot, but they can't offset price gravity indefinitely. And the basin's geology is maturing. Goldman notes that while output per rig keeps climbing, the underlying rock quality is deteriorating, meaning operators are running harder just to hold the line.
The Permian has always been the industry's best argument for American energy independence. In 2026, it's also becoming the best argument for operational intelligence, because when margins compress and rig counts drop, the only lever left is how smart you are with the wells you have.
Data ≠ Decisions. Context changes everything. DrillSense is the intelligence layer for drilling operations, built for the people who make the calls.
Know someone who should be reading this? Send them the archive, or subscribe from drillsense.com.